These aren't related in the way you think they are. Stock price reacts quickly to broader market trends, but more slowly for company-specific trends where revenue is likely stable. The impact of AI in engineering work will take months to show up in the product, probably a year after that for customers and the market to take notice. An AI product is a different thing entirely.
I believe the lack of quick evident profit increases are partly a failure of imagination or a failure of understanding that AI agents are different from people. More impressive or faster in some ways, but much much less reliable in others.
The evolution of harnesses like claude code or open cause, and metaharnesses like Ralph loops, gas town, claws, etc. Will progressively allow for gradually better results and abilities even if models stopped evolving, and if the Mythos eval numbers are to be believed, there is still no hard ceiling to be felt yet.
At the same time, small models that can run on PCs VRAM/UNIFIED RAM have like Qwen are becoming more useful.
I predict that having more and more loops within loops within loops and layers of cloud/local models of different capabilities will solve a great many limitations of LLMS today...at the cost of speed and token count.
We've never had a tool that is at the same time so unreliable and complicated as GenAI before. It will take us a minute to figure out how to use it properly.
Actually I think the opposite - we will learn that the most important thing is the ability to manage context & steer these models instead of using a rube goldberg machine. Some of the top performing agent harnesses on Terminal Bench provide literally one tool: tmux, which outperforms Claude Code et al. To me, the most important thing by far when getting reasonable output from these machines are what you put into it.
On one side, there is the usual process of figuring out how to properly use this new tech. It is to be expected that some experimentation is necessary to figure out what applications AI boosts productivity for and what applications it doesn't. There is unusually strong evangelism pushing AI into everything, so the negatives are going to be salient and may make it hard to spot some of the successes.
On the other side is something a little bit new: Deliberate enshittification. OpenAI and others no doubt saw the power crunch coming years in advance, yet it's still happening and is, ostensibly, the reason why prices are starting to go up. This was not unexpected. It's the business model. Build to the capacity that is cheaply available while offering your customers a sweetheart deal to get them addicted, and then jack up the prices when the competition has no cheap power to build upon. The result is locked in customers and locked out competition.
On one side, you have people learning when AI is appropriate and how to use it efficiently. On the other side, you have a small number of AI companies trying to extract every last bit of value so that any productivity gains wind up in their owners' pockets. Will the gains of more appropriately applying AI be entirely wiped out by enshittification?
Partially a contracting real economy following overhiring early in the decade, partially trying to discipline labor, partially a pretty profound disconnect from both market pressures and concrete metrics that comes from a business model more centered around stock value and funding raises than revenue per se
There’s alway a bit of that going on, but ironically if AI does result in mass labor replacement India and the Philippines are likely going to be ground zero where workforces get wiped out first. They’re ripe with the kind of things that AI is, in theory, getting very good at.
Typical bad management decisions that came home to roost. It’s a lot easier to say “AI productivity improvements” than for the CEO to say “I’m cleaning up terrible performance on my part and a lot of bad business decisions.”
To juice the next quarter. Extreme short-term thinking has become the norm at every business I've worked at and every business I'm aware of, so upper management has no issue cutting teams right down to the bone.
It's why software has become far more unstable. There's nobody around to actually maintain it.
AI isn’t going away, but it’s also clear the much promised impacts aren’t there and aren’t coming anytime soon. A bit like the claims a few years back that we’d all have self driving cars by now.
The most likely outcome is an AI bubble correction that will be somewhat painful and wipe out many/most AI startups, followed by AI settling into day to day in a way that’s useful and found in many places, but not world-as-we-know-it-ending like the AI bros predict.
If AI just means automation, then sure. We absolutely need more automation and if LLMs are not the mechanism then something else better be. More automation is the life blood of our industry. But are LLMs a game changer or today's fuzzy logic? [1] Time will tell...
P.S. I'm not saying fuzzy logic doesn't have applications, I know rice cookers are a thing, but I think it's safe to say we have other options for controlling non-linear systems these days.
WE do have self driving cars with Waymo data showing it is clearly better than human drivers in certain markets like Phoenix. It is human regulations, laws and the general societal unease that is preventing a total rapid change. In fact a Robotaxis only urban area which is continuously mapped might be feasible today and probably could even reduce the no of cars needed for the population making it accessible to many more.
Phoenix is probably about as good a location as you could get for a self driving car. It’s not yet clear how wide their success will be outside of that niche.
AI has the same problem. It’s not that it doesn’t work, but that folks just aren’t all that interested in adopting it at scale. Tech makes this “build it and they will come” error a lot. The tech is quite good, but it’s all the non tech aspects of this that are why it’s not getting impact at scale.
