This is the person they have running GitHub?
(Yes, Ad Hominem)
He will just follow Satya's movements blindly.
A world with less developers is likely to harm the business he runs. Regardless, his tweet adds no value to anyone including him.
Right now I think it might just be a 1.5x boost, and most of the people claiming it’s more are just shovel sellers overstating the quality of near by mines.
It's only trendy now to say those things publicly without PR and media training filter.
https://tomrenner.com/posts/llm-inevitabilism/
Of course, when phrased with adversarial language, these self-proclaimed gurus and "leaders" are basically saying: "We know everything and you know not too much, get on board or get behind", to boost their sales. Many times these futures didn't work out:
- Elon's self driving cars
- Elon's Boring Tunnel
- NFTs (mega flop)
- Web3/blockchain (it happened but wasn't the revolution and future they were proclaiming)
These are just salespeople, they have no idea what they are talking about, they just want to create a need for their product and sell it to you.
Would you listen to Linus Torvalds: https://www.theregister.com/2024/10/29/linus_torvalds_ai_hyp... or to some guy you've never heard of so far. Same for the creator of curl, a tool you use every day: https://www.theregister.com/2025/07/15/curl_creator_mulls_ni.... Nah, this guy is talking BS, as many before him.
I think he might be drunk
The quote from the article "if you 10x a single developer, then 10 developers can do 100x." implies that companies have 100x the things to be developed productively, what if that isn't the case?
What if companies actually have about as many systems as they need, if you really can 10x your existing developers, then that would predicate a cut not an increase.
Of course the follow on from that, for suppliers who have per-seat licensing is that they'll need to find some other way to monetize if there are fewer seats to be sold, I guess they could start charging AI Agents as "developer seats"....
Why wouldn't it be? I believe programming is very much susceptible to Jevons paradox. I'm coding things I never would have started working on a year ago because it either would have taken too long or would have been simply impossible without an LLM (like and MCP server).
For example, say you're the CEO of one of the many many companies that view development costs as an overhead. Your CTO comes to you and tells you "hey we've just 10x the productivity of all our developers", do you think that conversation leads to the CEO saying "great hire more developers"?
FWIW I'd agree that (for now at least) it's challenging to get LLM/AI created code to avoid security vulnerabilities without careful prompting and review. Whether that's a fixable problem or whether it'll just generate mountains of new CVEs, bug bounties and ransomware attacks, remains to be seen.
> What if companies actually have about as many systems as they need
They think ~"companies don't have as many systems as they need".
My opinion? Even without GenAI, our sector has too many cooks.
I've not valued most of the stuff that was added to any consumer facing software product or service in the last decade or so. Facebook has, if anything, regressed since 2007. YouTube (itself, no comment on the content) peaked sometime around 2016. SwiftUI (2019) still isn't an improvement over UIKit. Google Maps is very marginally worse than it was in 2018. MacOS has improved so slowly since it was named after cats (2012) that it's hard to even care about the aggregated changes since then (even if I ignore all the bugs*), and similarly for Google Translate since they bought and integrated Word Lens in 2014.
Entirely new things are sometimes (like ChatGPT or Stable Diffusion) genuinely interesting and make a novel departure from what came before, but it's very rare.
* https://benwheatley.github.io/blog/2025/06/19-15.56.44.html
Though we got dark mode, so, maybe actually it is better
One of the bugs I didn't list in that blog post, was that dark mode sometimes only partially activates — I can get some (not even all!) dark mode icons on an otherwise light mode display.
Good point. I think that's demonstrably already a problem. Software products are notoriously prone to enshittification in part because successful companies build up vast development teams that have to somehow justify their existence with continuous feature churn.
I think it's reasonable to be sceptic towards a guy who:
- sells you a product (Copilot) that depends on him selling you the story that LLMs are the future
- has a lot of vested interest in Microsoft's bet on LLMs to succeed
On his own platform developers that I much more look up to have gone the other direction and have disabled AI PRs and/or bug bounties on their repos:
- Daniel Stenberg (curl) https://www.theregister.com/2025/07/15/curl_creator_mulls_ni...
