I'm so frustrated by both the zealous AI bulls and the blind AI opposition.
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.
Nuance has been completely lost on our society. If it doesn't spark immediate outrage or joy it has no place in our attention economy. I really wish this wasn't the case and I'm not sure how we can reverse it. Most people just aren't interested in the 'balanced takes' because it's just not exciting enough it seems sadly.
I recently learned the term "thought-terminating cliché" and it's so incredibly prevalent in society these days. Always has been I'm sure, but short form text based anonymous communication must be making it worse.
It is good to remember that the discourse is not fully real.
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).
> To say otherwise tells me you haven't actually tried to make use of these tools.
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.
We're in a dangerous valley where AI is _just_ good enough to fool some otherwise very smart people. Similar to the old adage of "a little bit of information is a dangerous thing." Lots of CEOs got duped into thinking that model capabilities were far ahead of where they actually were. I'm actually not sure if we're going to get out of the valley without figuring out a surefire way to reliably evaluate these things.
They got duped into it partly because of the enthusiasm and demoware shown by folks like the ones that topped HN with their exhortations about the wonders of AI for coding.
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.
I don’t claim responsibility for any of that. I’m not part of that “we”. Time and time again I’ve been sceptical of AI on places like Hacker News and I’m hardly alone. As a community I think there is a great variation in opinion.
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 complicit. They were fooled when they shouldn’t have been and it’s their own fault.
I don’t know who you’re including in “we”, but since most CEOs are being paid more for their judgment than HN commenters are, it wouldn’t be unreasonable to hold them to higher standards.
What about the large contingent of resident wet rags? Can we still shit on the CEOs for not seeing through the BS and guiding their companies effectively?
I have a theory there’s a more nefarious problem that AI changes the incentives around what work is easy to do, in a way that can affect how work hits the bottom line. Aside from the slop problem- people create more documentation more people have to read, etc. it makes it easier for me to say do a bunch of bug fixes or refactors that aren’t in the critical path.
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.
> Since the start of the year, Chinese AI models have overtaken their US counterparts in token consumption, according to data from OpenRouter, an aggregation platform that allows users to access multiple AI models.
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.
I'm consistently burning over $100/day in Claude/OpenAI tokens using Codex and Claude - but I'm on the subscriptions so I'm only paying $100/month to each vendor.
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.
They believed and were financially incentivized to embrace the hype and operationalized it because technical leadership thought rapid PoCs would apply to the rest of the tech stack.
IMHO it’s just good old fomo. They fear that the landscape will shift overnight and they’ll be left with a ton of useless meat bags instead of a pay as you go models that you can upgrade instantly every quarter.
Out of touch leadership and management combined with a culture of self-congratulatory "innovation" where companies just plagiarize eachother in an endless loop... It doesn't surprise me at all, tbh.
Once you realize most execs are literal sociopaths driven by pure greed and power a lot of things start to make sense. Of course for simple minded folks like us none of it is remotely understandable
> The ride-hailing company has introduced usage caps, limiting employees to $1,500 in monthly token spending on individual AI tools, after blowing through its entire AI 2026 budget by April.
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.
I think this probably has more to do with companies switching to API pricing for enterprises, no?
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?
I think it's both. Last year even the most AI-hungry companies had employees who were primarily using ChatGPT and Claude. It's really hard to burn a noticeable number of tokens with those tools.
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.
Does anyone know the inside story of some of these AI adoptions that have been downsized? The company I’m at has only only recently gotten an enterprise license.
We’ve had our AI budget per Engineer cut twice now since the peak mania in early 2026 and Fable was banned even before it was removed by Anthropic due to cost. It’s still used a lot but I think maybe they’re not really seeing the ROI especially when some people spent thousands and thousands per month on dubious AI things.
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
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.
I knew this was going to be the end result after seeing so many companies reward employees based on how many tokens they were using. My company even pulled stats and gave physical awards out at an in person retreat. Ridiculous, and it was never going to last.
If your choices are “reduce productivity, make money” or “increase productivity, but still make the same money”, why would you ever choose the latter?
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.
I would imagine this is the route it will go too, once (if) hardware supply chains catch up and costs can start coming down.
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.
Companies are learning that even with mass layoffs, AI isn't worth what it costs for most use cases. This is an important inflection point because none of the AI companies are profitable, i.e. they're all still charging substantially less than what it costs to actually deliver the service. With things in that intermediate state, it's hard to know what a future stable state will be like.
Partially because they think that once employees put enough context and rules into markdown files they can be fired and replaced by an online subscription.
> 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.
I remember back when ChatGPT first came out, there was an article on HN about this AI researcher who worked for one of the big companies (I think Google) who came to believe the model was truly intelligent and that it was being abused by being locked in the machine. We all laughed as the guy had clearly lost his mind to AI psychosis. What we didn’t realize is he may have been patient zero.
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?
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.
I'm so frustrated by both the zealous AI bulls and the blind AI opposition.
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.
Nuance has been completely lost on our society. If it doesn't spark immediate outrage or joy it has no place in our attention economy. I really wish this wasn't the case and I'm not sure how we can reverse it. Most people just aren't interested in the 'balanced takes' because it's just not exciting enough it seems sadly.
I recently learned the term "thought-terminating cliché" and it's so incredibly prevalent in society these days. Always has been I'm sure, but short form text based anonymous communication must be making it worse.
It is what it is.
It is good to remember that the discourse is not fully real.
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).
> Am I able to write things faster with LLMs, yes.
LLM is great at identifying useless writing - if you have llm writing it, it should not exist.
