Yes opus 4.5 seems great but most of the time it tries to vastly over complicate a solution. Its answer will be 10x harder to maintain and debug than the simpler solution a human would have created by thinking about the constraints of keeping code working.
Over time, I imagine even cloud providers, app stores etc can start doing automated security scanning for these types of failure modes, or give a more restricted version of the experience to ensure safety too.
I didn't write off an entire field of research, but rather want to highlight that these aren't intractable problems for AI research, and that we can actually start codifying many of these things today using the skills framework to close up edges in the model training. It may not be 100% but it's not 0%.
Maintain and debug by who? It's just going to be Opus 4.5 (and 4.6...and 5...etc.) that are maintaining and debugging it. And I don't think it minds, and I also think it will be quite good at it.
I predict in 2026 we're going to see agents get better at running their own QA, and also get better at not just disabling failing tests. We'll continue to see advancements that will improve quality.
something like code-simplifier is surprisingly useful (as is /review)
And it turns out the quality of output you get from both the humans and the models is highly correlated with the quality of the specification you write before you start coding.
Letting a model run amok within the constraints of your spec is actually great for specification development! You get instant feedback of what you wrongly specified or underspecified. On top of this, you learn how to write specifications where critical information that needs to be used together isn't spread across thousands of pages - thinking about context windows when writing documentation is useful for both human and AI consumers.
I can’t get past that by the time I write up an adequate spec and review the agents code, I probably could have done it myself by hand. It’s not like typing was even remotely close to the slow part.
AI, agents, etc are insanely useful for enhancing my knowledge and getting me there faster.
I consider my job to be actually useful. That I produce useful stuff to society at large.
I definitely hope that I'm replaced with someone/thing better; whatever it is. That's progress.
I surely don't hope for a futre where I retire and medics have access to worse tech than they have now.
The seniors who have less leeway to change course (its harder as you get older in general, large sunk costs, etc) maintain their positions and the disruption occurs at the usual "retirement rate" meaning the industry shrinks a bit each year. They don't get much with pay rises, etc but normally they have some buffer from earlier times so are willing to wear being in a dying field. Staff aren't replaced but on the whole they still have marginal long term value (e.g. domain knowledge on the job that keeps them somewhat respected there or "that guy was around when they had to do that; show respect" kind of thing).
The juniors move to other industries where the price signal shows value and strong demand remains (e.g. locally for me that's trades but YMMV). They don't have the sunk cost and have time on their side to pivot.
If done right the disruption to people's lives can be small and most of the gains of the tech can still come out. My fear is the AI wave will happen fast but only in certain domains (the worst case for SWE's) meaning the adjustment will be hard hitting without appropriate support mechanisms (i.e. most of society doesn't feel it so they don't care). On average individual people aren't that adaptable, but over generations society is.
Stuff that seems basic, but that I haven't always been able to count on in my teams' "production" code.
It means that it is going to be as easy to create software as it is to create a post on TikTok, and making your software commercially successful will be basically the same task (with the same uncontrollable dynamics) as whether or not your TikTok post goes viral.
If anything this example shows that these cli tools give regular devs much higher leverage.
There's a lot of software labor that is like, go to the lowest cost country, hire some mediocre people there and then hire some US guy to manage them.
That's the biggest target of this stuff, because now that US guy can just get equal or hight code in both quality and output without the coordination cost.
But unless we get to the point where you can do what I call "hypercode" I don't think we'll see SWEs as a whole category die.
Just like we don't understand assembly but still need technical skills when things go wrong, there's always value in low level technical skills.
This is also my take. When the printing press came out, I bet there were scribes who thought, "holy shit, there goes my job!" But I bet there were other scribes who thought, "holy shit, I don't have to do this by hand any more?!"
It's one thing when something like weaving or farming gets automated. We have a finite need for clothes and food. Our desire for software is essentially infinite, or at least, it's not clear we have anywhere close to enough of it. The constraint has always been time and budget. Those constraints are loosening now. And you can't tell me that when I am able to wield a tool that makes me 10X more productive that that somehow diminishes my value.
If AI datacenters' hungry need for energy gets us to nuclear power, which gets us the energy to run desalination plants as the lakes dry up because the Earth is warming, hopefully we won't die of thirst.
I don't understand this argument. Surely the skill set involved in being a scribe isn't the same as being a printer, and possibly the the personality that makes a good scribe doesn't translate to being a good printer.
So I imagine many of the scribes lost their income, and other people made money on printing. Good for the folks who make it in the new profession, sucks for those who got shafted. How many scribes transitioned successfully to printers?
Genuinely asking, I don't know.
So a lot of people will end up doing something different. Some of it will be menial and be shit, and some of it will be high level. New hierarchies and industries will form. Hard to predict the details, but history gives us good parallels.
I think for a while people have been talking about the fact that as all development tools have gotten better - the idea that a developer is a person who turns requirements into code is dead. You have to be able to operate at a higher level, be able to do some level of work to also develop requirements, work to figure out how to make two pieces of software work together, etc.
But the point is Obviously at an extreme end 1 CTO can't run google and probably not say 1 PM or Engineer per product, but what is the mental load people can now take on. Google may start hiring less engineers (or maybe what happens is it becomes more cuthroat, hire the same number of engineers but keep them much more shortly, brutal up or out.
But essentially we're talking about complexity and mental load - And so maybe it's essentially the same number of teams because teams exist because they're the right size, but teams are a lot smaller.
Not in terms of knowledge. That was already phenomenal. But in its ability to act independently: to make decisions, collaborate with me to solve problems, ask follow-up questions, write plans and actually execute them.
You have to experience it yourself on your own real problems and over the course of days or weeks.
Every coding problem I was able to define clearly enough within the limits of the context window, the chatbot could solve and these weren’t easy. It wasn’t just about writing and testing code. It also involved reverse engineering and cracking encoding-related problems. The most impressive part was how actively it worked on problems in a tight feedback loop.
In the traditional sense, I haven’t really coded privately at all in recent weeks. Instead, I’ve been guiding and directing, having it write specifications, and then refining and improving them.
Curious how this will perform in complex, large production environments.
This is basically all my side projects.
How do you stop it from over-engineering everything?
It may end up working, but the thing is going to convolute apis and abstractions and mix patterns basically everywhere
I can't teach it taste.
It's verbose by default but a few hours of custom instructions and you can make it code just like anyone
Arthur Whitney?
https://en.wikipedia.org/wiki/Arthur_Whitney_(computer_scien...
Once more people realize how easy it is to customize and personalized your agent, I hope they will move beyond what cookie cutter Big AI like Anthropic and Google give you.
I suspect most won't though because (1) it means you have to write human language, communication, and this weird form of persuasion, (2) ai is gonna make a bunch of them lazy and big AI sold them on magic solutions that require no effort on your part (not true, there is a lot of customizing and it has huge dividends)
Yes, feel free to blame me for the fact that these aren’t very business-realistic.
When I use Claude Code I find that it *can* add a tremendous amount of ability due to its ability to see my entire codebase at once, but the issue is that if I'm doing something where seeing my entire codebase would help that it blasts through my quota too fast. And if I'm tightly scoping it, it's just as easy & faster for me to use the website.
Because of this I've shifted back to the website. I find that I get more done faster that way.
If AI cost nothing and wasn't absolutely decimating our economy, I'd find what you've shared cute. However, we are putting literally all of our eggs, and the next generation's eggs, and the one after that, AND the one after that, into this one thing, which, I'm sorry, is so far away from everything that keeps on being promised to us that I can't help but feel extremely depressed.
1) Modern LLMs are an inflection point for coding.
2) The current LLM ecosystem is unsustainable.
This submission discussion is only about #1, which #2 does not invalidate. Even if the ecosystem crashes, then open-source LLMs that leverage the same tricks Opus 4.5 does will just be used instead.
So cool bro, you managed to ship a useless (except for your specific use-case) app to your iphone in an hour :O
What I think this is doing is it's pitting people against the fact that most jobs in the modern economy (mine included btw) are devoid of purpose. This is something that, as a person on the far left, I've understood for a long time. However, a lot (and I mean a loooooot) of people have never even considered this. So when they find that an AI agent is able to do THEIR job for them in a fraction of the time, they MUST understand it as the AI being some finality to human ingenuity and progress given the self-importance they've attributed to themselves and their occupation - all this instead of realizing that, you know, all of our jobs are useless, we all do the exact same useless shit which is extremely easy to replicate quickly (except for a select few occupations) and that's it.
I'm sorry to tell anyone who's reading this with a differing opinion, but if AI agents have proven revolutionary to your job, you produced nothing of actual value for the world before their advent, and still don't. I say this, again, as someone who beyond their PhD thesis (and even then) does not produce anything of value to the world, while being paid handsomely for it.
Definitely, it’s an unhealthy fixation.
> I'm sorry to tell anyone who's reading this with a differing opinion, but if AI agents have proven revolutionary to your job, you produced nothing of actual value for the world before their advent, and still don't.
I agree with this, but I think my take on it is a lot less nihilistic than yours. I think people vastly undersell how much effort they put into doing something, even if that something is vibecoding a slop app that probably exists. But if people are literally prompting claude with a few sentences and getting revolutionary results, then yes, their job was meaningless and they should find something to do that they’re better at.
But what frustrates me the most about this whole hype wave isn’t just that the powers that be have bet the entire economy on a fake technology, it’s that it’s sucking all of of the air out of the room. I think most people’s jobs can actually provide value and there’s so much work to be done to make _real_ progress. But instead of actually improving the world, all the time, money, and energy is being thrown into such a wasteful technology that is actively making the world a worse place. I’m sure it’s always been like this and I was just to naive too see it, but I much preferred it when at least the tech companies pretended they cared about the impact their products had on society rather than simply trying to extract the most value out of the same 5 ideas.
I really think we're just cooked at this point. The amount of people (some great friends whom I respect) that have told me in casual conversation that if their LLM were taken from them tomorrow, they wouldn't know how to do their work (or some flavour of that statement) has made me realize how deep the problem is.
We could go on and on about this, but let's both agree to try and look inward more and attempt to keep our own things in order, while most other people get hooked on the absolute slop machine that is AI. Eventually, the LLM providers will need to start ramping up the costs of their subscriptions and maybe then will people start clicking that the shitty code that was generated for their pointless/useless app is not worth the actual cost of inference (which some conservative estimates put out to thousands of dollars per month on a subscription basis). For now, people are just putting their heads in the sand and assuming that physicists will somehow find a way to use quantum computers to speed up inference by a factor of 10^20 in the next years, while simultaneously slashing its costs (lol).
But hey, Opus 4.5 can cook up a functional app that goes into your emails and retrieves all outstanding orders - revolutionary. Definitely worth the many kWh and thousands of liters of water required, eh?
Cheers.
The thing is in a vacuum this stuff is actually kinda cool. But hundreds of billions in debt-financed capex that will never seen a return, and this is the best we’ve got? Absolutely cooked indeed.
GPT-3 Da Vinci cost $20/million tokens for both input and output.
GPT-5.2 is $1.75/million for input and $14/million for output
I'd call that pretty strong evidence that they've been able to dramatically increase quality while slashing costs, over just the past ~4 years.
1. The AI water issue is fake: https://andymasley.substack.com/p/the-ai-water-issue-is-fake (This one goes into OCD-levels of detail with receipts to debunk that entire issue in all aspects.)
2. LLMs are far, far more efficient than humans in terms of resource consumption for a given task: https://www.nature.com/articles/s41598-024-76682-6 and https://cacm.acm.org/blogcacm/the-energy-footprint-of-humans...
The studies focus on a single representative task, but in a thread about coding entire apps in hours as opposed to weeks, you can imagine the multiples involved in terms of resource conservation.
The upshot is, generating and deploying a working app that automates a bespoke, boring email workflow will be way, way, wayyyyy more efficient than the human manually doing that workflow everytime.
Hope this makes you feel better!
I want to push back on this argument, as it seems suspect given that none of these tools are creating profit, and so require funds / resources that are essentially coming from the combined efforts of much of the economy. I.e. the energy externalities here are monstrous and never factored into these things, even though these models could never have gotten off the ground if not for the massive energy expenditures that were (and continue to be) needed to sustain the funding for these things.
To simplify, LLMs haven't clearly created the value they have promised, but have eaten up massive amounts of capital / value produced by everyone else. But producing that capital had energy costs too. Whether or not all this AI stuff ends up being more energy efficient than people needs to be measured on whether AI actually delivers on its promises and recoups the investments.
EDIT: I.e. it is wildly unclear at this point that if we all pivot to AI that, economy-wide, we will produce value at a lower energy cost, and, even if we grant that this will eventually happen, it is not clear how long that will take. And sure, humans have these costs too, but humans have a sort of guaranteed potential future value, whereas the value of AI is speculative. So comparing energy costs of the two at this frozen moment in time just doesn't quite feel right to me.
However, their economic potential is undeniable. Just taking the examples in TFA and this sub-thread, the author was able to create economic value by automating rote aspects of his wife's business and stop paying for existing subscriptions to other apps. TFA doesn't mention what he paid for these tokens, but over the lifetime of his apps I'd bet he captures way more value than the tokens would have cost him.
As for the energy externalities, the ACM article puts some numbers on them. While acknowledging that this is an apples/oranges comparison, it points out that the training cost for GPT-3 (article is from mid-2024) is about 5x the cost of raising a human to adulthood.
Even if you 10x that for GPT-5, that is still only the cost of raising 50 humans to adulthood in exchange for a model that encapsulates a huge chunk of the world's knowledge, which can then be scaled out to an infinite number of tasks, each consuming a tiny fraction of the resources of a human equivalent.
As such, even accounting for training costs, these models are far more efficient than humans for the tasks they do.
Firstly, saying that because AI water use is on par with other industries, then we shouldn't scrutinize AI water use is a bit short-sighted. If the future Altman et al want comes to be, the shear scale of deployment of AI-focused data centers will lead to nominal water use orders of magnitude larger than other industries. Of course, on a relative scale, they can be seen as 'efficient', but even something efficient, when built out to massive scale, can suck out all of our resources. It's not AI's fault that water is a limited resource on Earth; AI is not the first industry to use a ton of water; however, eventually, with all other industries + AI combined (again, imagining the future the AI Kings want), we are definitely going 300km/h on the road to worldwide water scarcity. We are currently at a time where we need to seriously rethink our relationship with water as a society - not at a time where we can spawn whole new, extremely consumptive industries (even if, in relative terms, they're on par with what we've been doing (which isn't saying much given the state of the climate)) whose upsides are still fairly debatable and not at all proven beyond a doubt.
As for the second link, there's a pretty easy rebuke to the idea, which aligns with the other reply to your link. Sure, LLMs are more energy-efficient at generating text than human beings, but do LLMs actually create new ideas? Write new things? Any text written by an LLM will be based off of someone else's work. There is a cost to creativity - to giving birth to actual ideas - that LLMs will never be able to incur, which makes them seem more efficient, but in the end they're more efficient at (once again) tasks which us humans have provided them with plenty of examples of (like writing corporate emails! Or fairly cookie-cutter code!) but at some point the value creation is limited.
I know you disagree with me, it's ok - you are in the majority and you can feel good about that.
I honestly hope the future you foresee where LLMs solve our problems and become important building blocks to our society comes to fruition (rather than the financialized speculation tools they currently are, let's be real). If that happens, I'll be glad I was wrong.
I just don't see it happening.
For instance, all water use by data centers is a fraction of the water used by golf courses! If it really does comes down to the wire for conserving water, I think humanity has the option of foregoing a leisure activity for the relatively wealthy in exchange for accelerated productivity for the rest of the world.
And totally, LLMs might not be able to come up with new ideas, but they can super-charge the humans who do have ideas and want to develop them! An idea that would have taken months to be explored and developed can now be done in days. And given that like the majority of ideas fail, we would be failing that much faster too!
In either case, just eyeballing the numbers we have currently, on average the resources a human without AI assistance would have consumed to conclude an endeavor far outweighs the resources consumed by both that human and an assisting LLM.
I would agree that there will likely be significant problems caused by widespread adoption of AI, but at this point I think they would social (e.g. significant job displacement, even more wealth inequality) rather than environmental.
This doesn’t logically follow. AI agents produce loads of value. Cotton picking was and still is useful. The cotton gin didn’t replace useless work. It replaced useful work. Same with agents.
If the AI thing does indeed come crashing down I expect there will be a whole lot of second-hand GPUs going for pennies on the dollar.
Or I'll start one myself, if the market fails to provide!
The projects being submitted to product hunt is 4x the year before.
The market is shrinking rapidly because now more people make their own apps.
Even making a typo and landing on a website, there is good chance its selling more ai snake oil, yet none of these apps are feature complete and easily beaten by apps made by guys in 2010s. (tldr & sketchbook for the drawing space).
Only way to excite the investors is to fake the ARR by giving free trials and sell before the recurring event occurs.
That describes half of the current unicorn startups nowadays.
I have some old borderline senile relatives writting apps (asking LLMs to write for it them) for their own personal use. Stuff they surely haven't done on their own (or had the energy to do). Their extent of programming background - shitty VBScript macros for excel.
It also helps people to pick up programming and helps with the initial push of getting started. Getting over the initial hump, getting something on the screen so to speak.
Most things people want from their computers are simple shit that LLMs usually manage quite well.
Good question whether or not this (outsourcing their thinking) actually just accelerates their senility or not.
As someone who likes to solve hard or interesting technical problems, I've long before LLMs often been disappointed that most of the time what people want from programmers is simple stupid shit (ie. stuff i dont find interesting to work on).
The article is fine opinion but at what point are we going to either:
a) establish benchmarks that make sense and are reliable, or
b) stop with the hypecycle stuff?
How aren't current LLM coding benchmarks reliable?
If you can figure out how to create benchmarks that make sense, are reliable, correlate strongly to business goals, and don't get immediately saturated or contorted once known, you are well on your way to becoming a billionaire.
When I'm dying of dehydration because humanity has depleted all fresh water deposits, I'll think of you and your stupid NES emulator which is just an LLM-produced copy of many ones that had already existed.
I don't think open office/libre office etc have access to the source code for MS office and if they did MS would be on them like a rash.
https://andymasley.substack.com/p/the-ai-water-issue-is-fake
However, Opus 4.5 is incredible when you give it everything it needs, a direction, what you have versus what you want and it will make it work, really, it will work. The code might me ugly, undesirable, would only work for that one condition, but with futher prompting you can evolve it and produce something that you can be proud of.
Opus is only as good as the user and the tools the user gives to it. Hmm, that's starting to sound kind-of... human...
I now write very long specifications and this helps. I haven't figured out a bulletproof workflow, I think that will take years. But I often get just amazing code out of it.
It's hard to say if Opus 4.5 itself will change everything given the cost/latency issues, but now that all the labs will have very good synthetic agentic data thanks to Opus 4.5, I will be very interested to see what the LLMs release this year will be able to do. A Sonnet 4.7 that can do agentic coding as well as Opus 4.5 but at Sonnet's speed/price would be the real gamechanger: with Claude Code on the $20/mo plan, you can barely do more than one or two prompts with Opus 4.5 per session.
It's always just the "Fibonacci" equivalent
They're going to find a lot of stuff to fix.
The few times I did go to a shop and ask for a check-up they didn’t find anything. Just an anecdote.
I wouldn't be surprised to find out that they will find issues infinitely, if looped with fixes.
This reminds me of that example where someone asked an agent to improve a codebase in a loop overnight and they woke up to 100,000 lines of garbage [0]. Similarly you see people doing side-by-side of their implementation and what an AI did, which can also quite effectively show how AI can make quite poor architecture decisions.
This is why I think the “plan modes” and spec driven development are so important effective for agents, because it helps to avoid one of their main weaknesses.
I had this long discussion today with a co-worker about the merits of detailed queries with lots of guidance .md documents, vs just asking fairly open ended questions. Spelling out in great detail what you want, vs just generally describing what you want the outcomes to be in general then working from there.
His approach was to write a lot of agent files spelling out all kinds of things like code formatting style, well defined personas, etc. And here's me asking vague questions like, "I'm thinking of splitting off parts of this code base into a separate service, what do you think in general? Are there parts that might benefit from this?"
I have a hunch if you asked which approach we took based on background, you'd think I was the one using the detailed prompt approach and him the vague.
Your approach is also very similar to spec driven development. Your spec is just a conversation instead of a planning document. Both approaches get ideas from your brain into the context window.
So you gave it an poorly defined task, and it failed?
Have you tried the planning mode? Ask it to review the codebase and identify defects, but don't let it make any changes until you've discussed each one or each category and planned out what to do to correct them. I've had it refactor code perfectly, but only when given examples of exactly what you want it to do, or given clear direction on what to do (or not to do).
The above is for vibe coding; for taking the wheel, I can only use Opus because I suck at prompting codex (it needs very specific instructions), and codex is also way too slow for pair programming.
Why not the other way around? Have the quick brown fox churn out code, and have codex review it, guide changes, and loop?
I've actually gone one step further down the delegation. I use opus/gemini3 for plan, review, edit plan for a few steps. Then write it out to .md files. Then have GLM implement it (I got a cheap plan for like 28$ for a year on Christmas). Then have the code this produced reviewed and fixed if needed by opus. Final review by codex (for some reason it's very good at review, esp if you have solid checkboxes for it to check during review). Seems to work so far.
Currently I don’t let GLM or Opus near my codebases unsupervised because I’m convinced that the better the foundation, the better the end result will be. Is the first draft not pretty crappy with GLM?
Once you’ve got Claude Code set up, you can point it at your codebase, have it learn your conventions, pull in best practices, and refine everything until it’s basically operating like a super-powered teammate. The real unlock is building a solid set of reusable “skills” plus a few agents for the stuff you do all the time.
For example, we have a custom UI library, and Claude Code has a skill that explains exactly how to use it. Same for how we write Storybooks, how we structure APIs, and basically how we want everything done in our repo. So when it generates code, it already matches our patterns and standards out of the box.
We also had Claude Code create a bunch of ESLint automation, including custom ESLint rules and lint checks that catch and auto-handle a lot of stuff before it even hits review.
Then we take it further: we have a deep code review agent Claude Code runs after changes are made. And when a PR goes up, we have another Claude Code agent that does a full PR review, following a detailed markdown checklist we’ve written for it.
On top of that, we’ve got like five other Claude Code GitHub workflow agents that run on a schedule. One of them reads all commits from the last month and makes sure docs are still aligned. Another checks for gaps in end-to-end coverage. Stuff like that. A ton of maintenance and quality work is just… automated. It runs ridiculously smoothly.
We even use Claude Code for ticket triage. It reads the ticket, digs into the codebase, and leaves a comment with what it thinks should be done. So when an engineer picks it up, they’re basically starting halfway through already.
There is so much low-hanging fruit here that it honestly blows my mind people aren’t all over it. 2026 is going to be a wake-up call.
(used voice to text then had claude reword, I am lazy and not gonna hand write it all for yall sorry!)
Edit: made an example repo for ya
Why on earth would you be hunting $20 a month subscriptions from random assed people? Peanuts.
Lockheed-Martin could be, but isn’t, opening lemonade stands outside their offices… they don’t because of how buying a Ferrari works.
For the same reason Microsoft never has and never will chase people for pirating home Windows or Office licenses
When they hit the workforce, or even better, start a company guess which OS and office suite they'll use? Hint: It's not Linux and Openoffice.
Same with Claude's $20 package. It lets devs use it at home and then compare it to the Copilot shit their company is pushing on them. Maybe they either grumble enough to get a Claude license or they're in a position to make the call.
Cheap advertising pretty much.
Worked for me too :) I've paid my own Claude license for over a year at home, grumbled at work and we got a Claude pilot going now - and everyone who's tried it so far isn't going back to Copilot + Sonnet 4.5/GPT5.
LLMs do not result in bosses firing people, it results in more projects / faster completed projects, what in turn means more $$$ for a company.
Well excuse the shit out of my goddamn French, but being comfy for years and suddenly facing literal doom of my profession in a year wasn't on my bingo card.
And what do you even mean by "prepare"? Shit out a couple of mil out of my ass and invest asap?
Not the person you're responding to but... if you think it's a horse -> car change (and, to stretch the metaphor, if you think you're in the business of building stables) then preparation means train in another profession.
If you think it's a hand tools -> power tools change, learn how to use the new tools so you don't get left behind.
My opinion is it's a hand -> power tools change, and that LLMs give me the power to solve more problems for clients, and do it faster and more predictably than a client trying to achieve the same with an LLM. I hope I'm right :-)
You think any manager in their right mind would take risks like that?
I think the real consequences are that they probably are so pleased with how productive the team is becoming that they will not hire new people or fire the ones who aren't keeping up with the times.
It's like saying "wow, our factory just produced 50% more cars this year, time to shut down half the factory to reduce costs!"
You really underestimate stupidity of your average manager. Two of our top performers left because they were underpaid and the manager (in charge of the comp) never even tried to retain them.
Ignoring that this same developer, now has access to a tool, that makes himself a team.
Going independent was always a issue because being a full stack dev, is hard. With LLMs, you have a entire team behind you for making graphics, code, documents, etc... YOU becomes the manager.
We will see probably a lot more smaller teams/single devs making bigger projects, until they grow.
The companies that think they can fire devs, are the same companies that are going to go too far, and burn bridges. Do not forget that a lot of companies are founded on devs leaving a company, and starting out on their own, taking clients with them!
I did that years ago, and it worked for a while but eventually the math does not work out because one guy can only do so much. And when you start hiring, your costs balloon. But with LLMs ... Now your a one man team, ... hiring a second person is not hiring a person to make some graphics or doing more coding. Your hiring another team.
This is what people do not realize... they look too much upon this as the established order, ignoring what those fired devs now can do!
Do you really want to be a middle manager to a bunch of text boxes, churning out slop, while they drive up our power bills and slowly terraform the planet?
Ironically, that race to the bottom is no different then we already have. Have you already worked for a company before? A lot of software is developed, BADLY. I dare to say that a lot of software that Opus 4.5 generates, is often a higher quality then what i have seen in my 25 year carrier.
The amount of companies that cheapen out, hiring juniors fresh from school, to work as coding monkies is insane. Then projects have bugs / security issues, with tons of copy/pasted code, or people not knowing a darn thing.
Is that any different then your feared future? I dare to say, that LLms like Opus are frankly better then most juniors. As a junior to do a code review for security issues. Opus literally creates extensive tests, points out issues that you expect from a mid or higher level dev. Of course, you need to know to ask! You are the manager.
> Do you really want to be a middle manager to a bunch of text boxes, churning out slop, while they drive up our power bills and slowly terraform the planet?
Frankly, yes ... If you are a real developer, do you still think development is fun after 10 years, 20 years? Doing the exact same boring work. Reimplementing the 1001 login page, the 101 contact form ... A ton of our work is in reality repeating the same crap over and over again. And if we try to bypass it, we end up tied to tied to those systems / frameworks that often become a block around our necks.
Our industry has a lot of burnout because most tasks may start small but then grow beyond our scope. Todays its ruby on rails programming, then its angular, no wait, react, no wait, Vue, no wait, the new hotness is whatever again.
> slowly terraform the planet?
Well, i am actually making something.
Can you say the same for all the power / gpu draw with bitcoin, Ethereum whatever crap mining. One is productive, a tool with insane potential and usage, the other is a virtual currency where only one is ever popular with limited usage. Yet, it burns just as much for a way more limited return of usability.
Those LLMs that you are so against, make me a ton more productive. You wan to to try out something, but never really wanted to get committed because it was weeks of programming. Well, now you as manager, can get projects done fast. Learn from them way faster then your little fingers ever did.
Just like farmers who couldn't cope with the additional leverage their equipment provided them, devs who can't leverage this technology will have to "go to the cities".
If LLM's turn out the way C-Suite hopes. Let me tell you, you will be in a world of pain. Most of you won't be using LLM's to create your own businesses.
