I think most of this is correct, in spite of potentially being built on a bad assumption.
The assumption is that LLMs should be writing the code and human engineers reviewing and verifying the LLM output. And that this pushes the cost of producing down. And I fundamentally disagree with that.
Every time I ask LLMs to write code, even with Opus 4.8 (haven't tried it with Opus 5 yet), what I get ends up being totally rewritten. LLMs still aren't good at writing maintainable code. Can they write plausibly functional code? Yes. But it won't survive the long term. People using LLMs to write all their code are gambling on them eventually getting to a point where the LLMs can fix their own code. It's possible, but I wouldn't necessarily bet on it.
Where I have found immense value from LLMs is in code review. Repeated review by LLMs catches an amazing amount of potential issues. They really shine on security review, but are very effective with any kind of review.
The other thing that the "LLMs write code camp" misunderstands is that writing was never the bottleneck. Understanding was. And understanding the code is still the bottleneck. But understanding is truly gained during the writing loop. The understanding you gain from pure reading or code review is marginal compared to the understanding you gain while writing.
Most of the time previously spent writing was actually spent updating and deepening our understanding of the system under development. There's no replacement for that understanding in a world where LLMs are doing the writing.
But if you flip it: humans write, LLMs review, then you still get a major gain -- not in speed, but in quality. And you keep the understanding loop intact. I would propose that this might be the best way to deploy LLMs.
" But understanding is truly gained during the writing loop. The understanding you gain from pure reading or code review is marginal compared to the understanding you gain while writing."
Debugging code step by step is how I understand complicated code.
I write code in Go in an established codebase. 90% of the time the LLM gets a basic api/feature right and fully e2e tested as long as I give it enough business context in the prompt. The other 10% of the time I have to do some follow up prompts to either change the behaviour or change the approach of a given step in the flow or just to point out that the wrong pattern was used and please rather use our codebase standard. I haven’t actually hand-written code in well over a year. Is this way faster? Absolutely. Does it lead to better code? Yes, because I have time to write all the regression tests that keep the behaviour as I wanted it when others come bungling around.
I do think the comment is right about understanding. I'm still feeling out what I think is the right level of understanding to invest in now. I think it is probably not the deep line by line understanding I used to have of the systems I worked on. But I also think it's easy to remain too aloof from how the system is being built, and that this is very bad. I think the right answer is somewhere in between, but I'm still working through my own process and forcing functions to strike that balance properly.
You're right about understanding. Where we might disagree is the conclusion. I think it's possible to build the understanding without having to type out syntax by hand (debugging, writing tests, etc). Maybe we're missing much better verification tools? LLMs will likely play a big part in those too (explain the codebase, walk me through A, B, C etc)
How do you handle the fact that an LLM can't seem to help itself from disgorging page after page of words no matter what it's asked to do? I've never seen an LLM say "this looks good as-is; I would not spend any more time on it; what's next?" it will always seem to suggest using another pattern or additional abstractions or other yak shaving.
But to be fair human code reviews have the same problem. It's like reviewers feel they have not done their job if they don't find something wrong.
Yeah, that's just a problem with review. I tend to keep it going until the stuff its finding is stuff that I look at and think "I can live with that". Usually by that point it's found the worthwhile stuff and what remains is minor or, like you said, yak shaving.
There used to be a bit of "wisdom" passed around that said "when you submit something to the client (or management, etc) for approval, leave an obvious mistake somewhere." Then they will find it and point it out, you can easily correct it, and everyone is happy. Otherwise they will find something to critique, just to demonstrate that they contributed something to the review.
I've tried and tried to get an LLM to delete code, it couldn't do it. I knew a file was 40% bad so what I ended up doing was deleting the file, then asking AI to create the missing file. That's how I got AI to delete code :)
I’ve found stating “only call out critical issues, report but ignore everything else” will basically get what you want. Put that in a review skill and you’re good to go.
To be clear I mostly only use Opus and Gemini Flash but this might work for others too.
My guess would be that this is a matter of fine-tuning the model to the problem. You would need to run the review experiment N times, observe the distribution of answers, and then fine tune with model operational parameters.
Perhaps someone more knowledgable could jump in here to clarify?
The verbosity itself seems to be a problem with post-4.6 Claude rather than a general issue with all LLMs. and IME the yak shaving is due to an overly generic prompt. We have a generic automated review bot but it’s prompted to only look for errors and never suggests refactoring; that was a conscious trade off to avoid what you’re talking about. If you’re reviewing interactively then you can give additional guidance about any code smells etc. that jump out
I personally rely on this behavior of LLMs. For example, for reviewing prose, I might ask the LLM to find five issues. If there are obvious issues I missed, then it'll find them. But if all the issues it lists are minor or hallucinated, then I can have some confidence there weren't any glaring issues.
I do the same for LLM code review comments: some changes are out of scope or could be moved to a separate PR; some edge cases don't happen in practice and should just fail noisily instead of writing more code to maintain. When these are the only issues it's raising, then I know it's done.
It is surprisingly hard to do in a single prompt. I’ve had good luck though with being very explicit about asking for review, then to think deeply about what actually matters, and then reduce to a very short extremely concise reply. —- the model NEEDS to output those pages of text as part of its thinking process. It’s used to doing it in the assistant reply but can be cajoled into doing it in the thinking tokens.
I've come to realize that the smarter LLMs get, the less they understand the point of abstractions.
I've been using it for a unity game for the past few years. Nowadays it will go sleuthing into packages and assembly and make decisions based upon what it sees there.
It will make comments about why it's doing something based upon a function call 3 methods deep.
God forbid any of these details change in a minor version update.
Try reducing thinking and use lighter models. For example I observed that using Sonnet works much better (compared to Opus) for tasks I want to be in control of architecture and just need a faster code input.
> The other thing that the "LLMs write code camp" misunderstands is that writing was never the bottleneck. Understanding was. And understanding the code is still the bottleneck. But understanding is truly gained during the writing loop. The understanding you gain from pure reading or code review is marginal compared to the understanding you gain while writing.
This was my stance a couple of years ago, but now I've given it up.
