The focus on "code" is sometimes too strong. This isn't just a code/quant issue. All of this applies too to any "answer" the LLM gives. Experts very often track their lineage - what teachers they had - because the details and slant of what they were taught guided their way of thinking and defined the "handle" they use to grab a problem. There is no "right way" to do most things - there are ways of approaching. Imagining that you are free-flowing and not in a channel is a rookie mistake! Encouraging this way of thinking is like the uptight enforcer of minor social mores: "Elbows!" no elbows on the table is how things are done and this is reified. Like you said "it takes no effort to produce and immense effort to debunk" and sometimes the stakes are high. I'm convinced that racism, for example, operates exactly like this: 0 friction go-with-the-flow insist it is natural...
If AI does anything, it amplifies asymmetries. Hacking, scamming, reducing code quality... its really good at making defending against those hard problems even harder.
My strategy is to skim the workshop to find suspicious sections, then prioritize debunking at least one part that makes the creator look bad.
Then respond with that addressed, with a firm but polite explanation that the document has some problems that need to be addressed carefully before we invest time on it.
None of this works if it’s coming from your boss, but it’s very effective at making people think twice before sending you slop.
People only do the workslop thing when they believe the benefits of showing the work outweigh the reputational risks. If they get caught every time they do it and exposed on an email chain, they start doing less.
Sadly not my experience. When I give this kind of feedback on a document, highlight a gap, or give constructive feedback, most of the time this feedback is fed back to the AI and a new revision of the document shows up.
BUT because it is AI it's not that section that's addressed, it's about 90% of the document that now has changed, and I need to spend another gargantuan effort to review it. It's like by trying to be helpful I actually lose more time.
The AI has made it so that the thinking before writing is mostly gone, and shifted that step to the reviewers.
The Bullshit Asymmetry Principle, a/k/a Brandolini’s Law: “the amount of energy needed to refute bullshit is an order of magnitude bigger than that needed to produce it".
It's worth noting that the context here is a system that was translated by an LLM which is the ideal use case for it. 90% of the thinking had already been done by developers of the existing system.
Yeah, I think most success stories in AI seem to come from ports, or from projects where a deliberate architecture has been laid by humans who understand it, and AI is sprinkling features on top. Adding features and drivers to a Linux distro seems like one of those places where AI is uniquely well positioned to bring value, because the foundation has already been laid.
I wrote a post the other day on how agentic code still needs this kind of architectural foundation to not end up in a tangled mess. It might interest someone:
https://ljtn.github.io/epiq/blog/the-soft-fabric-of-agentic-...
That's a really great point, and you can often start with documenting current behavior with tests and then carefully going through to figure out what bugs you've just enshrined in test cases…
“Ideal use case” such a wild take. The “10%” (assuming you mean all the issues pointed out in the article) are the things that take the most time and effort to do (or fix now), which you can’t do without understanding and fixing the “90%” of “translation”.
You need to understand the system you're working on enough to make well-informed decisions about its current and future states.
In some cases you can get away with not reading the code. And maybe in the future that will be more common. But for now, I personally prefer to read (or skim) the code, and not abdicate to AI.
Or an alternative: The whole background why in Dune thinking machines are outlawed and how it came to pass, that humanity had to fight the Butlerian Jihad.
Free-thinking, critical thinking, has been considered taboo, or an exclusive privilege of the elite, in most human societies since the dawn of civilization.
At work, all thinking has basically become discouraged.
If you are not using AI agents for everything, you're doing something wrong, and wasting company time. It has been emphasized that no one should be writing code by hand and if you think an AI has implemented something wrong you need solid reasoning to show why or else people just label you as some anti-AI troublemaker, who just becomes an obstacle standing in the way of things getting done.
If your human generated opinion is in any way wrong or simply not a true homerun then you get snubbed in future reviews, people stop listening to you.
I saw a post on here, basically they get ai to design the change and give them each each one at a time that they manually type in.
They argued this still allowed them to fully understand what has being implemented and how it fitted together.
I’ve been doing this since I read that and it also allows you to catch stupid stuff while your typing, you can reason about what each little change does and why it’s needed.
This also lets the LLM change the future of the plan if you fine something.
This turns out to save a whole bunch of time later because you already know how it works.
It’s not nearly as fast as just letting the agent do everything.
And more than anything, it doesn't take away the trade-offs of what to build to deliver value for customers, what to make the business improve, what avoids paging you on a vacation.
