> Reading through these commits sparked an idea: what if we treated prompts as the actual source code? Imagine version control systems where you commit the prompts used to generate features rather than the resulting implementation.
Please god, no, never do this. For one thing, why would you not commit the generated source code when storage is essentially free? That seems insane for multiple reasons.
> When models inevitably improve, you could connect the latest version and regenerate the entire codebase with enhanced capability.
How would you know if the code was better or worse if it was never committed? How do you audit for security vulnerabilities or debug with no source code?
My work has involved a project that is almost entirely generated code for over a decade. Not AI generated, the actual work of the project is in creating the code generator.
One of the things we learned very quickly was that having generated source code in the same repository as actual source code was not sustainable. The nature of reviewing changes is just too different between them.
Another thing we learned very quickly was that attempting to generate code, then modify the result is not sustainable; nor is aiming for a 100% generated code base. The end result of that was that we had to significantly rearchitect the project for us to essentially inject manually crafted code into arbitrary places in the generated code.
Another thing we learned is that any change in the code generator needs to have a feature flag, because someone was relying on the old behavior.
> One of the things we learned very quickly was that having generated source code in the same repository as actual source code was not sustainable
My rule of the thumb is to have both in same repo, but treat generated code like binary data. This was informed by when I was burned by a tooling regression that broke the generated code and the investigation was complicated by having to correlate commits across different repositories
I think the biggest difference here is that your code generator is probably deterministic and you likely are able to debug the results it produces rather than treating it like a black box.
If you can 100% reproduce the same generated code from the same prompts, even 5 years later, given the same versions and everything then I'd say "Sure, go ahead and don't saved the generated code, we can always regenerate it". As someone who spent some time in frontend development, we've been doing it like that for a long time with (MB+) generated code, keeping it in scm just isn't feasible long-term.
But given this is about LLMs, which people tend to run with temperature>0, this is unlikely to be true, so then I'd really urge anyone to actually store the results (somewhere, maybe not in scm specifically) as otherwise you won't have any idea about what the code was in the future.
> If you can 100% reproduce the same generated code from the same prompts, even 5 years later
Reproducible builds with deterministic stacks and local compilers are far from solved. Throwing in LLM randomness just makes for a spicier environment to not commit the generated code.
I didn't read it as that - If I understood correctly, generated code must be quarantined very tightly. And inevitably you need to edit/override generated code and the manner by which you alter it must go through some kind of process so the alteration is auditable and can again be clearly distinguished from generated code.
Tbh this all sounds very familiar and like classic data management/admin systems for regular businesses. The only difference is that the data is code and the admins are the engineers themselves so the temptation to "just" change things in place is too great. But I suspect it doesn't scale and is hard to manage etc.
> I feel like using a compiler is in a sense a code generator where you don't commit the actual output
Compilers are deterministic. Given the same input you always get the same output so there's no reason to store the output. If you don't get the same output we call it a compiler bug!
LLMs do not work this way.
(Aside: Am I the only one who feels that the entire AI industry is predicated on replacing only development positions? we're looking at, what, 100bn invested, with almost no reduce in customer's operating costs other than if the customer has developers).
I will guess that you are generating orders of magnitude more lines of code with your software than people do when building projects with LLMs - if this is true I don't think the analogy holds.
There’s a huge difference between deterministic generated code and LLM generated code. The latter will be different every time, sometimes significantly so. Subsequent prompts would almost immediately be useless. “You did X, but we want Y” would just blow up if the next time through the LLM (or the new model you’re trying) doesn’t produce X at all.
The idea as stated is a poor one, but a slight reshuffling and it seems promising:
You generate code with LLMs. You write tests for this code, either using LLMs or on your own. You of course commit your actual code: it is required to actually run the program, after all. However you also save the entire prompt chain somewhere. Then (as stated in the article), when a much better model comes along, you re-run that chain, presumably with prompting like "create this project, focusing on efficiency" or "create this project in Rust" or "create this project, focusing on readability of the code". Then you run the tests against the new codebase and if the suite passes you carry on, with a much improved codebase. The theoretical benefit of this over just giving your previously generated code to the LLM and saying "improve the readability" is that the newer (better) LLM is not burdened by the context of the "worse" decisions made by the previous LLM.