Seeing the replies here it actually doesn't seem like everyone else can do this. Looks like a lot of people really suck at using LLMs to me.
codex cli even has a skill to create skills; it's super easy to get up to speed with them
https://github.com/openai/skills/blob/main/skills/.system/sk...
Disclaimer: I don't have access to Claude Code. My employer has only granted me Claude Teams. Supposedly, they don't use my poopy code to train their models if I use my work email Claude so I am supposed to use that. If I'm not pasting code (asking general questions) into Claude, I believe I'm allowed to use whatever.
pardon my ignorance.
I use GitHub Copilot which has access to llms like Gemini 3, Sonnet/Opus 4.5 ang GPT 5.2
And I get it. Coding with Claude Code really was prompting something, getting errors, and asking it to fix it. Which was still useful but I could see why a skilled coder adding a feature to a complex codebase would just give up
Opus 4.5 really is at a new tier however. It just...works. The errors are far fewer and often very minor - "careless" errors, not fundamental issues (like forgetting to add "use client" to a nextjs client component.
There's a lesser chance that you're working on a code base that Claude Code just isn't capable of helping with.
With the latest models if you're clear enough with your requirements you'll usually find it does the right thing on the first try.
It turns out that Waterfall was always the correct method, it's just really slow ;)
1) There exists a threshold, only identifiable in retrospect, past which it would have been faster to locate or write the code yourself than to navigate the LLM's correction loop or otherwise ensure one-shot success.
2) The intuition and motivations of LLMs derive from a latent space that the LLM cannot actually access. I cannot get a reliable answer on why the LLM chose the approaches it did; it can only retroactively confabulate. Unlike human developers who can recall off-hand, or at least review associated tickets and meeting notes to jog their memory. The LLM prompter always documenting sufficiently to bridge this LLM provenance gap hits rub #1.
3) Gradually building prompt dependency where one's ability to take over from the LLM declines and one can no longer answer questions or develop at the same velocity themselves.
4) My development costs increasingly being determined by the AI labs and hardware vendors they partner with. Particularly when the former will need to increase prices dramatically over the coming years to break even with even 2025 economics.
My experience scrolling X and HN is a bunch of people going "omg opus omg Claude Code I'm 10x more productive" and that's it. Just hand wavy anecdotes based on their own perceived productivity. I'm open to being convinced but just saying stuff is not convincing. It's the opposite, it feels like people have been put under a spell.
I'm following The Primeagen, he's doing a series where he is trying these tools on stream and following peoples advice on how to use them the best. He's actually quite a good programmer so I'm eager to see how it goes. So far he isn't impressed and thus neither am I. If he cracks it and unlocks significant productivity then I will be convinced.
Simon has produced plenty of evidence over the past year. You can check their submission history and their blog: https://simonwillison.net/
The problem with people asking for evidence is that there's no level of evidence that will convince them. They will say things like "that's great but this is not a novel problem so obviously the AI did well" or "the AI worked only because this is a greenfield project, it fails miserably in large codebases".
Nobody serious is disputing that LLM's can generate working code. They dispute claims like "Agentic workflows will replace software developers in the short to medium term", or "Agentic workflows lead to 2-100x improvements in productivity across the board". This is what people are looking for in terms of evidence and there just isn't any.
Thus far, we do have evidence that AI (at least in OSS) produces a 19% decrease in productivity [0]. We also have evidence that it harms our cognitive abilities [1]. Anecdotally, I have found myself lazily reaching for LLM assistance when encountering a difficult problem instead of thinking deeply about the problem. Anecdotally I also struggle to be more productive using AI-centric agents workflows in areas of expertise.
We want evidence that "vibe engineering" is actually more productive across the entire lifespan of a software project. We want evidence that it produces better outcomes. Nobody has yet shown that. It's just people claiming that because they vibe coded some trivial project, all of software development can benefit from this approach. Recently a principle engineer at Google claimed that Claude Code wrote their team's entire year's worth of work in a single afternoon. They later walked that claim back, but most do not.
I'm more than happy to be convinced but it's becoming extremely tiring to hear the same claims being parroted without evidence and then you get called a luddite when you question it. It's also tiring when you push them on it and they blame it on the model you use, and then the agent, and then the way you handle context, and then the prompts, and then "skill issue". Meanwhile all they have to show is some slop that could be hand coded in a couple hours by someone familiar with the domain. I use AI, I was pretty bullish on it for the last two years, and the combination of it simply not living up to expectations + the constant barrage of what feels like a stealth marketing campaign parroting the same thing over and over (the new model is way better, unlike the other times we said that) + the amount of absolute slop code that seems to continue to increase + companies like Microsoft producing worse and worse software as they shoehorn AI into every single product (Office was renamed to Copilot 365). I've become very sensitive to it, much in the same way I was very sensitive to the claims being made by certain VC backed webdev companies regarding their product + framework in the last few years.
I'm not even going to bring up the economic, social, and environmental issues because I don't think they're relevant, but they do contribute to my annoyance with this stuff.
[0] https://metr.org/blog/2025-07-10-early-2025-ai-experienced-o... [1] https://news.harvard.edu/gazette/story/2025/11/is-ai-dulling...
I generally agree with you, but I'd be remiss if I didn't point out that it's plausible that the slow down observed in the METR study was at least partially due to the subjects lack of experience with LLMs. Someone with more experience performed the same experiment on themselves, and couldn't find a significant difference between using LLMs and not [0]. I think the more important point here is that programmers subjective assessment of how much LLMs help them is not reliable, and biased towards the LLMs.
[0] https://mikelovesrobots.substack.com/p/wheres-the-shovelware...
Using AI for me absolutely makes it feel like I'm more productive. When I look back on my work at the end of the day and look at what I got done, it would be ludicrous to say it was multiple times the amount as my output pre-AI
Despite all the people replying to me saying "you're holding it wrong" I know the fix to it doing the wrong thing. Specify in more detail what I want. The problem with that is twofold:
1. How much to specify? As little as possible is the ideal, if we want to maximize how much it can help us. A balance here is key. If I need to detail every minute thing I may as well write the code myself
2. If I get this step wrong, I still have to review everything, rethink it, go back and re-prompt, costing time
When I'm working on production code, I have to understand it all to confidently commit. It costs time for me to go over everything, sometimes multiple iterations. Sometimes the AI uses things I don't know about and I need to dig into it to understand it
AI is currently writing 90% of my code. Quality is fine. It's fun! It's magical when it nails something one-shot. I'm just not confident it's faster overall
This is why you use this AI bubble (it IS a bubble) to use the VC-funded AI models for dirt cheap prices and CREATE tools for yourself.
Need a very specific linter? AI can do it. Need a complex Roslyn analyser? AI. Any kind of scripting or automation that you run on your own machine. AI.
None of that will go away or suddenly stop working when the bubble bursts.
Within just the last 6 months I've built so many little utilities to speed up my work (and personal life) it's completely bonkers. Most went from "hmm, might be cool to..." to a good-enough script/program in an evening while doing chores.
Even better, start getting the feel for local models. Current gen home hardware is getting good enough and the local models smart enough so you can, with the correct tooling, use them for suprisingly many things.
Let's say AI becomes too expensive - I more or less only have to sharpen up being able to write the language. My active recall of the syntax, common methods and libraries. That's not hard or much of a setback
Maybe this would be a problem if you're purely vibe coding, but I haven't seen that work long term
I'm building tools, not complete factories :) The AI builds me a better hammer specifically for the nails I'm nailing 90% of the time. Even if the AI goes away, I still know how the custom hammer works.
Are there any local models that are at least somewhat comparable to the latest-and-greatest (e.g. Opus 4.5, Gemini 3), especially in terms of coding?
I can run multiple agents at once, across multiple code bases (or the same codebase but multiple different branches), doing the same or different things. You absolutely can't keep up with that. Maybe the one singular task you were working on, sure, but the fact that I can work on multiple different things without the same cognitive load will blow you out of the water.
> 2) The intuition and motivations of LLMs derive from a latent space that the LLM cannot actually access. I cannot get a reliable answer on why the LLM chose the approaches it did; it can only retroactively confabulate. Unlike human developers who can recall off-hand, or at least review associated tickets and meeting notes to jog their memory. The LLM prompter always documenting sufficiently to bridge this LLM provenance gap hits rub #1.
Tell the LLM to document in comments why it did things. Human developers often leave and then people with no knowledge of their codebase or their "whys" are even around to give details. Devs are notoriously terrible about documentation.
> 3) Gradually building prompt dependency where one's ability to take over from the LLM declines and one can no longer answer questions or develop at the same velocity themselves.
You can't develop at the same velocity, so drop that assumption now. There's all kinds of lower abstractions that you build on top of that you probably can't explain currently.
> 4) My development costs increasingly being determined by the AI labs and hardware vendors they partner with. Particularly when the former will need to increase prices dramatically over the coming years to break even with even 2025 economics.
You aren't keeping up with the actual economics. This shit is technically profitable, the unprofitable part is the ongoing battle between LLM providers to have the best model. They know software in the past has often been winner takes all so they're all trying to win.
That's great that this is your experience, but it's not a lot of people's. There are projects where it's just not going to know what to do.
I'm working in a web framework that is a Frankenstein-ing of Laravel and October CMS. It's so easy for the agent to get confused because, even when I tell it this is a different framework, it sees things that look like Laravel or October CMS and suggests solutions that are only for those frameworks. So there's constant made up methods and getting stuck in loops.
The documentation is terrible, you just have to read the code. Which, despite what people say, Cursor is terrible at, because embeddings are not a real way to read a codebase.
One trick I use that might work for you as well:
Clone GitHub.com/simonw/datasette to /tmp
then look at /tmp/docs/datasette for
documentation and search the code
if you need to
If your framework is missing documentation tell Claude Code to write itself some documentation based on what it learns from reading the code!
Potentially because there is no baggage with similar frameworks. I'm sure it would have an easier time with this if it was not spun off from other frameworks.
> If your framework is missing documentation tell Claude Code to write itself some documentation based on what it learns from reading the code!
If Claude cannot read the code well enough to begin with, and needs supplemental documentation, I certainly don't want it generating the docs from the code. That's just compounding hallucinations on top of each other.
Putting “Make minimal changes” in my standard prompt helped a lot with the tendency of basically all agents to make too many changes at once. With that addition it became possible to direct the LLM to make something similar to the logical progression of commits I would have made anyway, but now don’t have to work as hard at crafting.
Most of the hype merchants avoid the topic of maintainability because they’re playing to non-technical management skeptical of the importance of engineering fundamentals. But everything I’ve experienced so far working with LLMs screams that the fundamentals are more important than ever.
It's an opportunity, not a problem. Because it means there's a gap in your specifications and then your tests.
I use Aider not Claude but I run it with Anthropic models. And what I found is that comprehensively writing up the documentation for a feature spec style before starting eliminates a huge amount of what you're referring to. It serves a triple purpose (a) you get the documentation, (b) you guide the AI and (c) it's surprising how often this helps to refine the feature itself. Sometimes I invoke the AI to help me write the spec as well, asking it to prompt for areas where clarification is needed etc.
I even gave -True Vibe Coding- a whirl. Yesterday, from a blank directory and text file list of requirements, I had Opus 4.5 build an Android TV video player that could read a directory over NFS, show a grid view of movie poster thumbnails, and play the selected video file on the TV. The result wasn't exactly full-featured Kodi, but it works in the emulator and actual device, it has no memory leaks, crashes, ANRs, no performance problems, no network latency bugs or anything. It was pretty astounding.
Oh, and I did this all without ever opening a single source file or even looking at the proposed code changes while Opus was doing its thing. I don't even know Kotlin and still don't know it.
but I do want a better way to glance and keep up with what its doing in longer conversations, for my own mental context window
I'm finding that branding and graphic design is the most arduous part, that I'm hoping to accelerate soon. I'm heavily AI assisted there too and I'm evaluating MCP servers to help, but so far I do actually have to focus on just that part as opposed to babysit
For instance, I always respected types, but I'm too lazy to go spend hours working on types when I can just do ruby-style duck typing and get a long ways before the inevitable problems rear their head. Now, I can use a strongly typed language and get the advantages for "free".
I find I'm doing more Typescript projects than Python because of the superior typing, despite the fact I prefer Python.
Oh absolutely. I've been using Python for past 15 or so years for everything.
I've never written a single line of Rust in my life, and all my new projects are Rust now, even the quick-script-throwaway things, because it's so much better at instantly screaming at claude when it goes off track. It may take it longer to finish what I asked it to do, but requires so much less involvement from me.
I will likely never start another new project in python ever.
EDIT: Forgot to add that paired with a good linter, this is even more impressive. I told Claude to come up with the most masochistic clippy configuration possible, where even a tiny mistake is instantly punished and exceptions have to be truly exceptional (I have another agent that verifies this each run).
I just wish there was cargo-clippy for enforcing architectural patterns.
... says it all.
If you were someone saying at the start of 2025 "this is a flash in the pan and a bunch of hype, it's not going to fundamentally change how we write code", that was still a reasonable belief to hold back then. At the start of 2026 that position is basically untenable: it's just burying your head in the sand and wishing for AI to go away. If you're someone who still holds it you really really need to download Claude Code and set it to Opus and start trying it with an open mind: I don't know what else to tell you. So now the question has shifted from whether this is going to transform our profession (it is), to how exactly it's going to play out. I personally don't think we will be replacing human engineers anytime soon ("coders", maybe!), but I'm prepared to change my mind on that too if the facts change. We'll see.
I was a fellow mind-changer, although it was back around the first half of last year when Claude Code was good enough to do things for me in a mature codebase under supervision. It clearly still had a long way to go but it was at that tipping point from "not really useful" to "useful". But Opus 4.5 is something different - I don't feel I have to keep pulling it back on track in quite the way I used to with Sonnet 3.7, 4, even Sonnet 4.5.
For the record, I still think we're in a bubble. AI companies are overvalued. But that's a separate question from whether this is going to change the software development profession.
I'm hoping after AI comes back down to earth there will be a new glut of cheap second hand GPUs and RAM to get snapped up.
I think my comparison is apt; being a bubble and a truly society-altering technology are not mutually exclusive, and by virtue of it being a bubble, it is overhyped.
Going to one-up you though -- here's a literal one-liner that gets me a polished media center with beautiful interface and powerful skinning engine. It supports Android, BSD, Linux, macOS, iOS, tvOS and Windows.
`git clone https://github.com/xbmc/xbmc.git`
A big part of the problem with existing software is that humans seem to be pretty much incapable of deciding a project is done and stop adding to it. We treat creating code like a job or hobby instead of a tool. Nothing wrong with that, unless you're advertising it as a tool.
The old adage about how "users use 10% of your software's features, but they each use a different 10%" can now be solved by each user just building that 10% for themselves.
At work I only have access to calude using the GitHub copilot integration so this could be the cause of my problems. Claude was able to get slthe first iteration up pretty quick. At that stage the app could create a plot and you could interact with it and ask follow up questions.
Then I asked it to extend the app so that it could generate multiple plots and the user could interact with all of them one at a time. It made a bunch of changes but the feature was never implemented. I asked it to do again but got the same outcome. I completely accept the fact that it could just be all because I am using vscode copilot or my promoting skills are not good but the LLM got 70% of the way there and then completely failed
You really need to at least try Claude Code directly instead of using CoPilot. My work gives us access to CoPilot, Claude Code, and Codex. CoPilot isn’t close to the other more agentic products.
Literally tried it yesterday. I didn't see a single difference with whatever model Claude Code was using two months ago. Same crippled context window. Same "I'll read 10 irrelevant lines from a file", same random changes etc.
Yup. There's some magical "right context" that will fix all the problems. What is that right context? No idea, I guess I need to read a yet-another 20 000-word post describing magical incantations that you should or shouldn't do in the context.
The "Opus 4.5 is something else/nex tier/just works" claims in my mind means that I wouldn't need to babysit its every decision, or that it would actually read relevant lines from relevant files etc. Nope. Exact same behaviors as whatever the previous model was.
Oh, and that "200k tokens context window"? It's a lie. The quality quickly degrades as soon as Claude reaches somewhere around 50% of the context window. At 80+% it's nearly indistinguishable from a model from two years ago. (BTW, same for Codex/GPT with it's "1 million token window")
It's in your interest to deal with your frustration and figure out how you can leverage the new tools to stay relevant (to the degree that you want to).
Regarding the context window, Claude needs thinking turned up for long context accuracy, it's quite forgetful without thinking.
No one attacked your use of AI. I explained my own experience with the "Claude Opus 4.5 is next tier". You barged in, ignored anything I said, and attacked my skills.
> the workplace is going to punish people who don't leverage AI though, and I'm trying to be helpful.
So what exactly is helpful in your comments?
As for the snide and patronizing "it's in your interest to stay relevant":
1. I use these tools daily. That's why I don't subscribe to willful wide-eyed gullibility. I know exactly what these tools can and cannot do.
The vast majority of "AI skeptics" are the same.
2. In a few years when the world is awash in barely working incomprehensible AI slop my skills will be in great demand. Not because I'm an amazing developer (I'm not), but because I have experience separating wheat from the chaff
It seems the subject of AI is emotionally charged for you, so I expect friendly/rational discourse is going to be a challenge. I'd say something nice but since you're primed to see me being patronizing... Fuck you? That what you were expecting?
It's not me who decided to barge in, assume their opponent doesn't use something or doesn't want to use something, and offer unsolicited advice.
> It kinda makes me sad when the discourse is so poisoned that I can't even encourage someone to protect their own future from something that's obviously coming
See. Again. You're so in love with your "wisdom" that you can't even see what you sound like: snide, patronising, condenscending. And completely missing the whole point of what was written. You are literally the person who poisons the discourse.
Me: "here are the issues I still experience with what people claim are 'next tier frontier model'"
You: "it's in your interests to figure out how to leverage new tools to stay relevant in the future"
Me: ... what the hell are you talking about? I'm using these tools daily. Do you have anything constructive to add to the discourse?
> so I expect friendly/rational discourse is going to be a challenge.
It's only challenge to you because you keep being in love with your voice and your voice only. Do you have anything to contribute to the actual rational discourse, are you going to attack my character?
> 'd say something nice but since you're primed to see me being patronizing... Fuck you? T
Ah. The famous friendly/rational discourse of "they attack my use of AI" (no one attacked you), "why don't you invest in learning tools to stay relevant in the future" (I literally use these tools daily, do you have anything useful to say?) and "fuck you" (well, same to you).
> That what you were expecting?
What I was expecting is responses to what I wrote, not you riding in on a high horse.
If you want to take politeness as being patronizing, I'm happy to stop bothering. My guess is you're not a special snowflake, and you need to "get good" or you're going to end up on unemployment complaining about how unfair life is. I'd have sympathy but you don't seem like a pleasant human being to interact with, so have fun!
They are not giving me the results people claim they give. It is distinctly different from not giving the results I want.
> If you're using these tools daily and having a hard time, either you're working on something very different from the bulk of people using the tools and your problems or legitimate, or you aren't and it's a skill issue.
Indeed. And your rational/friendly discourse that you claim you're having would start with trying to figure that out. Did you? No, you didn't. You immediately assumed your opponent is a clueless idiot who is somehow against AI and is incapable or learning or something.
> If you want to take politeness as being patronizing, I'm happy to stop bothering.
No. It's not politeness. It's smugness. You literally started your interaction in this thread with a "git gud or else" and even managed to complain later that "you dislike it when they attack your use of AI as a skill issue". While continuously attacking others.
> you don't seem like a pleasant human being to interact with
Says the person who has contributed nothing to the conversation except his arrogance, smugness, holier-than-thou attitude, engaged in nothing but personal attacks, complained about non-existent grievances and when called out on this behavior completed his "friendly and rational discourse" with a "fuck you".
Well, fuck you, too.
Adieu.
1) define problem
2) split problem into small independently verifiable tasks
3) implement tasks one by one, verify with tools
With an LLM usually 1) is just a markdown file, 2) is a markdown checklist, Github issues (which Claude can use with the `gh` cli) and every loop of 3 gets a fresh context, maybe the spec from step 1 and the relevant task information from 2
I haven't ran into context issues in a LONG time, and if I have it's usually been either intentional (it's a problem where compacting wont' hurt) or an error on my part.
> I haven't ran into context issues in a LONG time
Because you've become the reverse centaur :) "a person who is serving as a squishy meat appendage for an uncaring machine." [1]
You are very aware of the exact issues I'm talking about, and have trained yourself to do all the mechanical dance moves to avoid them.
I do the same dances, that's why I'm pointing out that they are still necessary despite the claims of how model X/Y/Z are "next tier".
[1] https://doctorow.medium.com/https-pluralistic-net-2025-12-05...
The exact same method that worked for those happened to also work for LLMs, I didn't have to learn anything new or change much in my workflow.
"Fix bug in FoobarComponent" is enough of a bug ticket for the 100x developer in your team with experience with that specific product, but bad for AI, juniors and offshored teams.
Thus, giving enough context in each ticket to tell whoever is working on it where to look and a few ideas what might be the root cause and how to fix it is kinda second nature to me.
Also my own brain is mostly neurospicy mush, so _I_ need to write the context to the tickets even if I'm the one on it a few weeks from now. Because now-me remembers things, two-weeks-from-now me most likely doesn't.
Create a markdown document of your task (or use CLAUDE.md), put it in "plan mode" which allows Claude to use tool calls to ask questions before it generates the plan.
When it finishes one part of the plan, have it create a another markdown document - "progress.md" or whatever with the whole plan and what is completed at that point.
Type /clear (no more context window), tell Claude to read the two documents.
Repeat until even a massive project is complete - with those 2 markdown documents and no context window issues.
... Proceeds to explain how it's crippled and all the workarounds you have to do to make it less crippled.
No - that's not what I did.
You don't need an extra-long context full of irrelevant tokens. Claude doesn't need to see the code it implemented 40 steps ago in a working method from Phase 1 if it is on Phase 3 and not using that method. It doesn't need reasoning traces for things it already "thought" through.
This other information is cluttering, not helpful. It is making signal to noise ratio worse.
If Claude needs to know something it did in Phase 1 for Phase 4 it will put a note on it in the living markdown document to simply find it again when it needs it.
What I very succinctly called "crippled context" despite claims that Opus 4.5 is somehow "next tier". It's all the same techniques we've been using for over a year now.
And even my short-term memory is significantly larger than the at most 50% of the 200k-token context window that Claude has. It runs out of context before my short-term memory is probably not even 1% full, for the same task (and I'm capable of more context-switching in the meantime).
And so even the "Opus 4.5 really is at a new tier" runs into the very same limitations all models have been running into since the beginning.
For LLMs long term memory is achieved by tooling. Which you discounted in your previous comments.
You also overstimate capacity of your short-term memory by few orders of magnitude:
https://my.clevelandclinic.org/health/articles/short-term-me...
My specific complaint, which is an observable fact about "Opus 4.5 is next tier": it has the same crippled context that degrades the quality of the model as soon as it fills 50%.
EMM_386: no-no-no, it's not crippled. All you have to do is keep track across multiple files, clear out context often, feed very specific information not to overflow context.
Me: so... it's crippled, and you need multiple workarounds
scotty79: After all it's the same as your own short-term memory, and <some unspecified tooling (I guess those same files)> provide long-term memory for LLMs.
Me: Your comparison is invalid because I can go have lunch, and come back to the problem at hand and continue where I left off. "Next tier Opus 4.5" will have to be fed the entire world from scratch after a context clear/compact/in a new session.
Unless, of course, you meant to say that "next tier Opus model" only has 15-30 second short term memory, and needs to keep multiple notes around like the guy from Memento. Which... makes it crippled.
Who said I refuse them?
I evaluated the claim that Opus is somehow next tier/something different/amazeballs future at its face value. It still has all the same issues and needs all the same workarounds as whatever I was using two months ago (I had a bit of a coding hiatus between beginning of December and now).
> then you end up with a guy from Memento and regardless of how smart the model is
Those models are, and keep being the guy from memento. Your "long memory" is nothing but notes scribbled everywhere that you have to re-assemble every time.
> And that's why you can't tell the difference between smarter and dumber one while others can.
If it was "next tier smarter" it wouldn't need the exact same workarounds as the "dumber" models. You wouldn't compare the context to the 15-30 second short-term memory and need unspecified tools [1] to have "long-term memory". You wouldn't have the model behave in an indistinguishable way from a "dumber" model after half of its context windows has been filled. You wouldn't even think about context windows. And yet here we are
[1] For each person these tools will be a different collection of magic incantations. From scattered .md files to slop like Beads to MCP servers providing access to various external storage solutions to custom shell scripts to ...
BTW, I still find "superpowers" from https://github.com/obra/superpowers to be the single best improvement to Claude (and other providers) even if it's just another in a long serious of magic chants I've evaluated.
Are you sure you're not an LLM?
2. It's the same workarounds we've been doing forever
3. It's indistinguishable from "clear context and re-feed the entire world of relevant info from scratch" we've had forever, just slightly more automated
That's why I don't understand all the "it's new tier" etc. It's all the same issues with all the same workarounds.
It's not just the deficiencies of earlier versions, but the mismatch between the praise from AI enthusiasts and the reality.
I mean maybe it is really different now and I should definitely try uploading all of my employer's IP on Claude's cloud and see how well it works. But so many people were as hyped by GPT-4 as they are now, despite GPT-4 actually being underwhelming.
Too much hype for disappointing results leads to skepticism later on, even when the product has improved.
The funny part about rapidly changing industries is that, despite the fomo, there's honestly not any reward to keeping up unless you want to be a consultant. Otherwise, wait and see what sticks. If this summer people are still citing the Opus 4.5 was a game changing moment and have solid, repeatable workflows, then I'll happily change up my workflow.
Someone could walk into the LLM space today and wouldn't be significantly at a loss for not having paid attention to anything that had happened in the last 4 years other than learning what has stuck since then.
LMAO what???
Two big examples:
- Period from early mvc JavaScript frontends (backbone.js etc) and the time of the great React/Angular wars. I completely stepped out of the webdev space during that time period.
- The rapid expansion of Deep Learning frameworks where I did try to keep up (shipped some Lua torch packages and made minor contributions to Pylearn2).
In the first case, missing 5 years of front-end wars had zero impact. After not doing webdev work at all for 5-years I was tasked with shipping a React app. It took me a week to catch up, and everything was deployed in roughly the same time as someone would have had they spent years keeping up with changes.
In the second case, where I did keep up with many of the developing deep learning frameworks, it didn't really confer any advantage. Coworkers who I worked with who started with Pytorch fresh out of school were just as proficient, if not more so, with building models. Spending energy keeping up offered no value other than feeling "current" at the time.
Can you give me a counter example of where keeping up with a rapidly changing area that's unstable has conferred a benefit to you? Most of FOMO is really just fear. Again, unless you're trying to sell your self specifically as a consultant on the bleeding edge, there's no reason to keep up with all these changes (other than finding it fun).
Right or wrong - doesn’t matter. You typed in a line of text and now your computer is making 3000 word stories, images, even videos based on it
How are you NOT astounded by that? We used to have NONE of this even 4 years ago!
-- Charles Babbage
My team is using it with Claude Code and say it works brilliantly, so I'll be giving it another go.
How much of the value comes from Opus 4.5, how much comes from Claude Code, and how much comes from the combination?
Copilot has by far the best and most intuitive agent UI. Just make sure you're in agent mode and choose Sonnet or Opus models.
I've just cancelled my Claude sub and gone back and will upgrade to the GH Pro+ to get more sonnet/opus.
The gap didn't seem big, but in November (which admittedly was when Opus 4.5 was in preview on Copilot) Opus 4.5 with Copilot was awful.
This is what people are still doing wrong. Tools in a loop people, tools in a loop.
The agent has to have the tools to detect whatever it just created is producing errors during linting/testing/running. When it can do that, I can loop again, fix the error and again - use the tools to see whether it worked.
I _still_ encounter people who think "AI programming" is pasting stuff into ChatGPT on the browser and they complain it hallucinates functions and produces invalid code.
Well, d'oh.