> To say otherwise tells me you haven't actually tried to make use of these tools.
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.
No, we can't all be friends. Never happened, never will be.
We're in a dangerous valley where AI is _just_ good enough to fool some otherwise very smart people. Similar to the old adage of "a little bit of information is a dangerous thing." Lots of CEOs got duped into thinking that model capabilities were far ahead of where they actually were. I'm actually not sure if we're going to get out of the valley without figuring out a surefire way to reliably evaluate these things.
They got duped into it partly because of the enthusiasm and demoware shown by folks like the ones that topped HN with their exhortations about the wonders of AI for coding.
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.
I don’t claim responsibility for any of that. I’m not part of that “we”. Time and time again I’ve been sceptical of AI on places like Hacker News and I’m hardly alone. As a community I think there is a great variation in opinion.
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 complicit. They were fooled when they shouldn’t have been and it’s their own fault.
I don’t know who you’re including in “we”, but since most CEOs are being paid more for their judgment than HN commenters are, it wouldn’t be unreasonable to hold them to higher standards.
What about the large contingent of resident wet rags? Can we still shit on the CEOs for not seeing through the BS and guiding their companies effectively?
I have a theory there’s a more nefarious problem that AI changes the incentives around what work is easy to do, in a way that can affect how work hits the bottom line. Aside from the slop problem- people create more documentation more people have to read, etc. it makes it easier for me to say do a bunch of bug fixes or refactors that aren’t in the critical path.
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.
> Since the start of the year, Chinese AI models have overtaken their US counterparts in token consumption, according to data from OpenRouter, an aggregation platform that allows users to access multiple AI models.
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.
I genuinely have no idea how some of these companies got so far over their skis on AI. It simply does not make sense to me.
I'm consistently burning over $100/day in Claude/OpenAI tokens using Codex and Claude - but I'm on the subscriptions so I'm only paying $100/month to each vendor.
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.
These companies already have laid people off or are preparing to.
They believed and were financially incentivized to embrace the hype and operationalized it because technical leadership thought rapid PoCs would apply to the rest of the tech stack.
IMHO it’s just good old fomo. They fear that the landscape will shift overnight and they’ll be left with a ton of useless meat bags instead of a pay as you go models that you can upgrade instantly every quarter.
Simple answer is that these execs live in a big bubble and goad each other into bigger and bigger investments via groupthink.
Out of touch leadership and management combined with a culture of self-congratulatory "innovation" where companies just plagiarize eachother in an endless loop... It doesn't surprise me at all, tbh.
Once you realize most execs are literal sociopaths driven by pure greed and power a lot of things start to make sense. Of course for simple minded folks like us none of it is remotely understandable
> The ride-hailing company has introduced usage caps, limiting employees to $1,500 in monthly token spending on individual AI tools, after blowing through its entire AI 2026 budget by April.
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.
I think this probably has more to do with companies switching to API pricing for enterprises, no?
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?
I think it's both. Last year even the most AI-hungry companies had employees who were primarily using ChatGPT and Claude. It's really hard to burn a noticeable number of tokens with those tools.
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.
That’s fair. I mean like any other service it’s going to need some cost analysis.
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.
Does anyone know the inside story of some of these AI adoptions that have been downsized? The company I’m at has only only recently gotten an enterprise license.
We’ve had our AI budget per Engineer cut twice now since the peak mania in early 2026 and Fable was banned even before it was removed by Anthropic due to cost. It’s still used a lot but I think maybe they’re not really seeing the ROI especially when some people spent thousands and thousands per month on dubious AI things.
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
Also, unlike the other Anthropic models, Fable/Mythos couldn't be used with a zero data retention agreement.
Company is a fortune 100 and a client of mine.
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.
Probably looks like this:
- 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.
https://archive.ph/z24oE
CEOs laid people off to replace them with AI, but turns out AI is more expensive and does a worse job.
If I made a blunder of that scale, would I or would I not be put on a PIP?
CEOs generally suffer zero consequences for their decisions, unless that decision costs someone real money.
I knew this was going to be the end result after seeing so many companies reward employees based on how many tokens they were using. My company even pulled stats and gave physical awards out at an in person retreat. Ridiculous, and it was never going to last.
If your choices are “reduce productivity, make money” or “increase productivity, but still make the same money”, why would you ever choose the latter?
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.
Corporate AI will likely go on prem where costs can be fixed.
I would imagine this is the route it will go too, once (if) hardware supply chains catch up and costs can start coming down.
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.
Companies are learning that even with mass layoffs, AI isn't worth what it costs for most use cases. This is an important inflection point because none of the AI companies are profitable, i.e. they're all still charging substantially less than what it costs to actually deliver the service. With things in that intermediate state, it's hard to know what a future stable state will be like.
If AI isn't worth what it costs why are some of these companies allowing $1,000+/employee/month?
Partially because they think that once employees put enough context and rules into markdown files they can be fired and replaced by an online subscription.
I'm not convinced that's happening, any good evidence that it is?
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.
The two that I know of are in a phase of trying to figure out what the return is.
AI is Labubu for CEOs.
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)
Mass hysteria
I remember back when ChatGPT first came out, there was an article on HN about this AI researcher who worked for one of the big companies (I think Google) who came to believe the model was truly intelligent and that it was being abused by being locked in the machine. We all laughed as the guy had clearly lost his mind to AI psychosis. What we didn’t realize is he may have been patient zero.
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?
> what breaks the spell?
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.
https://www.scientificamerican.com/article/google-engineer-c...
it was before ChatGPT
https://archive.ph/AKMjS