Obviously it's not actually that simple, as tests don't catch everything (tho with fuzz testing and complete coverage and such they can catch most issues), but we programmers often treat them as if they do, so it might still be a worthwhile endeavor.
Your rephrasing better encompasses my idea, and I should have emphasized in the post that I do not think this is a good idea (nor possible) right now, it was more of a hand-wavy "how could we rethink source control in a post-LLM world" passing thought I had while reading through all the commits.
Clearly it struck a chord with a lot of the folks here though, and it's awesome to read the discourse.
Means the temperature should be set to 0 (which not every provider supports) so that the output becomes entirely deterministic. Right now with most models if you give the same input prompt twice it will give two different solutions.
Even at temp 0, you might get different answers, depending on your inference engine. There might be hardware differences, as well as software issues (e.g. vLLM documents this, if you're using batching, you might get different answers depending on where in the batch sequence your query landed).
> this assumes models can achieve strict prompt adherence
What does strict adherence to an ambiguous prompt even mean? It’s like those people asking Babbage if his machine would give the right answer when given the wrong figures. I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a proposition.
Prompts are like story on the board, and like engineers, depends on the understanding of the model the generated source code can vary. Saying the prompts could be the actual code is so wrong and dangerous thought
Nitpick - it's the ML system that is sampling from model predictions that has a temperature parameter, not the model itself. Temperature and even model aside, there are other sources of randomness like the underlying hardware that can cause the havoc you describe.
Plus, commits depend on the current state of the system.
What sense does “getting rid of vulnerabilities by phasing out {dependency}” make, if the next generation of the code might not rely on the mentioned library at all? What does “improve performance of {method}” mean if the next generation used a fully different implementation?
It makes no sense whatsoever except for a vibecoders script that’s being extrapolated into a codebase.
I'd say commit a comprehensive testing system with the prompts.
Prompts are in a sense what higher level programming languages were to assembly. Sure there is a crucial difference which is reproducibility. I could try and write down my thoughts why I think in the long run it won't be so problematic. I could be wrong of course.
I run https://pollinations.ai which servers over 4 million monthly active users quite reliably. It is mostly coded with AI. Since about a year there was no significant human commit. You can check the codebase. It's messy but not more messy than my codebases were pre-LLMs.
I think prompts + tests in code will be the medium-term solution. Humans will be spending more time testing different architecture ideas and be involved in reviewing and larger changes that involve significant changes to the tests.
Agreed with the medium-term solution. I wish I put some more detail into that part of the post, I have more thoughts on it but didn't want to stray too far off topic.
Apart from obvious non-reproducibility, the other problem is lack of navigable structure. I can't command+click or "show usages" or "show definition" any more.
I think the author is saying you commit the prompt with the resulting code. You said it yourself, storage is free, so comment the prompt along with the output (don’t comment that out that if I’m not being clear); it would show the developers(?) intent, and to some degree, almost always contribute to the documentation process.
Author here :). Right now, I think the pragmatic thing to do is to include all prompts used in either the PR description and/or in the commit description. This wouldn't make my longshot idea of "regenerating a repo from the ground up" possible, but it still adds very helpful context to code reviewers and can help others on your team learn prompting techniques.
I'm pretty sure most people aren't doing "software engineering" when they program. There's the whole world of WordPress and dream Weaver like programing out there too where the consequences of messing up aren't really important.
Llms can be configured to have deterministic output too
Yes, it's too early to be doing that now, but if you see the move to AI-assisted code as at least the same magnitude of change as the move from assembly to high level languages, the argument makes more sense.
Nobody commits the compiled code; this is the direction we are moving in, high level source code is the new assembly.
These posts are funny to me because prompt engineers point at them as evidence of the fast-approaching software engineer obsolescence but the years of experience in software engineering necessary to even guide an AI in this way is very high.
The reason he keeps adjusting the prompts is because he knows how to program. He knows what it should look like.
> It just blurs the line between engineer and tool.
I realise you meant it as “the engineer and their tool blend together”, but I read it like a funny insult: “that guy likes to think of himself as an engineer, but he’s a complete tool”.
The argument is that this stuff will so radically improve senior engineer productivity that the demand for junior engineers will crater. And without a pipeline of junior engineers, the junior-to-senior trajectory will radically atrophy
Essentially, the field will get frozen where existing senior engineers will be able to utilize AI to outship traditional senior-junior teams, even as junior engineers fail to secure employment
I don’t think anything in this article counters this argument
Right. I don’t understand why everyone thinks this will make it impossible for junior devs to learn. The people I had around to answer my questions when I was learning knew a whole lot less than Claude and also had full time jobs doing something other than answering my questions.
