One very unfortunate curse we are stuck with is LLM providers pulling the rug from under us and swapping models with quantized versions during peak demand or on monthly plans. You never know if that's the reason AI starts writing shitty code or it's just the context size that has grown. On top of that there is the queuing mechanism which makes tasks take longer. Sometimes AI is super fast and sometimes a simple task takes 10, 15 or even 20 minutes of "thinking". You may not realize this if you are just using your same subscription but if you switch between providers you can immediately tell. I advise everyone to invest $5 in openrouter, deepseek and qwencode. Just do an experiment and use same model (deepseek-4.1 flash) asking same prompt on same code, then look at the result. Spoiler: you will see trash coming out of openrouter at 1-5 tokens/second, you will see slightly faster and better quality output from qwencloud and when you get to deepseek API where you are paying per token you will not believe this is the same model. Because it isn't.
I also had similar feelings recently. I think the article is good and captures many of the issues I have with the current state of AI development, but I feel like the conclusion/reaction is somewhat exaggerated.
Of course if someone wants to stop using AI completely that's a completely valid decision[0], but I somewhat feel like AI is just a tool that can be easily misused.
I constantly have to review giant PRs and I noticed that I'm handwaving them more and more often. We went from almost no commit messages to walls of text that no one reads. We're starting to become bottlenecked on reviews because code is coming out too fast.
But at the same time, these are mostly issues stemming from a lack of understanding of why some of the standards/processes existed in the first place. If a developer thinks the commits have to be written just to tick a checkbox, they won't care about making them readable.
And at the same time, I'm getting a lot of value from AI, in tasks that do not necessarily have such adverse effects:
- I can create quick tools to test something, or parse/process some data. In these instances code quality is not important and I don't really want to spend hours on developing it myself (just to feel accomplished?)
- I can research issues in our codebase by just providing a log file. It's not always gonna be accurate or correct but it often gives me a very good starting point, almost always quicker than I could've done it myself
- While I do not use AI to completely generate ticket descriptions, asking it to generate me a body containing the relevant code snippets and references allows me to focus on verifying that what I'm writing is correct and understandable.
Etc etc.
So I don't know if it's just the nature of my work, the fact that I have a different skillset, or different priorities. But it somehow feels weird to me wanting to completely abandon AI just because in some cases it can lead to frustrating consequences.
[0]: I too just started a new project where I'm forcing myself to use absolutely no AI!
Yeah, I have never understood the over reliance on AI. Writing the code is not the challenge. The time it takes to push a new feature and test it out is often trivial, maybe a few hours.
The real challenge is forming the new ideas in the first place and most of those new ideas coming either from using the code as a product or time spent maintaining and refactoring large code.
Wriring the code is not the challenge, but it's what was taking up most of the time. Not the typing itself, but also because I had to think of how to implement it.
Now I can just say "add 2FA" and in 5 minutes, while I test something else, it is done.
It also made iterations a lot faster, you can try something out, see how it feels, if it doesn't work, you can just trash all the code and start again.
Haven't typed a line of code or read any code for over 6 months now.
And I used to love coding and be a competitive programmer, but this is how "coding" goes nowdays.
I have a mental model of what it would do, and how it would work, and I ask questions to confirm things and tell it to watch for specific gotchas. Then simply test the feature myself a bit.
Security-wise, I think the latest cyber models are better than me anyway at finding vulnerabilitates and pentesting such features.
Plus 2FA is a very common pattern, so it likely has in the training dataset many really good implementations.
> Security-wise, I think the latest cyber models are better than me anyway at finding vulnerabilitates and pentesting such features.
I don't doubt that, but they are equally good in making mistakes, over-engineering, or adding things you never asked for. They have all sorts of patterns in their training data from excellent to inadequate and I find them challenging to guide them consistently in one direction. Also with questions and tests, they can add something extra you didnt need and you dont know about, so your scrutinizing questions and test cases could miss that.
At least for myself, I didnt find them reliable enough yet to do what you describe and just not look at the code at all.
This app[0] is entirely "vibe-coded", started on bolt.new which was like the OG way to vibe code and slowly improved over 1-2 years using various models.
I have no idea how the code looks like, and barely even tested the app entirely, because it is still not released yet, but I do think a lot about new features, tweaks, improvements, etc. Years of programming and game development did help, but I don't think anymore that code is relevant, as long as it looks ok and feels good.
You are abdicating your responsibility, which is fine for toy apps for yourself, but less fine when people expect that 2fa implementation to protect their accounts.
