So even if in say 100 person engineering out 10 folks might get 2-3x critical path work. 50 folks Might just add non-critical path work, and the other forty might use it in a way that they end up doing g less critical path work.
But depending on your metrics productivity could look up while the bottom line is unaffected. In which case model quality is a red herring.
We’re not innocent bystanders here, and it’s important to recognize that. Our hype added to the hype. Our optimism added to the optimism. After layoffs due to Section 172 and interest rates going up, technologists were looking for a reason to be in-demand again, and generative AI as a platform specialization provided that.
We can’t now criticize CEOs for being taken in by the same enthusiasm we pushed for our own purposes.
If I worked for a larger company that isn't allowed to use those subscriptions and has to pay list price I'd be costing them ~$2,000/month. Now times that by a large engineering team.
Personally I truly can’t manage that many tasks (really 1 or 2 max) in parallel with these since you need to think very hard about everything the AI spits out because they’re such natural bullshitters and you end up in places where no one on the team understands anything in the project
- teams blitz through Jira tickets
- developers figure out they can't keep up with reviewing the code
- too many new features pushed at once, rushed work to develop training materials
- features don't work as good, many edge cases come out in support tickets
- tickers return to Jira board
- teams spend time triaging and fixing with less AI
- some features turned out to be mistake, but now have to stay
- once mess is cleared, return to "normal" pace, no AI agents, Cursor allowed with budget cap.
The thing about option 1 is that you still have “potential productivity” that you can tap into during critical times, where as in option 2, employees have already used up the “potential productivity” doing god knows what with AI, and you can’t push them more without breaking.
It's too useful to ignore, but too expensive to pay-as-you-go at current pricing for frontier models. Smaller, less capitalized companies will use the subscriptions (if they still exist by then), and larger firms will just build out their own compute for their dev teams.
If I made a blunder of that scale, would I or would I not be put on a PIP?
This is the delusion that went viral, or at least one version of it. It all leads back hijacking the human tendency to anthropomorphize, leading to the belief that an LLM is somehow something more than it actually is. So the question is - what breaks the spell? Failed attempts to automate that don’t work out? The realization that the return on money spent doesn’t make sense? Furthermore, how to we accelerate the eventual realization?
it was before ChatGPT
If throwing more compute at the problem keeps only resulting in incremental gains, I think that should do it. It goes one of 2 ways, really. Either we can throw enough compute at pre-training that results in infinitely more capable models to the point that the cost is now justified [1], or, we hit a scaling wall, get stuck with what we have now (or at that time) and the valuations crash knowing that "this is it" for the foreseeable future without a big breakthrough.
The labs go bankrupt or get acquired by the typical giants (Google, Microsoft, Amazon), the models get rolled into GCP, Azure, and AWS as a service, and that's it. It becomes another dev tool, much like a new IDE.
[1] cost being justified I'd rank as "your average non technical PM can now end to end develop robust, production software free of most serious vulnerabilities." model & tool capabilities that would allow you to hire a small team of non-techincal roles, for half the salary, that can produce the output of a large engineering org. If that doesn't happen, I don't see how the current buildout is sustainable.
Right, because they set their 2026 budget in 2025. And in 2025 nobody could predict how good (and token-hungry) coding agents would get after November 2025.
I'd be surprised if any company that set an AI budget for 2026 hasn't blown through it by now, assuming their staff have picked up Claude Code or Copilot or Cowork.
Regardless, the C-suite wouldn’t be performing due diligence if they weren’t at least attempting to perform the calculus of “what are we getting out of this spend?” and what we’re seeing now is them looking for the justification.
Didn’t Uber mention that they’re having a hard time tying all of that spend to any new or improved features?
The moment you have Claude Code or Codex running in a loop - or in multiple streams (something that people don't really do with chat because it returns fast enough there's no point running them in parallel) your token usage goes through the roof.
And then by May this year both OpenAI and Anthropic had migrated their enterprise pricing to API costs, not fixed subscription per month costs. That wouldn't have made much of a difference in the pre-coding-agent era, but today it means $100s or $1000s per employee per month.
I figure we’ll go from over usage to a shock leading to under usage and then land somewhere in the middle when companies decide what works for them.
That's a bit of a dodgy statistic. OpenRouter only tracks their own users - the vast majority of API customers for OpenAI and Anthropic presumably go straight to their APIs.
There's a lot of issues, ranging over technical, cultural, environmental, and moral problems. But there's also obvious value. To say otherwise tells me you haven't actually tried to make use of these tools.
It's one thing to get an AI google response and feel like it's dubious, it's another thing to know what you want and have an LLM find the APIs for a framework you're not familiar with yet and put the pieces together. The only way I use AI for programming still involves a large amount of rejecting the responses and a massive amount of reading and validating.
Am I able to write things faster with LLMs, yes. Am I missing out on the work involved in learning things I would be forced to otherwise, also yes. Are coworkers pushing stuff they don't understand more, surely.
It's a mixed bag, and we need more balanced takes in the discussion around this.
CEOs are very well compensated for the things they do. If they were led astray by snake oil salesman that does not make Hacker News as a community “not innocent bystanders”. They were fooled when they shouldn’t have been and it’s their own fault.
Since the switch to API pricing they’ve cut usage limits in half twice and are now saying that anyone with high usage is essentially going to be audited.
They’re down to about 500 a month per person.
There are fake accounts pushing AI everywhere, and people burned out by the marketing that react positioning too aggressively in the opposite direction (particularly when it's their boss who bought into the marketing and makes unreasonable demands).
There's also a lot of ideological opposition, which often tries to claim that the tech is useless etc.
> Am I missing out on the work involved in learning things I would be forced to otherwise, also yes.
Yes, but many of those things are things you might not really care about learning about. And if you want to learn about them, AI can be a big help, if you use it appropriately.
The "mixed bag" comes from the way people use it, mostly not from the tech itself.
josefritzishere•1h ago
simonw•1h ago
prymitive•58m ago
simonw•52m ago
The impact of AI on jobs appears to be more in terms of hiring slowdowns: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5425555
> Following adoption, junior employment declines in adopting firms relative to non-adopters, while senior employment trends remain largely unchanged. This decline is concentrated in occupations most exposed to GenAI and is driven primarily by slower hiring rather than increased separations.
ofjcihen•48m ago
Jtarii•22m ago
ares623•3m ago
The value is somewhere else (i.e. being able to brag at the next conference/dinner/awards show and to not feel like a loser)