> and if you think an AI has implemented something wrong you need solid reasoning to show why or else people just label you as some anti-AI troublemaker,
I'm going to read you charitably, and assume that what you really meant is "the burden of proof is entirely on you, and is set unreasonably high." Because of course, if you do think someone has done something wrong and want to say it publicly, you do need a solid reason.
Huh, I don't use AI much/or at all anywhere, apart from googling something and reading the Gemini answer before scrolling down, without facing any backlash. Though people are more likely to discuss and brainstorm their work/problems with me because of this, which I consider a positive. Maybe I've been lucky at where I work.
That is if I ignore the slop MRs with 60+ files changed that I have to painstakingly go through and then politely tell the author to fix the crater sized holes in it, while a vein in my forehead almost explodes. And some part of my sanity is lost forever.
Informative thought process here. Pre-committing benchmarks to measure real efficiency gains might have helped, but hands-off complex rewrites are still not shovel ready.
> hands-off complex rewrites are still not shovel ready
...which is a bit disappointing, actually. I mean, if you give an AI agent the complete code for a working application (ideally including tests), it should be able to translate it into an equivalent application in another language. I can understand if it makes mistakes because of ambiguous or incomplete prompts etc., but for this task you shouldn't even need a prompt, any information it might need is right there in the code?
Agreed. You would expect this to be a task that AI has a massive advantage over humans as it has both the testing suite to keep it on track and has the source implementation to reference. The classic failure case for agents is that as the task increases in complexity, it requires more steps to complete. With each step there's a chance for error. As errors accumulate, they become more likely. With verification functions, like test suites or comparison to some source, the agent is more likely to catch errors and recover.
So porting should be perfect conditions for an AI agent. It has the testing suite. It has the original implementation. Hell, it can A/B the original with its working copy and attach the original to a debugger.
So yeah I dunno. I guess despite some people's rhetoric on this site we shouldn't underestimate how difficult these tasks are "in the real world" (as opposed to like a game port which doesn't have real money on the line so far). And it is kind of a miracle these tools can even get in the ballpark and fool a lot of people.
My suspicion is DHH is intentionally creating controversy to drive interest to Omarchy
If he truly believed in the idea of never reading the code, he'd fire all his developers and have the designers do it all - create very high level feature descriptions and iterate on them. Put your money where your mouth is
I think attempting few-shot agentic ports of an app as an exercise and having the results not be perfect is really a referendum on thinking or using AI in coding in any particularly meaningful way.
I dunno. You still have to test, but in my experience if I noticed unreliability during high load I could just tell Claude "I noticed unreliability during high load" and it would easily have set up a high load benchmark, discovered this broadcast channel issue and fixed it.
It definitely improves the results if you do think but I don't think "you still have to think" is a safe space that's going to save us from unemployment.
Author is saying that some decisions aren't as obvious as "my code should be reliable under high load". Some decisions are tradeoffs that require you to have domain/technical expertise to answer.
And some do require design thinking. Like sndio or alsa, drm, caddy, … With no domain expertise and technical knowledge, a lot of software are out of scope.
> It definitely improves the results if you do think but I don't think "you still have to think" is a safe space that's going to save us from unemployment.
Who said anything about saving people from unemployment?
> I noticed unreliability during high load
This is pretty ambiguous. You still need to provide more context, such as which failure modes you are trying to benchmark.
This post resonates a lot with me and my approach to counter the idea that an LLM (as amazing as a tool it is) gives programmers permission to stop engineering. To me, executing a decision in the face of trade-offs is what the job is all about.
So, I want to keep pulling the thread: is it worth reading—meaning, understanding—the code in order to prevent p95 latency from skyrocketing, avoid OOM death, and keep our apps reliable for customers?
The cost of understanding the code is ostensibly very high relative to just using a clanker (citation needed), so does paying that cost translate to value—what is it worth? Concretely, if vibe coding increases bugs and enshittification, but decreases spending without decreasing revenue (read: customers suffer, but they don't leave), do we still need to pay the cost of reading code? Is it better to invest in stickiness, lobbying, and market capture?
I hope this comment isn't read as advocating for this; it's more of a thought experiment about pragmatism and trade-offs. I know that I personally care a lot about understanding the code, and I believe I'm paid to do exactly that (I very much enjoy understanding code and nobody is paying me to run a business, so I may be biased in this). In everyday SaaS-land—not talking about critical safety systems—we've seen production databases destroyed, personal data leaked, platforms unwittingly exploited, and UX bugs creep into our operating systems, and yet the companies involved keep on keeping on.