I think it's the opposite, with the right guideance, testing, frameworks, and using the top models today, implementing something with AI is most of the times better than what most developers would do.
I basically only manually test e2e myself, other tests are automated, code review is automated. I can tell the model to test for me too specific things or to add tests for specific potential issues, performance benchmarks, compared different implementations, etc.
The focus is a lot more around the code than on the code.
The limiting factor of development (for money) is good ideas, or valuable ideas. If you have 5 features you want, but only time in a month for 2, then that forces you to choose the best 2, which come with the tech debt, support and opportunity cost.
Removing the opportunity cost doesn't eliminate the other two costs of a feature
> you stop questioning, and start accepting as good some code you would have never accepted, just because you cannot tell why it’s bad. You have lost control
I have not lost control.
I my most prolific project I do not review the code, but I QA test extensively.
In other projects at work, I review the code.
I prompt to simplify, I challenge implementation that solves irrelevant edge cases, resulting in much smaller PRs.
In projects where I do not work alone, I still write two line PR descriptions myself.
Dumping paragraphs of AI output into the description of a MR where I ask others to review I consider disrespectful.
---
> If you turn off your brain, and relax babysitting AIs, you’re not getting any better. You’re losing value
I'm hardly turning off my brain here.
As the author notes, the context switching and so on takes concentration and effort too.
I can say without doubt that I am more productive than ever.
I am getting better by the month, and I am not currently losing value, until the AI fully replaces both me and the author.
I find it strange to see people writing articles like this as if everyone has used AI for decades. I've programmed for decades. I thought I retired three years ago but got an offer I couldn't refuse. Already there were little things I'd forgotten how to use.
Over the past six months I tried using Claude, chatgpt, Grok and Gemini. At best I got reminders of how things worked. People online say they use them to write their code. The code they supplied to me has NEVER worked or was so convoluted that I threw it away and did it myself.
At most, I use these tools as search engines. Even then some references are poor.
I'm starting to think this is becoming a sad, sad world and AI is just the new TV of the programming world.
“ There were tasks I could have done in 20 minutes easily, that took 5 minutes of an AI agent, and then 2 days for me to review.”
This is using AI for productivity in any domain, in a nutshell. I just wrote a book using Claude as an experiment, and while the thing got done and it was an amazing tool and a great experience, what I’m left with is a book where every line needs rewriting, there are logical inconsistencies throughout, and the style is so bad it should actually just be binned rather than rewritten.
I noticed lately that recent LLMs write very short sentences, sprinkling so many periods over one paragraph. I'm pretty sure that this is some new regression that we're having with new models.
That's a side effect of people bashing the em-dash and the labs rushing to correct. The period is the most common replacement (along with some minor tweaks to sentence structure), so now that models are being RL'd away from the em-dash the models are overusing a new construction.
Hmmm I was using Claude Opus for the most part. I didn’t notice any particular shortness in the length of the sentences, but that’s also an issue in itself – the whole thing is just average fine-ness: the sentences are fine, the paragraphs are fine, everything is generic and inoffensive and fine. But at the same time it was giving huge amounts of pushback when I was getting it to interrogate logical faults of certain domains, to the point where I was having to argue with and even convince an AI using data and even its own analysis that there was cause to talk about certain things from a certain angle (can appreciate I’m being vague about which things). It can’t write complex, deeply-claused sentences (or won’t). And it also wants to write snappy little concluding sentences that make a fairly well-argued piece of sociological analysis read like Sex and the City or something.
And just like that, the book got abandoned.
N.b. some of its analysis and laying out of faults in arguments was actually pellucid and brilliant, it can’t be denied. Just it comes with prose that can’t really be used for anything. And even on another occasion when I got it to help me redraft and extend a different book of mine, then it randomly and consistently started stripping out all the stylistic flourishes out of my sentences, to the point where it couldn’t notice that word choices were very deliberate and actually set up little punchlines and logical payoffs paragraphs or chapters hence. And even when I explained and showed it what it was doing, it was like “ahh that’s so clever and brilliant” but just continued to do the same thing.
The biggest problem with AI isn't being able to do things. You can do them. The question is "should you?" I had a peer dev this week put in a commit that was 1.2mb (this despite an AGENTS.md saying not to do exactly this) of plan documents and ~20 lines of code. The conversation on the PR review was probably one of the best learning experiences for both of us. It was all about the ramifications of a book's worth of AI pontificating about what it would do and what succeeded and what failed. And we were able to validate that that plan document had huge side effects as we prompted with it in the repo. As we read the document we realized that it was just 1.2MB of smart sounding gibberish and maybe two bullet points of signal.