You learn by doing.. eg typing the code. It's not just knowledge, it's the intuition you develop when you write code yourself. Just like physical exercise. Or playing an instrument. It's not enough to know the theory, practice is key.
AI makes it very easy to avoid typing and hence make learning this skill less attractive.
But I don't necessarily see it as doom and gloom, what I think will happen - juniors will develop advanced intuition about using AI and getting the functionality they need, not the quality of the code, while at the same time the AI models will get increasingly better and write higher quality code.
It won't make it impossible for junior engineers to learn.
It will simply reduce the amount of opportunities to learn (and not just for juniors), by virtue of companies' beancounters concluding "two for one" (several juniors) doesn't return the same as "buy one get one free" (existing staff + AI license).
I dread the day we all "learn from AI". The social interaction part of learning is just as important as the content of it, really, especially when you're young; none of that comes across yet in the pure "1:1 interaction" with AI.
I learnt programming on my own, without any social interaction involved. In fact, I loved programming because it does not involve any social interaction.
Programming has become more of a "social game" in the last 15 years or so. AI is a new superpower for people like me, bringing balance to the Force.
This is not necessarily true in practical terms when it comes to hiring or promoting. Often a senior dev becomes a senior because of having an advanced skillset, despite years on the job. Similarily, often developers who have been on the job for many years aren’t ready for senior because of their lack or soft and hard skills.
Maybe you could enlighten the rest of us then. According to your favorite definition, what does senior mean, what does seniority mean, and what's a term for someone who knows what they're doing?
You're letting Claude do your programming for you, and then sweeping up whatever it does afterwards. Bluntly, you're off-loading your cognition to the machine. If that's fine by you then that's fine enough, it just means that the quality of your work becomes a function of your tooling rather than your capabilities.
I think it makes sense that GP is skeptical of this article considering it contains things like:
> this tool is improving itself, learning from every interaction
which seem to indicate a fundamental misunderstanding of how modern LLMs work: the 'improving' happens by humans training/refining existing models offline to create new models, and the 'learning' is just filling the context window with more stuff, not enhancement of the actual model or the model 'learning' - it will forget everything if you drop the context and as the context grows it can 'forget' things it previously 'learned'.
When you consider the "tool" as more than just the LLM model, but the stuff wrapped around calling that model then I feel like you can make a good argument it's improving when it keeps context in a file on disk and constantly updates and edits that file as you work throguh the project.
I do this routinely for large initiatives I'm kicking off through Claude Code - it writes a long detailed plan into a file and as we work through the project I have it constantly updating and rewriting that document to add information we have jointly discovered from each bit of the work. That means every time I come back and fire it back up, it's got more information than when it started, which looks a lot more improvement from my perspective.
So, it means that you and the LLM together have managed to write SEVEN lines of trivial code per hour. On a protocol that is perfectly documented, where you can look at about one million other implementations when in doubt.
It is not my intention to hurt your feelings, but it sounds like you and/or the LLM are not really good at their job. Looking at programmer salaries and LLM energy costs, this appears to be a very very VERY expensive OAuth library.
Again: Not my intention to hurt any feelings, but the numbers really are shockingly bad.
I spent about 5 days semi-focused on this codebase (though I always have lots of people interrupting me all the time). It's about 5000 lines (if you count comments, tests, and documentation, which you should). Where do you get 7 lines per hour?
Yes, my brain got confused on who wrote the code and who just reported about it. I am truly sorry. I will go see my LLM doctor to get my brain repaired.
> Imagine version control systems where you commit the prompts used to generate features rather than the resulting implementation.
So every single run will result in different non-reproducible implementation with unique bugs requiring manual expert interventions. How is this better?
I did human notes -> had Claude condense and edit -> manually edit. A few of the sentences (like the stinky one below) were from Claude which I kept if it matched my own thoughts, though most were changed for style/prose.
I'm still experimenting with it. I find it can't match style at all, and even with the manual editing it still "smells like AI" as you picked up. But, it also saves time.