Is thinking, meaning taking the time to actually understand what these systems are doing, going to increase their shareholder value?
> (read: customers suffer, but they don't leave), do we still need to pay the cost of reading code? Is it better to invest in stickiness, lobbying, and market capture?
Translating this from Linkedin-speak: Is it okay to subject our users to shitty software and while we use the money saved to bribe politicians, circumvent laws and buy our competitors to kill them?
I agree. I'm using AI on a mid-complexity system (IoT devices, encryption, concurrency, etc) and the number of errors / bad decisions taken "at first" by the AI (or if I don't guide it precisely enough) is huge, and I can only imagine how feeble and unmaintainable an entire codebase done without review from a senior coder would be
I feel a bit sad DHH did the elixir one so dirty. It's the only language with the rule to follow rails to the T and of course Elixir can fly and be incredibly powerful without redis and other bullshit. It feels deliberately hamstrung. Almost maliciously because heavy hitters like Jose Valim have told him about this and he just ignores it.
This article just reminds me that workslop is so hard to counter because it takes no effort to produce and immense effort to debunk
This is why the adverse reaction must be proportional to the effort required to counter the destructive behavior. It must be strongly discouraged.
One of the worst things one can do is pull punches when you get "contributions" that are net-negative because they waste everyone's time.
The focus on "code" is sometimes too strong. This isn't just a code/quant issue. All of this applies too to any "answer" the LLM gives. Experts very often track their lineage - what teachers they had - because the details and slant of what they were taught guided their way of thinking and defined the "handle" they use to grab a problem. There is no "right way" to do most things - there are ways of approaching. Imagining that you are free-flowing and not in a channel is a rookie mistake! Encouraging this way of thinking is like the uptight enforcer of minor social mores: "Elbows!" no elbows on the table is how things are done and this is reified. Like you said "it takes no effort to produce and immense effort to debunk" and sometimes the stakes are high. I'm convinced that racism, for example, operates exactly like this: 0 friction go-with-the-flow insist it is natural...
Aka https://en.wikipedia.org/wiki/Brandolini%27s_law
Mark Twain is halfway around the world before Brandolini puts on his shoes.
If AI does anything, it amplifies asymmetries. Hacking, scamming, reducing code quality... its really good at making defending against those hard problems even harder.
My strategy is to skim the workshop to find suspicious sections, then prioritize debunking at least one part that makes the creator look bad.
Then respond with that addressed, with a firm but polite explanation that the document has some problems that need to be addressed carefully before we invest time on it.
None of this works if it’s coming from your boss, but it’s very effective at making people think twice before sending you slop.
People only do the workslop thing when they believe the benefits of showing the work outweigh the reputational risks. If they get caught every time they do it and exposed on an email chain, they start doing less.
Sadly not my experience. When I give this kind of feedback on a document, highlight a gap, or give constructive feedback, most of the time this feedback is fed back to the AI and a new revision of the document shows up.
BUT because it is AI it's not that section that's addressed, it's about 90% of the document that now has changed, and I need to spend another gargantuan effort to review it. It's like by trying to be helpful I actually lose more time.
The AI has made it so that the thinking before writing is mostly gone, and shifted that step to the reviewers.
The Bullshit Asymmetry Principle, a/k/a Brandolini’s Law: “the amount of energy needed to refute bullshit is an order of magnitude bigger than that needed to produce it".
Illuminati propaganda!
MJ 12
It's worth noting that the context here is a system that was translated by an LLM which is the ideal use case for it. 90% of the thinking had already been done by developers of the existing system.
Yeah, I think most success stories in AI seem to come from ports, or from projects where a deliberate architecture has been laid by humans who understand it, and AI is sprinkling features on top. Adding features and drivers to a Linux distro seems like one of those places where AI is uniquely well positioned to bring value, because the foundation has already been laid. I wrote a post the other day on how agentic code still needs this kind of architectural foundation to not end up in a tangled mess. It might interest someone: https://ljtn.github.io/epiq/blog/the-soft-fabric-of-agentic-...
LLMs are great at translation. That's why language translators were among the first professionals losing their jobs to AI.
It's foundations all the way down. The difference is people have reality as our reinforcement learning environment.