The point about multitasking getting out of control very much hits home for me. This summer my team tried going hard for agentic coding. Within a month we had 6 people each juggling at least two projects at a time, all the time.
Soon we rediscovered Little’s Law. WIP was piling up and we were getting overwhelmed at the integration phase, and realized that we had got really good at starting projects but actually finishing them was a struggle. Tickets were moving fine, of course. Our rate of generating code and committing PRs was through the roof. But getting actual projects to a point where the stakeholders and customers were happy with the result was just not happening.
So now we have gone back to strict WIP limits and requiring every non-trivial project to have at least two people collaborating on it. The rate at which we are churning out code has gone back down, along with the token bill, but the logjam is clearing. Better yet, the stakeholders, who never cared about our quantitative velocity metrics in the first place, have eased off on complaining that we aren’t getting anything done.
My workflow leans on AI for some tasks: review what I wrote, drop me a one-off script (this one is the generative exception), and so on.Or even initial scaffolding.
For the rest I do not. I do not place AI-generated code anywhere.
You lose all control AND UNDERSTANDING.
When things go wrong it gets very messy.
I will keep doing this, I think it works well, I emjoy programming and I think it is productive.
For testimg I tend to write randomized testing, which takes a bit of design but oncr you have it, well, it os test-generatove and increases the quality of checks.
It’s not just coding that’s affected by the skill degradation issue. As a sysadmin and jack of all trades IT engineer over reliance on AI has unsurprisingly led to a serious knowledge dip.
I see great engineers troubleshoot everything by pasting logs into the prompt and blindly accepting the answer. Zero added value while they ctrl-c ctrl-v themselves out of a job.
I think there might be (dare I say) a middle ground to get the productivity of the llm, esp as we evolve them, while still maintain a global and even fine-grain comprehension of a code base.
It is not a simple change, however, but a fundamental one.
Overall I think we are still living in the past and try to apply ourselves to the future. But if the ai craze is to be taken clear-headedly for what it is, it is a complete break from the von Neumann computer and all its resulting artifacts. So why should we use the same tools?
I think that comparing AI to cocaine is a really good analogy. It gives you that feeling that you are amazing and can do anything, but it's just a feeling, in the end you are just high and sooner or later will have to come back to reality and face a bad implementation or a hangover
Where I find LLMs most useful is maybe not even writing code (though helps with boilerplate and test cases, mechanical refactoring, translation from one form into another, rebases and conflicts), but debugging weird production issues, where I have to pull data from multiple places (with often horrible UX) and correlate the data and logs and code. Just need to be careful to follow its reasoning as it can easily go down the wrong route if it misses some important part of the contex, but an SME in the area spots that and directs the agent accordingly. This cut down time for investigations quite a bit.
I get the sentiment but really you can't argue you are nearly as productive from a business point of view without LLM agentic assistance. Sure do a hand code Friday to keep skills up similar to how a airplane pilot hand flies in the clouds every now and then to keep up skills, but autopilots most of the time.
T2 the war has already happened and Skynet sends the T-100 back in time.
T3 Skynet already exists and is hunting John Connor down
I wanna see the conference room meeting where they decide to push an unfinished, unstoppable technology. I guess T1 is the closest as it happens when Skynet "gains intelligence"
It's time travel, so "already happened" is not well defined.
> I wanna see the conference room meeting where they decide to push an unfinished, unstoppable technology. I guess T1 is the closest as it happens when Skynet "gains intelligence"
This is the mid-point of T3. From the Wikipedia plot summary:
General Brewster is supervising the development of Skynet for Cyber Research Systems (CRS), an autonomous weapons developer. The Chairman of the Joint Chiefs of Staff pressures him to activate Skynet to stop an anomalous computer virus from invading servers worldwide. Brewster fails to discover that the virus was Skynet becoming sentient. John, Kate, and the Terminator arrive too late to stop him from activating it. The T-X appears, fatally injures Brewster, and controls weaponized CRS T-1 and HK drones to kill other employees.
> I stopped reading those descriptions, like, I told it what to do, I assume it did what I said. The description is for others.
If you cannot read it as the author, what hope do I have to read and make sense of the wall of text which doesn’t seem to describe what I actually need to start reviewing.
I really really encourage everyone to write their own descriptions for PRs. If you cannot succinctly describe it in a way another human understands then you don’t understand your own change and you should withdraw your request.