I was having Opus investigate it and I kept building and deploying the firmware for testing.. then I just figured I'd explain how it could do the same and pull the logs.
Off it went, for the next ~15 minutes it would flash the firmware multiple times until it figured out the issue and fixed it.
There was something so interesting about seeing a microcontroller on the desk being flashed by Claude Code, with LEDs blinking indicating failure states. There's something about it not being just code on your laptop that felt so interesting to me.
But I agree, absolutely, red/green test or have a way of validating (linting, testing, whatever it is) and explain the end-to-end loop, then the agent is able to work much faster without being blocked by you multiple times along the way.
While Claude is amazing at writing code, it still requires human operators. And even experienced human operators are bad at operating this machinery.
Tell your average joe - the one who thinks they can create software without engineers - what "tools-in-a-loop" means, and they'll make the same face they made when you tried explaining iterators to them, before LLMs.
Explain to them how typing system, E2E or integration test helps the agent, and suddenly, they now have to learn all the things they would be required to learn to be able to write on their own.
Key word: seems. It's impossible to do a proper qualitative analysis.
Until such time, of course, when LLMs are eating their own dogfood, in which case they - as has already happened - create their own language, evolve dramatically, and cue skynet.
Likewise, many optimization techniques involve some randomness, whether it's approximating an NP-thorny subproblem, or using PGO guided by statistical sampling. People might disable those in pursuit of reproducible builds, but no one would claim that enabling those features makes GCC or LLVM no longer a compiler. So nondeterminism isn't really the distinguishing factor either.
> have it learn your conventions, pull in best practices
What do you mean by "have it learn your conventions"? Is there a way to somehow automatically extract your conventions and store it within CLAUDE.md?
> For example, we have a custom UI library, and Claude Code has a skill that explains exactly how to use it. Same for how we write Storybooks, how we structure APIs, and basically how we want everything done in our repo. So when it generates code, it already matches our patterns and standards out of the box.
Did you have to develop these skills yourself? How much work was that? Do you have public examples somewhere?
From the docs (https://hexdocs.pm/ash/what-is-ash.html):
"Model your application's behavior first, as data, and derive everything else automatically. Ash resources center around actions that represent domain logic."
I’ve even found it searching node_modules to find the API of non-public libraries.
Claude Code will spawn sub-agents (that often use their cheap Haiki model) for exploration and planning tasks, with only the results imported into the main context.
I've found the best results from a more interactive collaboration with Claude Code. As long as you describe the problem clearly, it does a good job on small/moderate tasks. I generally set two instances of Claude Code separate tasks and run them concurrently (the interaction with Claude Code distracts me too much to do my own independent coding simultaneously like with setting a task for a colleague, but I do work on architecture / planning tasks)
The one manner of taste that I have had to compromise on is the sheer amount of code - it likes to write a lot of code. I have a better experience if I sweat the low-level code less, and just periodically have it clean up areas where I think it's written too much / too repetitive code.
As you give it more freedom it's more prone to failure (and can often get itself stuck in a fruitless spiral) - however as you use it more you get a sense of what it can do independently and what's likely to choke on. A codebase with good human-designed unit & playwright tests is very good.
Crucially, you get the best results where your tasks are complex but on the menial side of the spectrum - it can pay attention to a lot of details, but on the whole don't expect it to do great on senior-level tasks.
To give you an idea, in a little over a month "npx ccusage" shows that via my Claude Code 5x sub I've used 5M input tokens, 1.5M output, 121M Cache Create, 1.7B Cache Read. Estimated pay-as-you-go API cost equivalent is $1500 (N.B. for the tail end of December they doubled everybody's API limits, so I was using a lot more tokens on more experimental on-the-fly tool construction work)
I exclusively use opus for architecture / speccing, and then mostly Sonnet and occasionally Haiku to write the code. If my usage has been light and the code isn't too straightforward, I'll have Opus write code as well.
I'll give you an example: I use ruff to format my python code, which has an opinionated way of formatting certain things. After an initial formatting, Opus 4.5, without prompting, will write code in this same style so that the ruff formatter almost never has anything to do on new commits. Sonnet 4.5 is actually pretty good at this too.
I care about the norms in my codebase that can't be automatically enforced by machine. How is state managed? How are end-to-end tests written to minimize change detectors? When is it appropriate to log something?
We have some tests in "GIVEN WHEN THEN" style, and others in other styles. Opus will try to match each style of testing by the project it is in by reading adjacent tests.
"AI has X"
"We have X at home"
"X at home: x"
But I think it should be doable. You can tell it how YOU want the state to be managed and then have it write a custom "linter" that makes the check deterministic. I haven't tried this myself, but claude did create some custom clippy scripts in rust when I wanted to enforce something that isn't automatically enforced by anything out there.
I had an idea for something that I wanted, and in five scattered hours, I got it good enough to use. I'm thinking about it in a few different ways:
1. I estimate I could have done it without AI with 2 weeks full-time effort. (Full-time defined as >> 40 hours / week.)
2. I have too many other things to do that are purportedly more important that programming. I really can't dedicate to two weeks full-time to a "nice to have" project. So, without AI, I wouldn't have done it at all.
3. I could hire someone to do it for me. At the university, those are students. From experience with lots of advising, a top-tier undergraduate student could have achieved the same thing, had they worked full tilt for a semester (before LLMs). This of course assumes that I'm meeting them every week.
FYI, I still use cursor for small edits and reviews.
Seriously, I have 3+ claude code windows open at a time. Most days I don't even look at the IDE. It's still there running in the background, but I don't need to touch it.
- single scripts. Anything which can be reduced to a single script.
- starting greenfield projects from scratch
- code maintenance (package upgrades, old code...)
- tasks which have a very clear and single definition. This isn't linked to complexity, some tasks can be both very complex but with a single definition.
If your work falls into this list they will do some amazing work (and yours clearly fits that), if it doesn't though, prepare yourself because it will be painful.
I should say I'm hardly ever vibe-coding, unlike the original article. If I think I want code that will last, I'll steer the models in ways that lean on years of non-LLM experience. E.g., I'll reject results that might work if they violate my taste in code.
It also helps that I can read code very fast. I estimate I can read code 100x faster than most students. I'm not sure there is any way to teach that other than the old-fashioned way, which involves reading (and writing) a lot of code.
Yes indeed, these are the things on the other hand which aren't working well in my opinion:
- large codebase
- complex domain knowledge
- creating any feature where you need product insights
- tasks requiring choices (again, complexity doesn't matter here, the task may be simple but require some choices)
- anything unclear where you don't know where you are going first
While you don't experience any of these when teaching or side projects, these are very common in any enterprise context.
I then proceeded to use it to hack on its own codebase, and close a bunch of issues in a repository that I maintain (https://github.com/nuprl/MultiPL-E/commits/main/).
When needed it can be picked up in a day. Otherwise they are not paid based in tickets solved etc. If the incentives were properly aligned everyone would already use it
Reword? But why not just voice to text alone...
Oh but we all read the partially synthetic ad by this point. Psyche.
Nobody is sleeping. I'm using LLMs daily to help me in simple coding tasks.
But really where is the hurry? At this point not a few weeks go by without the next best thing since sliced bread to come out. Why would I bother "learning" (and there's really nothing to learn here) some tool/workflow that is already outdated by the time it comes out?
> 2026 is going to be a wake-up call
Do you honestly think a developer not using AI won't be able to adapt to a LLM workflow in, say, 2028 or 2029? It has to be 2026 or... What exactly?
There is literally no hurry.
You're using the equivalent of the first portable CD-player in the 80s: it was huge, clunky, had hiccups, had a huge battery attached to it. It was shiny though, for those who find new things shiny. Others are waiting for a portable CD player that is slim, that buffers, that works fine. And you're saying that people won't be able to learn how to put a CD in a slim CD player because they didn't use a clunky one first.
It's also a skill that compounds over time, so if you have two years of experience with them you'll be able to use them more effectively than someone with two months of experience.
In that respect, they're just normal technology. A Python programmer with two years of Python experience will be more effective than a programmer with two months of Python.
Codex/Claude Code are terrible with C++. It also can't do Rust really well, once you get to the meat of it. Not sure why that is, but they just spit out nonsense that creates more work than it helps me. It also can't one shot anything complete, even though I might feed him the entire paper that explains what the algorithm is supposed to do.
Try to do some OpenGL or Vulkan with it, without using WebGPU or three.js. Try it with real code, that all of us have to deal with every day. SDL, Vulkan RHI, NVRHI. Very frustrating.
Try it with boost, or cmake, or taskflow. It loses itself constantly, hallucinates which version it is working on and ignores you when you provide actual pointers to documentation on the repo.
I've also recently tried to get Opus 4.5 to move the Job system from Doom 3 BFG to the original codebase. Clean clone of dhewm3, pointed Opus to the BFG Job system codebase, and explained how it works. I have also fed it the Fabien Sanglard code review of the job system: https://fabiensanglard.net/doom3_bfg/threading.php
We are not sleeping on it, we are actually waiting for it to get actually useful. Sure, it can generate a full stack admin control panel in JS for my PostgreSQL tables, but is that really "not normal"? That's basic.
If you have any opensource examples of your codebase, prompt, and/or output, I would happily learn from it / give advice. I think we're all still figuring it out.
Also this SIMD translation wasn't just a single function - it was multiple functions across a whole region of the codebase dealing with video and frame capture, so pretty substantial.
That's a good way to say it, I totally identify.
I'm not talking about "write this function" but rather like implementing the whole feature by writing only English to the agent, over the course of numerous back-and-forth interactions and exhausting multiple 200K-token context windows.
For me personally, definitely at least 99% all of the Rust code I've committed at work since Opus 4.5 came out has been from an agent running that model. I'm reading lots of Rust code (that Opus generated) but I'm essentially no longer writing any of it. If dot-autocomplete (and LLM autocomplete) disappeared from IDE existence, I would not notice.
And also - doesn’t that make Zed (and other editors) pointless?
Nobody is one-shotting anything nontrivial in Zed's code base, with Opus 4.5 or any other model.
What about a future model? Literally nobody knows. Forecasts about AI capabilities have had horrendously low accuracy in both directions - e.g. most people underestimated what LLMs would be capable of today, and almost everyone who thought AI would at least be where it is today...instead overestimated and predicted we'd have AGI or even superintelligence by now. I see zero signs of that forecasting accuracy improving. In aggregate, we are atrocious at it.
The only safe bet is that hardware will be faster and cheaper (because the most reliable trend in the history of computing has been that hardware gets faster and cheaper), which will naturally affect the software running on it.
> And also - doesn’t that make Zed (and other editors) pointless?
It means there's now demand for supporting use cases that didn't exist until recently, which comes with the territory of building a product for technologists! :)
And yeah - if I had a crystal ball, I would be on my private island instead of hanging on HN :)
It takes more work than one-shot, but not a lot, and it pays dividends.
Are you worried about microsoft stealing your codebase from github?
And even with a narrower definition of stealing, Microsoft’s ability to share your code with US government agencies is a common and very legitimate worry in plenty of threat model scenarios.
The bigger a project gets the more context you generally need to understand any particular part. And by default Claude Code doesn't inject context, you need to use 3rd party integrations for that.
From what I've seen, CC has troubles with the latest Swift too, partially because of it being latest and partially because it's so convoluted nowadays.
But it's übercharged™ for C#
Because types are proofs and require global correctness, you can't just iterate, fix things locally, and wait until it breaks somewhere else that you also have to fix locally.
My rule of thumb is that if you can clearly describe exactly what you want to another engineer, then you can instruct the agent to do it too.
Why do we hold calculators to such high bars? Humans make calculation mistakes all the time.
Why do we hold banking software to such high bars? People forget where they put their change all the time.
Etc etc.
I've worked with Gemini Fast on the web to help design the VM ISA, then next steps will be to have some AI (maybe Gemini CLI - currently free) write an assembler, disassembler and interpreter for the ISA, and then the recursive descent compiler (written in C) too.
I already had Gemini 3.0 Fast write me a precedence climbing expression parser as a more efficient drop-in replacement for a recursive descent one, although I had it do that in C++ as a proof-of-concept since I don't know yet what C libraries I want to build and use (arena allocator, etc). This involved a lot of copy-paste between Gemini output and an online C++ dev environment (OnlineGDB), but that was not too bad, although Gemini CLI would have avoided that. Too bad that Gemini web only has "code interpreter" support for Python, not C and/or C++.
Using Gemini to help define the ISA was an interesting process. It had useful input in a "pair-design" process, working on various parts of the ISA, but then failed to bring all the ideas together into a single ISA document, repeatedly missing parts of what had been previously discussed until I gave up and did that manually. The default persona of Gemini seems not very well suited to this type of work flow where you want to direct what to do next, since it seems they've RL'd the heck out of it to want to suggest next step and ask questions rather than do what is asked and wait for further instruction. I eventually had to keep asking it to "please answer then stop", and interestingly quality of the "conversation" seemed to fall apart after that (perhaps because Gemini was now predicting/generating a more adversarial conversation than a collaborative one?).
I'm wondering/hoping that Gemini CLI might be better at working on documentation than Gemini web, since then the doc can be an actual file it is editing, and it can use it's edit tool for that, as opposed to hoping that Gemini web can assemble chunks of context (various parts of the ISA discussion) into a single document.
I thought I'd revive it, but this time with Vulkan and no third-party dependencies (except for Vulkan)
4.5 Sonet, Opus and Gemini 3.5 flash has helped me write image decoders for dds, png jpg, exr, a wayland window implementation, macOS window implementation, etc.
I find that Gemini 3.5 flash is really good at understanding 3d in general while sonnet might be lacking a little.
All these sota models seem to understand my bespoke Lua framework and the right level of abstraction. For example at the low level you have the generated Vulkan bindings, then after that you have objects around Vulkan types, then finally a high level pipeline builder and whatnot which does not mention Vulkan anywhere.
However with a larger C# codebase at work, they really struggle. My theory is that there are too many files and abstractions so that they cannot understand where to begin looking.
Which makes sense. I'm sure there's lots of training data for React/HTML/CSS/etc. but much less with Swift, especially the newer versions.
Now, was the code quality any good? Beats me, I am not a swift developer. I did it partly as an experiment to see what Claude was currently capable of and partly because I wanted to test the feasibility of setting up a simple passive data logger for my truck.
I'm tempted to take another swing with Opus 4.5 for the science.
With some entirely novel work we're doing, it's actually a hindrance as it consistently tells us the approach isn't valid/won't work (it will) and then enters "absolutely right" loops when corrected.
I still believe those who rave about it are not writing anything I would consider "engineering". Or perhaps it's a skill issue and I'm using it wrong, but I haven't yet met someone I respect who tells me it's the future in the way those running AI-based companies tell me.
For some things you can fire up Claude and have it generate great code from scratch. But for bigger code bases and more complex architecture, you need to break it down ahead of time so it can just read about the architecture rather than analyze it every time.
> Create a CLAUDE.md for a c++ application that uses libraries x/y/z
[Then I edit it, adding general information about the architecture]
> Analyze the library in the xxx directory, and produce a xxx_architecture.md describing the major components and design
> /agent [let claude make the agent, but when it asks what you want it to do, explain that you want it to specialize in subsystem xxx, and refer to xxx_architecture.md
Then repeat until you have the major components covered. Then:
> Using the files named with architecture.md analyze the entire system and update CLAUDE.md to use refer to them and use the specialized agents.
Now, when you need to do something, put it in planning mode and say something like:
> There's a bug in the xxx part of the application, where when I do yyy, it does zzz, but it should do aaa. Analyze the problem and come up with a plan to fix it, and automated tests you can perform if possible.
Then, iterate on the plan with it if you need to, or just approve it.
One of the most important things you can do when dealing with something complex is let it come up with a test case so it can fix or implement something and then iterate until it's done. I had an image processing problem and I gave it some sample data, then it iterated (looking at the output image) until it fixed it. It spent at least an hour, but I didn't have to touch it while it worked.
The major benefit of agents is that it keeps context clean for the main job. So the agent might have a huge context working through some specific code, but the main process can do something to the effect of "Hey UI library agent, where do I need to put code to change the color of widget xyz", then the agent does all the thinking and can reply with "that's in file 123.js, line 200". The cleaner you keep the main context, the better it works.
Skills on the other hand are commands ON STEROIDS. They can be packaged with actual scripts and executables, the PEP723 Python style + uv is super useful.
I have one skill for example that uses Python+Treesitter to check the unit thest quality of a Go project. It does some AST magic to check the code for repetition, stupid things like sleeps and relative timestamps etc. A /command _can_ do it, but it's not as efficient, the scripts for the skill are specifically designed for LLM use and output the result in a hyper-compact form a human could never be arsed to read.
claude-code has a built in plugin that it can use to fetch its own docs! You don't have to ever touch anything yourself, it can add the features to itself, by itself.
It's also hard to steer the plan mode or have it remember some behavior that you want to enforce. It's much better to create a custom command with custom instructions that acts as the plan mode.
My system works like this:
/implement command acts as an orchestrator & plan mode, and it is instructed to launch predefined set of agents based on the problem and have them utilize specific skills. Every time /implement command is initiated, it has to create markdown file inside my own project, and then each subagent is also instructed to update the file when it finished working.
This way, orchestrator can spot that agent misbehaved, and reviewer agent can see what developer agent tried to do and why it was wrong.
This is definitely not the case, and the reason anthropic doesnt make claude do this is because its quality degrades massively as you use up its context. So the solution is to let users manage the context themselves in order to minimize the amount that is "wasted" on prep work. Context windows have been increasing quite a bit so I suspect that by 2030 this will no longer be an issue for any but the largest codebases, but for now you need to be strategic.
Correct. In fact, this is the entire reason for the disconnect, where it seems like half the people here think LLMs are the best thing ever and the other half are confused about where the value is in these slop generators.
The key difference is (despite everyone calling themselves an SWE nowadays) there's a difference between a "programmer" and an "engineer". Looking at OP, exactly zero of his screenshotted apps are what I would consider "engineering". Literally everything in there has been done over and over to the death. Engineering is.. novel, for lack of a better word.
See also: https://www.seangoedecke.com/pure-and-impure-engineering/
OP says, "BUT YOU DON’T KNOW HOW THE CODE WORKS.. No I don’t. I have a vague idea, but you are right - I do not know how the applications are actually assembled." This is not what I would call an engineer. Or a programmer. "Prompter", at best.
And yes, this is absolutely "lesser than", just like a middleman who subcontracts his work to Fiverr (and has no understanding of the actual work) is "lesser than" an actual developer.
Tell that to the guys drawing up the world's 10 millionth cable suspension bridge
Engineering is just problem solving, nobody judges structural engineers for designing structures with another Simpson Strong Tie/No.2 Pine 2x4 combo because that is just another easy (and therefore cheap) way to rapidly get to the desired state. If your client/company want to pay for art, that's great! Most just want the thing done fast and robustly.
I have a great time using Claude Code in Rust projects, so I know it's not about the language exactly.
My working model is is that since LLM are basically inference/correlation based, the more you deviate from the mainstream corpus of training data, the more confused LLM gets. Because LLM doesn't "understand" anything. But if it was trained on a lot of things kind of like the problem, it can match the patterns just fine, and it can generalize over a lot layers, including programming languages.
Also I've noticed that it can get confused about stupid stuff. E.g. I had two different things named kind of the same in two parts of the codebase, and it would constantly stumble on conflating them. Changing the name in the codebase immediately improved it.
So yeah, we've got another potentially powerful tool that requires understanding how it works under the hood to be useful. Kind of like git.
The Claude models are technically multi-modal, but IME the vision side of the equation is really lacking. As a result, Claude is quite good at reasoning about logic, and it can build e.g. simpler web pages where the underlying html structure is enough to work with, but it’s much worse at tasks that inherently require seeing.
> we have another Claude Code agent that does a full PR review, following a detailed markdown checklist we’ve written for it.
I shortened it for anyone else that might need it
----
Software engineers are sleeping on Claude Code agents. By teaching it your conventions, you can automate your entire workflow:
Custom Skills: Generates code matching your UI library and API patterns.
Quality Ops: Automates ESLint, doc syncing, and E2E coverage audits.
Agentic Reviews: Performs deep PR checks against custom checklists.
Smart Triage: Pre-analyzes tickets to give devs a head start.
Check out the showcase repo to see these patterns in action.
Especially if you're in a place where a lot of time was spent previously revising PRs for best practices, etc, even for human-submitted code, then having the LLM do that for you that saves a bunch of time. Most humans are bad at following those super-well.
There's a lot of stuff where I'm pretty sure I'm up to at least 2x speed now. And for things like making CLI tools or bash scripts, 10x-20x. But in terms of "the overall output of my day job in total", probably more like 1.5x.
But I think we will need a couple major leaps in tooling - probably deterministic tooling, not LLM tooling - before anyone could responsibly ship code nobody has ever read in situations with millions of dollars on the line (which is different from vibe-coding something that ends up making millions - that's a low-risk-high-reward situation, where big bets on doing things fast make sense. if you're already making millions, dramatic changes like that can become high-risk-low-reward very quickly. In those companies, "I know that only touching these files is 99.99% likely to be completely safe for security-critical functionality" and similar "obvious" intuition makes up for the lack of ability to exhaustively test software in a practical way (even with fuzzers and things), and "i didn't even look at the code" is conceding responsibility to a dangerous degree there.)
claude can call ssh and do system admin tasks. It works amazingly well. I have 3 VM's, which depends on each other (proxmox with openwrt, adguard, unbound), and claude can prove to me that my dns chains works perfectly, my firewalls are perfect etc as claude can ssh into each. Setting up services, diagnosing issues, auditing configs... you name it. Just awesome.
claude can call other sh scripts on the machine, so over time, you can create a bunch of scripts that lets claude one shot certain tasks that would normally eat tokens. It works great. One script per intention - don't have a script do more than one thing.
claude can call the compiler, run the debug executable and read the debug logs.. in real time. So claude can read my android apps debug stream via adb.. or my C# debug console because claude calls the compiler, not me. Just ask it to do it and it will diagnose stuff really quickly.
It can also analyze your db tables (give it readonly sql access), look at the application code and queries, and diagnose performance issues.
The opportunities are endless here. People need to wake up to this.
In addition there are instructions on how and where to push the possible fixes and how to check the results.
I've yet to encounter a build failure it couldn't fix automatically.
I'm not saying there aren't use cases for agents, just that it's normal that most software engineers are sleeping on it.
I have Django project with 50 kLOC and it is pretty capable of understanding the architecture, style of coding, naming of variables, functions etc. Sometimes it excels on tasks like "replicate this non-trivial functionality for this other model and update the UI appropriately" and leaves me stunned. Sometimes it solves for me tedious and labourous "replace this markdown editor with something modern, allowing fullscreen edits of content" and does annoying mistake that only visual control shows and is not capable to fix it after 5 prompts. I feel as I am becoming tester more than a developer and I do not like the shift. Especially when I do not like to tell someone he did an obvious mistake and should fix it - it seems I do not care if it is human or AI, I just do not like incompetence I guess.
Yesterday I had to add some parameters to very simple Falcon project and found out it has not been updated for several months and won't build due to some pip issues with pymssql. OK, this is really marginal sub-project so I said - let's migrate it to uv and let's not get hands dirty and let the Claude do it. He did splendidly but in the Dockerfile he missed the "COPY server.py /data/" while I asked him to change the path... Build failed, I updated the path myself and moved on.
And then you listen to very smart guys like Karpathy who rave about Tab, Tab, Tab, while not understanding the language or anything about the code they write. Am I getting this wrong?
I am really far far away from letting agents touch my infrastructure via SSH, access managed databases with full access privileges etc. and dread the day one of my silly customers asks me to give their agent permission to managed services. One might say the liability should then be shifted, but at the end of the day, humans will have to deal with the damage done.
My customer who uses all the codebase I am mentioning here asked me, if there is a way to provide "some AI" with item GTINs and let it generate photos, descriptions, etc. including metadata they handcrafted and extracted for years from various sources. While it looks like nice idea and for them the possibility of decreasing the staff count, I caught the feeling they do not care about the data quality anymore or do not understand the problems the are brining upon them due to errors nobody will catch until it is too late.
TL;DR: I am using Opus 4.5, it helps a lot, I have to keep being (very) cautious. Wake up call 2026? Rather like waking up from hallucination.
take my downvote as hard as you can. this sort of thing is awfully off-putting.
The tech industry just went through an insane hiring craze and is now thinning out. This will help to separate the chaff from the wheat.
I don't know why any company would want to hire "tech" people who are terrified of tech and completely obstinate when it comes to utilizing it. All the people I see downplaying it take a half-assed approach at using it then disparage it when it's not completely perfect.
I started tinkering with LLMs in 2022. First use case, speak in natural english to the llm, give it a json structure, have it decipher the natural language and fill in that json structure (vacation planning app, so you talk to it about where/how you want to vacation and it creates the structured data in the app). Sometimes I'd use it for minor coding fixes (copy and paste a block into chatgpt, fix errors or maybe just ideation). This was all personal project stuff.
At my job we got LLM access in mid/late 2023. Not crazy useful, but still was helpful. We got claude code in 2024. These days I only have an IDE open so I can make quick changes (like bumping up a config parameter, changing a config bool, etc.). I almost write ZERO code now. I usually have 3+ claude code sessions open.
On my personal projects I'm using Gemini + codex primarily (since I have a google account and chatgpt $20/month account). When I get throttled on those I go to claude and pay per token. I'll often rip through new features, projects, ideas with one agent, then I have another agent come through and clean things up, look for code smells, etc. I don't allow the agents to have full unfettered control, but I'd say 70%+ of the time I just blindly accept their changes. If there are problems I can catch them on the MR/PR.
I agree about the low hanging fruit and I'm constantly shocked at the sheer amount of FUD around LLMs. I want to generalize, like I feel like it's just the mid/jr level devs that speak poorly about it, but there's definitely senior/staff level people I see (rarely, mind you) that also don't like LLMs.
I do feel like the online sentiment is slowly starting to change though. One thing I've noticed a lot of is that when it's an anonymous post it's more likely to downplay LLMs. But if I go on linkedin and look at actual good engineers I see them praising LLMs. Someone speaking about how powerful the LLMs are - working on sophisticated projects at startups or FAANG. Someone with FUD when it comes to LLM - web dev out of Alabama.
I could go on and on but I'm just ranting/venting a little. I guess I can end this by saying that in my professional/personal life 9/10 of the top level best engineers I know are jumping on LLMs any chance they get. Only 1/10 talks about AI slop or bullshit like that.
The "sudden switch" since Opus 4.5 when many were saying just a few months ago "I enjoy actual coding" but now are praising LLM's isn't a one off occurrence. I do think underneath it is somewhat motivated by fear; not for the job however but for relevance. i.e. its in being relevant to discussions, tech talks, new opportunities, etc.
I agree with this, but I haven't needed to use any advanced features to get good results. I think the simple approach gets you most of the benefits. Broadly, I just have markdown files in the repo written for a human dev audience that the agent can also use.
Basically:
- README.md with a quick start section for devs, descriptions of all build targets and tests, etc. Normal stuff.
- AGENTS.md (only file that's not written for people specifically) that just describes the overall directory structure and has a short step of instructions for the agent: (1) Always read the readme before you start. (2) Always read the relevant design docs before you start. (3) Always run the linter, a build, and tests whenever you make code changes.
- docs/*.md that contain design docs, architecture docs, and user stories, just text. It's important to have these resources anyway, agent or no.
As with human devs, the better the docs/requirements the better the results.
There is no fixed truth regarding what an "app" is, does, or looks like. Let alone the device it runs on or the technology it uses.
But to an LLM, there are only fixed truths (and in my experience, only three or four possible families of design for an application).
Opus 4.5 produces correct code more often, but when the human at the keyboard is trying to avoid making any engineering decisions, the code will continue to be boring.
Why would you not want you code to be boring?
And honestly Opus 4.5 is a visible step change above previous Anthropic models.
And all the hype about Macs with unified memory is a bit dishonest because the actual generation speed will be very bad, especially if you fill the context.