That's a really great point, and you can often start with documenting current behavior with tests and then carefully going through to figure out what bugs you've just enshrined in test cases…
“Ideal use case” such a wild take. The “10%” (assuming you mean all the issues pointed out in the article) are the things that take the most time and effort to do (or fix now), which you can’t do without understanding and fixing the “90%” of “translation”.
smokes cigarette
Life, eh?
You need to understand the system you're working on enough to make well-informed decisions about its current and future states.
In some cases you can get away with not reading the code. And maybe in the future that will be more common. But for now, I personally prefer to read (or skim) the code, and not abdicate to AI.
I dread for the future where "you have to think" is considered a controversial statement.
It’s not the future.
> "as soon as we started thinking for you it really became our civilization"
remember this timeless Agent Smith quote
Or an alternative: The whole background why in Dune thinking machines are outlawed and how it came to pass, that humanity had to fight the Butlerian Jihad.
"Mr. Anderson, what good is a brain if you're unable to think?"
Thinking is expensive. A lot of people try to avoid it for that reason.
Free-thinking, critical thinking, has been considered taboo, or an exclusive privilege of the elite, in most human societies since the dawn of civilization.
At work, all thinking has basically become discouraged.
If you are not using AI agents for everything, you're doing something wrong, and wasting company time. It has been emphasized that no one should be writing code by hand and if you think an AI has implemented something wrong you need solid reasoning to show why or else people just label you as some anti-AI troublemaker, who just becomes an obstacle standing in the way of things getting done.
If your human generated opinion is in any way wrong or simply not a true homerun then you get snubbed in future reviews, people stop listening to you.
I'm probably one of the heavier AI users for agentic workflows. I'm really enjoying it.
However, it doesn't take away the concern of architecture, design, etc. and questioning if the current solution is well designed or not.
In other words, if you can't question the AI and refute it in a topic, you aren't expert enough to use AI in that area in a engineering manner.
Granted, not everything needs an engineer behind it. A shack to store some tools will survive long enough probably
I saw a post on here, basically they get ai to design the change and give them each each one at a time that they manually type in.
They argued this still allowed them to fully understand what has being implemented and how it fitted together.
I’ve been doing this since I read that and it also allows you to catch stupid stuff while your typing, you can reason about what each little change does and why it’s needed.
This also lets the LLM change the future of the plan if you fine something.
This turns out to save a whole bunch of time later because you already know how it works.
It’s not nearly as fast as just letting the agent do everything.
And more than anything, it doesn't take away the trade-offs of what to build to deliver value for customers, what to make the business improve, what avoids paging you on a vacation.
> and if you think an AI has implemented something wrong you need solid reasoning to show why or else people just label you as some anti-AI troublemaker,
I'm going to read you charitably, and assume that what you really meant is "the burden of proof is entirely on you, and is set unreasonably high." Because of course, if you do think someone has done something wrong and want to say it publicly, you do need a solid reason.
Huh, I don't use AI much/or at all anywhere, apart from googling something and reading the Gemini answer before scrolling down, without facing any backlash. Though people are more likely to discuss and brainstorm their work/problems with me because of this, which I consider a positive. Maybe I've been lucky at where I work.
That is if I ignore the slop MRs with 60+ files changed that I have to painstakingly go through and then politely tell the author to fix the crater sized holes in it, while a vein in my forehead almost explodes. And some part of my sanity is lost forever.
You know who else was thinking? Hitler!
Informative thought process here. Pre-committing benchmarks to measure real efficiency gains might have helped, but hands-off complex rewrites are still not shovel ready.
> hands-off complex rewrites are still not shovel ready
...which is a bit disappointing, actually. I mean, if you give an AI agent the complete code for a working application (ideally including tests), it should be able to translate it into an equivalent application in another language. I can understand if it makes mistakes because of ambiguous or incomplete prompts etc., but for this task you shouldn't even need a prompt, any information it might need is right there in the code?
Agreed. You would expect this to be a task that AI has a massive advantage over humans as it has both the testing suite to keep it on track and has the source implementation to reference. The classic failure case for agents is that as the task increases in complexity, it requires more steps to complete. With each step there's a chance for error. As errors accumulate, they become more likely. With verification functions, like test suites or comparison to some source, the agent is more likely to catch errors and recover.
So porting should be perfect conditions for an AI agent. It has the testing suite. It has the original implementation. Hell, it can A/B the original with its working copy and attach the original to a debugger.