I vehemently oppose delegating our brains to the AI. I agree with the article in spirit.
> let me tell you about this experience, and how it was turning me dumber, lazy, and a worse developer.
Though the article discusses from the point of using agents, I digress to the topic of building with AI in general.
My experience has been the exact opposite. A new idea (usually related to correctness or architecture) is discussed first with the LLM where it defaults to average Joe idiotic bullshit pushback.
This frustrates me and I abuse the LLM for being idiotic by explaining the how. This results in a more refined and concrete form of the abstraction leading me to even more insights.
The LLM remains an idiot. But a useful idiot nonetheless.
Of course if someone wants to stop using AI completely that's a completely valid decision[0], but I somewhat feel like AI is just a tool that can be easily misused.
I constantly have to review giant PRs and I noticed that I'm handwaving them more and more often. We went from almost no commit messages to walls of text that no one reads. We're starting to become bottlenecked on reviews because code is coming out too fast.
But at the same time, these are mostly issues stemming from a lack of understanding of why some of the standards/processes existed in the first place. If a developer thinks the commits have to be written just to tick a checkbox, they won't care about making them readable.
And at the same time, I'm getting a lot of value from AI, in tasks that do not necessarily have such adverse effects:
- I can create quick tools to test something, or parse/process some data. In these instances code quality is not important and I don't really want to spend hours on developing it myself (just to feel accomplished?)
- I can research issues in our codebase by just providing a log file. It's not always gonna be accurate or correct but it often gives me a very good starting point, almost always quicker than I could've done it myself
- While I do not use AI to completely generate ticket descriptions, asking it to generate me a body containing the relevant code snippets and references allows me to focus on verifying that what I'm writing is correct and understandable.
Etc etc.
So I don't know if it's just the nature of my work, the fact that I have a different skillset, or different priorities. But it somehow feels weird to me wanting to completely abandon AI just because in some cases it can lead to frustrating consequences.
[0]: I too just started a new project where I'm forcing myself to use absolutely no AI!
I use it to find reasoning gaps, add examples, add citations etc. The LLM/Agent can find them quicker than I.
The real challenge is forming the new ideas in the first place and most of those new ideas coming either from using the code as a product or time spent maintaining and refactoring large code.
Anyways, if you want to continue on the path towards regaining control and take it to the next level I wrote something similar here: https://blog.sharefile.systems/be-brave-go-low/
Now I can just say "add 2FA" and in 5 minutes, while I test something else, it is done.
It also made iterations a lot faster, you can try something out, see how it feels, if it doesn't work, you can just trash all the code and start again.
Haven't typed a line of code or read any code for over 6 months now.
And I used to love coding and be a competitive programmer, but this is how "coding" goes nowdays.
I have a mental model of what it would do, and how it would work, and I ask questions to confirm things and tell it to watch for specific gotchas. Then simply test the feature myself a bit.
Security-wise, I think the latest cyber models are better than me anyway at finding vulnerabilitates and pentesting such features.
Plus 2FA is a very common pattern, so it likely has in the training dataset many really good implementations.
I don't doubt that, but they are equally good in making mistakes, over-engineering, or adding things you never asked for. They have all sorts of patterns in their training data from excellent to inadequate and I find them challenging to guide them consistently in one direction. Also with questions and tests, they can add something extra you didnt need and you dont know about, so your scrutinizing questions and test cases could miss that.
At least for myself, I didnt find them reliable enough yet to do what you describe and just not look at the code at all.
I have no idea how the code looks like, and barely even tested the app entirely, because it is still not released yet, but I do think a lot about new features, tweaks, improvements, etc. Years of programming and game development did help, but I don't think anymore that code is relevant, as long as it looks ok and feels good.
[0]: https://ultimidi.com
I basically only manually test e2e myself, other tests are automated, code review is automated. I can tell the model to test for me too specific things or to add tests for specific potential issues, performance benchmarks, compared different implementations, etc.
The focus is a lot more around the code than on the code.
I dread to think what that means in security conscious code.
Removing the opportunity cost doesn't eliminate the other two costs of a feature
I have not lost control.
I my most prolific project I do not review the code, but I QA test extensively.
In other projects at work, I review the code.
I prompt to simplify, I challenge implementation that solves irrelevant edge cases, resulting in much smaller PRs.
In projects where I do not work alone, I still write two line PR descriptions myself.
Dumping paragraphs of AI output into the description of a MR where I ask others to review I consider disrespectful.