It turns out writing actually was the bottleneck. You can understand perfectly well what you want, but writing it is long and tedious to the point where you find excuses not to do it. Particularly with version 2, the step where you have an OK system and you want to improve it. Quite a lot of changing the code is just useless busywork: re-wiring old functions, moving imports around, searching for locations that benefit from extracting a common piece of code. And each time you do one of those, there's a decent chance you did something even more trivial like forgetting a semicolon or calling the wrong function.
Now that I have an LLM helping me, I can see why. The critical decision is a terse declarative like "we need to have several TCP connections instead of one, and just use the sequence number to arbitrate". A human junior programmer could perfectly well understand what this meant, but he would have to go through all of the above to get to the final product. Now, I can just tell the LLM and I will get what I want, even with the things I didn't explicitly state, without spending attention.
This means I can use my attention on the things that matter. So instead of spending today thinking about how to arbitrate between the TCP connections and tomorrow thinking about pre-calculating my outgoing orders, I can just do both today. I don't waste the good waking hours chasing minor bugs, I just think about the large structure.
I get the feeling the best programmers of years past were actually masters of the little things, which led them to be able to look at the big things. Essentially it was cheaper for them to get to the top of the mountain, where you can see the landscape. Kinda like how the kid who was good at mental arithmetic in primary school was also good at calculus at the end of high school: if you don't have to concentrate on the little things, you have time for the big things.
>It turns out writing actually was the bottleneck. You can understand perfectly well what you want, but writing it is long and tedious to the point where you find excuses not to do it.
Speak for yourself. Writing code has never been a bottleneck for some of us. I can't speak for everyone, and neither should you.
>Now, I can just tell the LLM and I will get what I want, even with the things I didn't explicitly state, without spending attention.
This should worry you. All too often the LLM invents things I didn't ask for and implements things I didn't need. YMMV, I guess. If slop gets the job done, and nobody notices, then who should care?
> And each time you do one of those, there's a decent chance you did something even more trivial like forgetting a semicolon or calling the wrong function.
how did you decide to pick the most trivial kind regression for this example? do you compile your code before checking it in?
> A human junior programmer could perfectly well understand what this meant, but he would have to go through all of the above to get to the final product. Now, I can just tell the LLM and I will get what I want, even with the things I didn't explicitly state, without spending attention.
the main efficiency you have described here is offloading the verification of a change onto the LLM. that is the bottleneck. readers can decide whether a non-deterministic statistical model is a good tool for this job
> I get the feeling the best programmers of years past were actually masters of the little things, which led them to be able to look at the big things
the best programmers understand that their job is to automate workflows, and that includes their own. if you're worried about missing a semicolon, I'm sorry to say that's a skill issue
I’ve found that doing the full requirements capture, planning, writing, reviewing, gardening loop with frontier models has worked quite well since last October, and phenomenally since Fable 5.
The key I’ve found is human peer review. The reviewer jumps on a live call with the developer, pulls up the PR with transcription on, and asks questions. At the end of the call, the transcript passes back into the coding agent and the PR is polished up, becoming more self-documenting, and the humans are left with some degree of common understanding of what’s going on.
I’ve been operating my team of ~15 this way for 9mo to great effect… there is simply no going back to the stone ages.
I have not experienced this "llms dont write maintainable code narrative" I just tell it the shape of the entities and apis I want vaguely and the mental model and the output is excellent.
I could do this in the beginning of 2025. Unfortunately, given the over-proactiveness of model these days, I find that they end up inferring what my original request was and implementing it anyway, and then writing unit tests, etc. even when all I asked for was to wire up a few components, as an example.
> LLMs still aren't good at writing maintainable code.
I find that depends on the target language. They can be good at writing maintainable code, but not consistently across every language.
The languages beginners usually gravitate towards are especially hard for LLMs to produce quality output for. Presumably this is due to the training data including all the unmaintainable codebases written by beginners in those languages, which hasn't allowed the LLM to converge on recognizing what a maintainable codebase looks like in those languages.
There is the flip side of where the code is no longer read by humans and is becoming the prevailing way software is shipped in tiny businesses. You don't need to code to be maintainable since you will never maintain it, the AI will and the quality will naturally improve as models improve. For example 5.6 sol and fable are showing signs where you can feed garbage in and it will spit out something pretty decent, definitely not the quality you'd expect from a senior developer with millenia of experience, but that of your average grunt worker turning words in an issue board into code.
However, sometimes then I tell it to write an app with detailed instructions and it spits out garbage so your mileage might vary.
Code was very rarely the bottleneck in the first place.
If programmer productivity was something we actively optimized for, we wouldn't have crammed programmers like sardines in warm and noisy open floor offices with 2000 ppm CO2 levels and then further constantly interrupt them with emails and slack pings and meetings all day long, Jira rigmarole wouldn't make up a significant portion of what they did, programmers would have instead mostly been thinking and programming.
We've always had the ability to 2X if not 10X the output of each and every one of those poor souls. You don't end with this sort of programming purgatory because it's a productivity optimum, it very clearly isn't, but because it's a billable hours optimum and/or an org chart clout optimum and/or because of Jevons paradox got hands even in business management and the IT department was allocated too many dollars.
> Code was very rarely the bottleneck in the first place.
I disagree. Kinda
What AI has made much simpler is that you don't have to waste time checking docs and have the best autocomplete system by a long shot - this was a bottleneck unless you were doing Java or some other language with "perfect" AC
What AI made "kinda easier": solving for usual problems. The stuff you would search Stack Overflow, or think a couple of minutes for an optimized solution - not a bottleneck but not 100% smooth neither
You still have to test and validate your code. AI made this easier-ish but this is still where I see manual work being needed (even if you are automating tests - you still have to think on what you want the code to do)
This just says that we can output code faster. I'm saying that the rate at which we output code wasn't the thing that was slowing down development. We've always been able to increase that even without AI by adjusting the working environment and removing obstacles to programming.
In larger organizations, quite often it's the business that is holding back development. They can only handle so much change and speed needs direction to be velocity. Drafting requirements is generally much slower than implementing them.
Like the number one complaint from programmers has been that they don't get to do programming. They want to write code, not update jiras or spend hours in meetings.