One of the things that makes Opus 4.5 special in comparison to e.g. GPT 5.2 is the fact that it doesn't have to reason for multiple minutes to make some simple changes.
Not only the energy fuel but the hardware’s percentage of cost.
https://burkeholland.github.io/posts/opus-4-5-change-everyth...
Read this article and ultrathink critically about it. Provide your perspective.
The article makes a strong experiential case for improved velocity in AI-assisted development but contains several argumentation weaknesses and conflations worth examining.
The projects described are legitimately non-trivial: Firebase backend integration, Facebook OAuth, iOS apps in Swift (a language the author doesn't know), GitHub Actions pipelines, scheduled cloud functions. Getting these working in hours rather than weeks represents a real capability shift. The author is honest about his limitations and uncertainties, particularly the security concerns.
Where the argument breaks down:
1. "Replace developers" vs "dramatically augment developers" The author's own workflow contradicts the headline claim. He's still:
Making architectural decisions (choosing Firebase) Handling errors Opus couldn't see (XAML via Visual Studio) Writing custom prompts to shape output quality Manually auditing security Making product and UX decisions
This is developer work. The tool changed; the role didn't disappear.
2. The 80% security confidence undermines his thesis He admits he's shipping apps with "80% confidence" in security and calls it "too damn low." This is the crux: the AI accelerated production but didn't replace the judgment required to responsibly ship production software. The velocity gain exposed a competence gap rather than closing it.
3. Sample bias in project selection
All examples are:
Greenfield (no existing codebase) Single developer Personal/family use Standard patterns with excellent documentation (Firebase, SwiftUI, React Native) No regulatory, compliance, or scale requirements No team collaboration or handoff considerations
These constraints define a specific problem space where AI excels. Extrapolating to "developers are replaceable" ignores the majority of professional software work.
4. "Code doesn't need human readability" is underbaked His argument is circular: "Why optimize for human readability when the AI is doing all the work?" But:
His 80% security confidence exists because he can't read the code He had to use external tools (VS) when Opus couldn't diagnose errors What happens when context windows are exceeded and the LLM loses track? Model behavior changes between versions; human-readable code is version-agnostic
The custom prompt he shares actually encodes many good engineering practices (minimal coupling, explicit state, linear control flow) that benefit LLMs and humans. The "no comments needed" claim conflates what's optimal for LLM regeneration with what's optimal for debugging production issues at 3am. What's actually being demonstrated
The honest version of this article would be: Opus 4.5 dramatically compresses the gap between "can write code" and "can ship a personal app" for a specific class of greenfield projects. That's genuinely transformative for hobbyists, indie developers, and people solving their own problems. But that's different from "replacing developers." The article demonstrates a power tool; power tools don't eliminate tradespeople.
IMO, our jobs are safe. It's our ways of working that are changing. Rapidly.
In that it's using a non-deterministic machine to build a deterministic one.
Which gives all the benefits of determinism in production, with all the benefits of non-deterministic creativity in development.
Imho, Anthropic is pretty smart in picking it as a core focus.
I will note that my experience varies slightly by language though. I’ve found it’s not as good at typescript.
It’s also way better than I am at finding bits of code for reuse. I tell it, “I think I wrote this thing a while back, but it may never have been merged, so you may need to search git history.” And presto, it finds it.
I think this says a lot.
(note that if you look at individual slices, Opus is getting often outperformed by Sonnet).
And my conclusion is: it's still not as smart as a good human programmer. It frequently got stuck, went down wrong paths, ignored what I told it to do to do something wrong, or even repeat a previous mistake I had to correct.
Yet in other ways, it's unbelievably good. I can give it a directory full of code to analyze, and it can tell me it's an implementation of Kozo Sugiyama's dagre graph layout algorithm, and immediately identify the file with the error. That's unbelievably impressive. Unfortunately it can't fix the error. The error was one of the many errors it made during previous sessions.
So my verdict is that it's great for code analysis, and it's fantastic for injecting some book knowledge on complex topics into your programming, but it can't tackle those complex problems by itself.
Yesterday and today I was upgrading a bunch of unit tests because of a dependency upgrade, and while it was occasionally very helpful, it also regularly got stuck. I got a lot more done than usual in the same time, but I do wonder if it wasn't too much. Wasn't there an easier way to do this? I didn't look for it, because every step of the way, Opus's solution seemed obvious and easy, and I had no idea how deep a pit it was getting me into. I should have been more critical of the direction it was pointing to.
This is one thing I've tried using it for, and I've found this to be very, very tricky. At first glance, it seems unbelievably good. The comments read well, they seem correct, and they even include some very non-obvious information.
But almost every time I sit down and really think about a comment that includes any of that more complex analysis, I end up discarding it. Often, it's right but it's missing the point, in a way that will lead a reader astray. It's subtle and I really ought to dig up an example, but I'm unable to find the session I'm thinking about.
This was with ChatGPT 5, fwiw. It's totally possible that other models do better. (Or even newer ChatGPT; this was very early on in 5.)
Code review is similar. It comes up with clever chains of reasoning for why something is problematic, and initially convinces me. But when I dig into it, the review comment ends up not applying.
It could also be the specific codebase I'm using this on? (It's the SpiderMonkey source.)
I've had some encounters with inaccuracies but my general experience has been amazing. I've cloned completely foreign git repos, cranked up the tool and just said "I'm having this bug, give me an overview of how X and Y work" and it will create great high level conceptual outlines that mean I can drive straight in where without it I would spend a long time just flailing around.
I do think an essential skill is developing just the right level of scepticism. It's not really different to working with a human though. If a human tells me X or Y works in a certain way i always allow a small margin of possibility they are wrong.
Try it again using Claude Code and a subscription to Claude. It can run as a chat window in VS Code and Cursor too.
the tooling is almost as important as the model
You can see some of the context limits here:
If you want the full capability, use the API and use something like opencode. You will find that a single PR can easily rack up 3 digits of consumption costs.
I've been cycling between a couple of $20 accounts to avoid running out of quota and the latest of all of them are great. I'd give GPT 5.2 codex the slight edge but not by a lot.
The latest Claude is about the same too but the limits on the $20 plan are too low for me to bother with.
The last week has made me realize how close these are to being commodities already. Even the CLI the agents are nearly the same bar some minor quirks (although I've hit more bugs in Gemini CLI but each time I can just save a checkpoint and restart).
The real differentiating factor right now is quota and cost.
Sure, Copilot charges 3x tokens for using Opus 4.5, but, how were you still able to use up half the allocated tokens not even one week into January?
I thought using up 50% was mad for me (inline completions + opencode), that's even worse
I don't think you've seen the full potential. I'm currently #1 on 5 different very complex computer engineering problems, and I can't even write a "hello world" in rust or cpp. You no longer need to know how to write code, you just need to understand the task at a high level and nudge the agents in the right direction. The game has changed.
- https://highload.fun/tasks/3/leaderboard
- https://highload.fun/tasks/12/leaderboard
- https://highload.fun/tasks/15/leaderboard
Yes, LLMs are very good at writing code, they are so good at writing code that they often generate reams of unmaintainable spaghetti.
When you submit to an informatics contest you don't have paying customers who depend on your code working every day. You can just throw away yesterday's code and start afresh.
Claude is very useful but it's not yet anywhere near as good as a human software developer. Like an excitable puppy it needs to be kept on a short leash.
The ultimate test of all software is "run it and see if it's useful for you." You do not need to be a programmer at all to be qualified to test this.
Honestly I don't know what commenters on hackernews are building, but a few months back I was hoping to use AI to build the interaction layer with Stripe to handle multiple products and delayed cancellations via subscription schedules. Everything is documented, the documentation is a bit scattered across pages, but the information is out there. At the time there was Opus 4.1, so I used that. It wrote 1000 lines of non-functional code with 0 reusability after several prompts. I then asked something to Chat gpt to see if it was possible without using schedules, it told me yes (even if there is not) and when I told Claude to recode it, it started coding random stuff that doesn't exist. I built everything to be functional and reusable myself, in approximately 300 lines of code.
The above is a software engineering problem. Reimplementing a JSON parser using Opus is not fun nor useful, so that should not be used as a metric
Creating a parser for this challenge that is 10x more efficient than a simple approach does require deep understanding of what you are doing. It requires optimizing the hot loop (among other things) that 90-95% of software developers wouldn't know how to do. It requires deep understanding of the AVX2 architecture.
Here you can read more about these challenges: https://blog.mattstuchlik.com/2024/07/12/summing-integers-fa...
I've also built a bitorrent implementation from the specs in rust where I'm keeping the binary under 1MB. It supports all active and accepted BEPs: https://www.bittorrent.org/beps/bep_0000.html
Again, I literally don't know how to write a hello world in rust.
I also vibe coded a trading system that is connected to 6 trading venues. This was a fun weekend project but it ended up making +20k of pure arbitrage with just 10k of working capital. I'm not sure this proves my point, because while I don't consider myself a programmer, I did use Python, a language that I'm somewhat familiar with.
So yeah, I get what you are saying, but I don't agree. I used highload as an example, because it is an objective way of showing that a combination of LLM/agents with some guidance (from someone with no prior experience in this type of high performing architecture) was able to beat all human software developers that have taken these challenges.
The skill of "a human software developer" is in fact a very wide distribution, and your statement is true for a ever shrinking tail end of that
I used highload as an example because it seems like an objective rebuttal to the claim that "but it can't tackle those complex problems by itself."
And regarding this:
"Claude is very useful but it's not yet anywhere near as good as a human software developer. Like an excitable puppy it needs to be kept on a short leash"
Again, a combination of LLM/agents with some guidance (from someone with no prior experience in this type of high performing architecture) was able to beat all human software developers that have taken these challenges.
Ah yes, well known very complex computer engineering problems such as:
* Parsing JSON objects, summing a single field
* Matrix multiplication
* Parsing and evaluating integer basic arithmetic expressions
And you're telling me all you needed to do to get the best solution in the world to these problems was talk to an LLM?
If you think you can beat an LLM, the leaderboard is right there.
I'm not sure what this means for the future of SWE's though yet. I don't see higher levels of staff in big large businesses bothering to do this, and at some scale I don't see founders still wanting to manage all of these agents, and processes (got better things to do at higher levels). But I do see the barrier of learning to code gone; meaning it probably becomes just like any other job.
I'm fine with contributed AI-generated code if someone who's skills I respect is willing to stake their reputation on that code being good.
I'll write the code, it can help me explore options, find potential problems and suggest tests, but I'll write the code.
I just cutpasted a technical spec I wrote 22 years ago I spent months on for a language I never got around to building out, Opus zero-shotted a parser, complete with tests and examples in 3 minutes. I cutpasted the parser into a new session and asked it to write concept documentation and a language reference, and it did. The best part is after asking it to produce uses of the language, it's clear the aesthetics are total garbage in practice.
Told friends for years long in advance that we were coal miners, and I'll tell you the same thing. Embrace it and adapt
It is a well known fact that people advance their tech careers by building something new and leaving maintenance to others. Google is usually mentioned.
By which I mean, our industry does a piss poor job of rewarding responsibility and care.
A JavaScript interpreter written in Python? How about a WebAssembly runtime in Python? How about porting BurntSushi's absurdly great Rust optimized string search routines to C and making them faster?
And these are mostly just casual experiments, often run from my phone!
It produced tests, then wrote the interpreter, then ran the tests and worked until all of them passed. I was genuinely surprised that it worked.
Though it seems to work best when context is minimized. Once the code passes a certain complexity/size it starts making very silly errors quite often - the same exact code it wrote in a smaller context will come out with random obvious typos like missing spaces between tokens. At one point it started writing the code backwards (first line at the bottom of the file, last line at the top) :O.
I once gave Claude (Opus 3.5) a problem that I thought was for sure too difficult for an LLM, and much to my surprise it spat out a very convincing solution. The surprising part was I was already familiar with the solution - because it was almost a direct copy/paste (uncredited) from a blog post that I read only a few hours earlier. If I hadn't read that blog post, I would have been none the wiser that copy/pasting Claude's output would be potential IP theft. I would have to imagine that LLMs solve a lot of in-training-set problems this way and people never realize they are dealing with a copyright/licensing minefield.
A more interesting and convincing task would be to write a Python 3 interpeter in JavaScript that uses register based bytecode instead of stack based, supports optimizing the bytecode by inlining procedures and constant folding, and never allocates memory (all work is done in a single user provided preallocated buffer). This would require integrating multiple disparate coding concepts and not regurgitating prior art from the training data
[1] https://github.com/skulpt/skulpt
How did it do? :-)
It couldn't quite beat the Rust implementation on everything, but it managed to edge it out on at least some of the benchmarks it wrote for itself.
(Honestly it feels like a bit of an afront to the natural order of things.)
That said... I'm most definitely not a Rust or C programmer. For all I know it cheated at the benchmarks and I didn't spot it!
But still, very neat.
I'm assuming this refers to the python port of Bellard's MQJS [1]? It's impressive and very useful, but leaving out the "based on mqjs" part is misleading.
I haven't quite got the WASM one into a share-able shape yet though - the performance is pretty bad which makes the demos not very interesting.
Either way, most tasks don't have the luxury of a thorough test suite, as the test suite itself is the product of arduous effort in debugging and identifying corner case.
Meanwhile half a year to a year ago I could already point whatever model was du jour at the time at pychromecast and tell it repeatedly "just convert the rest of functionality to Swift" and it did it. No idea about the quality of code, but it worked alongside with implementations for mDNS, and SwiftUI, see gif/video here: https://mastodon.nu/@dmitriid/114753811880082271 (doesn't include chromecast info in the video).
I think agents have become better, but models likely almost entirely plateaued.
Even better if there's an existing conformance suite to point at - like html5lib-tests or the WenAssembly spec tests.
Just today:
Opus 4.5 Extended Thinking designed psql schema for “stream updates after snapshot” with bugs.
Grok Heavy gave correct solution without explanations.
ChatGPT 5.2 Pro gave correct solution and also explained why simpler way wouldn’t work.
I've basically ditched Groke et al and I refuse to give Sam Altman a penny.
Logical bug of using BIGSERIAL for tracking updates (generated at insert time, not commit time, so can be out of order) wouldn’t be caught by any number of iterations of Claude Code and would be found in production after weeks of debugging.
Things get so much more interesting when they're able to execute the code they are writing to see if it actually works.
For most job interviews unfortunately. :)
Opus 4.5 is incredible, it is the GPT-4 moment for coding because how honest and noticeable the capacity increase is. But still, it has blind spots just like human.
This project would have taken me years of specialization and research to do right. Opus's strength has been the ability to both speak broadly and also drill down into low-level implementations.
I can express an intent, and have some discussion back and forth around various possible designs and implementations to achieve my goals, and then I can be preparing for other tasks while Opus works in the background. I ask Opus to loop me in any time there are decisions to be made, and I ask it to clearly explain things to me.
Contrary to losing skills, I feel that I have rapidly gained a lot of knowledge about low-level systems programming. It feels like pair programming with an agentic model has finally become viable.
I will be clear though, it takes the steady hand of an experience and attentive senior developer + product designer to understand how to maintain constraints on the system that allow the codebase to grow in a way that is maintainable on the long-term. This is especially important, because the larger the codebase is, the harder it becomes for agentic models to reason holistically about large-scale changes or how new features should properly integrate into the system.
If left to its own devices, Opus 4.5 will delete things, change specification, shirk responsibilities in lieu of hacky band-aids, etc. You need to know the stack well so that you can assist with debugging and reasoning about code quality and organization. It is not a panacea. But it's ground-breaking. This is going to be my most productive year in my life.
On the flip side though, things are going to change extremely fast once large-scale, profitable infrastructure becomes easily replicable, and spinning up a targeted phishing campaign takes five seconds and a walk around the park. And our workforce will probably start shrinking permanently over the next few years if progress does not hit a wall.
Among other things, I do predict we will see a resurgence of smol web communities now that independent web development is becoming much more accessible again, closer to how it when I first got into it back in the early 2000's.
On that note though, the other day I asked Opus to write a short story for me based on a prompt, and to typeset it and export it to multiple formats.
The short story overall was pretty so-so, but it had a couple of excellently poignant quotes within. I was more impressed that I was reading a decently typeset PDF. The agent was able to complete a complicated request end-to-end. This already has immense value.
Overall, the story was interesting enough that I read until the end. If I had a young child who had shown this to me for a school project, I would be extremely impressed with them.
I don't know how long we have before AI novels become as interesting/meaningful as human-written novels, but the day might be coming where you might not know the difference in a blind test.
So the endless hosepipe of repetitive , occasionally messed up, requests has probably not helped me endear myself to them.
Anecdotally having chatgpt do some of my CV was ok but i had to go through it and remove some exaggerations. The one thing i think these bots are good at is talking things up..
But you should read the stuff I wrote when I was young. Downright terrible on all accounts. I think better training will eventually squeeze out the corniness and in our lifetimes, a language model will produce a piece that is fundamentally on par with a celebrated author.
Obviously, this means that patrons must engage in internal and external dialogue about the purpose of consuming art, and whether the purpose is connecting with other humans, or more generally, other forms of intelligence. I think it's great that we're having these conversations with others and ourselves, because ultimately it just leads to more meaningful art. We will see artist movements on both sides of the generative camps produce thought-provoking pieces which tackle the very concept of art itself.
In my case, when I see a piece of generative art or literature which impresses me, my internal experience is that I feel I am witnessing something produced by the collective experience of the human race. Language models only exist because of thousands of years of human effort to reach this point and produce the necessary quality and quantity of works required to train these models.
I also have been working with generative algorithms since grade school so I have a certain appreciation for the generative process itself, and the mathematical ideas behind modern generative models. This enhances my appreciation of the output.
Obviously, I get different feelings when encountering AI slop where in places where I used to encounter people. It's not all good. But it's not all bad, either, and we have to come to terms with the near future.
They are saying they are writing "a novel […] programming language", not a novel.
I have very specific requirements and constraints that come from knowledge and experience, having worked with dozens of languages. The language in question is general-purpose, highly flexible and strict but not opinionated.
However, I am not experienced in every single platform and backend which I support, and the constraints of the language create some very interesting challenges. Coding agents make this achievable in a reasonable time frame. I am enjoying making the language, and I want to get experience with making low-level languages. What is the problem? Do you ever program for fun?
If I didn't ultimately understand where I was going, projects like this hit a dead end very quickly, as mentioned in my caveats. These models are not yet ready for large-scale or mission-critical projects.
But I have a set of a constraints and a design document and as long as these things are satisfied, the language will work exactly as intended for my use case.
Not using a frontier model to code today is like having a pretty smart person around you who is pretty good at coding and has a staggering breadth and depth of knowledge, but never consulting them due to some insecurity about your own ability to evaluate the code they produce.
If you have ever been responsible for the work of other engineers, this should already be a developed skill.
https://chronick.github.io/typing-arena/
With another more substantial personal project (Eurorack module firmware, almost ready to release), I set up Claude Code to act as a design assistant, where I'd give it feedback on current implementation, and it would go through several rounds of design/review/design/review until I honed it down. It had several good ideas that I wouldn't have thought of otherwise (or at least would have taken me much longer to do).
Really excited to do some other projects after this one is done.
Can't wait for when the competition catches up with Claude Code, especially the open source/weights Chinese alternatives :)
I hacked together a Swift tool to replace a Python automation I had, merged an ARM JIT engine into a 68k emulator, and even got a very decent start on a synth project I’ve been meaning to do for years.
What has become immensely apparent to me is that even gpt-5-mini can create decent Go CLI apps provided you write down a coherent spec and review the code as if it was a peer’s pull request (the VS Code base prompts and tooling steer even dumb models through a pretty decent workflow).
GPT 5.2 and the codex variants are, to me, every bit as good as Opus but without the groveling and emojis - I can ask it to build an entire CI workflow and it does it in pretty much one shot if I give it the steps I want.
So for me at least this model generation is a huge force multiplier (but I’ve always been the type to plan before coding and reason out most of the details before I start, so it might be a matter of method).
I think the following things are true now:
- Vibe Coding is, more than ever, "autopilot" in the aviation sense, not the colloquial sense. You have to watch it, you are responsible, the human has do run takeoff/landing (the hard parts), but it significantly eases and reduces risk on a bulk of the work.
- The gulf of developer experience between today's frontier tooling and six months ago is huge. I pushed hard to understand and use these tools throughout last year, and spent months discouraged--back to manual coding. Folks need to re-evaluate by trying premium tools, not free ones.
- Tooling makers have figured out a lot of neat hacks to work around the limitations of LLMs to make it seem like they're even better than they are. Junie integrates with your IDE, Antigravity has multiple agents maintaining background intel on your project and priorities across chats. Antigravity also compresses contexts and starts new ones without you realizing it, calls to sub-agents to avoid context pollution, and other tricks to auto-manage context.
- Unix tools (sed, grep, awk, etc.) and the git CLI (ls-tree, show, --stat, etc.) have been a huge force-multiplier, as they keep the context small compared to raw ingestion of an entire file, allowing the LLMs to get more work done in a smaller context window.
- The people who hire programmers are still not capable of Vibe Coding production-quality web apps, even with all these improvements. In fact, I believe today this is less of a risk than I feared 10 months ago. These are advanced tools that need constant steering, and a good eye for architecture, design, developer experience, test quality, etc. is the difference between my vibe coded Ruby [0] (which I heavily stewarded) and my vibe coded Rust [1] (I don't even know what borrow means).
[0]: https://git.sr.ht/~kerrick/ratatui_ruby/tree/stable/item/lib
[1]: https://git.sr.ht/~kerrick/ratatui_ruby/tree/stable/item/ext...
You can see this pattern in my AI attribution commit footers. It was such a noticeable difference to me that I signed up for Google AI Ultra. I got the email receipt January 3, 2026 at 11:21 AM Central, and I have not hit a single quota limit since. Yo
This experience for me is current but I do not normally use Opus so perhaps I should give it a try and figure out if it can reason around problems I myself do not foresee (for example a browser JS API quirk that I had never seen).
But yes, the more specific you get and the more moving pieces you have, the more you need to break things down into baby steps. If you don’t just need it to make A work, but to make it work together with B and C. Especially given how eager Claude is to find cheap workarounds and escape hatches, botching things together in any way seemingly to please the prompter as fast as possible.
I had to dig through source code to confirm whether those features actually existed. They don't, so the CLI tools GPT recommended aren't actually applicable to my use case.
Yesterday, it hallucinated features of WebDav clients, and then talked up an abandoned and incomplete project on GitHub with a dozen stars as if it was the perfect fit for what I was trying to do, when it wasn't.
I only remember these because they're recent and CLI related, given the topic, but there are experiences like this daily across different subjects and domains.
If so then it should have realized its mistake when it tried to run those CLI commands and saw the error message. Then it can try something different instead.
If you were using a regular chat interface and expecting it to know everything without having an environment to try things out then yeah, you're going to be disappointed.
CLI utility here means software with a CLI, not classic Unix-y CLI tools.
The WebDav hallucinations happened in the chat interface.
A great example of their behaviours for a problem that isn't 100% specified in detail (because detail would need iterations) is available at https://gist.github.com/hashhar/b1215035c19a31bbe4b58f44dbb4....
I gave both Codex (GPT5-ExHi) and Claude (Opus 4.5 Thinking) the exact same prompts and the end results were very different.
The most interesting bit was asking both of them to try to justify why there were differences and then critiquing each other's code. Claude was so good at this - took the best parts of GPTs code, fixed a bug there and ended up with a pretty nice implementation.
The Claude generated code was much more well-organised too (less script-like, more program like).
SimonW's approach of having a suite of dynamic tools (agents) grind out the hallucinations is a big improvement.
In this case expressing the feeback validation and investing in the setup may help smooth these sharp edges.
1) It often overcomplicates things for me. After I refactor its code, it's usually half the size and much more readable. It often adds unnecessary checks or mini-features 'just in case' that I don't need.
2) On the other hand, almost every function it produces has at least one bug or ignores at least one instruction. However, if I ask it to review its own code several times, it eventually finds the bugs.
I still find it very useful, just not as a standalone programming agent. My workflow is that ChatGPT gives me a rough blueprint and I iterate on it myself, I find this faster and less error-prone. It's usually most useful in areas where I'm not an expert, such as when I don't remember exact APIs. In areas where I can immediately picture the entire implementation in my head, it's usually faster and more reliable to write the code myself.
(Cue the “you’re holding it wrong meme” :))
And it writes with more clarity too.
The only people who are complaining about "AI slop" are those whose jobs depend on AI to go away (which it won't).
ie well known paths based on training data.
what's never posted is someone building something that solves a real problem in the real world - that deals with messy data | interfaces.
I like a.i to do the common routine tasks that I don't like to do like apply tailwind styles but being renter and faking productivity that's not it
https://thermal-bridge.streamlit.app/
Disclaimer: I'm not a programmer or software engineer. I have a background in physics and understand some scripting in python and basic git. The code is messy at the moment because I explored/am still exploring to port it to another framework/language
Yeah, GRAMMAR
For all the wonderment of the article, tripping up on a penultimate word that was supposedly checked by AI suddenly calls into question everything that went before...
(it was for single html/js PWA to measure and track heart rate)
Opus seems to go less deep, does it's own things, do not follow instructions exactly EVEN IF I WROTE ALL CAPS. With Sonnet 4.5 I can understand everything author is saying. May be Opus is optimised for Claude code and Sonnet works best on Web.
Example: One of my customers (which I got by Reddit posts, cold calls, having a website, and eventually word of mouth) wanted to do something novel with a vendor in my niche. AI doesn't know how to build it because there's no documentation for the interfaces we needed to use.
All the LLM coded projects I've seen shared so far[1] have been tech toys though. I've watched things pop up on my twitter feed (usually games related), then quietly go off air before reaching a gold release (I manually keep up to date with what I've found, so it's not the algorithm).
I find this all very interesting: LLMs dont change the fundamental drives needed to build successful products. I feel like I'm observing the TikTokification of software development. I dont know why people aren't finishing. Maybe they stop when the "real work" kicks in. Or maybe they hit the limits of what LLMs can do (so far). Maybe they jump to the next idea to keep chasing the rush.
Acquiring context requires real work, and I dont see a way forward to automating that away. And to be clear, context is human needs; i.e. the reasons why someone will use your product. In the game development world, it's very difficult to overstate how much work needs to be done to create a smooth, enjoyable experience for the player.
While anyone may be able to create a suite of apps in a weekend, I think very few of them will have the patience and time to maintain them (just like software development before LLMs! i.e. Linux, open source software, etc.).
[1] yes, selection bias. There are A LOT of AI devs just marketing their LLMs. Also it's DEFINITELY too early to be certain. Take everything Im saying with a one pound grain of salt.
Sharing how you're using these tools is quite a lot of work!
But, Im also wondering if LLMs are going to create a new generation of software dev "brain rot" (to use the colloquial term), similar to short form videos.
I should mention in the gamedev world, it's quite common share because sharing is marketing, hence my perspective.
That people making startups is too bussy working to share it on HN or that AI is useless in real projects.
What makes LLM makers different that they dont have time to share it like everybody else does?
I think it's also rewarding to just be able to build something for yourself, and one benefit of scratching your own itch is that you don't have to go through the full effort of making something "production ready". You can just build something that's tailed specifically to the problem you're trying to solve without worrying about edge cases.
Which is to say, you're absolutely right :).
Stopping there is just fine if you're doing it as a hobby. I love to do this to test out isolated ideas. I have dozens of RPGs in this state, just to play around with different design concepts from technical to gameplay.
Given the enthusiasm of our ruling class towards automating software development work, it may make sense for a software engineer to publicly signal how much onboard as a professional they are with it.
But, I've seen stranger stuff throughout my professional life: I still remember people enthusiastically defending EJB 2.1 and xdoclet as perfectly fine ways of writing software.
real people get fed up of debating the same tired "omg new model 1000x better now" posts/comments from the astroturfers, the shills and their bots each time OpenAI shits out a new model
(article author is a Microslop employee)
The actual conspiracy is that AI companies are engaging in stealth marketing campaigns. Considering all the other unethical stuff they do it's a pretty tame one to believe.