- Linus (linux) https://www.theregister.com/2024/10/29/linus_torvalds_ai_hyp...
- or socket.dev's quest to prevent copilot generated PRs (https://socket.dev/blog/oss-maintainers-demand-ability-to-bl...) this might have struck a nerve
So who should I believe? A guy who I've never heard the name of (apparently the Github CEO's name is Thomas Dohmke, I just learned), or a bunch of legends whose tools I have been using since a child?
No to slop! Down with slop.
The developers resisting AI have a hard time understanding that "artisanal" is not real label for software developers. We are not glas blowers. We are not artists. We solve problems and there are infinitely many sufficiently good ways to solve a problem. Having opinions about how is only helpful in so far it helps solves the problem. It has no other purpose to the people paying for the product. If your opinion makes the good enough product more expensive then you make your work less desirable.
As long as people are competing in the labor marketplace, that is just not a winning tactic.
1. opensource devs who ran on 246 (iirc) problems with ai assistance all projected that they would save around 20% of their estimated time, but all went over by about the same margin - salaries cost money, people are paid per _time unit_, not per deliverable (though if we were, I think there'd be a lot fewer ai-first devs out there - see sunken time cost above)
2. studies of opensource code monitoring error rates for ai-assisted repos showed a much higher defect rate than those "artisinal" coders you disdain so much - this is likely a corollary to (1) where in that case, devs have to debug ai-generated code that they could have written faster themselves
And the whole argument forgets about the upskill issue that will become more prevalent in the future: if we're replacing low-skill workers with ai agents, even assuming that those agents can do a better job (and when you actually _measure_ it, instead of just _thinking_ it, it turns out they don't), we end up with a massive skill gap as no-one upskills when they're replaced.
It's good for me - I'll probably still be able to get work when I'm 80 (anyone remember y2k?) - but it's bad for the industry as a whole, and definitely bad for any entry-level coders who were going to become great, but couldn't, because they couldn't learn on the job they didn't have.
As for competing in the labor market - we'll see how that goes as more and more companies are starting to realise that the "savings" promised to them by ai code agents not only don't materialise, but are often actually _losses_. Again, see (1) and (2) above. I don't have links, but they should be relatively easy to find - I found them all right here.
"ai" in it's current incarnation (glorified token predictors) will never surplant skilled devs, simply because it cannot understand anything. It can't understand the domain, it can't understand the users, it can't understand the code. It has no concept of understanding. All it knows is "sequence a,b,c likely is followed by d", just on a larger scale. If it can do a fully-functional, unbuggy bit of code for you, it's simply because it's seen that exact code before, and in that case, it's robbed you of learning anything because the journey to finding that code, or the docs that spawned it, is part of the learning process. Much like we always scolded devs who copy-pasta'd from Stack Overflow, we should hold these agents to the same standard.
Please, please, at least if you're going to use these tools, _don't trust them_. Scrutinise everything. Personally, having to double check a confidently wrong junior dev all day (one which won't learn from its mistakes either) sounds like a step down from actually creating stuff.
> They were all about increasing ambition. We believe that means that we should update how we talk about (and measure) success when using these tools
I’m struggling to understand what this means.
no, you don't have to "get out of this career", but it's probably prudent to get out of these "ai-first" companies before the company (or at least department) implodes and you don't have a job anyway
That’s just a majority of people in the industry. If there any reason to leave, that’s it.
Thanks to what Microsoft did to Github, I moved to Gitlab. So I doubt I will ever take advice from a CEO, never mind this person. FWIW, most CEOs come out of marketing, so that alone tells you how much he really knows about development.
If I could find something that even pays half of what I made as a developer, I’d consider it.
I'm not taking advice from a person who thinks computer science is about, uh, memorizing syntax and APIs.
It's also hilarious how he downplays fundamental flaws of LLMs as something AI zealots, the truly smart people, can overcome by producing so much AI slop that they turn from skeptics into ...drumroll... AI strategists . lol
I use arch Linux btw
OK. I quit. Bye!
How it's going: platform wants to stop people from sharing the code they wrote with each other.
1oooqooq•6mo ago