My prompt was essentially "here are my old blog posts, here's my notes on reading a bunch of AI generated commits, help me condense this into a coherent article about the insights I learned"
I don’t like to accuse, and the article is fine overall, but this stinks: “This transparency transforms git history from a record of changes into a record of intent, creating a new form of documentation that bridges human reasoning and machine implementation.”
Can we please stop using the em-dash as a metric to “detect” LLM writing? It’s lazy and wrong. Plenty of people use em-dashes, it’s a useful punctuation mark. If humans didn’t use them, they wouldn’t be in the LLM training data.
There are better clues, like the kind of vague pretentious babble bad marketers use to make their products and ideas seem more profound than they are. It’s a type of bad writing which looks grandiose but is ultimately meaningless and that LLMs heavily pick up on.
That’s not true at all. Apple’s OS by default have smart punctuation enabled and convert -- (two hyphens) into — (“em-dash”; not an “en-dash”, which has a different purpose), " " (dumb quotes) into “ ” (smart quotes), and so forth.
Furthermore, on macOS there are simple key combinations (e.g. with ⌥) to make all sort of smart punctuation even if you don’t have the feature enabled by default, and on iOS you can long press on a key (such as the hyphen) to see alternates.
The majority of people may not use correct punctuation marks, but enough do that assuming a single character immediately means they used an LLM is just plain wrong. I have never used an LLM to write a blog post, internet comment, or anything of the sort, and I have used smart punctuation in all my writing for over a decade. Same with plenty of other HN commenters, journalists, writers, editors, and on and on. You don’t need to be a literal machine to care about correct character use.
> I don’t like to accuse, and the article is fine overall, but this stinks:
Now consider your reasonable instinct to not accuse other people coupled with the possibility setting AI lose with “write a positive article about AI where you have some paragraphs about the current limitations based on this link. write like you are just following the evidence.” Meanwhile we are supposed to sit here and weigh every word.
This reminds to write a prompt for a blogpost. How AI could be used for making personal-looking tech-guy who meditates and runs websites. (Do we have the technology? Yes we do)
It's an interesting review but I really dislike this type of techno-utopian determinism: "When models inevitably improve..." Says who? How is it inevitable? What if they've actually reached their limits by now?
Models are improving every day. People are figuring out thousands of different optimizations to training and to hardware efficiency. The idea that right now in early June 2025 is when improvement stops beggars belief. We might be approaching a limit, but that's going to be a sigmoid curve, not a sudden halt in advancement.
I think at this point we're reaching more incremental updates, which can score higher on some benchmarks but then simultaneously behave worse with real-world prompts, most especially if they were prompt engineered for a specific model. I recall Google updating their Flash model on their API with no way to revert to the old one and it caused a lot of people to complain that everything they've built is no longer working because the model is just behaving differently than when they wrote all the prompts.
Isn't it quite possible they replaced that Flash model with a distilled version, saving money rather than increasing quality? This just speaks to the value of open-weights more than anything.
5 years ago a person would be blown away by today’s LLMs. But people today will merely say “cool” at whatever LLMs are in use 5 years from now. Or maybe not even that.
Most of the developers I know personally who have been radicalized by coding agents, it happened within the past 9 months. It does not feel like we are in a phase of predictable boring improvement.
Deciding that AI is going nowhere to suddenly deciding that coding agents are how they will work going forward is a radical change. That is what they meant.
It is copium that it will suddenly stop and the world they knew before will return.
ChatGPT came out in Nov 2022. Attention Was All There Was in 2017, we were already 5 years in the past. Or 5 years of research to catch up to, and then from 2022 to now ... papers and research have been increasing exponentially. Even in if SOTA models were frozen, we still have years of research to apply and optimize in various ways.
The copium is I think many people got comfortable post financial crisis with nothing much changing or happening. I think many people really liked a decade stretch with not much more than web framework updates and smart phone versioning.
We are just back on track.
I just read Oracular Programming: A Modular Foundation for Building LLM-Enabled Software the other day.
We don't even have a new paradigm yet. I would be shocked that in 10 years I don't look back at this time of writing a prompt into a chatbot and then pasting the code into an IDE as completely comical.
The most shocking thing to me is we are right back on track to what I would have expected in 2000 for 2025. In 2019 those expectations seemed like science fiction delusions after nothing happening for so long.
I think it's equally copium that people keep assuming we're just going to compound our way into intelligence that generalizes enough to stop us from handholding the AI, as much as I'd genuinely enjoy that future.