I disagree. Coordinating 100 folks to align on a problems statement with "agile" and sprints was always the biggest bottleneck. Difference in opinion, internal politics is always a bottleneck.
I think now, code is the bottleneck. Just because you can generate million lines of code, people with different skill level think they are accomplishing the task, testing, merge conflicts, trust has become the bottleneck.
The “old way” would be lots of debate (both bike shedding and useful) among engineers during design phase, and then you’d implement.
Now it’s shifted so there are no design docs and there is only the generated prototype. People trying to do their design review while there’s already a functional-ish prototype and it goes nowhere. There’s an anchoring effect in place because the first thing already exists and management says “this seems to work, just use it and move on”. The result is that useful debates about substantive issues don’t happen and bikeshedding is all way get to do
The distance between substantive issues vs bikeshed sized issues is non-linear and ill-defined. Sometimes: yeah, we're just bikeshedding. Other times, it's important to suss out unshared implied context that doesn't match between various stakeholders and will have outsized ramifications later. What's new is the speed of executing changes. If there was no substantive debate about what programming language to even use so everyone is equally happy (read: sad), so what? Have the LLM rewrite the entire project in another language over the span of a couple of days after that discussion is had because the language used has specific shortcomings that have been unmined.
What is the right, best software organization in the current era of AI coding? This question is critical and wholly unanswered in comprehensive research along the same axis as Accelerate (2018, Forsgren, Humble, Kim).
There are a lot of (excruciatingly) long-form posts about what folks are pioneering but not a whole lot of follow up about what failed. Where are the short posts on the negative space? How did halving your staff work out? Flattening your org? All those dark factories, what haven't they produced? How about all the other things tried, failed, and unceremoniously scrapped?
We need to explore and communicate the negative space more efficiently. Don't repeat the same mistakes, and don't make me read 2653 words when 300 do it better.
A tech comm course I took in college was graded on two 20-page papers and accompanying 5-minute presentations. That was like 10k written words in a single semester. It was a challenging class, and gave me substantial sense of accomplishment, just to hand in completed work.
Similar to a functioning side project in the 5-10k LOC range. Announcing something that worked a year ago, was laudable, even if not profitable.
I vibe coded 15k LOC this morning and read 20k words of AI generated text while doing so. No longer are either noteworthy or valuable public contributions just by virtue of having been done. I don't think that's widely recognized yet.
> and don't make me read 2653 words when 300 do it better.
A million times this. Can we please RL the next models to learn the “if I had more time I would’ve written a shorter letter” method please.
I see it every day in tickets, many communication channels, PR descriptions, comments, documentation. All have at least 70% verbose fluff which is so taxing and makes it very hard to keep track of the one important thing they’re trying to communicate in the message
We still pay our software developers the same amount, but now an AI is the middleman taking more and more money (in the form of tokens) every day. With every new model, more tokens are consumed so the new model can "think" more. Nobody's even really reading or reviewing the code the AI produces, so software quality goes down, and we're paying more to produce it. And when there's a serious problem with the code? No human is going to dig through the mess the AI created - so burn even more tokens/money trying to get the AI to fix it, or just start over from scratch again with the AI, hoping for better results. It's the definition of insanity. I'm looking for a new job, maybe even a new industry, a new career path - after 30 years in this industry, software development is jumping the shark.
"Shield the team from the business." The assumption underneath this one is that attention is finite and context switching is expensive. That assumption is intact. What changed is the cost of starving the team of context. Engineers prompting AI tools without business context just produce fluent, plausible, wrong work, at scale."
Too many teams and organizations have business types, mostly PM's who seek to lord over their area of know how and see themselves as delegators and mini CEOs, actively avoid looping engineers in to validate themselves. Engineers need to take on PM roles, and the PM role needs to be 1:50+ eng or go.
Coding used to be an expensive task. Seeing how quickly repositories have grown with the rise of AI coding, it's clear how much people wanted to build things but were thirsting for the means to do so. Even languages like R, which were mostly used by graduate students and experts, have seen a massive increase in usage since vibe coding became popular.
Honestly, when people say AI code quality is bad, Linus himself has said it's now genuinely useful. AI is useful and writes better code than most people. Even in competitive coding, tourist lost to AI. And in the most logical field of all, mathematics, AI is churning out an enormous number of theorems.
Looking at all this, it's fair to say AI is at least at a PhD level of technical ability, and most people would admit they don't have PhD level skills. Of course, there are still many people who code better than AI. But at least when it comes to unfolding logical structures, AI has a higher chance of being more logical than humans. Within a given framework, AI constructs much more logical structures.
That's why I think the article's use of the word 'semantic' is right. It's humans who form the framework, and that's the semantic, while AI fills the empty spaces inside it. If you feed it a flawed framework, it fails.
And the fact that AI is more logical than humans is paradoxically a greater risk. Human developers can rely on tacit knowledge to make reasonable compromises even when the requirements, the framework, are sloppy. AI can't do that. If there's a logical gap in the framework humans design, AI will exploit that weakness and expand the state space into regions we can't cognitively grasp.
Programming is ultimately about how you occupy state space. The problem is that as the program grows, the cognitively inaccessible territory keeps expanding. So we distribute trust across reliable points, libraries, frameworks, and for my own code, once it exceeds tens of thousands of lines, I rely on tests and gates.
Honestly, the idea of understanding everything in a program is a purely academic claim. Once the program gets large, it's impossible. No one can know every external factor, test bug, or unexpected interaction.
The issue is that with LLMs, when the prompt input goes deeper into the semantic space, it also reaches into areas I don't understand, producing code at a depth that's untestable.
For example, I might be an expert in domain A but a beginner in domain B. If I inject expert level knowledge for domain A into the AI, the AI will try to match that level in domain B as well. That results in code I can't understand or modify, and eventually, I'm left with no choice but to replace all the code with AI generated code.
So I'm wondering what to do about this. Should I focus on gaining empirical experience in handling black boxes? Or should I stick with smaller, human written codebases?
But realistically, the current situation, where I can build bigger and touch more things, is more enjoyable to me. I think what I actually enjoyed wasn't programming itself, but the act of creating something.