---
> If you turn off your brain, and relax babysitting AIs, you’re not getting any better. You’re losing value
I'm hardly turning off my brain here.
As the author notes, the context switching and so on takes concentration and effort too.
I can say without doubt that I am more productive than ever.
I am getting better by the month, and I am not currently losing value, until the AI fully replaces both me and the author.
Over the past six months I tried using Claude, chatgpt, Grok and Gemini. At best I got reminders of how things worked. People online say they use them to write their code. The code they supplied to me has NEVER worked or was so convoluted that I threw it away and did it myself.
At most, I use these tools as search engines. Even then some references are poor.
I'm starting to think this is becoming a sad, sad world and AI is just the new TV of the programming world.
This is using AI for productivity in any domain, in a nutshell. I just wrote a book using Claude as an experiment, and while the thing got done and it was an amazing tool and a great experience, what I’m left with is a book where every line needs rewriting, there are logical inconsistencies throughout, and the style is so bad it should actually just be binned rather than rewritten.
Have you noticed it, too?
And just like that, the book got abandoned.
N.b. some of its analysis and laying out of faults in arguments was actually pellucid and brilliant, it can’t be denied. Just it comes with prose that can’t really be used for anything. And even on another occasion when I got it to help me redraft and extend a different book of mine, then it randomly and consistently started stripping out all the stylistic flourishes out of my sentences, to the point where it couldn’t notice that word choices were very deliberate and actually set up little punchlines and logical payoffs paragraphs or chapters hence. And even when I explained and showed it what it was doing, it was like “ahh that’s so clever and brilliant” but just continued to do the same thing.
Soon we rediscovered Little’s Law. WIP was piling up and we were getting overwhelmed at the integration phase, and realized that we had got really good at starting projects but actually finishing them was a struggle. Tickets were moving fine, of course. Our rate of generating code and committing PRs was through the roof. But getting actual projects to a point where the stakeholders and customers were happy with the result was just not happening.
So now we have gone back to strict WIP limits and requiring every non-trivial project to have at least two people collaborating on it. The rate at which we are churning out code has gone back down, along with the token bill, but the logjam is clearing. Better yet, the stakeholders, who never cared about our quantitative velocity metrics in the first place, have eased off on complaining that we aren’t getting anything done.
For the rest I do not. I do not place AI-generated code anywhere.
You lose all control AND UNDERSTANDING.
When things go wrong it gets very messy.
I will keep doing this, I think it works well, I emjoy programming and I think it is productive.
For testimg I tend to write randomized testing, which takes a bit of design but oncr you have it, well, it os test-generatove and increases the quality of checks.
I see great engineers troubleshoot everything by pasting logs into the prompt and blindly accepting the answer. Zero added value while they ctrl-c ctrl-v themselves out of a job.
I think there might be (dare I say) a middle ground to get the productivity of the llm, esp as we evolve them, while still maintain a global and even fine-grain comprehension of a code base.
It is not a simple change, however, but a fundamental one.
Overall I think we are still living in the past and try to apply ourselves to the future. But if the ai craze is to be taken clear-headedly for what it is, it is a complete break from the von Neumann computer and all its resulting artifacts. So why should we use the same tools?
https://youtu.be/Jt0OoXluC8g?t=248 at 4:08
T3 Skynet already exists and is hunting John Connor down
I wanna see the conference room meeting where they decide to push an unfinished, unstoppable technology. I guess T1 is the closest as it happens when Skynet "gains intelligence"
> I wanna see the conference room meeting where they decide to push an unfinished, unstoppable technology. I guess T1 is the closest as it happens when Skynet "gains intelligence"
This is the mid-point of T3. From the Wikipedia plot summary:
https://www.youtube.com/watch?v=JXJiSZhA5cg
If you cannot read it as the author, what hope do I have to read and make sense of the wall of text which doesn’t seem to describe what I actually need to start reviewing.
I really really encourage everyone to write their own descriptions for PRs. If you cannot succinctly describe it in a way another human understands then you don’t understand your own change and you should withdraw your request.
> let me tell you about this experience, and how it was turning me dumber, lazy, and a worse developer.
Though the article discusses from the point of using agents, I digress to the topic of building with AI in general.
My experience has been the exact opposite. A new idea (usually related to correctness or architecture) is discussed first with the LLM where it defaults to average Joe idiotic bullshit pushback.
This frustrates me and I abuse the LLM for being idiotic by explaining the how. This results in a more refined and concrete form of the abstraction leading me to even more insights.
The LLM remains an idiot. But a useful idiot nonetheless.