Lately I spend all day post-training models for my product, and I want to say 99% of the research specific to LLMs doesn't reproduce and/or matter once you actually dig in.
We're getting exponentially more papers on the topics and they're getting worse on average.
Every day there's a new paper claiming an X% gain by post-training some ancient 8B parameter model and comparing it to a bunch of other ancient models after they've overfitted on the public dataset of a given benchmark and given the model a best of 5.
And benchmarks won't ever show it, but even ChatGPT 3.5-Turbo has better general world knowledge than a lot models people consider "frontier" models today because post-training makes it easy to cover up those gaps with very impressive one-prompt outputs and strong benchmark scores.
-
It feels like things are getting stuck in a local maxima: we are making forward progress, the models are useful and getting more useful, but the future people are envisioning takes reaching a completely different goal post that I'm not at all convinced we're making exponential progress towards.
There maybe exponential number of techniques claiming to be ground breaking, but what has actually unlocked new capabilities that can't just as easily be attributed to how much more focused post-training has become on coding and math?
Test time compute feels like the only one and we're already seeing the cracks form in terms of its effect on hallucinations, and there's a clear ceiling for the performance the current iteration unlocks as all these models are converging on pretty similar performance after just a few model releases.
What is ironic, if we buy in to the theory that AI will write majority of the code in the next 5-10 years, what is it going to train on after? ITSELF? Seems this theoretic trajectory of "will inevitably get better" is is only true if humans are producing quality training data. The quality of code LLMs create is very well proportionate on how mature and ubiquitous the langues/projects are.
It is "inevitable" in the sense that in 99% of the cases, tomorrow is just like yesterday.
LLMs have been continually improving for years now. The surprising thing would be them not improving further. And if you follow the research even remotely, you know they'll improve for a while, because not all of the breakthroughs have landed in commercial models yet.
It's not "techno-utopian determinism". It's a clearly visible trajectory.
Meanwhile, if they didn't improve, it wouldn't make a significant change to the overall observations. It's picking a minor nit.
The observation that strict prompt adherence plus prompt archival could shift how we program is both true, and it's a phenomenon we observed several times in the past. Nobody keeps the assembly output from the compiler around anymore, either.
There's definitely valid criticism to the passage, and it's overly optimistic - in that most non-trivial prompts are still underspecified and have multiple possible implementations, not all correct. That's both a more useful criticism, and not tied to LLM improvements at all.
Models have improved significantly over the last 3 months. Yet people have been saying 'What if they've actually reached their limits by now?' for pushing 3 years.
For me, improvement means no hallucination, but that only seems to have gotten worse and I'm interested to find out whether it's actually solvable at all.
The benchmarks also claim random 32B parameter models beat Claude 4 at coding, so we know just how much they matter.
It should be obvious to anyone who with a cursory interest in model training, you can't trust benchmarks unless they're fully private black-boxes.
If you can get even a hint of the shape of the questions on a benchmark, it's trivial to synthesize massive amounts of data that help you beat the benchmark. And given the nature of funding right now, you're almost silly not to do it: it's not cheating, it's "demonstrably improving your performance at the downstream task"
Why do you care about hallucination for coding problems? You're in an agent loop; the compiler is ground truth. If the LLM hallucinates, the agent just iterates. You don't even see it unless you make the mistake of looking closely.
If the LLM hallucinates, then the code it produces is wrong. That wrong code isn't obviously or programmatically determinable as wrong, the agent has no way to figure out that it's wrong, it's not as if the LLM produces at the same time tests that identify that hallucinated code as being wrong. The only way that this wrong code can be identified as wrong is by the human user "looking closely" and figuring out that it is wrong.
You seem to have this fundamental belief that the code that's produced by your LLM is valid and doesn't need to be evaluated, line-by-line, by a human, before it can be committed?? I have no idea how you came to this belief but it certainly doesn't match my experience.
No, what's happening here is we're talking past each other.
An agent lints and compiles code. The LLM is stochastic and unreliable. The agent is ~200 lines of Python code that checks the exit code of the compiler and relays it back to the LLM. You can easily fool an LLM. You can't fool the compiler.
I didn't say anything about whether code needs to be reviewed line-by-line by humans. I review LLM code line-by-line. Lots of code that compiles clean is nonetheless horrible. But none of it includes hallucinated API calls.