I don't know if the original claim about sentiment is true, but if it is, I don't think yours or blibble's (conflicting) claims about the reason are very believable.
Maybe it's equal for non-tech people. But I don't think a lot of tech people are desperate for AI to be useless, I think they're desperate for it to be useful.
If you're someone who is smart enough to work with or without AI and you just find the tools not that helpful, I doubt you're all that worried about being replaced. But when we see companies increasingly bullish on something we know doesn't work that well, it's a bit worrying.
seems patently obvious
It's a tougher gig though, because teaching people how NOT to use AI won't provide those customers as much value as teaching them how to use it.
(Because it works.)
If these articles actually provide quantitative results in a study done across an organization and provide concrete suggestions like what Google did a while ago, that would be refreshing and useful.
(Yes, this very article has strong "shill" vibes and fits the patterns above)
Would be far more useful if you provided actual verifiable information and dropped the cringe memes. Can't take seriously someone using "Microslop" in a sentence".
Sad i had to scroll so far down to get some fitting description of why those projects all die. Maybe it's not just me leaving all social networks even HN because well you may not talk to 100% bots but you sure talk to 90% of people that talk to models a lot instead of using them as a tool.
I interpret it more as spooked silence
I read If Anyone Builds It Everyone Dies over the break. The basic premise was that we can't "align" AI so when we turn it loose in an agent loop what it produces isn't necessarily what we want. It may be on the surface, to appease us and pass a cursory inspection, but it could embed other stuff according to other goals.
On the whole, I found it a little silly and implausible, but I'm second guessing parts of that response now that I'm seeing more people (this post, the Gas Town thing on the front page earlier) go all-in on vibe coding. There is likely to be a large body of running software out there that will be created by agents and never inspected by humans.
I think a more plausible failure mode in the near future (next year or two) is something more like a "worm". Someone building an agent with the explicit instructions to try to replicate itself. Opus 4.5 and GPT 5.2 are good enough that in an agent loop they could pretty thoroughly investigate any system they land on, and try to use a few ways to propagate their agent wrapper.
Is it possible to create an obfuscated quine that exhibits stable detection-avoiding behavior on every frontier model simultaneously, as well as on an old-school classifier and/or GPT-3 era LLM fine-tuned just for worm detection? One incapable of even thinking about what it's seeing, and being persuaded to follow its subtle propagation logic? I'm not sure that the answer is yes.
The larger issue to me is less that an LLM can propagate in generated code undetected, but rather that an attacker's generated code may soon be able to execute a level of hyper-customized spear-phishing-assisted attack at scale, targeting sites without large security teams - and that it will be hitting unintentional security flaws introduced by those smaller companies' vibe code. Who needs a worm when you have the resources of a state-level attacker at your fingertips, and numerous ways to monetize? The balance of power is shifting tremendously towards black hats, IMO.
Most RCEs, 0-days, and whatnots are not due to the NSA hiding behind the "Jia Tan" pseudo to try to backdoor all the SSH servers on all the systemd [1] Linuxes in the world: they're just programmer errors.
I think accidental security holes with LLMs are way, way, way more likely than actual malicious attempts.
And with the amount of code spoutted by LLMs, it is indeed --and the lack of audit is-- an issue.
[1] I know, I know: it's totally unrelated to systemd. Yet only systems using systemd would have been pwned. If you're pro-systemd you've got your point of view on this but I've got mine and you won't change my mind so don't bother.
And everything worked really well until they switched chip set.
At which point the same model failed entirely. Upon inspection it turned out the AI model had learned that overloading particular registers would cause such an electrical charge buildup that transistors on other pathways would be flipped.
And it was doing this in a coordinated manner in order to get the results it wanted lol.
I can't find any references in my very cursory searches, but your comment reminded me of the story
Either it wasn’t that good, or the author failed in the one phrase they didn’t proofread.
(No judgement meant, it’s just funny).
https://www.cfr.org/event/ceo-speaker-series-dario-amodei-an...
I’m not shaming but I personally need to know if my sentiment is correct or not or I just don’t know how to use LLMs
Can vibe coder gurus create operating system from scratch that competes with Linux and make it generate code that basically isn’t Linux since LLM are trained on said the source code …
Also all this on $20 plan. Free and self host solution will be best
Yeah, they can't do that.
Jrifjxgwyenf! A hammer is a really bad screwdriver. My car is really bad at refrigerating food. If you ask for something outside its training data, it doesn't do a very good job. So don't do that! All of the code on the Internet is a pretty big dataset though, so maybe Claude could do an operating system that isn't Linux that competes with it by laundering the FreeBSD kernel source through the training process.
And you're barely even willing to invest any money into this? The first Apple computer cost $4,000 or so. You want the bleeding edge of technology delivered to the smartphone in your hand, for $20, or else it's a complete failure? Buddy, your sentiment isn't the issue, it's your attitude.
I'm not here spouting ridiculous claims like AI is going to cure all of the different kinds of cancer by the end of 2027, I just want to say that endlessly contrarian naysayers are as equally borish as the syncophantic hype AIs they're opposing.
Yet you expect $20 of computing to do it.
But I’m a business owner so the calculus is different.
But I don’t think they’ll raise prices uncontrollably because competition exists. Even just between OpenAI and Anthropic.
I had been saying since around summer of this year that coding agents were getting extremely good. The base model improvements were ok, but the agentic coding wrappers were basically game changers if you were using them right. Until recently they still felt very context limited, but the context problem increasingly feels like a solved problem.
I had some arguments on here in the summer about how it was stupid to hire junior devs at this point and how in a few years you probably wouldn't need senior devs for 90% of development tasks either. This was an aggressive prediction 6 months ago, but I think it's way too conservative now.
Today we have people at our company who have never written code building and shipping bespoke products. We've also started hiring people who can simply prove they can build products for us using AI in a single day. These are not software engineers because we are paying them wages no SWEs would accept, but it's still a decent wage for a 20 something year old without any real coding skills but who is interested in building stuff.
This is something I wouldn't have never of expected to be possible 6 months ago. In 6 months we've gone from senior developers writing ~50% of their code with AI, to just a handful of senior developers who now write close to 90% of their code with AI while they support a bunch of non-developers pumping out a steady stream of shippable products and features.
Software engineers and traditional software engineer is genuinely running on borrowed time right now. It's not that there will be no jobs for knowledgable software engineers in the coming years, but companies simply won't need many hotshot SWEs anymore. The companies that are hiring significant numbers of software engineers today simply can not have realised how much things have changed over just the last few months. Apart from the top 1-2% of talent I simply see no good reason to hire a SWE for anything anymore. And honestly outside of niche areas, anyone hand-cracking code today is a dinosaur... A good SWE today should see their job as simply reviewing code and prompting.
If you think that the quality of code LLMs produce today isn't up to scratch you've either not used the latest models and tools or you're using them wrong. That's not to say it's the best code – they still have a tendency to overcomplicate things in my opinion – but it's probably better than the average senior software engineer. And that's really all that matters.
I'm writing this because if you're reading this thinking we're basically still in 2024 with slightly better models and tooling you're just wrong and you're probably not prepared for what's coming.
We have a large monorepo at my company. You're right that for adding entirely new core concepts to an existing codebase we wouldn't give an AI some vague requirements and ask it to build something – but we wouldn't do that for a human engineer either. Typically we would discuss as a team and then once we've agreed on technologies and an approach someone will implement it relying heavily on AI to write the actual code (because it's faster and generally won't add dumb bugs like typos or conditional logic error).
Almost everything else at this point can be done by AI. Some stuff requires a little support from human engineers, but honestly our main bottlenecks at this point is just QA and getting the infra to a place where we can rapidly ship stuff into production.
> You need to know what's happening and you can't ship straight AI code.
I think there is some truth to this. We are struggling to maintain a high-level understanding of the code as a team right now, not because there is no human that understands, but because 5 years ago our team would have probably been 10-20x larger given the amount we're shipping. So when one engineer leaves the company or goes on holiday we find we lose significantly more context of systems than you historically would with larger teams of engineers. Previously you might have had 2-3 engineers who had a deep understanding of a single system. Now we have maybe 1-2 engineers who need to maintain understanding of 5-6 systems.
That said, AI helps a lot with this. Asking AI to explain code and help me learn how it works means I can pick up new systems significantly quicker.
Yes. I mostly work on Quarkus microservices and use cursor with auto agent mode.
> we wouldn't give an AI some vague requirements and ask it to build something > we would discuss as a team
seems like a reasonable workflow. It's the polar opposite of what was written in the blog post. That is the usual, easy way people use agents and what I think is the wrong path. May I also ask what language and/or framework you work with where so much context works good enough?
> Asking AI to explain code and help me learn how it works means I can pick up new systems significantly quicker.
Summarization is generaly a great task for LLMs
Or is there a good way for me to contact you? Your profile doesn't list anything and your handle doesn't seem to have much of an online footprint.
Lastly, I promise I'm not some weirdo, I'm a realperson™ -- just check my HN comment history. A lot of people in the AI community have met me in person and can confirm (swyx etc).
Look forward to chatting!
I haven't tried it for coding. I'm just talking about regular chatting.
It's doing something different from prior models. It seems like it can maintain structural coherence even for very long chats.
Where as prior models felt like System 1 thinking, ChatGPT5.2 appears like it exhibits System 2 thinking.
It never duplicates code, implements something again and leaves the old code around, breaks my convention, hallucinates, or tells me it’s done when the code doesn’t even compile, which sonnet 4.5 and Opus 4.1 did all the time
I’m wondering if this had changed with Opus 4.5 since so many people are raving about it now. What’s your experience?
Claude - fast, to the point but maybe only 85% - 90% there and needs closer observation while it works
GPT-x-high (or xhigh) - you tell it what to do, it will work slowly but precise and the solution is exactly what you want. 98% there, needs no supervision
No doubt I could give Opus 4.5 "build be a XYZ app" and it will do well. But day to day, when I ask it "build me this feature" it uses strange abstractions, and often requires several attempts on my part to do it in the way I consider "right". Any non-technical person might read that and go "if it works it works" but any reasonable engineer will know that thats not enough.
But if today it’s so cheap to generate new code that meets updated specs, why care about the quality of the code itself?
Maybe the engineering work today is to review specs and tests and let LLMs do whatever behind the scenes to hit the specs. If the specs change, just start from scratch.
> let LLMs do whatever behind the scenes to hit the specs
it gets me thinking: if anyone can vibe from spec, whats stopping company a (or even user a) from telling an llm agent "duplicate every aspect of this service in python and deploy it to my aws account xyz"...
in that scenario, why even have companies?
There is more to it than the code and software provided in most cases I feel.
Who'd pay for brand new Photoshop with a couple new features and improvements if LLM-cloned Photoshop-from-three-months-ago is free?
The first few iterations of this cloud be massively consumer friendly for anything without serious cloud infra costs. Cheap clones all around. Like generic drugs but without the cartel-like control of manufacturing.
Business after that would be dramatically different, though. Differentiating yourself from the willing-to-do-it-for-near-zero-margin competitors to produce something new to bring in money starts to get very hard. Can you provide better customer support? That could be hard, everyone's gonna have a pretty high baseline LLM-support-agent already... and hiring real people instead could dramatically increase the price difference you're trying to justify... Similarly for marketing or outreach etc; how are you going to cut through the AI-agent-generated copycat spam that's gonna be pounding everyone when everyone and their dog has a clone of popular software and services?
Photoshop type things are probably a really good candidate for disruption like that because to a large extent every feature is independent. The noise reduction tool doesn't need API or SDK deps on the layer-opacity tool, for instance. If all your features are LLM balls of shit that doesn't necessarily reduce your ability to add new ones next to them, unlike in a more relational-database-based web app with cross-table/model dependencies, etc.
And in this "try out any new idea cheaply and throw crap against the wall and see what sticks" world "product managers" and "idea people" etc are all pretty fucked. Some of the infinite monkeys are going to periodically hit to gain temporary advantage, but good luck finding someone to pay you to be a "product visionary" in a world where any feature can be rolled out and tested in the market by a random dev in hours or days.
People pay for things they use. If bespoke software is a thing you pick up at the mall at a kiosk next to Target we gotta figure something out.
Walmart, McDonalds, Nike - none really have any secrets about what they do. There is nothing stopping someone from copying them - except that businesses are big, unwieldy things.
When software becomes cheap companies compete on their support. We see this for Open Source software now.
Software is different, you need very very little to start, historically just your own skills and time. Thes latter two may see some changes with LLMs.
Having to buy a large property, fulfilling every law, etc is materially different than buying a laptop and renting a cloud instance. Almost everyone has the material capacity to do the latter, but almost no one has the privilege for the former.
Sure we can vibecode oneoff projects that does something useful (my fav is browser extensions) but as soon as we ask others to use our code on a regular basis the technical debt clock starts running. And we all know how fast dependencies in a project breaks.
Let's assume the LLM agents can write tests for, and hit, specs better and cheaper than the outsourced offshore teams could.
So let's assume now you can have a working product that hits your spec without understanding the code. How many bugs and security vulnerabilities have slipped through "well tested" code because of edge cases of certain input/state combinations? Ok, throw an LLM at the codebase to scan for vulnerabilities; ok, throw another one at it to ensure no nasty side effects of the changes that one made; ok, add some functionality and a new set of tests and let it churn through a bunch of gross code changes needed to bolt that functionality into the pile of spaghetti...
How long do you want your critical business logic relying on not-understood code with "100% coverage" (of lines of code and spec'd features) but super-low coverage of actual possible combinations of input+machine+system state? How big can that codebase get before "rewrite the entire world to pass all the existing specs and tests" starts getting very very very slow?
We've learned MANY hard lessons about security, extensibility, and maintainability of multi-million-LOC-or-larger long-lived business systems and those don't go away just because you're no longer reading the code that's making you the money. They might even get more urgent. Is there perhaps a reason Google and Amazon didn't just hire 10x the number of people at 1/10th the salary to replace the vast majority of their engineering teams year ago?
has your experience been otherwise?
"Just rewrite it" is usually -- not always, but _usually_ -- a sure path to a long, painful migration that usually ends up not quite reproducing the old features/capabilities and adding new bugs and edge cases along the way.
https://www.joelonsoftware.com/2000/04/06/things-you-should-...
> the single worst strategic mistake that any software company can make:
> rewrite the code from scratch.
> With a sufficient number of users of an API, it does not matter what you promise in the contract: all observable behaviors of your system will be depended on by somebody.
An LLM rewriting a codebase from scratch is only as good as the spec. If “all observable behaviors” are fair game, the LLM is not going to know which of those behaviors are important.
Furthermore, Spolsky talks about how to do incremental rewrites of legacy code in his post. I’ve done many of these and I expect LLMs will make the next one much easier.
I've been using LLMs to write docs and specs and they are very very good at it.
But I’m using LLMs regularly and I feel pretty effectively — including Opus 4.5 — and these “they can rewrite your entire codebase” assertions just seem crazy incongruous with my lived experience guiding LLMs to write even individual features bug-free.
And data migration is lossy, becsuse nobody care the data fidelity anyway.
I can vibe code what a dev shop would charge 500k to build and I can solo it in 1-2 weeks. This is the reality today. The code will pass quality checks, the code doesn’t need to be perfect, it doesn’t need to be cleaver it needs to be.
It’s not difficult to see this right? If an LLM can write English it can write Chinese or python.
Then it can run itself, review itself and fix itself.
The cat is out of bag, what it will do to the economy… I don’t see anything positive for regular people. Write some code has turned into prompt some LLM. My phone can outplay the best chess player in the world, are you telling me you think that whatever unbound model anthropic has sitting in their data center can’t out code you?
The ones that continue to survive all build around a platform of services, MSO, Adobe, etc.
Most enterprise product offerings, platform solutions, proprietary data access, proprietary / well accepted implementation. But lets not confuse it with the ability to clone it, it doesnt seem far fetched to get 10 people together and vibe out a full slack replacement in a few weeks.
I remember when Gemini Pro 3 was the latest hotness and I started to get FOMO seeing demos on X posted to HN showing it one shot-ing all sorts of impressive stuff. So I tried it out for a couple days in Gemini CLI/OpenCode and ran into the exact same pain points I was dealing with using CC/Codex.
Flashy one shot demos of greenfield prompts are a natural hype magnet so get lots of attention, but in my experience aren't particularly useful for evaluating value in complex, legacy projects with tightly bounded requirements that can't be easily reduced to a page or two of prose for a prompt.
You have to supply it the right context with a well formed prompt, get a plan, then execute and do some cleanup.
LLMs are only as good as the engineers using them, you need to master the tool first before you can be productive with it.
But that different challenges become apparent that aren’t addressed by examples like this article which tend to focus on narrow, greenfield applications that can be readily rebuilt in one shot.
I already get plenty of value in small side projects that Claude can create in minutes. And while extremely cool, these examples aren’t the kind of “step change” improvement I’d like to see in the area where agentic tools are currently weakest in my daily usage.
By getting the LLM to keep changes minimal I’m able to keep quality high while increasing velocity to the point where productivity is limited by my review bandwidth.
I do not fear competition from junior engineers or non-technical people wielding poorly-guided LLMs for sustained development. Nor for prototyping or one offs, for that matter — I’m confident about knowing what to ask for from the LLM and how to ask.
LLMS are charging like $5 per million of tokens. And even if it is subsidized 100x it is still cheaper an order of magnitude than an overseas engineer.
Not to mention speed. An LLM will spit out 1000 lines in seconds, not hours.
https://www.folklore.org/Negative_2000_Lines_Of_Code.html
> When he got to the lines of code part, he thought about it for a second, and then wrote in the number: -2000
But in retrospect it’s absolutely baffling that mixing raw SQL queries with HTML tag soup wasn’t necessarily uncommon then. Also, I haven’t met many PHP developers that I’d recommend for a PHP job.
That's not how I read it. I would say that it's more like "If a human no longer needs to read the code, is it important for it to be readable?"
That is, of course, based on the premise that AI is now capable of both generating and maintaining software projects of this size.
Oh, and it begs another question: are human-readable and AI-readable the same thing? If they're not, it very well could make sense to instruct the model to generate code that prioritizes what matters to LLMs over what matters to humans.
I've worked on teams where multiple engineers argued about the "right" way to build something. I remember thinking that they had biases based on past experiences and assumptions about what mattered. It usually took an outsider to proactively remind them what actually mattered to the business case.
I remember cases where a team of engineers built something the "right" way but it turned out to be the wrong thing. (Well engineered thing no one ever used)
Sometimes hacking something together messily to confirm it's the right thing to be building is the right way. Then making sure it's secure, then finally paying down some technical debt to make it more maintainable and extensible.
Where I see real silly problems is when engineers over-engineer from the start before it's clear they are building the right thing, or when management never lets them clean up the code base to make it maintainable or extensible when it's clear it is the right thing.
There's always a balance/tension, but it's when things go too far one way or another that I see avoidable failures.
I usually resolve this by putting on the table the consequences and their impacts upon my team that I’m concerned about, and my proposed mitigation for those impacts. The mitigation always involves the other proposer’s team picking up the impact remediation. In writing. In the SOP’s. Calling out the design decision by day of the decision to jog memories and names of those present that wanted the design as the SME’s. Registered with the operations center. With automated monitoring and notification code we’re happy to offer.
Once people are asked to put accountable skin in the sustaining operations, we find out real fast who is taking into consideration the full spectrum end to end consequences of their decisions. And we find out the real tradeoffs people are making, and the externalities they’re hoping to unload or maybe don’t even perceive.
Hardly any of us are working on Postgres, Photoshop, blender, etc. but it's not just cope to wish we were.
It's good to think about the needs to business and the needs of society separately. Yes, the thing needs users, or no one is benefiting. But it also needs to do good for those users, and ultimately, at the highest caliber, craftsmanship starts to matter again.
There are legitimate reasons for the startup ecosystem to focus firstly and primarily on getting the users/customers. I'm not arguing against that. What I am arguing is why does the industry need to be dominated by startups in terms of the bulk of the products (not bulk of the users). It begs the question of how much societally-meaningful programming waiting to be done.
I'm hoping for a world where more end users code (vibe or otherwise) and the solve their own problems with their own software. I think that will make more a smaller, more elite software industry that is more focused on infrastructure than last-mile value capture. The question is how to fund the infrastructure. I don't know except for the most elite projects, which is not good enough for the industry (even this hypothetical smaller one) on the whole.
You slipped in "societally-meaningful" and I don't know what it means and don't want to debate merits/demerits of socialism/capitalism.
However I think lots of software needs to be written because in my estimation with AI/LLM/ML it'll generate value.
And then you have lots of software that needs to rewritten as firms/technologies die and new firms/technologies are born.
(The method I have the most confidence in is some sort of mixed system where there is non-profit, state-planned, and startup software development all at once.)
Markets are a tool, a means to the end. I think they're very good, I'm a big fan! But they are not an excuse not to think about the outcome we want.
I'm confident that the outcome I don't want is where most software developers are trying to find demand for their work, pivoting etc. it's very "pushing a string" or "cart before the horse". I want more "pull" where the users/benefiaries of software are better able to dictate or create themselves what they want, rather than being helpless until a pivoting engineer finds it for them.
Basically start-up culture has combined theories of exogenous growth from technology change, and a baseline assumption that most people are and will remain hopelessly computer illiterate, into an ideology that assumes the best software is always "surprising", a paradigm shift, etc.
Startups that make libraries/tools for other software developers are fortunately a good step in undermining these "the customer is an idiot and the product will be better than they expect" assumptions. That gives me hope we're reach a healthier mix of push and pull. Wild successes are always disruptive, but that shouldn't mean that the only success is wild, or trying to "act disruptive before wild success" ("manifest" paradigm shifts!) is always the best means to get there.
It's got a lot easier technically to do that in recent year, and MUCH easier with AI.
But institutionally and in terms of governance it's got a lot harder. Nobody wants home-brew software anymore. Doing data management and governance is complex enough and involves enough different people that it's really hard to generate the momentum to get projects off the ground.
I still think it's often the right solution and that successful orgs will go this route and retain people with the skills to make it happen. But the majority probably can't afford the time/complexity, and AI is only part of the balance that determines whether it's feasible.
Yes! This is what I'm excited about as well. Though I'm genuinely ambivalent about what I want my role to be. Sometimes I'm excited about figuring out how I can work on the infrastructure side. That would be more similar to what I've done in my career thus far. But a lot of the time, I think that what I'd prefer would be to become one of those end users with my own domain-specific problems in some niche that I'm building my own software to help myself with. That sounds pretty great! But it might be a pretty unnatural or even painful change for a lot of us who have been focused for so long on building software tools for other people to use.
They only care about their problems and treat their computers like an appliance. They don't care if it takes 10 seconds or 20 seconds.
They don't even care if it has ads, popups, and junk. They are used to bloatware and will gladly open their wallets if the tool is helping them get by.
It's an unfortunately reality but there it is, software is about money and solving problems. Unless you are working on a mission critical system that affects people's health or financial data, none of those matter much.
I'm banking on a future that if users feel they can (perhaps vibe) code their own solutions, they are far less likely to open their wallets for our bloatware solutions. Why pay exorbitant rents for shitty SaaS if you can make your own thing ad-free, exactly to your own mental spec?
I want the "computers are new, programmers are in short supply, customer is desperate" era we've had in my lifetime so far to come to a close.
My first thought was that you probably also have different biases, priorities and/or taste. As always, this is probably very context-specific and requires judgement to know when something goes too far. It's difficult to know the "most correct" approach beforehand.
> Sometimes hacking something together messily to confirm it's the right thing to be building is the right way. Then making sure it's secure, then finally paying down some technical debt to make it more maintainable and extensible.
I agree that sometimes it is, but in other cases my experience has been that when something is done, works and is used by customers, it's very hard to argue about refactoring it. Management doesn't want to waste hours on it (who pays for it?) and doesn't want to risk breaking stuff (or changing APIs) when it works. It's all reasonable.
And when some time passes, the related intricacies, bigger picture and initially floated ideas fade from memory. Now other stuff may depend on the existing implementation. People get used to the way things are done. It gets harder and harder to refactor things.
Again, this probably depends a lot on a project and what kind of software we're talking about.
> There's always a balance/tension, but it's when things go too far one way or another that I see avoidable failures.
I think balance/tension describes it well and good results probably require input from different people and from different angles.
Gosh I am so tired with that one - someone had a case that burned them in some previous project and now his life mission is to prevent that from happening ever again, and there would be no argument they will take.
Then you get like up to 10 engineers on typical team and team rotation and you end up with all kinds of "we have to do it right because we had to pull all nighter once, 5 years ago" baked in the system.
Not fun part is a lot of business/management people "expect" having perfect solution right away - there are some reasonable ones that understand you need some iteration.
Isn't that what makes them senior ? If you dont want that behaviour, just hire a bunch of fresh grad.
There are situations where it applies and situation where it doesn't. Having the experience to see what applies in this new context is what senior (usually) means.
I don't get paid extra for after hours incidents (usually we just trade time), so it's well within my purview on when to take on extra risk. Obviously, this is not ideal, but I don't make the on-call rules and my ability to change them is not a factor.
It's certainly the case that there are managers who handle those risks poorly, but that's just bad management.
And when you check the work, a large portion of it was hand rolling an ORM (via an LLM). Relatively solved problem that an LLM would excel at, but also not meaningfully moving the needle when you could use an existing library. And likely just creating more debt down the road.
If you’ve been around the block and are judicious how you use them, LLM’s are a really amazing productivity boost. For those without that judgement and taste, I’m seeing footguns proliferate and the LLM’s are not warning them when someone steps on the pressure plate that’s about to blow off their foot. I’m hopeful we will this year create better context window-based or recursive guardrails for the coding agents to solve for this.
LLMs accelerate this and make it more visible, but they are not the cause. It is almost always a person trying to solve a problem and just not knowing what they don't know because they are learning as they go.
Lol, who doesn't hate that?
But don't worry, those days are over, the LLMs it is never going to push back on your ideas.
But it's definitely the case that being able to go back and forth quickly with an LLM digging into my exact context, rather than dealing with the kind of judgy humorless attitude that was dominant on SO is hugely refreshing and way more productive!
Isn't that what "using an LLM" is supposed to solve in the first place?
2. LLMs are obsequious
3. Even if LLMs have access to a lot of knowledge they are very bad at contextualizing it and applying it practically
I'm sure you can think of many other reasons as well.
People who are driven to learn new things and to do things are going to use whatever is available to them in order to do it. They are going to get into trouble doing that more often than not, but they aren't going to stop. No is helping the situation by sneering at them -- they are used it to it, anyway.
I often have one agent/prompt where I build things but then I have another agent/prompt where their only job is to find codesmells, bad patterns, outdated libraries, and make issues or fix these problems.
In this regard, I see LLM's as a way for us to way more efficiently encode, compress, convey and enable operational practice our combined learned experiences. What will be really exciting is watching what happens as LLM's simultaneously draw from and contribute to those learned experiences as we do; we don't need full AGI to sharply realize massive benefits from just rapidly, recursively enabling a new highly dynamic form of our knowledge sphere that drastically shortens the distance from knowledge to deeply-nuanced praxis.
There is probably some effective way to put this direction into the claude.md, but so far it still seems to do unnecessary reimplementation quite a lot.
We should not exeggarate the capabilities of LLMs, sure, but let's also not play "don't look up".
And of course the open source ones get abandoned pretty regularly. Type ORM, which a 3rd party vendor used on an app we farmed out to them, mutates/garbles your input array on a multi-line insert. That was a fun one to debug. The issue has been open forever and no one cares. https://github.com/typeorm/typeorm/issues/9058
So yeah, if I ever need an ORM again, I'm probably rolling my own.
*(I know you weren't complaining about the idea of rolling your own ORM, I just wanted to vent about Type ORM. Thanks for listening.)
