CIO published a very interesting article about how the use of AI by the best engineers actually is slowing them down, and quite not delivering on the promised speed up of production code:
We've all heard the pitch. By now, it's practically background noise in every tech conference: AI coding is solved. We are told that large language models (LLMs) will soon write 80% of all code, leaving human engineers to merely supervise the output.
For a CIO, this narrative is quite seductive. It promises a massive drop in the cost of software production while increasing the engineering speed. It suggests that the bottleneck of writing code is about to vanish.
But as someone who spends his days building mission-critical financial infrastructure and autonomous agent platforms, I have to be the bearer of bad news: it's not working out that way. At least, not for your best engineers.
The deployment of AI copilots into the workflows of experienced engineers isn't producing the frictionless acceleration promised in the brochures. Instead, I'm seeing the emergence of a productivity trap — a hidden tax on velocity that is disproportionately hitting your most valuable technical talent.
[...]
For the first few years of the generative AI boom, we operated on vibes. We had anecdotal evidence and vendor-sponsored studies claiming massive productivity gains. And for junior developers working on simple tasks, those gains were real. If you just need a basic react component for a login button, using AI feels like a miracle.
But we got a reality check in mid-2025. A randomized controlled trial by METR (Model Evaluation & Threat Research) analyzed the impact on senior engineering talent. Unlike previous studies that used toy problems, this one watched experienced developers working on their own mature codebases — the kind of messy, complex legacy systems that actually power your business.
The results were stark. When experienced developers used AI tools to complete real-world maintenance tasks, they took 19% longer than when they worked without them.
[...]
It comes down to what I call the illusion of velocity. In the study, developers felt faster. They predicted the AI would save them huge amounts of time. Even after they finished — and were objectively recorded as being slower — they still believed the AI had been a timesaver.
The AI gives you a dopamine hit. Text appears on the screen at superhuman speed and the blank page problem vanishes. But the engineer's role has shifted from being a creator to being a reviewer and that is where the trap snaps shut.
According to the 2025 Stack Overflow Developer Survey, the single greatest frustration for developers is dealing with AI solutions that look correct but are slightly wrong. Nearly half of developers explicitly stated that debugging AI-generated code takes more time than writing it themselves.
In software engineering, blatantly broken code is fine. The compiler screams, the app crashes upon launch, the red squiggly lines appear. You know it's wrong immediately.
Almost-right code is insidious. It compiles. It runs. It passes the basic unit tests. But it contains subtle logical flaws or edge-case failures that aren't immediately obvious.
When I use an AI, I am forced into reverse-engineering. I get a block of code I didn't write. I have to read it, decipher the intent of the model and then map that intent against the requirements of my system.
I saw this firsthand when building financial systems for enterprise logistics. The logic required to calculate net revenue was sophisticated with bespoke business rules. If I asked an LLM to generate the billing code, it would give me something that looked mathematically perfect. It would sum the line items correctly.
[...]
There is also the cost of context switching. Deep work, or flow state, is the essence of high-level engineering. It takes time to load the context of a distributed system into your brain.
AI tools, in their current chat-based forms, encourage a fragmented workflow. You stop coding, you prompt the bot, you wait, you review, you reject, you re-prompt. The flow is gone.
[...]
So, if the current copilot model is a trap for your best talent, what do we do? We certainly don't ban AI. That would be like banning calculators because you sometimes hit the wrong button.
We need to move from AI-assisted coding to AI-enabled architecture. The goal isn't to make your senior engineers type faster, but to enable them to build systems that are robust enough to handle the chaos of AI-generated code.
[...]
The popular 80/20 split — where AI does 80% of the work and humans do the 20% — is misleading. It implies the human part is just a finishing touch. In reality, that 20% is 100% of the value. It's the architecture, the security model and the business logic.
To escape the productivity trap, you need to direct your engineering leaders to focus entirely on this human 20%.
My own work has shifted away from writing features and toward defining the physics of our codebase. When I was at Uber, I spent a huge amount of time migrating our systems to use strict types and schemas.
[...]
This is the strategic shift. The role of the senior engineer is to build the compiler for the AI. They need to create the schemas, the type systems and the automated rules that constrain what the AI can do.
This transforms the almost-right problem. Instead of me manually reviewing code to find errors, the system rejects the code automatically if it doesn't fit the architecture. I stop being a reviewer and start being a legislator.
[...]
The AI productivity trap is real, but it's not inevitable. It's a symptom of applying a new technology using an old workflow. The path forward is rigorous, architectural and deeply human. It requires us to value the design and the constraint-setting as the true core of engineering.
As Brian Kernighan said, "Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it". And now, a corollary: "If AI is smarter than you, who the hell is going to debug the code?"
(Score: 5, Funny) by turgid on Monday February 09, @09:09AM (2 children)
According to the 2025 Stack Overflow Developer Survey, the single greatest frustration for developers is dealing with AI solutions that look correct but are slightly wrong. Nearly half of developers explicitly stated that debugging AI-generated code takes more time than writing it themselves.