You’re generalizing too much here. One of the biggest problems with LLM’s today is in-fact that they are not at the level being advertised. This is not solely a case of regulation standing in the way of a «revolution».
No, it’s actually the same issue with AI in a lot of cases. In perfect conditions it can work reliably, but outside of that it falls apart in a way humans don’t.
This has not been my experience with Waymo. I drove a total of about ~3.5 hours in Waymos in LA when I was visiting and their robustness to very unusual situations absolutely floored me.
I am sure you can find truly out-of-distribution cases where the car will make a mistake, but the data shows that this is more rare than a human driver making a mistake.
depends if post-correction it is worth anyone's money to keep training new frontier models. It could be that it isn't, so we are left with models that were trained in the bubble, but are now increasingly out of date, or (open?) models that are trained much more cheaply somehow with consequent lack of utility.
Did the hype cycle not have an impact on employment with the various layoffs? Or is this and admission that the layoffs were for other reasons and were just attributed to AI?
I’m not surprised about productivity though. Efficiency gains are limited by the actual bottlenecks. And truthfully, I think people are deluding themselves a bit about how effective vibe coding is and how much faster they are actually moving when you consider developers still need to form an understanding of the codebase and its systems.
Outside of coding, is there really a use case for LLMs that has the potential to make big efficiency gains? Idk.
I've found the best way for me to wield it is the tool to build tools. I would have never in a million years been able to code. But I've used it to replace things I was paying hefty monthly subscriptions for....
So I'm not actually being more productive, but I've cut my costs significantly to do the same things I could do before.
I thought I would do this, but of all the vibe coded tools I've built, I think I still use...one. The rest are just not worth the upkeep relative to the utility, or are either broken functionally or in their UX and I can't be arsed to put the effort into making them good. Which brings up why these tools didn't really exist in the first place.
Of course ymmv, and if you find yourself paying subscriptions for stuff you can replace with vibe coded apps, all power to you.
This article is underlining the stark contrast between the viewpoints of “AI Enthusiasts” and everyone else.
Don’t get me wrong, I use these tools daily. That being said I’m having a very hard time finding where the productivity gains are.
I imagine I’m far from alone in that search and when you pair that with the constant marketing and glowing “analysis” from some of the enthusiasts about how this technology is “solving coding” or “changing the face of security” or even leading to AGI it starts to tickle that part of my brain where I keep blockchain, NFTs and copper bracelets.
So TLDR the tech is good but the hype-slaves and their masters are killing it with overpromising and under delivering.
Not the OP, but there are likely many tens/hundreds of thousands of people using AI daily because their management requires it. Management tracks AI usage by employee and uses it as a KPI. You want to keep your job, you use AI. You want a bonus, you use AI a lot.
This is simultaneously one of the easier management KPIs for employees to hit and one of the least meaningful.
I think a lot of the disconnect in the programming world is we treat all programming as equivalent and it's not.
There really are many programming jobs that are rote and I have no problem believing that an LLM based tool can learn the pattern and regurgitate with the tweak de jour. In those jobs LLMs probably do increase productivity.
But there are other programming jobs that are not rote and there is no pattern to learn because you haven't done the thing yet. LLMs aren't any more useful than a normal base library would be, and if you're already good at using a library of code, they're not a productivity booster and often, in my experience, a hinderance.
I think another point is the prompt actually forces the engineer to spend a moment to actually think about what they're doing and make some kind of plan. Pre-AI tools way too many programmers just jumped straight into problems without thinking what they were doing figuring they could code their way out of anything and ending up stuck in some cul de sac and having to back track. And if they just stopped and made a basic plan they wouldn't have that issue. Forcing engineers to make a plan, who wouldn't otherwise do so, before they start, could definitely be a productivity booster for them.
> Don’t get me wrong, I use these tools daily. That being said I’m having a very hard time finding where the productivity gains are.
So why are you using the tools? Personal curiosity? Workplace mandate?
I've made measurably more and faster progress on both professional and personal projects since adopting these tools. Sometimes assisted is less productive than unassisted, but the net gain is pretty obvious to me.
I use the tools, but I'm under no delusions that I'm not just being lazy. I could just do it myself, and in some cases it would take roughly the same amount of time, but I can scroll TikTok while it dutifully churns out code.
I don’t like the tools personally, and find the reversion of any sort of interface to a chat interface a huge loss to UI - but for the love of all things holy why are using them if they don’t provide any benefit?