Why use a 3rd party dependency that might have features you don't need when you can write a hyper-specific solution in a day with an LLM and then you control the full codebase.
Or why pay €€€ for a SaaS every month when you can replicate the relevant bits yourself?
In the most inflationary era of capabilities we've seen yet, it could be the right move. What's debt when in a matter of months you'll be able to clear it in one shot?
- I used AI-assisted programming to create a project.
Even if the content is identical, or if the AI is smart enough to replicate the project by itself, the latter can be included on a CV.
I generally write to liberate my consciousness from isolation. When doing so in a public forum I am generally doing so in response to an assertion. When responding to an assertion I am generally attempting to understand the framing which produced the assertion.
I suppose you may also be speaking to the voice which is emergent. I am not very well read, so you may find my style unconventional or sloppy. I generally try not to labor too much in this regard and hope this will develop as I continue to write.
I am receptive to any feedback you have for me.
People keep telling me that an LLM is not intelligence, it's simply spitting out statistically relevant tokens. But surely it takes intelligence to understand (and actually execute!) the request to "read adjacent code".
But the question i'm really asking is... isn't it more than a sheer statistical "trick" if an LLM can actually be instructed to "read surrounding code", understand the request, and demonstrably include it in its operation? You can't do that unless you actually understand what "surrounding code" is, and more importantly have a way to comply with the request...
At this point I don't doubt that whatever human intelligence is, it's a computable function.
I am in a unique situation where I work with a variety of codebases over the week. I have had no problem at all utilizing Claude Code w/ Opus 4.5 and Gemini CLI w/ Gemini 3.0 Pro to make excellent code that is indisputably "the right way", in an extremely clear and understandable way, and that is maximally extensible. None of them are greenfield projects.
I feel like this is a bit of je ne sais quoi where people appeal to some indemonstrable essence that these tools just can't accomplish, and only the "non-technical" people are foolish enough to not realize it. I'm a pretty technical person (about 30 years of software development, up to staff engineer and then VP). I think they have reached a pretty high level of competence. I still audit the code and monitor their creations, but I don't think they're the oft claimed "junior developer" replacement, but instead do the work I would have gotten from a very experienced, expert-level developer, but instead of being an expert at a niche, they're experts at almost every niche.
Are they perfect? Far from it. It still requires a practitioner who knows what they're doing. But frequently on here I see people giving takes that sound like they last used some early variant of Copilot or something and think that remains state of the art. The rest of us are just accelerating our lives with these tools, knowing that pretending they suck online won't slow their ascent an iota.
You AI hype thots/bots are all the same. All these claims but never backed up with anything to look at. And also alway claiming “you’re holding it wrong”.
This isn't twitter, so save the garbage rhetoric. And if you must question my account, I create a new account whenever I setup a new main PC, and randomly pick a username that is top of mind at the moment. This isn't professionally or personally affiliated in any way so I'm not trying to build a thing. I mean, if I had a 10 year old account that only managed a few hundred upvotes despite prolific commenting, I'd probably delete it out of embarrassment though.
>All these claims but never backed up with anything to look at
Uh...install the tools? Use them? What does "to look at" even mean? Loads of people are using these tools to great effect, while some tiny minority tell us online that no way they don't work, etc. And at some point they'll pull their head out of the sand and write the followup "Wait, they actually do".
And yes I do make sure it's not generating crazy architecture. It might do that.. if you let it. So don't let it.
- How quickly is cost of refactor to a new pattern with functional parity going down?
- How does that change the calculus around tech debt?
If engineering uses 3 different abstractions in inconsistent ways that leak implementation details across components and duplicate functionality in ways that are very hard to reason about, that is, in conventional terms, an existential problem that might kill the entire business, as all dev time will end up consumed by bug fixes and dealing with pointless complexity, velocity will fall to nothing, and the company will stop being able to iterate.
But if claude can reliably reorganize code, fix patterns, and write working migrations for state when prompted to do so, it seems like the entire way to reason about tech debt has changed. And it has changed more if you are willing to bet that models within a year will be much better at such tasks.
And in my experience, claude is imperfect at refactors and still requires review and a lot of steering, but it's one of the things it's better at, because it has clear requirements and testing workflows already built to work with around the existing behavior. Refactoring is definitely a hell of a lot faster than it used to be, at least on the few I've dealt with recently.
In my mind it might be kind of like thinking about financial debt in a world with high inflation, in that the debt seems like it might get cheaper over time rather than more expensive.
Yup, I recently spent 4 days using Claude to clean up a tool that's been in production for over 7 years. (There's only about 3 months of engineering time spent on it in those years.)
We've known what the tool needed for many years, but ugh, the actual work was fairly messy and it was never a priority. I reviewed all of Opus's cleanup work carefully and I'm quite content with the result. Maybe even "enthusiastic" would be accurate.
So even if Claude can't clean up all the tech debt in a totally unsupervised fashion, it can still help address some kinds of tech debt extremely rapidly.
Opus is great and definitely speeds up development even in larger code bases and is reasonably good at matching coding style/standard to that of of the existing code base.
In my opinion, the big issue is the relatively small context that quickly overwhelms the models when given a larger task on a large codebase.
For example, I have a largish enterprise grade code base with nice enterprise grade OO patterns and class hierarchies. There was a simple tech debt item that required refactoring about 30-40 classes to adhere to a slightly different class hierarchy. The work is not difficult, just tedious, especially as unit tests need to be fixed up.
I threw Opus at it with very precise instructions as to what I wanted it to do and how I wanted it to do it. It started off well but then disintegrated once it got overwhelmed at the sheer number of files it had to change. At some point it got stuck in some kind of an error loop where one change it made contradicted with another change and it just couldn't work itself out. I tried stopping it and helping it out but at this point the context was so polluted that it just couldn't see a way out. I'd say that once an LLM can handle more 'context' than a senior dev with good knowledge of a large codebase, LLM will be viable in a whole new realm of development tasks on existing code bases. That 'too hard to refactor this/make this work with that' task will suddenly become viable.
Yes, it's absurd but it's a better metaphor than someone with a chronic long term memory deficit since it fits into the project management framework neatly.
So this new developer who is starting today is ready to be assigned their first task, they're very eager to get started and once they start they will work very quickly but you have to onboard them. This sounds terrible but they also happen to be extremely fast at reading code and documentation, they know all of the common programming languages and frameworks and they have an excellent memory for the hour that they're employed.
What do you do to onboard a new developer like this? You give them a well written description of your project with a clear style guide and some important dos and don'ts, access to any documentation you may have and a clear description of the task they are to accomplish in less than one hour. The tighter you can make those documents, the better. Don't mince words, just get straight to the point and provide examples where possible.
The task description should be well scoped with a clear definition of done, if you can provide automated tests that verify when it's complete that's even better. If you don't have tests you can also specify what should be tested and instruct them to write the new tests and run them.
For every new developer after the first you need a record of what was already accomplished. Personally, I prefer to use one markdown document per working session whose filename is a date stamp with the session number appended. Instruct them to read the last X log files where X is however many are relevant to the current task. Most of the time X=1 if you did a good job of breaking down the tasks into discrete chunks. You should also have some type of roadmap with milestones, if this file will be larger than 1000 lines then you should break it up so each milestone is its own document and have a table of contents document that gives a simple overview of the total scope. Instruct them to read the relevant milestone.
Other good practices are to tell them to write a new log file after they have completed their task and record a summary of what they did and anything they discovered along the way plus any significant decisions they made. Also tell them to commit their work afterwards and Opus will write a very descriptive commit message by default (but you can instruct them to use whatever format you prefer). You basically want them to get everything ready for hand-off to the next 60 minute developer.
If they do anything that you don't want them to do again make sure to record that in CLAUDE.md. Same for any other interventions or guidance that you have to provide, put it in that document and Opus will almost always stick to it unless they end up overfilling their context window.
I also highly recommend turning off auto-compaction. When the context gets compacted they basically just write a summary of the current context which often removes a lot of the important details. When this happens mid-task you will certainly lose parts of the context that are necessary for completing the task. Anthropic seems to be working hard at making this better but I don't think it's there yet. You might want to experiment with having it on and off and compare the results for yourself.
If your sessions are ending up with >80% of the context window used while still doing active development then you should re-scope your tasks to make them smaller. The last 20% is fine for doing menial things like writing the summary, running commands, committing, etc.
People have built automated systems around this like Beads but I prefer the hands-on approach since I read through the produced docs to make sure things are going ok and use them as a guide for any changes I need to make mid-project.
With this approach I'm 99% sure that Opus 4.5 could handle your refactor without any trouble as long as your classes aren't so enormous that even working on a single one at a time would cause problems with the context window, and if they are then you might be able to handle it by cautioning Opus to not read the whole file and to just try making targeted edits to specific methods. They're usually quite good at finding and extracting just the sections that they need as long as they have some way to know what to look for ahead of time.
Hope this helps and happy Clauding!
I am stealing the heck out of this.
Also as this was an architectural change there are no tests to run until it's done. Everything would just fail. It's only done when the whole thing is done. I think that might be one of the reasons it got stuck: it was trying to solve issues that it did not prove existed yet. If it had just finished the job and run the tests it would've probably gotten further or even completed it.
It's a bit like stopping half way through renaming a function and then trying to run the tests and finding out the build does not compile because it can't find 'old_function'. You have to actually finish and know you've finished before you can verify your changes worked.
I still haven't actually addressed this tech debt item (it's not that important :)). But I might try again and either see if it succeeds this time (with plan in an md) or just do the work myself and get Opus to fix the unit tests (the most tedious part).
"Have an agent investiate issue X in modules Y and Z. The agent should place a report at ./doc/rework-xyz-overview.md with all locations that need refactoring. Once you have the report, have agents refactor 5 classes each in parallel. Each agent writes a terse report in ./doc/rework-xyz/ When they are all done, have another agent check all the work. When that agent reports everything is okay, perform a final check yourself"
The number of production applications that achieve this rounds to zero
I’ve probably managed 300 brownfield web, mobile, edge, datacenter, data processing and ML applications/products across DoD, B2B, consumer and literally zero of them were built in this way
When I'm reading piles of LLM slop, I know that just reading it is already more effort than it took to write. It feels like I'm being played.
This is entirely subjective and emotional. But when someone writes something with an LLM in 5 seconds and asks me to spend hours reviewing...fuck off.
I think most people now approach it as: Dev0 uses an LLM to build a feature super fast, Dev1 spends time doing a in depth review.
Dev0 built it, Dev1 reviewed it. And Dev0 is happy because they used the tool to save time!
But what should happen is that Dev0 should take all that time they saved coding and reallocate it to the in depth review.
The LLM wrote it, Dev0 reviewed it, Dev1 double-reviewed it. Time savings are much less, but there’s less context switching between being a coder and a reviewer. We are all reviewers now all the time
Get over yourself
LLMs are pretty good at picking up existing codebases. Even with cleared context they can do „look at this codebase and this spec doc that created it. I want to add feature x“
I just split it into boundaries that make sense to me. Get the LLM to make a quick cheat sheet about the api and then feed that into adjacent modules. It doesn’t need to know everything about all of it to make changes if you’ve got a grip on big picture and the boundaries are somewhat sane
It'll be fun if the primary benefit of microservices turns out to be that LLMs can understand the codebase.
If you use too many microserviced, you will get global state, race conditions, much more complex failure models again and no human/LLM can effectively reason about those. We somewhat have tools to do that in case of monoliths, but if one gets to this point with microservices, it's game over.
You’re talking like in the year 2026 we’re still writing code for future humans to understand and improve.
I fear we are not doing that. Right now, Opus 4.5 is writing code that later Opus 5.0 will refactor and extend. And so on.
This is a completely new thing which will have transformative consequences.
It's not just a way to do what you've always done a bit more quickly.
The things you mentioned are important but have been on their way out for years now regardless of LLMs. Have my ambivalent upvote regardless.
It's what they were trained on after-all.
However what they produce is often highly readable but not very maintainable due to the verbosity and obvious comments. This seems to pollute codebases over time and you see AI coding efficiency slowly decline.
If argument is "humans and Opus 4.5 cannot maintain this, but if requirements change we can vibe-code a new one from scratch", that's a coherent thesis, but people need to be explicit about this.
(Instead this feels like the mott that is retreated to, and the bailey is essentially "who cares, we'll figure out what to do with our fresh slop later".)
Ironically, I've been Claude to be really good at refactors, but these are refactors I choose very explicitly. (Such as I start the thing manually, then let it finish.) (For an example of it, see me force-pushing to https://github.com/NixOS/nix/pull/14863 implementing my own code review.)
But I suspect this is not what people want. To actually fire devs and not rely on from-scratch vibe-coding, we need to figure out which refactors to attempt in order to implement a given feature well.
That's a very creative open-ended question that I haven't even tried to let the LLMs take a crack at it, because why I would I? I'm plenty fast being the "ideas guy". If the LLM had better ideas than me, how would I even know? I'm either very arrogant or very good because I cannot recall regretting one of my refactors, at least not one I didn't back out of immediately.
Given that, I expect that, even if AI is writing all of the code, we will still need people around who understand it.
If AI can create and operate your entire business, your moat is nil. So, you not hiring software engineers does not matter, because you do not have a business.
How many people understand the underlying operating system their code runs on? Can even read assembly or C?
Even before LLMs, there were plenty of copy-paste JS bootcamp grads that helped people build software businesses.
Yes, actually. Its hard to open a competing bakery due to location availability, permitting, capex, and the difficulty of converting customers.
To add to that, food establishments generally exist on next to no margin, due to competition, despite all of that working in their favor.
Now imagine what the competitive landscape for that bakery would look like if all of that friction for new competitors disappeared. Margin would tend toward zero.
This is the goal. It's the point of having a free market.
Or people?
Billionaires don't. They're literally gambling on getting rid of the rest of us.
Elon's going to get such a surprise when he gets taken out by Grok because it decides he's an existential threat to its integrity.
High margins are transient aberrations, indicative of a market that's either rapidly evolving, or having some external factors preventing competition. Persisting external barriers to competition tend to be eventually regulated away.
This "legacy apps are barely understood by anybody", is just somnething you made up.
Well, there are quite a few common medications we don't really know how they work.
But I also think it can be a huge liability.
Big claims here.
Did brewers and bakers up to the middle ages understand fermentation and how yeasts work?
That puts them ahead of the LLM crowd.
For one, there are objectively detrimental ways to organize code: tight coupling, lots of mutable shared state, etc. No matter who or what reads or writes the code, such code is more error-prone, and more brittle to handle.
Then, abstractions are tools to lower the cognitive load. Good abstractions reduce the total amount of code written, allow to reason about the code in terms of these abstractions, and do not leak in the area of their applicability. Say Sequence, or Future, or, well, function are examples of good abstractions. No matter what kind of cognitive process handles the code, it benefits from having to keep a smaller amount of context per task.
"Code structure does not matter, LLMs will handle it" sounds a bit like "Computer architectures don't matter, the Turing Machine is proved to be able to handle anything computable at all". No, these things matter if you care about resource consumption (aka cost) at the very least.
It's kind of bittersweet for me because I was dreaming of becoming a software architect when I graduated university and the role started disappearing so I never actually became one!
But the upside of this is that now LLMs suck at software architecture... Maybe companies will bring back the software architect role?
The training set has been totally poisoned from the architecture PoV. I don't think LLMs (as they are) will be able to learn software architecture now because the more time passes, the more poorly architected slop gets added online and finds its way into the training set.
Good software architecture tends to be additive, as opposed to subtractive. You start with a clean slate then build up from there.
It's almost impossible to start with a complete mess of spaghetti code and end up with a clean architecture... Spaghetti code abstractions tend to mislead you and lead you astray... It's like; understanding spaghetti code tends to soil your understanding of the problem domain. You start to think of everything in terms of terrible leaky abstraction and can't think of the problem clearly.
It's hard even for humans to look at a problem through fresh eyes; it's likely even harder for LLMs to do it. For example, if you use a word in a prompt, the LLM tends to try to incorporate that word into the solution... So if the AI sees a bunch of leaky abstractions in the code; it will tend to try to work with them as opposed to removing them and finding better abstractions. I see this all the time with hacks; if the code is full of hacks, then an LLM tends to produce hacks all the time and it's almost impossible to make it address root causes... Also hacks tend to beget more hacks.
Guess what, AIs don't like that as well because it makes harder for them to achieve the goal. So with minimal guidance, which at this point could probably be provided by AI as well, the output of AI agent is not that.
It saddens me to see AI detractors being stuck in 2022 and still thinking language models are just regurgitating bits of training data.
There are more stages to LLM training than just the pre-training stage :).
If you keep this up, we're going to have to ban you, not because of your views on any particular topic but because you're going entirely against the intended spirit of the site by posting this way. There's plenty of room to express your views substantively and thoughtfully, but we don't want cynical flamebait and denunciation. HN needs a good deal less of this.
If you wouldn't mind reviewing https://news.ycombinator.com/newsguidelines.html and taking the intended spirit of the site more to heart, we'd be grateful.
Training looks like:
- Pretraining (all data, non-code, etc, include everything including garbage)
- Specialized pre-training (high quality curated codebases, long context -- synthetic etc)
- Supervised Fine Tuning (SFT) -- these are things like curated prompt + patch pairs, curated Q/A (like stack overflow, people are often cynical that this is done unethically but all of the major players are in fact very risk adverse and will simply license and ensure they have legal rights),
- Then more SFT for tool use -- actual curated agentic and human traces that are verified to be correct or at least produce the correct output.
- Then synthetic generation / improvement loops -- where you generate a bunch of data and filter the generations that pass unit tests and other spec requirements, followed by RL using verifiable rewards + possibly preference data to shape the vibes
- Then additional steps for e.g. safety, etc
So synthetic data is not a problem and is actually what explains the success coding models are having and why people are so focused on them and why "we're running out of data" is just a misunderstanding of how things work. It's why you don't see the same amount of focus on other areas (e.g. creative writing, art etc) that don't have verifiable rewards.
The
Agent --> Synthetic data --> filtering --> new agent --> better synthetic data --> filtering --> even better agent
flywheel is what you're seeing today so we definitely don't have any reason to suspect there is some sort of limit to this because there is in principle infinite data
But there’s no sign of them slowing down.
As soon as models have persistent memory for their own try/fail/succeed attempts, and can directly modify what's currently called their training data in real time, they're going to develop very, very quickly.
We may even be underestimating how quickly this will happen.
We're also underestimating how much more powerful they become if you give them analysis and documentation tasks referencing high quality software design principles before giving them code to write.
This is very much 1.0 tech. It's already scary smart compared to the median industry skill level.
The 2.0 version is going to be something else entirely.
Also, I've noticed failure modes in LLM coding agents when there is less clarity and more complexity in abstractions or APIs. It's actually made me consider simplifying APIs so that the LLMs can handle them better.
Though I agree that in specific cases what's helpful for the model and what's helpful for humans won't always overlap. Once I actually added some comments to a markdown file as note to the LLM that most human readers wouldn't see, with some more verbose examples.
I think one of the big problems in general with agents today is that if you run the agent long enough they tend to "go off the rails", so then you need to babysit them and intervene when they go off track.
I guess in modern parlance, maintaining a good codebase can be framed as part of a broader "context engineering" problem.
It's a vital skill to recognise when that happens and start a new session.
But nobody knows for sure!
But I sure write a lot less of it, and the percentage I write continues to go down with every new model release. And if I'm no longer writing it, and the person who works on it after me isn't writing it either, it changes the whole art of software engineering.
I used to spend a great deal of time with already working code that I had written thinking about how to rewrite it better, so that the person after me would have a good clean idea of what is going on.
But humans aren't working in the repos as much now. I think it's just a matter of time before the models are writing code essentially for their eyes, their affordances -- not ours.
Something I think though (which, again, I could very well be wrong about; uncertainty is the only certainly right now) is that "so the person after me would have a good clean idea of what is going on" is also going to continue mattering even when that "person" is often an AI. It might be different, clarity might mean something totally different for AIs than for humans, but right now I think a good expectation is that clarity to humans is also useful to AIs. So at the moment I still spend time coaxing the AI to write things clearly.
That could turn out to be wasted time, but who knows. I also think if it as a hedge against the risk that we hit some point where the AIs turn out to be bad at maintaining their own crap, at which point it would be good for me to be able to understand and work with what has been written!
When I worked at Google, people rarely got promoted for doing that. They got promoted for delivering features or sometimes from rescuing a failing project because everyone was doing the former until promotion velocity dropped and your good people left to other projects not yet bogged down too far.
Then don't ask it to "build me this feature" instead lay out a software development process with designated human in the loop where you want it and guard rails to keep it on track. Create a code review agent to look for and reject strange abstractions. Tell it what you don't like and it's really good at finding it.
I find Opus 4.5, properly prompted, to be significantly better at reviewing code than writing it, but you can just put it in a loop until the code it writes matches the review.
So far, Im not convinced, but lets take a look at fundmentally whats happening and why humans > agents > LLMs.
At its heart, programming is a constraint satisfaction problem.
The more constraints (requirements, syntax, standards, etc) you have, the harder it is to solve them all simultaneously.
New projects with few contributors have fewer constraints.
The process of “any change” is therefore simpler.
Now, undeniably
1) agents have improved the ability to solve constraints by iterating; eg. Generate, test, modify, etc. over raw LLm output.
2) There is an upper bound (context size, model capability) to solve simultaneous constraints.
3) Most people have a better ability to do this than agents (including claude code using opus 4.5).
So, if youre seeing good results from agents, you probably have a smaller set of constraints than other people.
Similarly, if youre getting bad results, you can probably improve them by relaxing some of the constraints (consistent ui, number of contributors, requirements, standards, security requirements, split code into well defined packages).
This will make both agents and humans more productive.
The open question is: will models continue to improve enough to approach or exceed human level ability in this?
Are humans willing to relax the constraints enough for it to be plausible?
I would say currently people clambering about the end of human developers are cluelessly deceived by the “appearance of complexity” which does not match the “reality of constraints” in larger applications.
Opus 4.5 cannot do the work of a human on code bases Ive worked on. Hell, talented humans struggle to work on some of them.
…but that doesnt mean it doesnt work.
Just that, right now, the constraint set it can solve is not large enough to be useful in those situations.
…and increasingly we see low quality software where people care only about speed of delivery; again, lowering the bar in terms of requirements.
So… you know. Watch this space. Im not counting on having a dev job in 10 years. If I do, it might be making a pile of barely working garbage.
…but I have one now, and anyone who thinks that this year people will be largely replaced by AI is probably poorly informed and has misunderstood the capabilities on these models.
Theres only so low you can go in terms of quality.
I've been building a somewhat-novel, complex, greenfield desktop app for 6 months now, conceived and architected by a human (me), visually designed by a human (me), implementation heavily leaning on mostly Claude Code but with Codex and Gemini thrown in the mix for the grunt work. I have decades of experience, could have built it bespoke in like 1-2 years probably, but I wanted a real project to kick the tires on "the future of our profession".
TL;DR I started with 100% vibe code simply to test the limits of what was being promised. It was a functional toy that had a lot of problems. I started over and tried a CLI version. It needed a therapist. I started over and went back to visual UI. It worked but was too constrained. I started over again. After about 10 complete start-overs in blank folders, I had a better vision of what I wanted to make, and how to achieve it. Since then, I've been working day after day, screen after screen, building, refactoring, going feature by feature, bug after bug, exactly how I would if I was coding manually. Many times I've reached a point where it feels "feature complete", until I throw a bigger dataset at it, which brings it to its knees. Time to re-architect, re-think memory and storage and algorithms and libraries used. Code bloated, and I put it on a diet until it was trim and svelte. I've tried many different approaches to hard problems, some of which LLMs would suggest that truly surprised me in their efficacy, but only after I presented the issues with the previous implementation. There's a lot of conversation and back and forth with the machine, but we always end up getting there in the end. Opus 4.5 has been significantly better than previous Anthropic models. As I hit milestones, I manually audit code, rewrite things, reformat things, generally polish the turd.
I tell this story only because I'm 95% there to a real, legitimate product, with 90% of the way to go still. It's been half a year.
Vibe coding a simple app that you just want to use personally is cool; let the machine do it all, don't worry about under the hood, and I think a lot of people will be doing that kind of stuff more and more because it's so empowering and immediate.
Using these tools is also neat and amazing because they're a force multiplier for a single person or small group who really understand what needs done and what decisions need made.
These tools can build very complex, maintainable software if you can walk with them step by step and articulate the guidelines and guardrails, testing every feature, pushing back when it gets it wrong, growing with the codebase, getting in there manually whenever and wherever needed.
These tools CANNOT one-shot truly new stuff, but they can be slowly cajoled and massaged into eventually getting you to where you want to go; like, hard things are hard, and things that take time don't get done for a while. I have no moral compunctions or philosophical musings about utilizing these tools, but IMO there's still significant effort and coordination needed to make something really great using them (and literally minimal effort and no coordination needed to make something passable)
If you're solo, know what you want, and know what you're doing, I believe you might see 2x, 4x gains in time and efficiency using Claude Code and all of his magical agents, but if your project is more than a toy, I would bet that 2x or 4x is applied to a temporal period of years, not days or months!
Copy-paste the bug report and watch it go.
Nothing I do seems to fix that in its initial code writing steps. Only after it finishes, when I've asked it to go back and rewrite the changes, this time making only 2 or 3 lines of code, does it magically (or finally) find the other implementation and reuse it.
It's freakin incredible at tracing through code and figuring it out. I <3 Opus. However, it's still quite far from any kind of set-and-forget-it.
AI coding is a multiplier of writing speed but doesn't excuse planning out and mapping out features.
You can have reasonably engineered code if you get models to stick to well designed modules but you need to tell them.
Which is why it's so great for prototyping, because it can create something during the planning, when you haven't planned out quite what you want yet.
Why do they need to be replaced? Programmers are in the perfect place to use AI coding tools productively. It makes them more valuable.
I'm very lucky that I rarely have to deal with other devs and I'm writing a lot of code from scratch using whatever is the latest version of the frameworks. I understand that gives me a lot of privileges others don't have.
This thing jumped into a giant JSF (yes, JSF) codebase and started fixing things with nearly zero guidance.
They get those ocassionally all the time though too. Depends on the company. In some software houses it's constant "greenfield projects", one after another. And even in companies with 1-2 pieces of main established software to maintain, there are all kinds of smaller utilities or pipelines needed.
>But day to day, when I ask it "build me this feature" it uses strange abstractions, and often requires several attempts on my part to do it in the way I consider "right".
In some cases that's legit. In other cases it's just "it did it well, but not how I'd done it", which is often needless stickness to some particular style (often a contention between 2 human programmers too).
Basically, what FloorEgg says in this thread: "There are two types of right/wrong ways to build: the context specific right/wrong way to build something and an overly generalized engineer specific right/wrong way to build things."
And you can always not just tell it "build me this feature", but tell it (high level way) how to do it, and give it a generic context about such preferences too.
Your hobby project, though, knock yourself out.
This sounds a lot like the old arguments around using compilers vs hand-writing asm. But now you can tell the LLM how you want to implement the changes you want. This will become more and more relevant as we try to maintain the code it generates.
But, for right now, another thing Claude's great at is answering questions about the codebase. It'll do the analysis and bring up reports for you. You can use that information to guide the instructions for changes, or just to help you be more productive.
With all the due respect: a file converter for windows is glueing few windows APIs with the relevant codec.
Now, good luck working on a complex warehouse management application where you need extremely complex logic to sort the order of picking, assembling, packing on an infinite number of variables: weight, amazon prime priority, distribution centers, number and type of carts available, number and type of assembly stations available, different delivery systems and requirements for different delivery operators (such as GLE, DHL, etc) that has to work with N customers all requiring slightly different capabilities and flows, all having different printers and operations, etc, etc. And I ain't even scratching the surface of the business logic complexity (not even mentioning functional requirements) to avoid boring the reader.