In my experience, management is mostly pissing away the gains made by AI by either:
1. Pursuing polish and quality beyond previous norms
2. Replacing $100/mo/seat SAAS with something coded by a junior costing $200/day to develop over months.
The cost of code approaches zero, but the cost of having accountability, and hosting remains the same, and so individuals need to only coordinate to the extent that those things remain finite resources. Management needs to stop insisting that their directs adopt each others vibe coded tooling.
I don't think it changes much for good managers. It should always be able setting people and processes up so the team can land durable measurable impact. The managers that thought the job of software engineers was to write code were bad managers. PRs or LoC were never good metrics.
AI reduces the cost of writing code, but it makes adding things that nobody uses even cheaper. The bottleneck then becomes deciding what deserves to exist—and having the discipline to remove the rest. So this will imply more time spent in code reviews that lead to more iterations in PR's.
AI is doing a good job on writting code these days! Nothing against it; I use it every day, but the context switching is costing us a lot!
You don't build credibility by putting out slop and correcting it every time someone points out that it's wrong. You think it's our job to proofread and fact-check?
A genuine mistake. I make these posts for myself, mainly to structure & share my thoughts. Let me know if you find other errors, I appreciate the feedback.
If you wrote an initial draft you should consider using a better model or just posting what you wrote.
The smaller models can be sufficient for coding but for document writing not highly specific I've yet to be satisfied with AI output. I certainly wouldn't expect gemma to produce good outputs.
Makes sense, perhaps I'll post the raw notes + more material with the next article, prior to any edit. I like to use local models, even if they are not that powerful precisely because of that, I'm forced to put in more thought and it's my signed off article in the end.
Generative language models have helped me most by drawing my attention to the importance of context, communicative compression, prompting, comprehension, coherence, and coordination in the domain of human groups.
I’m struggling to imagine a project that would require more than one talented engineer and a bucket of tokens anymore.
I think perhaps the assumption that engineering managers should have any employees may be outdated.
I can imagine average and mediocre engineers equipped with tokens could create chaos and debt on a scale never before imaginable, so it’s easy to see how orgs who still have these employees around are struggling with the transition.
The reality is you need to get rid of them all, and replace them with the most experienced highest paid person you can find. In the near future that person will become obsolete too.
I'm really open to feedback. I checked your past comments and you posted:
> AI is good at coding if there's an oracle. If the system is ancient, unreadable, untestable, that's exactly the opposite. It won't get the exact set of corner cases.
I sort of mention this in the article, so I'm sure we're somewhat aligned on the core. How would you have worded things?
Pangram's marketing always reminds me of Anchorman's Sex Panther cologne: "They've done studies, you know. Sixty percent of the time, it works every time."
I did, the problem is that you need a sort of standardized feature change to compare similar repos. So the metric is relative only to same projects lacking that same feature. So no, its a useless thing. No architecture score comparisson between apples and oranges.
Software development has always evolved. Sometimes slowly, sometimes quicker.
LLMs have brought a different unlock, and for everything we're seeing become easier, it allows people learn to use the tools to take on solving problems that couldn't be approached before.
Your short post speaks of exactly what people should be talking about. Due to the panic related to job replacement, not many are talking about the new frontiers. Vibe coding is boring because it does the same faster/cheaper. I'm more interested in the things we couldn't do before but we now can because there's a crazy savant a few keystrokes away.
As I understand it, the purpose of management is to match financial resources with material+human resources to perform feasible tasks. There is nothing here I see that can't be done by an experienced token generator. If anything, automating management seems easier than automating engineering.
As for leadership, it can be done by the investors.
Code was always expensive. Entire industries of design tools cropped up simply to avoid writing the wrong code since it is so expensive. There are entire journals dedicated to this. Organizations are fixated on making sure that the right code is written since the cost of getting that wrong is so great.
I think most of this is correct, in spite of potentially being built on a bad assumption.
The assumption is that LLMs should be writing the code and human engineers reviewing and verifying the LLM output. And that this pushes the cost of producing down. And I fundamentally disagree with that.
Every time I ask LLMs to write code, even with Opus 4.8 (haven't tried it with Opus 5 yet), what I get ends up being totally rewritten. LLMs still aren't good at writing maintainable code. Can they write plausibly functional code? Yes. But it won't survive the long term. People using LLMs to write all their code are gambling on them eventually getting to a point where the LLMs can fix their own code. It's possible, but I wouldn't necessarily bet on it.
Where I have found immense value from LLMs is in code review. Repeated review by LLMs catches an amazing amount of potential issues. They really shine on security review, but are very effective with any kind of review.
The other thing that the "LLMs write code camp" misunderstands is that writing was never the bottleneck. Understanding was. And understanding the code is still the bottleneck. But understanding is truly gained during the writing loop. The understanding you gain from pure reading or code review is marginal compared to the understanding you gain while writing.
Most of the time previously spent writing was actually spent updating and deepening our understanding of the system under development. There's no replacement for that understanding in a world where LLMs are doing the writing.
But if you flip it: humans write, LLMs review, then you still get a major gain -- not in speed, but in quality. And you keep the understanding loop intact. I would propose that this might be the best way to deploy LLMs.
" But understanding is truly gained during the writing loop. The understanding you gain from pure reading or code review is marginal compared to the understanding you gain while writing."
Debugging code step by step is how I understand complicated code.
I write code in Go in an established codebase. 90% of the time the LLM gets a basic api/feature right and fully e2e tested as long as I give it enough business context in the prompt. The other 10% of the time I have to do some follow up prompts to either change the behaviour or change the approach of a given step in the flow or just to point out that the wrong pattern was used and please rather use our codebase standard. I haven’t actually hand-written code in well over a year. Is this way faster? Absolutely. Does it lead to better code? Yes, because I have time to write all the regression tests that keep the behaviour as I wanted it when others come bungling around.