> Don’t get me wrong, I use these tools daily. That being said I’m having a very hard time finding where the productivity gains are
I'm really struggling to understand why you would use them that much if you aren't sure they are more productive. Is it just a more enjoyable workflow for you?
I ask because I find AI assisted workflows extremely painful. Constantly pulling me out of flow, like driving in gridlock traffic.
A lot of organizations live in some game theoretic equilibrium that prevents cost improvements from being metabolized by the org without burning the cost elsewhere.
For example, consider a commodity business for software product X. All vendors of this product had their costs reduced by a factor of 100 over night for developing new product. They could increase their profits, lower their price, or re-invest the dividend. In software, the buyer usually buys on quality - so they all re-invest.
Now they are spending the same amount on product development, for the same price tag, and earning the same profit - but they might be shipping much faster.
My company has pushed engineering all-in for AI in the last few months
Our stock price has also gone down 70% in the last few months
Naturally, we're pivoting our platform to put AI front and center
Did they try and be an blockchain-first company when that was the rage? Making NFTs and whatnot. Is your CEO just a trend-follower?
These aren't related in the way you think they are. Stock price reacts quickly to broader market trends, but more slowly for company-specific trends where revenue is likely stable. The impact of AI in engineering work will take months to show up in the product, probably a year after that for customers and the market to take notice. An AI product is a different thing entirely.
The beatings will continue until moral improves
Are businesses all running on sunk cost fallacy now? These findings have been coming out for a while but it doesn't seem to change anything.
They’ll say no but really… you know.
I believe the lack of quick evident profit increases are partly a failure of imagination or a failure of understanding that AI agents are different from people. More impressive or faster in some ways, but much much less reliable in others.
The evolution of harnesses like claude code or open cause, and metaharnesses like Ralph loops, gas town, claws, etc. Will progressively allow for gradually better results and abilities even if models stopped evolving, and if the Mythos eval numbers are to be believed, there is still no hard ceiling to be felt yet.
At the same time, small models that can run on PCs VRAM/UNIFIED RAM have like Qwen are becoming more useful.
I predict that having more and more loops within loops within loops and layers of cloud/local models of different capabilities will solve a great many limitations of LLMS today...at the cost of speed and token count.
We've never had a tool that is at the same time so unreliable and complicated as GenAI before. It will take us a minute to figure out how to use it properly.
Actually I think the opposite - we will learn that the most important thing is the ability to manage context & steer these models instead of using a rube goldberg machine. Some of the top performing agent harnesses on Terminal Bench provide literally one tool: tmux, which outperforms Claude Code et al. To me, the most important thing by far when getting reasonable output from these machines are what you put into it.
https://archive.ph/ztkLP
I wish anytime someone used the word "productivity" there was an accompanying definition.
repost from february; many referencing the same NBER report.
Some related discussions recently and months ago:
90% of CEOs Say AI Changed Nothing. The Other 10% Have a PR Team
https://news.ycombinator.com/item?id=47766164
Majority of CEOs report zero payoff from AI splurge
https://news.ycombinator.com/item?id=46696636
There's an interesting race happening here.
On one side, there is the usual process of figuring out how to properly use this new tech. It is to be expected that some experimentation is necessary to figure out what applications AI boosts productivity for and what applications it doesn't. There is unusually strong evangelism pushing AI into everything, so the negatives are going to be salient and may make it hard to spot some of the successes.
On the other side is something a little bit new: Deliberate enshittification. OpenAI and others no doubt saw the power crunch coming years in advance, yet it's still happening and is, ostensibly, the reason why prices are starting to go up. This was not unexpected. It's the business model. Build to the capacity that is cheaply available while offering your customers a sweetheart deal to get them addicted, and then jack up the prices when the competition has no cheap power to build upon. The result is locked in customers and locked out competition.
On one side, you have people learning when AI is appropriate and how to use it efficiently. On the other side, you have a small number of AI companies trying to extract every last bit of value so that any productivity gains wind up in their owners' pockets. Will the gains of more appropriately applying AI be entirely wiped out by enshittification?
Dutch AI would just demand a 3 day workweek.
Spanish AI would require all AI systems to pause for 2 hours after lunch hour
Then why the layoffs???
Partially a contracting real economy following overhiring early in the decade, partially trying to discipline labor, partially a pretty profound disconnect from both market pressures and concrete metrics that comes from a business model more centered around stock value and funding raises than revenue per se
We've been moving to faith-based markets for decades - markets where belief and hope almost entirely replace quantifiable economic activity.
