METR reran the study early this year and, while they caveat it, this time they found a speedup, which is consistent with subjective estimates of productivity also having increased -- the simplest explanation is that subjective estimates exaggerate, but there's still a speedup with current models: https://metr.org/blog/2026-02-24-uplift-update/#wider-adopti...

(Nobody seems to cite the followup since it's not such a fun counterintuitive finding.)

At least the author admit he is biased..

I could use current models forever. LLM does things for me that couldn't be done in 2023. I don't want to go back to a world without it, but I'm also ok with progress stalling right here.

As for the productivity benefits, I find there is a kind of skeptic that just has his eyes closed. Open an AI page, ask it to write you an app that renames all your scientific papers so they have their title as the filename instead of the weird long number.

Without AI you could easily waste an hour looking for the right PDF library and working out how to use it. Or fiddling with the file search mechanism so that you only get the articles and not every file in your downloads. Now you will have the files renamed in a few minutes, and a tool that keeps them correct every time you have download some more.

How is this not a productivity win? It clearly is. Do we really need a double blind test to check this kind of thing?

Whether it gets squandered in modern orgs is another matter, but the core win is clear for everyone to see.

That kills all arguments from the article instantly.

December 2025 through today GenAI has become massively better. I’ve built things with Claude in contracted spans of time that would have been man-years in the before times.

Would you commission a study about whether cars are really a faster form of locomotion before deciding to purchase one?

There is empirical data right in front of your face. You're just choosing to exclude it because of where you already stand.

That's exactly what Feynman was warning about.

Here are my personal anecdotes for the week. I occasionally ask AI to write some SystemVerilog test bench code. This is an area where I have deep and long experience. AI has never once saved me time with that. Just yesterday, however, I was writing a tricky Makefile, something I rarely do, and AI helped a lot.

Aside from productivity, this article mentions ethical and financial issues with AI. That doesn't stick with the subject that the title of the essay implies, but it's a more concrete thing to discuss and I wish we'd spend more time on that. I think that's where people really are fooling themselves.

> Meanwhile, a study (late 2025) seems to report that although participating developers felt they completed tasks faster using AI, they where around 19% slower.

I think you might be fooling yourself if you build your entire worldview concerning the productivity benefits of AI-assisted programming around that one study from one organization that confirms your priors.

1. I am not yet convinced that I am more productive with AI for complex tasks. But, I am fairly certain for trivial tasks, or bureaucratic tasks AI has lessened my cognitive load (data point of one).

2. Does AI run on money, or does it run on energy? Money is the means by which we facilitate exchange. But in reality, the energy in the universe is ginormous. The fundamental thing to keep AI going is available in excess, all other things are negotiation.

3. Is AI bad for the environment or is it the way we obtain energy bad for the environment? The trend seems to be that energy is getting cheaper and cleaner. The world is not static, but will move to accommodate a world with AI.

But, I don't understand the position of that has no value or that is going to be "turned off", that is just doomsday prophecy at this point.

We focus so much in code, but AI have changed many if not most of the work. It made possible to take a chance in challenges that would once be dismissed of the bat.

I do think we are still get used to the new ways of work, which IMO is a moving target. Once that we overcome that, results may come clear, but that may not be in companies making 10x more revenue.

If investors are looking for a 50% return, they need the investment to be worth about $1.5 trillion. If they're looking for a 100% return, it needs to be worth about $2 trillion. That's still an enormous hurdle, but it's much smaller than suggesting AI needs to somehow generate trillions of dollars in cash simply to recoup the original investment.

Thats a pretty strong outcome. It implies that not only are GPUs that power AI too expensive long term, but they cost too much to operate even if they were free.

Seems to me the more likely outcome is a wave of dotCom style bankruptcies wiping out equity holders for companies who contracted to buy chips and datacenters at MSRP, and a second wave for the groups that step in after to operate whats left without the absurd financing charges and lower capex.

Ironically, the author would have been better asking an LLM to write the article. At least it would have found better sources and developed more convincing arguments.

I wish this were how things are going but Uber is still surviving. I figure at least one AI company is going to come out of this alive and kicking.

And if you don't know anything about programming, or don't care that the solution it'll come up with is potentially flawed, then it's a lot quicker to vibe code up some sort of hack job answer than learn how to do everything manually. I'm pretty sure that my non-technical parents or grandparents could ask Claude or ChatGPT to build a simple website for their small business or charity, and have something workable within an hour or two.

For a software engineer, AI might not make them any more productive. For a non-engineer, it's basically magic.

The unsustainability thing is more of a concern though, and the external effects of AI are a huge problem. It's just that pure productivity wise, a lot of people will be significantly more productive with these systems than without them.

Sure - it was light on the meat, But it's an opinion blog post, not a full article. the author ventures to cite their refs, acknowledges potential bias.. way more than you get with a tonne of front page articles (many of which probably should have been short opinion pieces like this)

Sure - it's not particularly notable or high entropy enough to save, but it's far from offensive or bothersome ...yet all I'm seeing (at time of posting) is defensive dismissals, ad hominems, and echo chamber pablum

Some people still suggest that AI flat out isn't useful though, and that angle to me doesn't hold much water. I was a slow AI adopter, but I have found plenty of benefit from using AI in various applications recently. I haven't come close to using on the level of some people who orchestrate agents and rarely read or touch the code they are writing, but there have been clear times when I popped into Claude and had it help me very successfully with something that would have taken longer to do without that help. So increasingly I am skeptical of anyone positing that there is no potential benefit to appropriately using AI assistance. I think the author could rewrite around the angle of their perceived negative externalities outweighing potential advantages, but to deny or ignore those potential benefits ultimately distracts or actively takes away from a serious discussion of said negative externalities.

I don't put much stock in the very very frequently cited study about AI productivity included in this blog. It's a single study. The study itself cites multiple other studies that did find a productivity boost in writing software with AI assistance. I don't see the methodology as being very sound, the sample size isn't very big, some of those developers had likely never used AI before so of course they were slower as they explored using a new tool for the first time (which the study more or less writes off as a non-factor). They also used "Claude 3.5/3.7 Sonnet", which is not the strongest model. I personally find the Sonnet family of models not that useful and much more inaccurate than something like Opus or Fable.

One thing I've noticed is, that coding manually is exponentially more stressful than the slow AI approach. I'm not going any faster than I normally would, but the burnout doesn't show up on the fifth day of the week. The hyper-focus is kind of fun but utterly exhausting.

With the fast approach I can create impressive demos and POCs that are nowhere near ready for production. But sometimes that's exactly what I need.

Anybody actually using AI in anger (to get paid) will tell you AI was pretty bad early 2025. It wasn’t usable till June. By December it was writing most of my code - with significant reviews by me. Now it’s basically getting it right. YMMV.

This is largely an infrastructure problem, and it will be solved just as similar problems were solved with storage, network capacity, and computing power.

There will be bankruptcies, but the technology will prevail, just as the web did after the dot com bubble. Personally, I see the loss of craft and practical skill, which can only be built through doing, as a much greater danger. I still practise LeetCode from time to time so I do not lose the ability to write code myself.