I do think the comment is right about understanding. I'm still feeling out what I think is the right level of understanding to invest in now. I think it is probably not the deep line by line understanding I used to have of the systems I worked on. But I also think it's easy to remain too aloof from how the system is being built, and that this is very bad. I think the right answer is somewhere in between, but I'm still working through my own process and forcing functions to strike that balance properly.
many of us work with swarms of cheap off-shore contractors and have to rewrite or finish their code already as it is. The llm is much lower friction
IME, the LLM when used well also gets closer to correct.
You're right about understanding. Where we might disagree is the conclusion. I think it's possible to build the understanding without having to type out syntax by hand (debugging, writing tests, etc). Maybe we're missing much better verification tools? LLMs will likely play a big part in those too (explain the codebase, walk me through A, B, C etc)
How do you handle the fact that an LLM can't seem to help itself from disgorging page after page of words no matter what it's asked to do? I've never seen an LLM say "this looks good as-is; I would not spend any more time on it; what's next?" it will always seem to suggest using another pattern or additional abstractions or other yak shaving.
But to be fair human code reviews have the same problem. It's like reviewers feel they have not done their job if they don't find something wrong.
Yeah, that's just a problem with review. I tend to keep it going until the stuff its finding is stuff that I look at and think "I can live with that". Usually by that point it's found the worthwhile stuff and what remains is minor or, like you said, yak shaving.
There used to be a bit of "wisdom" passed around that said "when you submit something to the client (or management, etc) for approval, leave an obvious mistake somewhere." Then they will find it and point it out, you can easily correct it, and everyone is happy. Otherwise they will find something to critique, just to demonstrate that they contributed something to the review.
I wonder if the same trick works with AI?
I think so, if you can inject synthetic errors with the same distribution as the errors you care about.
I've tried and tried to get an LLM to delete code, it couldn't do it. I knew a file was 40% bad so what I ended up doing was deleting the file, then asking AI to create the missing file. That's how I got AI to delete code :)
I’ve found stating “only call out critical issues, report but ignore everything else” will basically get what you want. Put that in a review skill and you’re good to go.
To be clear I mostly only use Opus and Gemini Flash but this might work for others too.
> I’ve found stating “only call out critical issues, report but ignore everything else” will basically get what you want.
Disabling your C compiler warnings works too. You get to ship then leave work early!
My guess would be that this is a matter of fine-tuning the model to the problem. You would need to run the review experiment N times, observe the distribution of answers, and then fine tune with model operational parameters.
Perhaps someone more knowledgable could jump in here to clarify?
The verbosity itself seems to be a problem with post-4.6 Claude rather than a general issue with all LLMs. and IME the yak shaving is due to an overly generic prompt. We have a generic automated review bot but it’s prompted to only look for errors and never suggests refactoring; that was a conscious trade off to avoid what you’re talking about. If you’re reviewing interactively then you can give additional guidance about any code smells etc. that jump out
I personally rely on this behavior of LLMs. For example, for reviewing prose, I might ask the LLM to find five issues. If there are obvious issues I missed, then it'll find them. But if all the issues it lists are minor or hallucinated, then I can have some confidence there weren't any glaring issues.
I do the same for LLM code review comments: some changes are out of scope or could be moved to a separate PR; some edge cases don't happen in practice and should just fail noisily instead of writing more code to maintain. When these are the only issues it's raising, then I know it's done.
It is surprisingly hard to do in a single prompt. I’ve had good luck though with being very explicit about asking for review, then to think deeply about what actually matters, and then reduce to a very short extremely concise reply. —- the model NEEDS to output those pages of text as part of its thinking process. It’s used to doing it in the assistant reply but can be cajoled into doing it in the thinking tokens.
I've come to realize that the smarter LLMs get, the less they understand the point of abstractions.
I've been using it for a unity game for the past few years. Nowadays it will go sleuthing into packages and assembly and make decisions based upon what it sees there.
It will make comments about why it's doing something based upon a function call 3 methods deep.
God forbid any of these details change in a minor version update.
Try reducing thinking and use lighter models. For example I observed that using Sonnet works much better (compared to Opus) for tasks I want to be in control of architecture and just need a faster code input.
> The other thing that the "LLMs write code camp" misunderstands is that writing was never the bottleneck. Understanding was. And understanding the code is still the bottleneck. But understanding is truly gained during the writing loop. The understanding you gain from pure reading or code review is marginal compared to the understanding you gain while writing.
This was my stance a couple of years ago, but now I've given it up.
It turns out writing actually was the bottleneck. You can understand perfectly well what you want, but writing it is long and tedious to the point where you find excuses not to do it. Particularly with version 2, the step where you have an OK system and you want to improve it. Quite a lot of changing the code is just useless busywork: re-wiring old functions, moving imports around, searching for locations that benefit from extracting a common piece of code. And each time you do one of those, there's a decent chance you did something even more trivial like forgetting a semicolon or calling the wrong function.
Now that I have an LLM helping me, I can see why. The critical decision is a terse declarative like "we need to have several TCP connections instead of one, and just use the sequence number to arbitrate". A human junior programmer could perfectly well understand what this meant, but he would have to go through all of the above to get to the final product. Now, I can just tell the LLM and I will get what I want, even with the things I didn't explicitly state, without spending attention.
This means I can use my attention on the things that matter. So instead of spending today thinking about how to arbitrate between the TCP connections and tomorrow thinking about pre-calculating my outgoing orders, I can just do both today. I don't waste the good waking hours chasing minor bugs, I just think about the large structure.
I get the feeling the best programmers of years past were actually masters of the little things, which led them to be able to look at the big things. Essentially it was cheaper for them to get to the top of the mountain, where you can see the landscape. Kinda like how the kid who was good at mental arithmetic in primary school was also good at calculus at the end of high school: if you don't have to concentrate on the little things, you have time for the big things.
>It turns out writing actually was the bottleneck. You can understand perfectly well what you want, but writing it is long and tedious to the point where you find excuses not to do it.
Speak for yourself. Writing code has never been a bottleneck for some of us. I can't speak for everyone, and neither should you.
>Now, I can just tell the LLM and I will get what I want, even with the things I didn't explicitly state, without spending attention.
This should worry you. All too often the LLM invents things I didn't ask for and implements things I didn't need. YMMV, I guess. If slop gets the job done, and nobody notices, then who should care?