Outsourcing to India and the Philippines
There’s alway a bit of that going on, but ironically if AI does result in mass labor replacement India and the Philippines are likely going to be ground zero where workforces get wiped out first. They’re ripe with the kind of things that AI is, in theory, getting very good at.
Trend following - everyone's jumping. And bad economy.
Because there was bloat and AI was a good scape goat.
Typical bad management decisions that came home to roost. It’s a lot easier to say “AI productivity improvements” than for the CEO to say “I’m cleaning up terrible performance on my part and a lot of bad business decisions.”
To juice the next quarter. Extreme short-term thinking has become the norm at every business I've worked at and every business I'm aware of, so upper management has no issue cutting teams right down to the bone.
It's why software has become far more unstable. There's nobody around to actually maintain it.
AI isn’t going away, but it’s also clear the much promised impacts aren’t there and aren’t coming anytime soon. A bit like the claims a few years back that we’d all have self driving cars by now.
The most likely outcome is an AI bubble correction that will be somewhat painful and wipe out many/most AI startups, followed by AI settling into day to day in a way that’s useful and found in many places, but not world-as-we-know-it-ending like the AI bros predict.
If AI just means automation, then sure. We absolutely need more automation and if LLMs are not the mechanism then something else better be. More automation is the life blood of our industry. But are LLMs a game changer or today's fuzzy logic? [1] Time will tell...
[1] https://www.electronicdesign.com/technologies/embedded/digit...
P.S. I'm not saying fuzzy logic doesn't have applications, I know rice cookers are a thing, but I think it's safe to say we have other options for controlling non-linear systems these days.
WE do have self driving cars with Waymo data showing it is clearly better than human drivers in certain markets like Phoenix. It is human regulations, laws and the general societal unease that is preventing a total rapid change. In fact a Robotaxis only urban area which is continuously mapped might be feasible today and probably could even reduce the no of cars needed for the population making it accessible to many more.
As a counterpoint, Waymo conducted a pilot in NYC then abandoned the permit for it:
https://www.thecity.nyc/2026/04/06/waymo-driverless-cars-tes...
Phoenix is probably about as good a location as you could get for a self driving car. It’s not yet clear how wide their success will be outside of that niche.
AI has the same problem. It’s not that it doesn’t work, but that folks just aren’t all that interested in adopting it at scale. Tech makes this “build it and they will come” error a lot. The tech is quite good, but it’s all the non tech aspects of this that are why it’s not getting impact at scale.
You’re generalizing too much here. One of the biggest problems with LLM’s today is in-fact that they are not at the level being advertised. This is not solely a case of regulation standing in the way of a «revolution».
Ever driven in Bali?
No, it’s actually the same issue with AI in a lot of cases. In perfect conditions it can work reliably, but outside of that it falls apart in a way humans don’t.
This has not been my experience with Waymo. I drove a total of about ~3.5 hours in Waymos in LA when I was visiting and their robustness to very unusual situations absolutely floored me.
I am sure you can find truly out-of-distribution cases where the car will make a mistake, but the data shows that this is more rare than a human driver making a mistake.
No, it’s actually the same issue with AI in a lot of cases. In perfect conditions it can work reliably, but outside of that it falls apart.
> certain markets like Phoenix
So, basically the easiest robotaxi market on the planet? Call me when it works in Bucharest, Mumbai, Istanbul, Cairo, etc.
For software the last 80% of effort needed to finish the 20% remaining items is the hardest and hardware is even harder.
Was there any recent technology that really delivered what was the general promise?
depends if post-correction it is worth anyone's money to keep training new frontier models. It could be that it isn't, so we are left with models that were trained in the bubble, but are now increasingly out of date, or (open?) models that are trained much more cheaply somehow with consequent lack of utility.
Good point. At some point there will be a reality check for the giant pile of burning cash that is new model training.
they’re holding it wrong.
They should have asked AI CEOs
Ah yes, first the return to office, now being forced to use AI in 50%+ of projects. Will the ingenuity of modern executives never cease?
Did the hype cycle not have an impact on employment with the various layoffs? Or is this and admission that the layoffs were for other reasons and were just attributed to AI?
I’m not surprised about productivity though. Efficiency gains are limited by the actual bottlenecks. And truthfully, I think people are deluding themselves a bit about how effective vibe coding is and how much faster they are actually moving when you consider developers still need to form an understanding of the codebase and its systems.
Outside of coding, is there really a use case for LLMs that has the potential to make big efficiency gains? Idk.