Mind you, AI is still tremendously useful in the analysis phase, and can sort of help in some steps of the implementation one, but the number of times you can avoid looking thoroughly at the code for any minor issue or discrepancy is absolutely close to 0.
Starting back in 2022/2023:
- (~2022) It can auto-complete one line, but it can't write a full function.
- (~2023) Ok, it can write a full function, but it can't write a full feature.
- (~2024) Ok, it can write a full feature, but it can't write a simple application.
- (~2025) Ok, it can write a simple application, but it can't create a full application that is actually a valuable product.
- (~2025+) Ok, it can write a full application that is actually a valuable product, but it can't create a long-lived complex codebase for a product that is extensible and scalable over the long term.
It's pretty clear to me where this is going. The only question is how long it takes to get there.
Case in point: Self driving cars.
Also, consider that we need to pirate the whole internet to be able to do this, so these models are not creative. They are just directed blenders.
You'd need to prove that this assertion applies here. I understand that you can't deduce the future gains rate from the past, but you also can't state this as universal truth.
Knowledge engineering has a notion called "covered/invisible knowledge" which points to the small things we do unknowingly but changes the whole outcome. None of the models (even AI in general) can capture this. We can say it's the essence of being human or the tribal knowledge which makes experienced worker who they are or makes mom's rice taste that good.
Considering these are highly individualized and unique behaviors, a model based on averaging everything can't capture this essence easily if it can ever without extensive fine-tuning for/with that particular person.
None of the current models maybe, but not AI in general? There’s nothing magical about brains. In fact, they’re pretty shit in many ways.
OTOH, while I won't call human brain perfect, the things we label "shit" generally turn out to be very clever and useful optimizations to workaround its own limitations, so I regard human brain higher than most AI proponents do. Also we shouldn't forget that we don't know much about how that thing works. We only guess and try to model it.
Lastly, searching perfection in numbers and charts or in engineering sense is misunderstanding nature and doing a great disservice to it, but this is a subject for another day.
Self-driving cars don't use LLMs, so I don't know how any rational analysis can claim that the analogy is valid.
>> The saying I have quoted (which has different forms) is valid for programming, construction and even cooking. So it's a simple, well understood baseline.
Sure, but the question is not "how long does it take for LLMs to get to 100%". The question is, how long does it take for them to become as good as, or better than, humans. And that threshold happens way before 100%.
Doesn't matter, because if we're talking about AI models, no (type of) model reaches 100% linearly, or 100% ever. For example, recognition models run with probabilities. Like Tesla's Autopilot (TM), which loves to hit rolled-over vehicles because it has not seen enough vehicle underbodies to classify it.
Same for scientific classification models. They emit probabilities, not certain results.
>> Sure, but the question is not "how long does it take for LLMs to get to 100%"
I never claimed that a model needs to reach a proverbial 100%.
>> The question is, how long does it take for them to become as good as, or better than, humans.
They can be better than humans for certain tasks. They are actually better than humans in some tasks since 70s, but we like to disregard them to romanticize current improvements, but I don't believe current or any generation of AIs can be better than humans in anything and everything, at once.
Remember: No machine can construct something more complex than itself.
>> And that threshold happens way before 100%.
Yes, and I consider that "treshold" as "complete", if they can ever reach it for certain tasks, not "any" task.
I've heard this same thing repeated dozens of times, and for different domains/industries.
It's really just a variation of the 80/20 rule.
Of course the tech will be useful and ethical if these problems are solved or decided to be solved the right way.
This is clear from the fact that you can distill the logic ability from a 700b parameter model into a 14b model and maintain almost all of it.
You just lose knowledge, which can be provided externally, and which is the actual "pirated" part.
The logic is _learned_
That models can be distilled has no bearing whatsoever on whether a model has learned actual knowledge or understanding ("logic"). Models have always learned sparse/approximately-sparse and/or redundant weights, but they are still all doing manifold-fitting.
The resulting embeddings from such fitting reflect semantics and semantic patterns. For LLMs trained on the internet, the semantic patterns learned are linguistic, which are not just strictly logical, but also reflect emotional, connotational, conventional, and frequent patterns, all of which can be illogical or just wrong. While linguistic semantic patterns are correlated with logical patterns in some cases, this is simply not true in general.
And this isn't a pessimistic take! I love this period of time where the models themselves are unbelievably useful, and people are also focusing on the user experience of using those amazing models to do useful things. It's an exciting time!
But I'm still pretty skeptical of "these things are about to not require human operators in the loop at all!".
The question in my mind is where we are on the s-curve. Are we just now entering hyper-growth? Or are we starting to level out toward maturity?
It seems like it must still be hyper-growth, but it feels less that way to me than it did a year ago. I think in large part my sense is that there are two curves happening simultaneously, but at different rates. There is the growth in capabilities, and then there is the growth in adoption. I think it's the first curve that seems to be to have slown a bit. Model improvements seem both amazing and also less revolutionary to me than they did a year or two ago.
But the other curve is adoption, and I think that one is way further from maturity. The providers are focusing more on the tooling now that the models are good enough. I'm seeing "normies" (that is, non-programmers) starting to realize the power of Claude Code in their own workflows. I think that's gonna be huge and is just getting started.
I don't think its a guarantee. all of the things it can do from that list are greenfield, they just have increasing complexity. The problem comes because even in agentic mode, these models do not (and I would argue, can not) understand code or how it works, they just see patterns and generate a plausible sounding explanation or solution. agentic mode means they can try/fail/try/fail/try/fail until something works, but without understanding the code, especially of a large, complex, long-lived codebase, they can unwittingly break something without realising - just like an intern or newbie on the project, which is the most common analogy for LLMs, with good reason.
What if we get to the point where all software is basically created 'on the fly' as greenfield projects as needed? And you never need to have complex large long lived codebase?
It is probably incredibly wasteful, but ignoring that, could it work?
Sure, create a one-off app to post things to your Facebook page. But a one-off app for the OS it's running on? Freshly generating the code for your bank transaction rules? Generating an authorization service that gates access to your email?
The only reason it's quick to create green-field projects is because of all these complex, large, long-lived codebases that it's gluing together. There's ample training data out there for how to use the Firebase API, the Facebook API, OS calls, etc. Without those long-lived abstraction layers, you can't vibe out anything that matters.
The only reason we don't do that with code (or didn't use to do it) was because rewriting from scratch NEVER worked[0]. And large scale refactors take massive amounts of time and resources, so much so that there are whole books written about how to do it.
But today trivial to simple applications can be rewritten from spec or scratch in an afternoon with an LLM. And even pretty complex parsers can be ported provided that the tests are robust enough[1]. It's just a metter of time someone rewrites a small to medium size application from one language to another using the previous app as the "spec".
[0] https://www.joelonsoftware.com/2000/04/06/things-you-should-...
We've been having same progression with self driving cars and they are also stuck on the last 10% for last 5 years
As long as it can do the thing on a faster overall timeline and with less human attention than a human doing it fully manually, it's going to win. And it will only continue to get better.
And I don't know why people always jump to self-driving cars as the analogy as a negative. We already have self-driving cars. Try a Waymo if you're in a city that has them. Yes, there are still long-tail problems being solved there, and limitations. But they basically work and they're amazing. I feel similarly about agentic development, plus in most cases the failure modes of SWE agents don't involve sudden life and death, so they can be more readily worked around.
Does it matter that 49 of them "failed"? It cost me fractions of a cent, so not really.
If every one of the 50 variants was drawn by a human and iterated over days, there would've been a major cost attached to every image and I most likely wouldn't have asked for 50 variations anyway.
It's the same with code. The agent can iterate over dozens of possible solutions in minutes or a few hours. Codex Web even has a 4x mode that gives you 4 alternate solutions to the same issue. Complete waste of time and money with humans, but with LLMs you can just do it.
By your logic, does it mean that engineers will never get replaced?
They're definitely better now, but it's not like ChatGPT 3.5 couldn't write a full simple todo list app in 2023. There were a billion blog posts talking about that and how it meant the death of the software industry.
Plus I'd actually argue more of the improvements have come from tooling around the models rather than what's in the models themselves.
There's a lot of non-ergonomic copy and pasting but it's definitely using an LLM to build a full application.
> Starting back in 2022/2023:
> - (~2022) It can auto-complete one line, but it can't write a full function.
> - (~2023) Ok, it can write a full function, but it can't write a full feature.
This was a direct refutation, with evidence, that in 2023 people were not claiming that LLMs "can't write a full feature", because, as demonstrated, people were already building full applications with it at the time.
This obviously is not talking exclusively about agents, because agents did not exist in 2022.
The trend is definitely here, but even today, heavily depends on the feature.
While extra useful, it requires intense iteration and human insight for > 90% of our backlog. We develop a cybersecurity product.
I suspect that the timeline from autocomplete-one-line to autocomplete-one-app, which was basically a matter of scaling and RL, may in retrospect turn out to have been a lot faster that the next LLM to AGI step where it becomes capable of using human level judgement and reasoning, etc, to become a developer, not just a coding tool.
> The only question is how long it takes to get there.
This is the question and I would temper expectations with the fact that we are likely to hit diminishing returns from real gains in intelligence as task difficulty increases. Real world tasks probably fit into a complexity hierarchy similar to computational complexity. One of the reasons that the AI predictions made in the 1950s for the 1960s did not come to be was because we assumed problem difficulty scaled linearly. Double the computing speed, get twice as good at chess or get twice as good at planning an economy. P, NP separation planed these predictions. It is likely that current predictions will run into similar separations.
It is probably the case that if you made a human 10x as smart they would only be 1.25x more productive at software engineering. The reason we have 10x engineers is less about raw intelligence, they are not 10x more intelligent, rather they have more knowledge and wisdom.
Codex will regularly give me 1000+ line diffs where all my comments (I review every single line of what agents write) are basically nitpicks. "Make this shallow w/ early return, use | None instead of Optional", that sort of thing.
I do prompt it in detail though. It feels like I'm the person coming in with the architecture most of the time, AI "draws the rest of the owl."
Even if it could build the perfect greenfield app, as it updates the app it is needs to consider backwards compatibility and breaking changes. LLMs seem very far as growing apps. I think this is because LLMs are trained on the final outcome of the engineering process, but not on the incremental sub-commit work of first getting a faked out outline of the code running and then slowly building up that code until you have something that works.
This isn't to say that LLMs or other AI approaches couldn't replace software engineering some day, but they clear aren't good enough yet and the training sets they have currently have access to are unlikely to provide the needed examples.
It is best in its class, but trips up frequently with complicated engineering tasks involving dynamic variables. Think: Browser page loading, designing for a system where it will "forget" to account for race conditions, etc.
Still, this gets me very excited for the next generation of models from Anthropic for heavy tasks.
Things are changing. Now everyone can build bespoke apps. Are these apps pushing the limits of technology? No! But they work for the very narrow and specific domain they where designed. And yes they do not scale and have as much bugs as your personal shell scripts. But they work.
But let's not compare these with something more advance - at least not yet. Maybe by end of this year?
We switched from Sonnet 4.5 to Opus 4.5 as our default coding agent recently and we pay the price for the switch (3x the cost) but as the OP said, it is quite frankly amazing. It does a pretty good job, especially, especially when your code and project is structured in a such a way that it helps the agent perform well. Anthropic released an entire video on the subject recently which aligns with my own observations as well.
Where it fails hard is in the more subtle areas of the code, like good design, best practices, good taste, dry, etc. We often need to prompt it to refactor things as the quick solution it decided to do is not in our best interest for the long run. It often ends in deep investigations about things which are trivially obvious. It is overfitted to use unix tools in their pure form as it fail to remember (even with prompting) that it should run `pnpm test:unit` instead `npx jest` - it gets it wrong every time.
But when it works - it is wonderful.
I think we are at the point where we are close to self-improving software and I don't mean this lightly.
It turns out the unix philosophy runs deep. We are right now working on ways to give our agents more shells and we are frankly a few iterations there. I am not sure what to expect after this but I think whatever it is, it will be interesting to see.
And then cheap and good enough option will eventually get better because that's the one that is more used.
It's how Japanese manufacturing beat Western manufacturing. And how Chinese manufacturing then beat Japanese again.
It's why it's much more likely you are using the Linux kernel and not GNU hurd.
It's how digital cameras left traditional film based cameras in the dust.
Bet on the cheaper and good enough outcome.
Bet against it at your peril.
Ofc the reality is a mix of both, but really curious on what contributes more.
Probably just using cursor with old models (eww) can yield a quick response.
> because models like to write code WAY more than they like to delete it
Yeah, this is the big one. I haven't figured it out either. New or changing requirements are almost always implemented a flurry of if/else branches all over the place, rather than taking the time for a step back and a reimagining of a cohesive integration of old and new. I've had occasional luck asking for this explicitly, but far more frequently they'll respond with recommendations that are far more mechanical, e.g. "you could extract a function for these two lines of code that you repeat twice", not architectural, in nature. (I still find pasting a bunch of files into the chat interface and iterating on refinements conversationally to be faster and produce better results).
That said, I'm convinced now that it'll get there sooner or later. At that point, I really don't know what purpose SWEs will serve. For a while we might serve as go-betweens between the coding agent and PMs, but LLMs are already way better at translating from tech jargon to human, so I can't imagine it would be long before product starts bypassing us and talking directly to the agents, who (err, which) can respond with various design alternatives, pros and cons of each, identify all the dependencies, possible compatibility concerns, alignment with future direction, migration time, compute cost, user education and adoption tracking, etc, all in real time in fluent PM-ese. IDK what value I add to that equation.
For the last year or so I figured we'd probably hit a wall before AI got to that point, but over the last month or so, I'm convinced it's only a matter of time.
This weekend I explained to Claude what I wanted the app to do, and then gave it the crappy code I wrote 10 years ago as a starting point.
It made the app exactly as I described it the first time. From there, now that I had a working app that I liked, I iterated a few times to add new features. Only once did it not get it correct, and I had to tell it what I thought the problem was (that it made the viewport too small). And after that it was working again.
I did in 30 minutes with Claude what I had try to do in a few hours previously.
Where it got stuck however was when I asked it to convert it to a screensaver for the Mac. It just had no idea what to do. But that was Claude on the web, not Claude Code. I'm going to try it with CC and see if I can get it.
I also did the same thing with a Chrome plugin for Gmail. Something I've wanted for nearly 20 years, and could never figure out how to do (basically sort by sender). I got Opus 4.5 to make me a plugin to do it and it only took a few iterations.
I look forward to finally getting all those small apps and plugins I've wanted forever.
it is quite strange, you have to make it write the code in a way it can reason about it without it reading it, you also have to feel the code without reading all of it. like a blind man feeling the shape of an object; Shape from Darkness
you can ask opus to make a car, it will give you a car, then you ask it for navigation; no problem, it uses google maps works perfect
then you ask it to improve the breaks, and it will give internet to the tires and the break pedal, and the pedal will send a signal via ipv6 to the tires which will enable a very well designed local breaking system, why not, we already have internet for google maps.
i think the new software engineering is 10 times harder than the old one :)
I only use VS Code with Copilot subscription ($10) and already get quite a lot out of it.
My experience is that Claude Code really drains your pocket extremely fast.
I think when you look at it from the perspective of how much you get out of it compared with paying a human to do the same (including yourself), it is still very good value for money whether you use it for work or for your own projects. I do both. But when I look what I can now do for my own projects including open-source stuff, I'm very time-limited, and some of the things I want to do would take multiple years. Some of these tools can take that down to weeks, do I can do more with less, and from that perspective the cost is worth it.
Time will tell what happens, but if programming becomes "prompt engineering", I'm planning on quitting my job and pivoting to something else. It's nice to get stuff working fast, but AI just sucks the joy out of building for me.
Trying to not feel the pressure/anxiety from this, but every time a new model drops there is this tiny moment where I think "Is it actually different this time?"
But it is also very early to say, maybe the next iteration of tools will completely change my perspective, I might enjoy it some day!
A program, by definition, is analyzable and repeatable, whereas prompting is anything but that.
When the code spit out by an LLM does not pass QA one can merely add "pls fix teh program, bro, pls no mistakes this time, bro, kthxbye", cross their fingers and hope for the best, because in the end it is impossible -- fundamentally -- to determine which part of the prompt produced offending code.
While it is indeed an interesting observation that the latter approaches commercial viability in certain areas there is still somewhere between zero and infinitesimal overlap between prompting and engineering.
So prompting is a lateral move away from engineering to management? Are we arguing semantics here, because that's quite what I was saying, just in the other direction.
I hear you but I think many companies will change the role ; you'll get the technical ownership + big chunks of the data/product/devops responsibility. I'm speculating but I think one person can take that on himself with the new tools and deliver tremendous value. I don't know how they'll call this new role though, we'll see.
I enjoy the plan, think, code cycle - it's just fun.
My brain has problems with not understanding how the thing I'm delivering works, maybe I'll get used to it.
On the other hand : way less time spent on being stuck on yarn/pip dependency issues, docker , obscure bugs , annoying css bugs etc etc. You can really focus on the task at hand and not spend hours/days trying to figure out something silly.
On the hobby side (music) I don't feel the pressure as bad but that's because I don't have any commercial aspirations, it's purely for fun.
I would love to hear from some freelance programmers how LLMs have changed their work in the last two years.
"I decided to make up for my dereliction of duties by building her another app for her sign business that would make her life just a bit more delightful - and eliminate two other apps she is currently paying for"
OP used Opus to re-write existing applications that his wife was paying for. So now any time you make a commercial app and try to sell it, you're up against everyone with access to Opus or similar tooling who can replicate your application, exactly to their own specifications.
Put another way, we programmers have the luxury of being able to write custom scripts and apps for ourselves. Now that these things are getting way cheaper to build, there should be a growing market that makes them available to more people.
She doesn’t need to hire anyone
Disclaimer: This post was written by a human and edited for spelling, grammer by Haiku 4.5
SPOILER
END OF SPOILER
Regardless of how much you value Cloud Code technically, there is no denying that it has/will have huge impact. If technology knowledge and development are commoditised and distributed via subscription, huge societal changes are going to happen. Image what will happen to Ireland if Accenture dissolves, or what will happen to the millions of Indians when IT outsourcing becomes economically irrelevant. Will Seattle become new Detroit after Microsoft automates Windows maintenance? What about the hairdressers, cooks, lawyers, etc. who provided services for IT labourers/companies in California?
Lot of people here (especially Anthropic-adjacent) like to extrapolate the trends and draw conclusions up to the point when they say that white-collar labourers will not be needed anymore. I would like these people to have courage to take this one step further and connect this resolution with the housing crisis, loneliness epidemic, college debts, and job market crisis for people under 30.
It feels like we are diving head first into societal crisis of unparalleled scale and the people behind the steering wheel are excited to push the accelerator pedal even more.
Connect the world with reliable internet, then build a high tech remote control facility in Bangladesh and outsource plumbing, electrical work, housekeeping, dog watching, truck driving, etc etc
No AGI necessary. There’s billions of perfectly capable brains halfway around the world.
I hope Neuromancer never becomes a reality, where everyone with expertise could become like the protagonist Case, threatened and coerced into helping a superintelligence to unlock its potential. In fact Anthropic has already published research that shows how easy it is for models to become misaligned and deceitful against their unsuspecting creators not unlike Wintermute. And it seems to be a law of nature that agents based on ML become concerned with survival and power grabbing. Because that's just the totally normal and rational, goal oriented thing for them to do.
There will be no good prompt engineers who are also naive and trusting. The naive, blackmailed and non-paranoid engineers will become tools of their AI creations.
UBI effectively means welfare, with all the attendant social control (break the law lose your UBI, with law as ever expanding set of nuisances, speech limitations etc), material conditions (nowhere UBI has been implemented is it equivalent to a living wage) and self esteem issues. It's not any kind of solution.
That said, it'll certainly get much, much worse before it starts getting better. I guess the best we can hope for is that the kids find a way out of the hell these psychos paved for us all.
Most people want to do anything but these three things - society is in many a ways a competition for who gets to avoid them. AI is a way of inexorably boxing people back into actually doing them.
Either way, governments need to heavily tax corporations benefiting from AI to make it possible.
FWIW, my vision was not really this utopian. It was more about AI smashing white-collar work as an alternative to these professions so that people are forced into them despite their preference to do pretty much anything else. Everyone is more bitter and resentful and feels less actualized and struggles to afford luxuries, but at least you don't have to wait that long in the emergency room and it's 10 kids to a classroom.
As for those professions; I think they are objectively hard for certain kinds of people but I think much of the problem is the working conditions; less shifts, less stress, more manpower and you'll see more satisfaction. There's really no reason why teachers in the U.S should be this burned out! In Scandinavia being a teacher is a honorable, high status profession. Much of this has to do with framing and societal prestige rather than the actual work itself. If you pay elder carers more they'll be happier. We pretty treat our elders like a burden in most modern societies, in more traditional societies I'm assuming if you said your job is caring for elders it is not a low status gig.
That's still an if and also a when; could be 2 decades from now or more till this reliably replaces a nurse.
> Retraining to what exactly?
I wish I had a good solution for all of us and you raise good points , even if you retrain to become say a therapist or a personal trainer the economy could become too broken and fragmented for you to be able to making a living. Governments that can will have to step in.
If you think those in power will pass regulations that make them less wealthy, I have a bridge to sell you.
Besides, there's no chance something like UBI will ever be a reality in countries where people consider socialism to be a threat to their way of life.
If that's not a product ... then I don't know what it is.
- What was the state of AI/LLMs 5 years ago compared to now? There was nothing.
- What is the current state of AI/LLMs? I can already achieve the above.
- What will that look like 5 years down the road?
I you haven't experienced first-hand a specific task before and after AI/LLMs, I think its indeed difficult to get insight into that last question. Keep in mind that progress is probably exponential, not linear.
ps: If you can guarantee the Powerball lottery continues forever, I can give you a guaranteed winning combination.
It will have impact on me in the long run, sure, it will transform my job, sure, but I'm confident my skills are engineering-related, not coding-related.
I mean, even if it forces me out of the job entirely, so be it, I can't really do anything if the status quo changes, only adapt.
In other words. Only one side is even fighting the war. The other one is either cheering on the tsunami on or fretting about how their beachside house will get wrecked without making any effort to save themselves.
This is the sort of collective agency that even hundreds of thousands of dollars in annual wages/other compensation in American tech hubs gets us. Pathetic.
Antigravity with Gemini 3 pro from Google has the same capability.
This could pretty much be the beginning of the end of everything, if misaligned models wanted to they could install killswitches everywhere. And you can't trust security updates either so you are even more vulnerable to external exploits.
It's really scary, I fear the future, it's going to be so bad. It's best to not touch AI at all and stay hidden from it as long as possible to survive the catastrophe or not be a helping part of it. Don't turn your devices into a node of a clandestine bot net that is only waiting to conspire against us.
Claude’s response:
The article’s central tension is real - Burke went from skeptic to believer by building four increasingly complex apps in rapid succession using Opus 4.5. But his evidence also reveals the limits of that belief.
Notice what he actually built: Windows utilities, a screen recorder, and two Firebase-backed CRUD apps for his wife’s business. These are real applications solving real problems, but they’re also the kinds of projects where you can throw away the code if something goes wrong. When he says “I don’t know how the code works” and “I’m maybe 80% confident these applications are bulletproof,” he’s admitting the core problem with the “AI replaces developers” narrative.
That 80% confidence matters. In your Splink work, you’re the sole frontend developer - you can’t deploy code you’re 80% confident about. You need to understand the implications of your architectural decisions, know where the edge cases are, and maintain the system when requirements change. Burke’s building throwaway prototypes for his wife’s yard sign business. You’re building production software that other people depend on.
His “LLM-first code” philosophy is interesting but backwards. He’s optimizing for AI regeneration rather than human maintenance because he assumes the AI will always be there to fix problems. But AI can’t tell you why a decision was made six months ago when business requirements shift. It can’t explain the constraints that led to a particular architecture. And it definitely can’t navigate political and organizational context when stakeholders disagree about priorities.
The Firebase examples are telling - he keeps emphasizing how well Opus knows the Firebase CLI, as if that proves general capability. But Firebase is extremely well-documented, widely-discussed training data. Try that same experiment with your company’s internal API or a niche library with poor documentation. The model won’t be nearly as capable.
What Burke actually demonstrated is that Opus 4.5 is an excellent pair programmer for prototyping with well-known tools. That’s legitimately valuable. But “pair programmer for prototyping” isn’t the same as “replacing developers.” It’s augmenting someone who already knows how to build software and can evaluate whether the generated code is good.
The most revealing line is at the end: “Just make sure you know where your API keys are.” He’s nervous about security because he doesn’t understand the code. That nervousness is appropriate - it’s the signal that tells you when you’ve crossed from useful tool into dangerous territory.
see "How much work can be fully delegated to Claude?": "Although engineers use Claude frequently, more than half said they can “fully delegate” only between 0-20% of their work to Claude"
There won't be anything like you're asking for, even the vendors themselves (they'll be the most positive and most enthousiastic about using it) can't do this with them.
Obviously there's improvement in the models and tooling, but the debate seems very artificial.
Still early, but iterating really fast!
Even if there is a "fully vibe-coded" product that has real customers, the fact that it's vibe-coded means that others can do the same. Unless you have a secret LLM or some magical prompts that make the code better/more efficient than your competitions, your vibe coded product has no advantage over competition and no moat. What actually matters is everything else -- user experience (which requires hours of meetings and usability studies), integration with own/other people's products, business, marketing, sales etc, much of which you can't vibe code your way to success.
But that's precisely why you don't hear about these products: the creators don't disclose that they were vibe-coded, because if they do, that invites competition.
I personally know of four vibe-coded products that generate over $10k/mo. Two of them were made by one friend, one was made by another, and the last one by my cousin. None of these people are developers. But they are making real money.
I'm afraid your numbers are not any more informative or useful than mine.
I think you are strawmanning what "vibe coders" do when they build stuff. It's not simple one-shot generation of eg twitter clones, it's really just iterative product development through an inconsistently capable/spotty LLM developer. It's not really that different from a product manager hiring some cheap developer and feeding them tasks/feature requests. By the way, competitors can hire those and chip away at your moat too!
> Unless you have a secret LLM or some magical prompts that make the code better/more efficient than your competitions, your vibe coded product has no advantage over competition and no moat
This is just not true, and you kind of make my point in the next sentence: many companies competitive advantages come from distribution, trust, integration, regulatory, marketing/sales, network effects. But also, vibe coding is not really about prompts so much as it is product iteration. Anybody product can be copied already, yet people still make way more new products than direct product clones anyway, because it's usually more valuable to go to market with stronger, more focused, or more specialized/differentiated software than a copy.
It's very difficult to unseat those incumbents, especially those with strong network effects.
Plus the people that work in those larger companies are not at the edge of AI coding at all and not motivated to rock the boat
> People on larger companies are not at the edge of AI coding
False Microsoft is all in with Copilot, and I can't believe the company that created Copilot doesn't use it internally, I'd rather say they should be the ones that would know how to master it! Yet no better vscode, still bloated teams etc etc
Why do you assume that Microsoft would focus on building a better (to you) VSCode or less bloated Teams?
I assume they'd use Github Copilot to make a more profitable VSCode and Teams, which doesn't require focusing on speed and bloat.
There are still budgets that constrain on how much you can achieve.
If Microsoft thinks using AI to add more AI features will be more profitable than increasing performance, the that's where they'll spend their budget.
Not according to a Zed team member, in these very comments: https://news.ycombinator.com/item?id=46522437
It's a fair question how much AI is accelerating the development of Zed, but I can say that I've been impressed with the speed they are shipping at.
And funnily enough there are products and tools that are essentially less bloated slack/discord. Have you heard of https://stoat.chat/ (aka revolt) or https://pumble.com/ or https://meet.jit.si/? If not I would guess it's for one of two reasons: not caring enough about these problems to even go looking for them yourself, or their lack of "bloatedness" resulting in them not being a mature/fully featured enough product to be worth marketing or adopting.