> And each time you do one of those, there's a decent chance you did something even more trivial like forgetting a semicolon or calling the wrong function.
how did you decide to pick the most trivial kind regression for this example? do you compile your code before checking it in?
> A human junior programmer could perfectly well understand what this meant, but he would have to go through all of the above to get to the final product. Now, I can just tell the LLM and I will get what I want, even with the things I didn't explicitly state, without spending attention.
the main efficiency you have described here is offloading the verification of a change onto the LLM. that is the bottleneck. readers can decide whether a non-deterministic statistical model is a good tool for this job
> I get the feeling the best programmers of years past were actually masters of the little things, which led them to be able to look at the big things
the best programmers understand that their job is to automate workflows, and that includes their own. if you're worried about missing a semicolon, I'm sorry to say that's a skill issue
I’ve found that doing the full requirements capture, planning, writing, reviewing, gardening loop with frontier models has worked quite well since last October, and phenomenally since Fable 5.
The key I’ve found is human peer review. The reviewer jumps on a live call with the developer, pulls up the PR with transcription on, and asks questions. At the end of the call, the transcript passes back into the coding agent and the PR is polished up, becoming more self-documenting, and the humans are left with some degree of common understanding of what’s going on.
I’ve been operating my team of ~15 this way for 9mo to great effect… there is simply no going back to the stone ages.
Here's an odd request for you, as a fellow New Jersey-an...
Can I watch/observe one of your review sessions?
Every few weeks, I hear the beginnings of a great approach towards working with LLMs but I rarely see it in practice.
If you're open to this, remote or in person, ping my username at gmail.
I have not experienced this "llms dont write maintainable code narrative" I just tell it the shape of the entities and apis I want vaguely and the mental model and the output is excellent.
I could do this in the beginning of 2025. Unfortunately, given the over-proactiveness of model these days, I find that they end up inferring what my original request was and implementing it anyway, and then writing unit tests, etc. even when all I asked for was to wire up a few components, as an example.
> LLMs still aren't good at writing maintainable code.
I find that depends on the target language. They can be good at writing maintainable code, but not consistently across every language.
The languages beginners usually gravitate towards are especially hard for LLMs to produce quality output for. Presumably this is due to the training data including all the unmaintainable codebases written by beginners in those languages, which hasn't allowed the LLM to converge on recognizing what a maintainable codebase looks like in those languages.
There is the flip side of where the code is no longer read by humans and is becoming the prevailing way software is shipped in tiny businesses. You don't need to code to be maintainable since you will never maintain it, the AI will and the quality will naturally improve as models improve. For example 5.6 sol and fable are showing signs where you can feed garbage in and it will spit out something pretty decent, definitely not the quality you'd expect from a senior developer with millenia of experience, but that of your average grunt worker turning words in an issue board into code.
However, sometimes then I tell it to write an app with detailed instructions and it spits out garbage so your mileage might vary.
Code was very rarely the bottleneck in the first place.
If programmer productivity was something we actively optimized for, we wouldn't have crammed programmers like sardines in warm and noisy open floor offices with 2000 ppm CO2 levels and then further constantly interrupt them with emails and slack pings and meetings all day long, Jira rigmarole wouldn't make up a significant portion of what they did, programmers would have instead mostly been thinking and programming.
We've always had the ability to 2X if not 10X the output of each and every one of those poor souls. You don't end with this sort of programming purgatory because it's a productivity optimum, it very clearly isn't, but because it's a billable hours optimum and/or an org chart clout optimum and/or because of Jevons paradox got hands even in business management and the IT department was allocated too many dollars.
> Code was very rarely the bottleneck in the first place.
I disagree. Kinda
What AI has made much simpler is that you don't have to waste time checking docs and have the best autocomplete system by a long shot - this was a bottleneck unless you were doing Java or some other language with "perfect" AC
What AI made "kinda easier": solving for usual problems. The stuff you would search Stack Overflow, or think a couple of minutes for an optimized solution - not a bottleneck but not 100% smooth neither
You still have to test and validate your code. AI made this easier-ish but this is still where I see manual work being needed (even if you are automating tests - you still have to think on what you want the code to do)
This just says that we can output code faster. I'm saying that the rate at which we output code wasn't the thing that was slowing down development. We've always been able to increase that even without AI by adjusting the working environment and removing obstacles to programming.
In larger organizations, quite often it's the business that is holding back development. They can only handle so much change and speed needs direction to be velocity. Drafting requirements is generally much slower than implementing them.
Like the number one complaint from programmers has been that they don't get to do programming. They want to write code, not update jiras or spend hours in meetings.
I disagree. Coordinating 100 folks to align on a problems statement with "agile" and sprints was always the biggest bottleneck. Difference in opinion, internal politics is always a bottleneck.
I think now, code is the bottleneck. Just because you can generate million lines of code, people with different skill level think they are accomplishing the task, testing, merge conflicts, trust has become the bottleneck.
A pattern I have seen:
The “old way” would be lots of debate (both bike shedding and useful) among engineers during design phase, and then you’d implement.
Now it’s shifted so there are no design docs and there is only the generated prototype. People trying to do their design review while there’s already a functional-ish prototype and it goes nowhere. There’s an anchoring effect in place because the first thing already exists and management says “this seems to work, just use it and move on”. The result is that useful debates about substantive issues don’t happen and bikeshedding is all way get to do
The distance between substantive issues vs bikeshed sized issues is non-linear and ill-defined. Sometimes: yeah, we're just bikeshedding. Other times, it's important to suss out unshared implied context that doesn't match between various stakeholders and will have outsized ramifications later. What's new is the speed of executing changes. If there was no substantive debate about what programming language to even use so everyone is equally happy (read: sad), so what? Have the LLM rewrite the entire project in another language over the span of a couple of days after that discussion is had because the language used has specific shortcomings that have been unmined.
What is the right, best software organization in the current era of AI coding? This question is critical and wholly unanswered in comprehensive research along the same axis as Accelerate (2018, Forsgren, Humble, Kim).