I've found the best way for me to wield it is the tool to build tools. I would have never in a million years been able to code. But I've used it to replace things I was paying hefty monthly subscriptions for....
So I'm not actually being more productive, but I've cut my costs significantly to do the same things I could do before.
I thought I would do this, but of all the vibe coded tools I've built, I think I still use...one. The rest are just not worth the upkeep relative to the utility, or are either broken functionally or in their UX and I can't be arsed to put the effort into making them good. Which brings up why these tools didn't really exist in the first place.
Of course ymmv, and if you find yourself paying subscriptions for stuff you can replace with vibe coded apps, all power to you.
This article is underlining the stark contrast between the viewpoints of “AI Enthusiasts” and everyone else.
Don’t get me wrong, I use these tools daily. That being said I’m having a very hard time finding where the productivity gains are.
I imagine I’m far from alone in that search and when you pair that with the constant marketing and glowing “analysis” from some of the enthusiasts about how this technology is “solving coding” or “changing the face of security” or even leading to AGI it starts to tickle that part of my brain where I keep blockchain, NFTs and copper bracelets.
So TLDR the tech is good but the hype-slaves and their masters are killing it with overpromising and under delivering.
Not the OP, but there are likely many tens/hundreds of thousands of people using AI daily because their management requires it. Management tracks AI usage by employee and uses it as a KPI. You want to keep your job, you use AI. You want a bonus, you use AI a lot.
This is simultaneously one of the easier management KPIs for employees to hit and one of the least meaningful.
https://www.wsj.com/tech/ai/ai-work-use-performance-reviews-...
you know a tool is good when your boss hinges your career on you using it
I think a lot of the disconnect in the programming world is we treat all programming as equivalent and it's not.
There really are many programming jobs that are rote and I have no problem believing that an LLM based tool can learn the pattern and regurgitate with the tweak de jour. In those jobs LLMs probably do increase productivity.
But there are other programming jobs that are not rote and there is no pattern to learn because you haven't done the thing yet. LLMs aren't any more useful than a normal base library would be, and if you're already good at using a library of code, they're not a productivity booster and often, in my experience, a hinderance.
I think another point is the prompt actually forces the engineer to spend a moment to actually think about what they're doing and make some kind of plan. Pre-AI tools way too many programmers just jumped straight into problems without thinking what they were doing figuring they could code their way out of anything and ending up stuck in some cul de sac and having to back track. And if they just stopped and made a basic plan they wouldn't have that issue. Forcing engineers to make a plan, who wouldn't otherwise do so, before they start, could definitely be a productivity booster for them.
> Don’t get me wrong, I use these tools daily. That being said I’m having a very hard time finding where the productivity gains are.
So why are you using the tools? Personal curiosity? Workplace mandate?
I've made measurably more and faster progress on both professional and personal projects since adopting these tools. Sometimes assisted is less productive than unassisted, but the net gain is pretty obvious to me.
Honestly? It allows me to be lazy.
This.
An AI is like delegating it to the junior programmer you don't have. You spend 5 minutes writing the spec rather than an hour coding.
It's usually something you could do yourself, and just can't get motivated to type out the code in the moment.
That's a productivity gain in my book.
Yep...same
I use the tools, but I'm under no delusions that I'm not just being lazy. I could just do it myself, and in some cases it would take roughly the same amount of time, but I can scroll TikTok while it dutifully churns out code.
I don’t like the tools personally, and find the reversion of any sort of interface to a chat interface a huge loss to UI - but for the love of all things holy why are using them if they don’t provide any benefit?
tbf, there is some benefit there but its much more nuanced than the hype suggests (as usual)
> Don’t get me wrong, I use these tools daily. That being said I’m having a very hard time finding where the productivity gains are
I'm really struggling to understand why you would use them that much if you aren't sure they are more productive. Is it just a more enjoyable workflow for you?
I ask because I find AI assisted workflows extremely painful. Constantly pulling me out of flow, like driving in gridlock traffic.
It allows me to be lazy honestly.
That and using it like a search engine feels a little like having good Google back.
A lot of organizations live in some game theoretic equilibrium that prevents cost improvements from being metabolized by the org without burning the cost elsewhere.
For example, consider a commodity business for software product X. All vendors of this product had their costs reduced by a factor of 100 over night for developing new product. They could increase their profits, lower their price, or re-invest the dividend. In software, the buyer usually buys on quality - so they all re-invest.
Now they are spending the same amount on product development, for the same price tag, and earning the same profit - but they might be shipping much faster.