So yeah I dunno. I guess despite some people's rhetoric on this site we shouldn't underestimate how difficult these tasks are "in the real world" (as opposed to like a game port which doesn't have real money on the line so far). And it is kind of a miracle these tools can even get in the ballpark and fool a lot of people.
With all these ports, I always wonder: Who implements new features, and in which code base?
It's easy to port a codebase (at least, if you have automated tests), but once you've done that, do you
- Implement features in the old code and port again?
- Implement them in the new, unfamiliar codebase?
- Just write a Jira ticket and let the agent YOLO it?
Really appreciate the time taken to dig into this with solid improvements to boot, great work
My suspicion is DHH is intentionally creating controversy to drive interest to Omarchy
If he truly believed in the idea of never reading the code, he'd fire all his developers and have the designers do it all - create very high level feature descriptions and iterate on them. Put your money where your mouth is
I think attempting few-shot agentic ports of an app as an exercise and having the results not be perfect is really a referendum on thinking or using AI in coding in any particularly meaningful way.
I dunno. You still have to test, but in my experience if I noticed unreliability during high load I could just tell Claude "I noticed unreliability during high load" and it would easily have set up a high load benchmark, discovered this broadcast channel issue and fixed it.
It definitely improves the results if you do think but I don't think "you still have to think" is a safe space that's going to save us from unemployment.
Author is saying that some decisions aren't as obvious as "my code should be reliable under high load". Some decisions are tradeoffs that require you to have domain/technical expertise to answer.
And some do require design thinking. Like sndio or alsa, drm, caddy, … With no domain expertise and technical knowledge, a lot of software are out of scope.
> It definitely improves the results if you do think but I don't think "you still have to think" is a safe space that's going to save us from unemployment.
Who said anything about saving people from unemployment?
> I noticed unreliability during high load
This is pretty ambiguous. You still need to provide more context, such as which failure modes you are trying to benchmark.
This post resonates a lot with me and my approach to counter the idea that an LLM (as amazing as a tool it is) gives programmers permission to stop engineering. To me, executing a decision in the face of trade-offs is what the job is all about.
So, I want to keep pulling the thread: is it worth reading—meaning, understanding—the code in order to prevent p95 latency from skyrocketing, avoid OOM death, and keep our apps reliable for customers?
The cost of understanding the code is ostensibly very high relative to just using a clanker (citation needed), so does paying that cost translate to value—what is it worth? Concretely, if vibe coding increases bugs and enshittification, but decreases spending without decreasing revenue (read: customers suffer, but they don't leave), do we still need to pay the cost of reading code? Is it better to invest in stickiness, lobbying, and market capture?
I hope this comment isn't read as advocating for this; it's more of a thought experiment about pragmatism and trade-offs. I know that I personally care a lot about understanding the code, and I believe I'm paid to do exactly that (I very much enjoy understanding code and nobody is paying me to run a business, so I may be biased in this). In everyday SaaS-land—not talking about critical safety systems—we've seen production databases destroyed, personal data leaked, platforms unwittingly exploited, and UX bugs creep into our operating systems, and yet the companies involved keep on keeping on.
Is thinking, meaning taking the time to actually understand what these systems are doing, going to increase their shareholder value?
> (read: customers suffer, but they don't leave), do we still need to pay the cost of reading code? Is it better to invest in stickiness, lobbying, and market capture?
Translating this from Linkedin-speak: Is it okay to subject our users to shitty software and while we use the money saved to bribe politicians, circumvent laws and buy our competitors to kill them?
I agree. I'm using AI on a mid-complexity system (IoT devices, encryption, concurrency, etc) and the number of errors / bad decisions taken "at first" by the AI (or if I don't guide it precisely enough) is huge, and I can only imagine how feeble and unmaintainable an entire codebase done without review from a senior coder would be
I had to google “clanker”.
You're one of today's lucky 10,000 [0]
[0] https://xkcd.com/1053/
There's a youtuber I watch who's playing through Zero Company, and I was very amused to hear NPCs yell "clanker" at enemy droids.
I feel a bit sad DHH did the elixir one so dirty. It's the only language with the rule to follow rails to the T and of course Elixir can fly and be incredibly powerful without redis and other bullshit. It feels deliberately hamstrung. Almost maliciously because heavy hitters like Jose Valim have told him about this and he just ignores it.
Not a good look tbh