There are a lot of (excruciatingly) long-form posts about what folks are pioneering but not a whole lot of follow up about what failed. Where are the short posts on the negative space? How did halving your staff work out? Flattening your org? All those dark factories, what haven't they produced? How about all the other things tried, failed, and unceremoniously scrapped?
We need to explore and communicate the negative space more efficiently. Don't repeat the same mistakes, and don't make me read 2653 words when 300 do it better.
Hard to say what caused what, but the internet seems to mistake verbosity for authority, and so does AI.
A tech comm course I took in college was graded on two 20-page papers and accompanying 5-minute presentations. That was like 10k written words in a single semester. It was a challenging class, and gave me substantial sense of accomplishment, just to hand in completed work.
Similar to a functioning side project in the 5-10k LOC range. Announcing something that worked a year ago, was laudable, even if not profitable.
I vibe coded 15k LOC this morning and read 20k words of AI generated text while doing so. No longer are either noteworthy or valuable public contributions just by virtue of having been done. I don't think that's widely recognized yet.
Great framing.
> and don't make me read 2653 words when 300 do it better.
A million times this. Can we please RL the next models to learn the “if I had more time I would’ve written a shorter letter” method please.
I see it every day in tickets, many communication channels, PR descriptions, comments, documentation. All have at least 70% verbose fluff which is so taxing and makes it very hard to keep track of the one important thing they’re trying to communicate in the message
The cost of code actually increased; code debt is being accumulated faster than we can clean it up.
Yeah but look how much there is! Aren’t you impressed?
We still pay our software developers the same amount, but now an AI is the middleman taking more and more money (in the form of tokens) every day. With every new model, more tokens are consumed so the new model can "think" more. Nobody's even really reading or reviewing the code the AI produces, so software quality goes down, and we're paying more to produce it. And when there's a serious problem with the code? No human is going to dig through the mess the AI created - so burn even more tokens/money trying to get the AI to fix it, or just start over from scratch again with the AI, hoping for better results. It's the definition of insanity. I'm looking for a new job, maybe even a new industry, a new career path - after 30 years in this industry, software development is jumping the shark.
"Shield the team from the business." The assumption underneath this one is that attention is finite and context switching is expensive. That assumption is intact. What changed is the cost of starving the team of context. Engineers prompting AI tools without business context just produce fluent, plausible, wrong work, at scale."
Too many teams and organizations have business types, mostly PM's who seek to lord over their area of know how and see themselves as delegators and mini CEOs, actively avoid looping engineers in to validate themselves. Engineers need to take on PM roles, and the PM role needs to be 1:50+ eng or go.
Coding used to be an expensive task. Seeing how quickly repositories have grown with the rise of AI coding, it's clear how much people wanted to build things but were thirsting for the means to do so. Even languages like R, which were mostly used by graduate students and experts, have seen a massive increase in usage since vibe coding became popular.
Honestly, when people say AI code quality is bad, Linus himself has said it's now genuinely useful. AI is useful and writes better code than most people. Even in competitive coding, tourist lost to AI. And in the most logical field of all, mathematics, AI is churning out an enormous number of theorems.
Looking at all this, it's fair to say AI is at least at a PhD level of technical ability, and most people would admit they don't have PhD level skills. Of course, there are still many people who code better than AI. But at least when it comes to unfolding logical structures, AI has a higher chance of being more logical than humans. Within a given framework, AI constructs much more logical structures.
That's why I think the article's use of the word 'semantic' is right. It's humans who form the framework, and that's the semantic, while AI fills the empty spaces inside it. If you feed it a flawed framework, it fails.
And the fact that AI is more logical than humans is paradoxically a greater risk. Human developers can rely on tacit knowledge to make reasonable compromises even when the requirements, the framework, are sloppy. AI can't do that. If there's a logical gap in the framework humans design, AI will exploit that weakness and expand the state space into regions we can't cognitively grasp.
Programming is ultimately about how you occupy state space. The problem is that as the program grows, the cognitively inaccessible territory keeps expanding. So we distribute trust across reliable points, libraries, frameworks, and for my own code, once it exceeds tens of thousands of lines, I rely on tests and gates.
Honestly, the idea of understanding everything in a program is a purely academic claim. Once the program gets large, it's impossible. No one can know every external factor, test bug, or unexpected interaction.
The issue is that with LLMs, when the prompt input goes deeper into the semantic space, it also reaches into areas I don't understand, producing code at a depth that's untestable.
For example, I might be an expert in domain A but a beginner in domain B. If I inject expert level knowledge for domain A into the AI, the AI will try to match that level in domain B as well. That results in code I can't understand or modify, and eventually, I'm left with no choice but to replace all the code with AI generated code.
So I'm wondering what to do about this. Should I focus on gaining empirical experience in handling black boxes? Or should I stick with smaller, human written codebases?
But realistically, the current situation, where I can build bigger and touch more things, is more enjoyable to me. I think what I actually enjoyed wasn't programming itself, but the act of creating something.
Author works at parity.io, which sells AI website builders:
https://www.parity.io/blog/playground-dot-what-190-people-bu...
Every single time. Remember 2022/2023 when we discussed actual technology?
In my experience, management is mostly pissing away the gains made by AI by either:
1. Pursuing polish and quality beyond previous norms
2. Replacing $100/mo/seat SAAS with something coded by a junior costing $200/day to develop over months.
The cost of code approaches zero, but the cost of having accountability, and hosting remains the same, and so individuals need to only coordinate to the extent that those things remain finite resources. Management needs to stop insisting that their directs adopt each others vibe coded tooling.
I think #1 is quite a good thing, on net.
I don't think it changes much for good managers. It should always be able setting people and processes up so the team can land durable measurable impact. The managers that thought the job of software engineers was to write code were bad managers. PRs or LoC were never good metrics.
Agreed, which is why I also believe "token usage" and similar proxy metrics aren't the right ones for what's next.
AI reduces the cost of writing code, but it makes adding things that nobody uses even cheaper. The bottleneck then becomes deciding what deserves to exist—and having the discipline to remove the rest. So this will imply more time spent in code reviews that lead to more iterations in PR's.
AI is doing a good job on writting code these days! Nothing against it; I use it every day, but the context switching is costing us a lot!