Also, from where did this "you seem to have a fundamental belief" stuff come from? You had like 35 words to go on.
> If the LLM hallucinates, then the code it produces is wrong. That wrong code isn't obviously or programmatically determinable as wrong, the agent has no way to figure out that it's wrong, it's not as if the LLM produces at the same time tests that identify that hallucinated code as being wrong. The only way that this wrong code can be identified as wrong is by the human user "looking closely" and figuring out that it is wrong
The LLM can easily hallucinate code that will satisfy the agent and the compiler but will still fail the actual intent of the user.
> I review LLM code line-by-line. Lots of code that compiles clean is nonetheless horrible.
Indeed most code that LLMs generate compiles clean and is nevertheless horrible! I'm happy that you recognize this truth, but the fact that you review that LLM-generated code line-by-line makes you an extraordinary exception vs. the normal user, who generates LLM code and absolutely does not review it line-by-line.
> But none of [the LLM generated code] includes hallucinated API calls.
Hallucinated API calls are just one of many many possible kinds of hallucinated code that an LLM can generate, by no means does "hallucinated code" describe only "hallucinated API calls" -- !
When you say "the LLM can easily hallucinate code that will satisfy the compiler but still fail the actual intent of the user", all you are saying is that the code will have bugs. My code has bugs. So does yours. You don't get to use the fancy word "hallucination" for reasonable-looking, readable code that compiles and lints but has bugs.
I think at this point our respective points have been made, and we can wrap it up here.
> When you say "the LLM can easily hallucinate code that will satisfy the compiler but still fail the actual intent of the user", all you are saying is that the code will have bugs. My code has bugs. So does yours. You don't get to use the fancy word "hallucination" for reasonable-looking, readable code that compiles and lints but has bugs.
There is an obvious and categorical difference between the "bugs" that an LLM produces as part of its generated code, and the "bugs" that I produce as part of the code that I write. You don't get to conflate these two classes of bugs as though they are equivalent, or even comparable. They aren't.
The parent seems to be, in part, referring to "reward hacking", which tends to be used as a super category to what many refer to as slop, hallucination, cheating, and so on.
You seem to be using "hallucinate" to mean "makes mistakes".
That's not how I use it. I see hallucination as a very specific kind of mistake: one where the LLM outputs something that is entirely fabricated, like a class method that doesn't exist.
The agent compiler/linter loop can entirely eradicate those. That doesn't mean the LLM won't make plenty of other mistakes that don't qualify as hallucinations by the definition I use!
It's newts and salamanders. Every newt is a salamander, not every salamander is a newt. Every hallucination is a mistake, not every mistake is a hallucination.
Linting isn't verification of correctness, and yes, you can fool the compiler, linters, etc. Work with some human interns, they are great at it. Agents will do crazy things to get around linting errors, including removing functionality.
Just checking off the list of things you've determined to be irrelevant. Compiler? Nope. Linter? Nope. Test suite? Nope. How about TLA+ specifications?
I truly don't know what you're trying to communicate, with all of your recent comments related to AI and LLM and codegen and etc., the only thing I can guess is that you're just cynically throwing sand into the wind. It's unfortunate, your username used to carry some clout and respect.
TLA+ specs don’t verify code. They verify design. Such design can be expressed in whatever, including pseudocode (think algorithms notation in textbooks). Then you write the TLA specs that will judge if invariants are truly respected. Once you’re sure of the design, you can go and implement it, but there’s no hard constraints like a type system.
My guy didn't you spend like half your life in the field where your job was to sift through code that compiled but nonetheless had bugs that you tried to exploit? How can you possibly have this belief about AI generated code?
I don't understand this question. Yes, I spent about 20 years learning the lesson that code is profoundly knowable; to start with, you just read it. What challenge do you believe AI-generated code presents to me?
> You seem to have this fundamental belief that the code that's produced by your LLM is valid and doesn't need to be evaluated, line-by-line, by a human, before it can be committed??
This is a mistaken understanding. The person you responded to has written on these thoughts already and they used memorable words in response to this proposal:
> Are you a vibe coding Youtuber? Can you not read code? If so: astute point. Otherwise: what the fuck is wrong with you?
It should be obvious that one would read and verify the code before they commit it. Especially if one works on a team.
If you want a model that's getting better at helping you as a tool (which for the record, I do), then you'd say in the last 3 months things got better between Gemini's long context performance, the return of Claude Opus, etc.