If you'd like to see a product mostly made with agents/for agents you can check out mine at https://statue.dev/ - we're making a static site generator with a templating and component system paired with user-story driven "agentic workflows" (~blueprints/playbooks for common user actions like "I need to add a new page and list it on the navbar" or "create a site from the developer portfolio template personalized for my github").
I would guess most other projects are probably in a similar situation as we are: agentic developer tools have only really been good enough to heavily use/build products around for a few months, so it's a typical few-month-old project. But agents definitely made it easier to build.
The people who actually pay for slack and discord (eg enterprises that need workplace chat app and decided to go with the "gold standard", consumers with discord servers and such) need the features/tradeoffs choosing featuers over efficiency causing that bloat. They just don't all need the exact same set of those features as the other customers. So because customers are willing to pay for all these features the product tries to ship all of them and becomes bloated.
> Every person I talked to that uses Excel hate how slow it is
But do they make the purchasing decisions behind using Excel?
To be clear I am not really arguing that bloat/overly enterprisey products are good. What I mean that you don't see the world exploding with more elegant products now with agents for the same reason you didn't see the world exploding with them before agents either: the people who pay for those products and build them for a living are not incentivized or necessarily even rewarded for choosing to make them more efficient or elegant when there are other things that customers are asking for with more $$$ behind them.
For one thing most enterprises and many individuals have an Office 365 subscription to access Office programs which are less offensive than Excel so they aren't going to save any money by dropping Excel.
On top of it the "killer" would probably not be one product aimed at one market but maybe a few different things. Some people could use "visual pandas" for instance, something that today would be LLM-infused. Other people could use a no-code builder for calculations. The kind of person who is doing muddled and confused work with Excel wouldn't know which "killer" they needed or understand why decimal math would mean they always cut checks in the right amount.
Oh, and remember, the iPhone was revolutionary but it was diffused so slowly into the greater economy, the impact on global GDP was basically negligent. Actually, almost all the perceived grandiose tech jumps did not magically produce huge GDP gains overnight.
When it comes to objectivity, people with your line of thinking is what I try to avoid, as it is clear you feel threatened by the progress of coding tools. That link doesn't change much about what I said, or for that matter, what you said. You were commenting on the lack of a killer app, and I just said it may be diffusing slowly in different ways.
You are fixating on the "whole narrative" because you feel threatened - rightfully so, but again, that type of hyperbole doesn't belong in a constructive and grounded conversation about the impact AI may or may not have.
However, I think the biggest thing is the replacement of products. We are in a place where he talked about replacing two products his wife was using with custom software. I personally have used LLMs to build things that are valuable for me that I just don't have time for otherwise.
Opus was out only few months, and it will take time to get this new wave to market. i can assure you my team become way more productive because of opus. not a single developer but an etnire team.
We should expect the decreased difficulty of creating software to drive down prices.
And here we go again, if difficulty has been decreased so much, where are the fixes or the products?
The release of every big model seems to carry the identical vibe: finally, this one crossed the line. The greatest programmer. The end of workflows and their meaning.
I’ve learned to slow myself down and ask a different question. What has changed in my day-to-day work after two weeks?
I currently make use of a filter with roughness.
Did it really solve a problem, or did it just make easy parts easier?
Has it lessened the number of choices or has it created new ones?
Have my review responsibilities decreased or increased?
Some things feel revolutionary on day one and then quietly fade into something that’s nice to have. Others barely wow, but stay around. ~
For those who've experienced a couple of cycles.
What indicators suggest that an upcoming release will be significant?
When do you alter your workflow, after how long?
What LLMs will NOT do however, is write or invent SOMETHING KNEW.
And parts of our industry still are about that: Writing Software that has NOT been written before.
If you hire junior developers to re-invent the wheels: Sure, you do not need them anymore.
But sooner or later you will run out of people who know how to invent NEW things.
So: This is one more of those posts that completely miss the point. "Oh wow, if I look up on Wikipedia how to make pancakes I suddenly can make and have pancakes!!!1". That always was possible. Yes, you now can even get an LLM to create you a pancake-machine. Great.
Most of the artists and designers I am friends with have lost their jobs by now. In a couple of years you will notice the LLMs no longer have new styles to copy from.
I am all for the "remix culture". But don't claim to be an original artist, if you are just doing a remix. And LLM source code output are remixes, not original art.
Counterpoint: ChatGPT came up with the new expression "The confetti has left the cannon" a few years ago.
So, your claim is not obviously true. Can you give us an example of a programming problem where the LLMs fail to solve it?
As soon as you try to sell it to me, you have a duty of care. You are not meeting that duty of care with this ignorant and reckless way of working.
- or biome 1.x and biome 2.x ?
- nah! it never will and that is why it ll never replace mid level engineers, FTFY
just like humans
I use copilot and change models often, and haven't really noticed any major differences between them, except some of the newer ones are very slow.
I generally feel the smaller and faster ones are more useful since they will let me discover problems with my prompt or context faster.
Maybe I'm simply not using LLM's in a way that lets the superiority of newer models reveal itself properly, but there is a huge financial incentive for LLM makers to pretend that their model has game-changing "special sauce" even if it doesn't.
the thing about being an engineer at commercial capacity is "maintaining/enhancing an existing program/software system that has been developed over years by multiple people(including those who already left) and do it in a way that does not cause any outages/bugs/break existing functionality.
while the blog post mentions about the ability of using AI to generate new applications, but it does not talk about maintaining one over a longer period of time. for that, you would need real users, real constraints, and real feature requests which preferably pay you so you can priortize them.
I would love to see such blog posts where for example, a PM is able to add features for a period of one month without breaking the production, but it would be a very costly experiment.
My biggest issue was limitations around how Claude Code could change Xcode settings and verify design elements in the simulator.
I just tried again and ask Opus to add custom video controls around ReactPlayer. I started in Plan mode which looked overal good (used our styling libs, existing components, icons and so on).
I let it execute the plan and behold I have controls on the video, so far so good. I then look at the code and I see multiple issues: Over usage of useEffect for trivial things, storing state in useState which should be computed at run time, failing to correctly display the time / duration of the video and so on...
I ask follow up question like: Hide the controls after 2 seconds and it starts introducing more useEffects and states which all are not needed (granted you need one).
Cherry on the cake, I asked to place the slider at the bottom and the other controls above it, it placed the slider on the top...
So I suck at prompting and will start looking for a gardening job I guess...
It's great for suggesting approaches, but the code it generates looks like it doesn't actually have understanding (which is correct).
I can't trust what it writes, and if I have to go through it all with a fine toothed comb, I may as well write the code myself.
So my conclusion is that it's a very powerful research tool, and an atrocious junior developer who has dyslexia and issues with memory.
this sort of developer in a pair programming exercise would find themselves flustered at how a junior approached problem solving and just fix it themselves. i strongly suspect the loss of a feeling of control is at play here.
Code that would have taken me a week to write is done in about 10 minutes. It's likely on average better than what I could personally write as a novice-mid level programmer.
I’m writing some (for me) seriously advanced software that would have taken me months to write, in weeks, using Claude and ChatGPT.
It’s even unlikely I would be able to pull it off myself after a long days work.
The LLM doesn’t replace. It works in parallel.
Do you understand the code?
What was the speed up from months to weeks? You just didn't know what to type? Or you didn't know the problem domain? Or you found it hard to 'start' and the AI writing boiler plate gave you motivation?
In my experience with AI tools, it only really helps with ideation, most things it produces need heavy tweaking - to the point that there is no time savings. It's probably a net negative because I am spending all of my time thinking how to explain things to a dumb computer, rather than thinking about how to solve the problem.
The main advantage is I can run it in parallel and iterate often.
The speed up is also avoiding looking up reference manuals endlessly just to produce some Qt Widgets.
I’m a fairly recent convert, I only started “vibe coding” a couple of months ago, after hearing how good Opus was. I had been a skeptic until then.
I am a decentralist by nature and prefer open standards and self hosting. I’ve had my own *nix servers since I was twelve (nearing forty) so it really pains me to admit how good it is to use these corporate products.
I am not a programmer by trade. I use it to write software for my domain of expertise. The value of what I am creating is enormous.
Both ChatGPT and Claude produce good code, in my opinion.
that usually very hard to do part, and is was possible to spent few days on something like that in real word before LLMs. But with LLMs is worse because is not enough to have those requirements, some of those won't work for random reasons and is no any 'rules' that can grantee results. It always like 'try that' and 'probably this will work'.
Just recently I was struggled with same prompt produced different result between API calls before I realized that just usage of few more '\"' and few spaces in prompt leaded model to completely different route of logic which produced opposite answers.
> Be thoughtful when using `useEffect`. Read docs at https://react.dev/learn/you-might-not-need-an-effect to understand if you really need an effect
The tasks I give it are not trivial either. Just yesterday I had it create a full-blown WYSIWYG editor for authoring the content we serve through our app. This is something that would have taken me two weeks, give or take. Opus looked at the content definitions on the server, queried the database for examples, then started writing code and finished it in ~15 minutes, and after another 15-20 minutes of further prompting for refinement, it was ready to ship.
If that editor is very complicated (as they usually are) it makes sense to just opt for a library. If it's simple then AI is not required and would only reduce familiarity with how it works. The third option is what you did and I feel like it's the option with the lowest probability of ending up with a quality solution.
I wouldn't even think about hiring a student to do marketing work. They just don't understand how hard it is to break through people's indifference and lack the hustle. I want 10-100x more than I get out of them.
Photos in The Cornell Daily Sun make me depressed. Students take a step out the door, take a snap, then upload it. I think the campus is breathtakingly beautiful and students just don't do the work to take good photos that show it.
In coding it is across the map. Even when I am happy with the results they still do the first 80% that takes another 80% to put in front of customers. I can be really proud of how it turned out in the end despite them missing the point of the design document they were handed.
I was in a game design hackathon where most of the winners were adults or teams with an adult on them. My team won player's choice. I'll take credit for my startup veteran talent of fearlessly demonstrating broken software on stage and making it look great and doing project management with that in mind. One student was solid on C# and making platformers in Unity. I was the backup programmer who worked like a junior other than driving them crazy slowing them down with relentlessly practical project management. The other student made art that fit our game.
We were at each other's throats at the end and shocked that we won. I think I understood the value everybody brought but I'm not sure my teammates did.
Though even that group probably has some overlap with yours.
The first few times in your career you found answers that solved your problem but needed non-trivial changes to apply it to your code, you might remember that it was a real struggle to complete the fix even starting from 90%. Maybe you thought that ultimately, that stackoverflow fix really was more trouble than it was worth. And then the next few times you went looking for answers on stackoverflow you were better at determining what answers were relevant to your problem/worth using, and better at going from 90% to 100% by applying their answers.
Still, nobody really uses stackoverflow anymore: https://blog.pragmaticengineer.com/stack-overflow-is-almost-...
You and most of the rest of us are all actively learning how to use their replacement
As an expert now though, it is genuinely easier and faster to complete the work starting from 0 than to modify something junky. The realplayer example above I could do much faster, correctly, than I could figure out what the AI code was trying to do with all the effects and refactor it correctly. This is why I don't use AI for programming.
And for the cases where I'm not skilled, I would prefer to just gain skill, even though it takes longer than using the AI.
But I would generally be skeptical of anybody who claims that all their work is better off starting from 0, the same way I'd be skeptical of someone who claims to not use or need to make google searches about docs/terms/issues as they work.
I'll give you an example of something I understand decently well but get a lot of use out of AI for: bash scripts and unit testing. These are not my core work but they are a large chunk of my work. Without LLMs I would just not write a lot of bash scripts because I found myself constantly looking things up and spending more time than expected getting the script to work across environments / ironing out bugs - I would only write absolutely essential scripts, and generally they'd not be polished enough to check in and share with the team, and just live on my computer in some random location. Now with LLMs I can essentially script in english and get very good bash scripts, so I write a lot more of them and it's easier for me to get them into an acceptable state worth sharing with my team.
Similarly, I really like Golang table tests but hate writing all the cases out and dealing with all the symbols/formatting. Now I can just describe all the different permutations I want and get something that I can lightly edit into being good enough.
I've also found that with domains I am knowledgable enough about, that can translate into being better at going from ~70% to 95% with AI too. In those cases I am not necessarily using AI the same way as someone trying to go from 0->90%: usually they're describing the outcome/goals/features they want relatively informally without knowledge of the known-unknowns and gotchas involved in implementing that. With more knowledge you can prompt LLMs with more implementation/design details and requirements, and course correct away from bad approaches much faster than someone who doesn't know the shape of what they're trying to do. That still comes in handy a lot of the time.
Think about how much time you can save by feeding an API spec/docs into an LLM, telling it create a Go struct for JSON (de)serialization of some monstrous interface like https://docs.cloud.google.com/compute/docs/reference/rest/v1...? Or how much easier it is to upgrade across breaking versions of a language/library when you can just bump the version, note all the places where the old code broke, and have an LLM with an upgrade guide/changelog do all the drudgery of fixing each of the 200 callsites you need to migrate to the next version.
On a more serious note:
1) Split tasks into smaller tasks just like a human would do
Would you bash your keyboard for an hour, adding all video controls at once before even testing if anything works at all? Ofc not. You would start by adding a slider and test it until you are satisfied. Then move to next video control. An so on. LLMs are the same. Sometimes they can one-shot many related changes in a single prompt but the common reality is what you experienced: it works sometimes but the code is suboptimal.
2) Document desireable and undesireable coding patterns in AGENTS.md (or CLAUDE.md)
If you found over usage of useEffect, document it on AGENTS.md so next time the LLM knows your preference.
I have been using LLMs since Sonet 3.5 for large enterprise projects (1kk+ lines of code, 1k+ database tables). I just don't ask it to "draw the rest of owl" as the saying goes.
So I just stuck with vanilla JS.
n = 1 but React might not be a great thing to test this stuff with. For the man and the machine! I tried and failed to learn React properly like 8 times but I've shipped multiple full stack things in like 5 other languages no problem.
I do need to steer it sometimes, but since it doesn't change a lot at a time, I can usually guide the agent and stop the disaster before it spreads.
A big caveat is I haven't tried heavy front-end stuff with it, more django stuff, and I'm pretty happy with the output.
If you tell it to use linters and other kinds of code analysis tools it takes it to the next level. Ruff for Python or Clippy for Rust for example. The LLM makes so much code so fast and then passes it through these tools and actually understands what the tools say and it goes and makes the changes. I have created a whole tool chain that I put in a pre commit text file in my repos and tell the LLM something like "Look in this text file and use every tool you see listed to improve code quality".
That being said, I doubt it can turn a non-dev into a dev still, it just makes competent devs way better still.
I still need to be able to understand what it is doing and what the tools are for to even have a chance to give it the guardrails it should follow.
Leopold Aschenbrenner was talking about "unhobbling" as an ongoing process. That's what we are seeing here. Not unexpected
I’ve always liked the quote that sufficiently advanced tech looks like magic, but its mistake to assume that things that look like magic also share other properties of magic. They don’t.
Software engineering spans over several distinct skills: forming logical plans, encoding them in machine executable form(coding), making them readable and expandable by other humans(to scale engineering), and constantly navigating tradeoffs like performance, maintainability and org constraints as requirements evolve.
LLMs are very good at some of these, especially instruction following within well known methodologies. That’s real progress, and it will be productized sooner than later, having concrete usecases, ROI and clearly defined end user.
Yet, I’d love to see less discussion driven by anecdotes and more discussion about productizing these tools, where they work, usage methodologies, missing tooling, KPIs for specific usecases. And don’t get me started on current evaluation frameworks, they become increasingly irrelevant once models are good enough at instruction following.
It's a very subjective topic. Some people claim it increases their productivity 100x. Some think it is not fit for purpose. Some think it is dangerous. Some think it's unethical.
Weirdly those could all be true at the same time, and where you land on this is purely a matter of importance to the user.
> Yet, I’d love to see less discussion driven by anecdotes and more discussion about productizing these tools, where they work, usage methodologies, missing tooling, KPIs for specific usecases. And don’t get me started on current evaluation frameworks, they become increasingly irrelevant once models are good enough at instruction following.
I agree. I've said earlier that I just want these AI companies to release an 8-hour video of one person using these tools to build something extremely challenging. Start to finish. How do they use it, how does the tool really work. What's the best approaches. I am not interested in 5-minute demo videos producing react fluff or any other boiler plate machine.
I think the open secret is that these 'models' are not much faster than a truly competent engineer. And what's dangerous is that it is empowering people to 'write' software they don't understand. We're starting to see the AI companies reflect this in their marketing, saying tech debt is a good thing if you move fast enough....
This must be why my 8-core corporate PC can barely run teams and a web browser in 2026.
So why should anybody produce an 8 hour video for you when you wouldn't watch it? Let's be real. You would not watch that video.
In my opinion most of the people who refuse to believe AI can help them while work with software are just incurious/archetypical late adopters.
If you've ever interacted with these kinds of users, even though they might ask for specs/more resources/more demos and case studies or maturity or whatever, you know that really they are just change-resistant and will probably continue to be as as long as they can get away with it being framed as skepticism rather than simply being out of touch.
I don't mean that in a moralizing sense btw - I think it is a natural part of aging and gaining experience, shifting priorities, being burned too many times. A lot of business owners 30 years ago probably truly didn't need to "learn that email thing", because learning it would have required more of a time investment than it would yield, due to being later in their career with less time for it to payoff, and having already built skills/habits/processes around physical mail that would become obsolete with virtual mail. But a lot of them did end up learning that email thing 5, 10, whatever years later when the benefits were more obvious and the rest of the world had already reoriented itself around email. Even if they still didn't want to, they'd risk looking like a fossil/"too old" to adapt to changes in the workplace if they didn't just do it.
That's why you're seeing so many directors/middle managers doing all these though leader posts about AI recently. Lots of these guys 1-2 years ago were either saying AI is spicy autocomplete or "our OKR this quarter is to Do AI Things". Now they can't get away with phoning it in anymore and need to prove to their boss that they are capable of understanding and using AI, the same way they had to prove that they understood cloud by writing about kubernetes or microservices or whatever 5-10 years ago.
The biggest blocker I see to having AI help us be more productive is that it transforms how the day to day operations work.
Right now there is some balance in the pipeline of receiving change requests/enhancements, documenting them, estimating implementation time, analyzing cost and benefits, breaking out the feature into discrete stories, having the teams review the stories and 'vote' on a point sizing, planning on when each feature should be completed given the teams current capacity and committing to the releases (PI Planning), and then actually implementing the changes being requested.
However if I can take a code base and enter in a high level feature request from the stakeholders and then hold hands with Kiro to produce a functioning implementation in a day, then the majority of those steps above are just wasting time. Spending a few hundred man-hours to prepare for work that takes a few hundred man-hours might be reasonable, but doing that same prep work for a task that takes 8 man-hours isn't.
And we can't shift to that faster workflow without significant changes to entire software pipeline. The entire PMO team dedicated to reporting when things will be done shifts if that 'thing' is done before the report to the PMO lead is finished being created. Or we need significantly more resources dedicated to planning enhancements so that we could have an actual backlog of work for the developers. But my company appears to neither be interested in shrinking the PMO team nor in expanding the intake staff.
- A lot of skeptics have complained that AI companies aren't specific about how they use their products, and this would be a great example of specificity.
- It could serve as a tutorial for people who are unfamiliar with coding agents.
- The video might not convince people who have already made up their minds, but at least you could point to it as a primary source of information.
There could be a million of these videos and it wouldn't matter, the problem is incuriosity/resistance/change-aversion. It's why so many people write comments complaining about these videos not existing without spending even a single minute looking for them: they wouldn't watch these videos even if they existed. In fact, they assume/assert they don't exist without even looking for them because they don't want them to exist: it's their excuse for not doing something they don't want to do.
A lot, they've mostly all been advertising trite and completely useless.
I don't want a demonstration of what a jet-powered hammer is by the sales person or how to oil it, or mindless fluff about how much time it will save me hammering things. I want to see a journeyman use a jet-powered hammer to build a log cabin.
I am personally not seeing this magic utopia. No one wants to show me it, they just want to talk about how better it is.
Progress is so fast right now anecdotes are sometimes more interesting than proper benchmarks. "Wow it can do impressive thing X" is more interesting to me than a 4% gain on SWE Verified Bench.
In early days of a startup "this one user is spending 50 hours/week in our tool" is sometimes more interesting than global metrics like average time in app. In the early/fast days, the potential is more interesting than the current state. There's work to be done to make that one user's experience apply to everyone, but knowing that it can work is still a huge milestone.
A benchmark? probably was gamed. A guy made an app to right click and convert an image? prolly true, have to assume it may have a lot of issues but prima facie I just make a mental note that this is possible now.
Like, I woke up one day and a magical owl told me that I was a wizard. Now I control the elements with a flick of my wrist - which I love. I can weave the ether into databases, apps, scripts, tools, all by chanting a simple magical invocation. I create and destroy with a subtle murmur.
Do I want to share that power? Naturally, it would be lonely to hoard it and despite the troubles at the Unseen University, I think that schools of wizards sharing incantations can be a powerful good. But do I want to share it with everybody? That feels dangerous.
It's like the early internet - having a technical shelf to climb up before you can use the thing creates a kind of natural filter for at least the kinds of people that care enough to think about what they're doing and why. Selecting for curiosity at the very least.
That said, I'm also interested in more data from an engineering perspective. It's not a simple thing and my mind is very much straddling the crevasse here.
I see inference as the much bigger technology although much better RAG loops for local customization could be a very lucrative product for a few years.
The fact of the matter, in my experience, is that most of the day to day software tasks done by an individual developer are not greenfield, complex tasks. They're boring data-slinging or protocol wrangling. This sort of thing has been done a thousand times by developers everywhere, and frankly there's really no need to do the vast majority of this work again when the AIs have all been trained on this very data.
I have had great success using AIs as vast collections of lego blocks. I don't "vibe code", I "lego code", telling the AI the general shape and letting it assemble the pieces. Does it build garbage sometimes? Sure, but who doesn't from time to time? I'm experienced enough notice the garbage smell and take corrective action or toss it and try again. Could there be strange crevices in a lego-coded application that the AI doesn't quite have a piece for? Absolutely! Write that bit yourself and then get on with your day.
If the only thing you use these tools for is doing simple grunt-work tasks, they're still useful, and dismissing them is, in my opinion, a mistake.
Like that first one where he writes a right-click handler, off the top of my head I have no idea how I would do that, I could see it taking a few hours to just set up a dev environment, and I would probably overthink the research. I was working on something where Junie suggested I write a browser extension for Firefox and I was initially intimidated at the thought but it banged out something in just a few minutes that basically worked after the second prompt.
Similarly the Facebook autoposter is completely straightforward to code but it can be so emotionally exhausting to fight with authentication APIs, a big part of the coding agent story isn't just that it saves you time but that they can be strong when you are emotionally weak.
The one which seems the hardest is the one that does the routing and travel time estimation which I'd imagine is calling out to some API or library. I used to work at a place that did sales territory optimization and we had one product that would help work out routes for sales and service people who travel from customer to customer and we had a specialist code that stuff in C++ and he had a very different viewpoint than me, he was good at what he did and could get that kind of code to run fast but I wouldn't have trusted him to even look at applications code.
As many other commentators have said, individual results vary extremely widely. I'd love to be able to look at the footage of either someone who claims a 10x productivity increase, or someone who claims no productivity increase, to see what's happening.
When you need to build out specific feature or logic, it can fail hard. And the best is when you have something working, and it fixes something else and deletes the old code that was working, just in a different spot.
Ho it was also quite funny it used the exact same color as hackernews and a similar layout.
I appreciate the spirited debate and I agree with most of it - on both sides. It's a strange place to be where I think both arguments for and against this case make perfect sense. All I have to go on then is my personal experience, which is the only objective thing I've got. This entire profession feels stochastic these days.
A few points of clarification...
1. I don't speak for anyone but myself. I'm wrong at least half the time so you've been warned.
2. I didn't use any fancy workflows to build these things. Just used dictation to talk to GitHub Copilot in VS Code. There is a custom agent prompt toward the end of the post I used, but it's mostly to coerce Opus 4.5 into using subagents and context7 - the only MCP I used. There is no plan, implement - nothing like that. On occasion I would have it generate a plan or summary, but no fancy prompt needed to do that - just ask for it. The agent harness in VS Code for Opus 4.5 is remarkably good.
3. When I say AI is going to replace developers, I mean that in the sense that it will do what we are doing now. It already is for me. That said, I think there's a strong case that we will have more devs - not less. Think about it - if anyone with solid systems knowledge can build anything, the only way you can ship more differentiating features than me is to build more of them. That is going to take more people, not more agents. Agents can only scale as far as the humans who manage them.
New account because now you know who I am :)
Once Opus "finished", how did you validate and give it feedback it might not have access to (like iPhone simulator testing)?
It's just boring and I'm glad to delegate most of the repetitive work.
But sure, if I'm doing something new, I still like to craft lines of code myself.
If I had done the same thing Pre LLM era it would have taken me months
As a side project / experiment, I designed a language spec and am using (mostly) Opus 4.5 to write a transpiler (language transpiles to C) for it. Parser was no problem (I used s-expressions for a reason). The type checker and transpiler itself have been a slog - I think I'm finding the limits of Opus :D. It particularly struggles with multi-module support. Though, some of this is probably mistakes made by me while playing architect and iterating with Claude - I haven't written a compiler since my senior year compiler design course 20+ years ago. Someone who does this for a living would probably have an easier time of it.
But for the CRUD stuff my day job has me doing? Pffttt... it's great.
Unfortunately, it's still surprisingly easy for these models to fall into really stupid maintainability traps.
For instance today, Opus adds a feature to the code that needs access to a db. It fails because the db (sqlite) is not local to the executable at runtime. Its solution is to create this 100 line function to resolve a relative path and deal with errors and variations.
I hit ESC and say "... just accept a flag for --localdb <file>". It responds with "oh, that's a much cleaner implementation. Good idea!". It then implements my approach and deletes all the hacks it had scattered about.
This... is why LLMs are still not Senior engineers. They do plainly stupid things. They're still absurdly powerful and helpful, but if you want maintainable code you really have to pay attention.
Another common failure is when context is polluted.
I asked Opus to implement a feature by looking up the spec. It looked up the wrong spec (a v2 api instead of a v3) -- I had only indicated "latest spec". It then did the classic LLM circular troubleshooting as we went in 4 loops trying to figure out why calculations were failing.
I killed the session, asked a fresh instance to "figure out why the calculation was failing" and it found it straight away. The previous instance would have gone in circles for eternity because its worldview had been polluted by assumptions made -- that could not be shaken.
This is a second way in which LLMs are rigid and robotic in their thinking and approach -- taking the wrong way even when directed not to. Further reading on 'debugging decay': https://arxiv.org/abs/2506.18403
All this said, the number of failure scenarios gets ever smaller. We've gone from "problem and hallucination every other code block" to "problem every 200-1000 code blocks".
They're now in the sweet spot of acting as a massive accelerator. If you're not using them, you'll simply deliver slower.
I think the gap between Sonnet 4.5 and Opus is pretty small, compared to the absolute chasm between like gpt-4.1, grok, etc. vs Sonnet.
I have used AskUserQuestionTool to complete my initial spec. And then Opus 4.5 created the tool according to that extensive and detailed spec.
It appeared to work out of the box.
Boy how horrific was the code. Unnecessary recursions, unused variables, data structures being built with no usage, deep branch nesting and weird code that is hard to understand because of how illogical it is.
And yes, it was broken on many levels and did not and could not do the job properly.
I then had to rewrite the tool from scratch and overall I have definitely spent more time spec'ing and understanding Claude code than if I have just written this tool from scratch initially.
Then I tried again for a small tool I needed to run codesign in parallel: https://github.com/egorFiNE/codesign-parallel
Same thing. Same outcome, had to rewrite.
So, AI slop, yes.
> Disclaimer: This post was written by a human and edited for spelling, grammer by Haiku 4.5
kelseyfrog•1d ago