Sorry, "Gemini 4 helped with editing"?
dang it i tried to check for a dupe but there were too many comments!
> Gemini 4 helped with the editing.
Does this guy have access to Gemini 4 already?
I'm guessing Gemma 4 was happy to be mistaken for Gemini and didn't catch this mistake.
“Helping” is doing some heavy lifting in that sentence! It appears to be 100% AI.
Yes correct, Gemma. Will correct it shortly.
You don't build credibility by putting out slop and correcting it every time someone points out that it's wrong. You think it's our job to proofread and fact-check?
A genuine mistake. I make these posts for myself, mainly to structure & share my thoughts. Let me know if you find other errors, I appreciate the feedback.
If you wrote an initial draft you should consider using a better model or just posting what you wrote.
The smaller models can be sufficient for coding but for document writing not highly specific I've yet to be satisfied with AI output. I certainly wouldn't expect gemma to produce good outputs.
Makes sense, perhaps I'll post the raw notes + more material with the next article, prior to any edit. I like to use local models, even if they are not that powerful precisely because of that, I'm forced to put in more thought and it's my signed off article in the end.
Thank you for the suggestion.
My take, as a non-coder (well, not software engineering, I write 'code' but it's infra, and utilities in go/bash/pythong)...
I work at a company where the biggest problems are not 'writing code', they are:
- Organising teams
- Designing the system
- Prioritisation of work
The fuckups that we make on a daily bases are not 'code errors' they are failures in THOSE three things. I'll go into detail if anyone cares.
don't forget 'actually making decisions'
Same as it ever was.
Generative language models have helped me most by drawing my attention to the importance of context, communicative compression, prompting, comprehension, coherence, and coordination in the domain of human groups.
I’m struggling to imagine a project that would require more than one talented engineer and a bucket of tokens anymore.
I think perhaps the assumption that engineering managers should have any employees may be outdated.
I can imagine average and mediocre engineers equipped with tokens could create chaos and debt on a scale never before imaginable, so it’s easy to see how orgs who still have these employees around are struggling with the transition.
The reality is you need to get rid of them all, and replace them with the most experienced highest paid person you can find. In the near future that person will become obsolete too.
> I’m struggling to imagine a project that would require more than one talented engineer and a bucket of tokens anymore
Try air traffic control?
Gemini 4 eh????
Pangram reports this post was 100% AI generated.
It's depressingly hard to find one that isn't. You see the title, think this might be interesting, and puke by the second paragraph.
I'm really open to feedback. I checked your past comments and you posted:
> AI is good at coding if there's an oracle. If the system is ancient, unreadable, untestable, that's exactly the opposite. It won't get the exact set of corner cases.
I sort of mention this in the article, so I'm sure we're somewhat aligned on the core. How would you have worded things?
Pangram's marketing always reminds me of Anchorman's Sex Panther cologne: "They've done studies, you know. Sixty percent of the time, it works every time."
Pangram's "100% AI generated" claims are right 65% of the time. https://link.springer.com/article/10.1007/s40979-026-00226-w
Can a llm predict the price of a change to a codebase in tokens and predict the origin of the price, aka cam it see good and bad architecture?
Ask it to do so, would love to know your results.
I did, the problem is that you need a sort of standardized feature change to compare similar repos. So the metric is relative only to same projects lacking that same feature. So no, its a useless thing. No architecture score comparisson between apples and oranges.
https://chatgpt.com/s/t_6a64f4a6e86c8191b41cb6a810359208
Some code bases better than others but the top models can.
Sure and having some in repo documentation markdown makes it a lot easier
The worse problem is blog posts after the cost of writing collapsed.
Not everything has to be written as though it’s a middle manager’s idea of what makes for a good TED talk.
I read this entire article and didn’t sniff AI, plus it had some good insights…?
Feels like a human-authored outline that was run through AI to make it into an article.
Well it contains a typo, so it made a cock up.
More than that, the AI disclaimer itself is a hallucination:
> Gemini 4 helped with the editing.
There is no Gemini 4, unless the author is writing from the future.
You are correct, my bad. I meant to write Gemma 4. I'll update.
I've written this myself, Gemini 4 did a bit of editing:
> What follows is a cleaned-up version of notes I accumulated over the past year. Gemini 4 helped with the editing.
The image is made by an AI image generator on fal.ai. It's better I spare you all my design skills :)
AI image, though.
I closed the tab after seeing the AI hero image. Good to know I didn't miss anything.
Reading the domain-name had the same effect for me.
You discount people's blogs based on them having a non-western name? Maybe I've misunderstood your comment, but the domain is just their name.
I can't change my name, but happy to change the hero picture :))
You did miss something. But the picture doesn’t help…why would you start an article with an image that telegraphs “low effort”?
Yes the article took a while to get going, but once it did it was thoughtful and well reasoned.
Thank you very much for the kind words :)
Software development has always evolved. Sometimes slowly, sometimes quicker.
LLMs have brought a different unlock, and for everything we're seeing become easier, it allows people learn to use the tools to take on solving problems that couldn't be approached before.
Your short post speaks of exactly what people should be talking about. Due to the panic related to job replacement, not many are talking about the new frontiers. Vibe coding is boring because it does the same faster/cheaper. I'm more interested in the things we couldn't do before but we now can because there's a crazy savant a few keystrokes away.
Engineers do not need management. Investors do.
As I understand it, the purpose of management is to match financial resources with material+human resources to perform feasible tasks. There is nothing here I see that can't be done by an experienced token generator. If anything, automating management seems easier than automating engineering.
As for leadership, it can be done by the investors.
> The cost of producing plausible code has collapsed
Who wants merely plausible code?
Code was never expensive.
Code was always expensive. Entire industries of design tools cropped up simply to avoid writing the wrong code since it is so expensive. There are entire journals dedicated to this. Organizations are fixated on making sure that the right code is written since the cost of getting that wrong is so great.
I assume GP meant exactly this (the people designing and writing good code are expensive).
Building the wrong thing is what is expensive.
Once you know what you are building and can clearly describe it, the code isn't the hard part. Or at least that's how it has always seemed to me.
At least not the writing of it