Who'd have thought it? Blindly adopting AI-generated code is pretty similar to blindly copying-and-pasting code from Stack Overflow.
Why did nobody think of this before?
I refuse to engage in a battle of wits with an unarmed opponent [wikipedia.org].
(Score: 3, Touché) by c0lo on Monday February 09, @01:15PM
Parting the fool from his money requires the fool not thinking straight. Not thinking at all is sufficient.
Fortunately, one can't fool everybody all the time.
https://www.youtube.com/@ProfSteveKeen https://soylentnews.org/~MichaelDavidCrawford
(Score: 4, Insightful) by JoeMerchant on Monday February 09, @01:42PM
My human colleagues excel at "almost right code" and have been doing so for decades.
AI just writes more of it, faster.
For what it's worth, AI also reviews more code, faster, and finds more flaws when you task it to do that. It still misses things, just like human reviewers and testers, but maybe if the Vibe Coders would refocus 15-20-30-35 on coding-documenting-reviewing-testing they might start to produce better quality code, just like human teams.
🌻🌻🌻🌻 [google.com]
(Score: 5, Interesting) by khallow on Monday February 09, @10:57AM (1 child)
It seems like one of those subjects where the same failed narratives play out each time the subject is introduced. Each time I have to remind people of the same two economic phenomena: comparative advantage [wikipedia.org] and Jevons paradox [wikipedia.org]. There's been a lot of bluster over the years about how these concepts don't work or don't scale, but they remain after each wave of hype washes over us.
(Score: 3, Interesting) by JoeMerchant on Monday February 09, @01:46PM
Never forget the comparative advantage of a pump and dump scheme - just because it's illegal doesn't mean that the risk of getting called out on it outweighs the reward, and then there are the "too big to fail" and "enough money to buy a private island and operate a few jets for the rest of my life" threshold-phenomena...
🌻🌻🌻🌻 [google.com]
(Score: 4, Insightful) by c0lo on Monday February 09, @11:31AM (2 children)
Once the investment in creating it is over, the code is strictly a liability
The faster, cheaper and more code you create (and throw at its creation an one-off investment), the more maintenance cost you incur (usually on an on-going basis).
Now, add to the above the Jevons paradox that khallow mentioned [soylentnews.org]
https://www.youtube.com/@ProfSteveKeen https://soylentnews.org/~MichaelDavidCrawford
(Score: 5, Insightful) by JoeMerchant on Monday February 09, @01:49PM (1 child)
Good software engineers write the fewest lines of code to achieve the desired objectives.
The best software engineers craft the objectives to meet all user needs as simply as possible.
🌻🌻🌻🌻 [google.com]
(Score: 3, Insightful) by c0lo on Monday February 09, @08:19PM
Indeed. "Everything should be made as simple as possible, but not simpler"
https://www.youtube.com/@ProfSteveKeen https://soylentnews.org/~MichaelDavidCrawford
(Score: 4, Funny) by looorg on Monday February 09, @11:43AM (12 children)
Next hype wave -- AI Debuggers!
If it can slop out code for people that can't really program I can't see the harm in it also debugging their code for them. It should really be going places ...
(Score: 3, Informative) by c0lo on Monday February 09, @01:12PM (11 children)
Half way there. Use the search query "AI code review".
https://www.youtube.com/@ProfSteveKeen https://soylentnews.org/~MichaelDavidCrawford
(Score: 2) by JoeMerchant on Monday February 09, @01:52PM (10 children)
AI doesn't just debug code, it reviews it, writes (sloppy long) reports, trims those reports down when/as directed, crafts design documents when directed to do so. This is the thing about the current crop of LLMs: they dont' say "I'm sorry, Dave...."
And they do all of it imperfectly, even poorly. And it is up to YOU, dear AI agent operator, to tell them what they got wrong and have them fix it.
https://every.to/@mike_2114 [every.to]
🌻🌻🌻🌻 [google.com]
(Score: 4, Insightful) by turgid on Monday February 09, @02:47PM (9 children)
And they do all of it imperfectly, even poorly. And it is up to YOU, dear AI agent operator, to tell them what they got wrong and have them fix it.
Right, so instead of actually doing our jobs, we now have to train AI agents to do them, badly?
I refuse to engage in a battle of wits with an unarmed opponent [wikipedia.org].
(Score: 2) by JoeMerchant on Monday February 09, @03:53PM (8 children)
>Right, so instead of actually doing our jobs, we now have to train AI agents to do them, badly?
Between managing people of limited abilities being asked to perform above their capabilities and managing AI agents, I must say I prefer managing AI agents.
🌻🌻🌻🌻 [google.com]
(Score: 3, Touché) by turgid on Monday February 09, @03:56PM (1 child)
We could always try training the people. Is that too radical a concept?
I refuse to engage in a battle of wits with an unarmed opponent [wikipedia.org].
(Score: 3, Insightful) by JoeMerchant on Monday February 09, @04:24PM
>We could always try training the people. Is that too radical a concept?
It sounds good in an HR meeting... in practice, over half the software engineers I have encountered at the professional level who were "over their head" in their current role didn't need more time to come up to speed or additional training, they needed a different role altogether - one that didn't involve writing code.
Students at university? Interns? Sure, most of those who are struggling just need a little guidance - the best ones just need time to figure it out for themselves. The ones who are still struggling after a Masters' degree and 3 years working experience? Give me Claude over them, every time, and give them some career counseling; they have talents, it's just that their talents are best applied elsewhere.
🌻🌻🌻🌻 [google.com]
(Score: 2) by c0lo on Monday February 09, @08:41PM (5 children)
And provide your work to the LLM owner for free. While you (or your employer) pay a subscription to use the said LLM.
Training AI for free instead of the naturally stupid human is the next dumb-fucking** step after sacrificing privacy for convenience.
** "They 'trust me'. Dumb fucks." -Mark Zuckerberg
https://www.youtube.com/@ProfSteveKeen https://soylentnews.org/~MichaelDavidCrawford
(Score: 2) by JoeMerchant on Monday February 09, @09:28PM (4 children)
It's at a stage, if I don't train it, others will. At least this way I am near the party instead of standing outside dissing it.
My megacorp employer has already decided they are "in."
🌻🌻🌻🌻 [google.com]
(Score: 2) by c0lo on Tuesday February 10, @12:18AM (3 children)
Let them, what do you have to lose?
That's in the personal gain column - staying aware/connected. Assuming you weren't paid for it, would it worth the cost for you?
At least you get paid for. The OSS guys don't.
In the context of my employer, there's too much proprietary/specialized code to maintain/enhance and the cost of training a specific in-house LLM is too high, while use of general AI coders doesn't make sense, they don't help much.
https://www.youtube.com/@ProfSteveKeen https://soylentnews.org/~MichaelDavidCrawford
(Score: 3, Informative) by JoeMerchant on Tuesday February 10, @12:55AM (2 children)
>>It's at a stage, if I don't train it, others will
>Let them, what do you have to lose?
The training is bi-directional, I'm learning how to use it as well. There's also the slight delusion that my use influences the direction of development (ever so slightly) to serve my kinds of needs.
>Assuming you weren't paid for it, would it worth the cost for you?
It's apparently worth $20 per month to me... I upped to $200 a month for a couple of months, but I really can't utilize that much on my personal time and have any kind of life at the same time, so I'm back at $20 per month - and I feel it's "worth" at least as much to me as my Netflix subscription, for now at least.
>while use of general AI coders doesn't make sense, they don't help much.
Back around April 2025, none of them did much for me. Starting with Claude Sonnet 4 I felt somewhat assisted in some areas, and since then they have improved noticeably several times. You might give Claude Opus 4.6 a try and see what it does with your code - Opus 4.6 is "making sense" of things that Sonnet 4 was blundering around cluelessly in. The tools like Claude Code or the VS Code plugin also make the LLMs quite a bit more helpful because they can check their own work and fix their own (obvious) mistakes.
🌻🌻🌻🌻 [google.com]
(Score: 2) by c0lo on Tuesday February 10, @03:19AM (1 child)
It doesn't come as a surprise to me.
Now, there're some parallels to draw with Wikipedia/StackExchange/Reddit - in which what makes them useful is "community contribution" (including the treatment of moderators, which occasionaly, may go into the "appalling exploitation" side).
But there are some differences that bother me in principle:
* unlike Wikipedia/StackExchange/Reddit, there's no community the LLM "sucks" from, just individuals. That include the lack (or very opaque) moderation of the content. Take this in the context of the Sturgeon's law
* unlike Wikipedia/StackExchange/Reddit, the (eventual) benefit of "growing up" the LLM is not shared publicly/freely. No warranty that your effort will benefit you, the contributor, outside the context of an employer who can pay the the subscription to access it.
+Informative
https://www.youtube.com/@ProfSteveKeen https://soylentnews.org/~MichaelDavidCrawford
(Score: 3, Informative) by JoeMerchant on Tuesday February 10, @03:58PM
>in the context of the Sturgeon's law * unlike Wikipedia/StackExchange/Reddit, the (eventual) benefit of "growing up" the LLM is not shared publicly/freely. No warranty that your effort will benefit you, the contributor, outside the context of an employer who can pay the the subscription to access it.
I find it interesting that Anthropic and some other players are registered as "Public Benefit Corporations" required to produce a report detailing how it has pursued its public benefit purpose (developing safe, steerable AI systems for the long-term benefit of humanity), "making its commitment legally binding and transparent."
How binding? How transparent? WTF does "the long-term benefit of humanity" even mean? The fact the corporation is based in Delaware gives me the opposite of the warm fuzzies right out of the gate, but mostly from a shareholder perspective (Delaware is a great place for majority shareholders to screw the minorities), and I'm not a shareholder in any of these things, except by proxy via some mutual funds.
In the meantime, my corporate owners are encouraging me to play with the toys, on the clock, on their dime. Who am I to say no?
🌻🌻🌻🌻 [google.com]