But if your goal post is replacing SWEs entirely... then it's not hard to argue we definitely didn't overcome any new foundational issues in the last 3 months, and not too many were solved in the last 3 years even.
In the last year the only real foundational breakthrough would be RL-based reasoning w/ test time compute delivering real results, but what that does to hallucinations + even Deepseek catching up with just a few months of post-training shows in its current form, the technique doesn't completely blow up any barriers that were standing the way people were originally touting it.
Overall models are getting better at things we can trivially post-train and synthesize examples for, but it doesn't feel like we're breaking unsolved problems at a substantially accelerated rate (yet.)
>Around the 40-commit mark, manual commits became frequent
This matches my experience: some shiny (even sometimes impressive) greenspace demos but dramatically less useful maintaining a codebase - which for any successful product is 90% of the work.
The documentation angle is really good. I've noticed it with the mdc files and llm.txt semi-standard. Documentation is often treated as just extra cost and a chore. Now, good description of the project structure and good examples suddenly becomes something devs want ahead of time. Even if the reason is not perfect, I appreciate this shift we'll all benefit from.
If/when to commit prompts has been fascinating as we have been doing similarly to build Louie.ai. I now have several categories with different handling:
- Human reviewed: Code guidelines and prompt templates are essentially dev tool infra-as-code and need review
- Discarded: Individual prompt commands I write, and implementation plan progress files the AI write, both get trashed, and are even part of my .gitignore . They were kept by Cloudflare, but we don't keep these.
- Unreviewed: Claude Code does not do RAG in the usual sense, so it is on us to create guides for how we do things like use big frameworks. They are basically indexes for speeding up AI with less grepping + hallucinating across memory compactions. The AI reads and writes these, and we largely stay out of it.
There are weird cases I am still trying to figure out. Ex:
- feature impl might start with an AI coming up with the product spec, so having that maintained as the AI progresses and committed in is a potentially useful artifact
- how prompt templates get used is helpful for their automated maintenance.
This only works if the model and its context are immutable. None of us really control the models we use, so I'd be sceptical about reproducing the artifacts later.
I have documented my experience using an Agent for a slightly different task -- upgrading framework version -- I had to abandon the work, but, my learning has been similar what is in the post.
Another way to phrase this is LLM-as-compiler and Python (or whatever) as an intermediate compiler artefact.
Finally, a true 6th generation programming language!
I've considered building a toy of this with really aggressive modularisation of the output code (eg. python) and a query-based caching system so that each module of code output only changes when the relevant part of the prompt or upsteam modules change (the generated code would be committed to source control like a lockfile).
I think that (+ some sort of WASM encapsulated execution environment) would one of the best ways to write one off things like scripts which don't need to incrementally get better and more robust over time in the way that ordinary code does.
> Don't be afraid to get your hands dirty. Some bugs and styling issues are faster to fix manually than to prompt through. Knowing when to intervene is part of the craft.
This has been my experience as well. to always run the cli tool in the bottom pane of an IDE and not in a standalone terminal.
That's exactly what I thought, too, before I tried it!
Turns out it feels very different than I expected. I really recommend trying it rather than assuming. There's no learning curve, you just install Claude Code and run it in your repo and ask it for things.
(I am the author of the code being discussed. Or, uh, the author of the prompts at least.)
I used almost 100% AI to build a SCUMM-like parser, interpreter, and engine (https://github.com/fpgaminer/scumm-rust). It was a fun workflow; I could generally focus on my usual work and just pop in occasionally to check on and direct the AI.
I used a combination of OpenAI's online Codex, and Claude Sonnet 4 in VSCode agent mode. It was nice that Codex was more automated and had an environment it could work in, but its thought-logs are terrible. Iteration was also slow because it takes awhile for it to spin the environment up. And while you _can_ have multiple requests running at once, it usually doesn't make sense for a single, somewhat small project.
Sonnet 4's thoughts were much more coherent, and it was fun to watch it work and figure out problems. But there's something broken in VSCode right now that makes its ability to read console output inconsistent, which made things difficult.
The biggest issue I ran into is that both are set up to seek out and read only small parts of the code. While they're generally good at getting enough context, it does cause some degradation in quality. A frequent issue was replication of CSS styling between the Rust side of things (which creates all of the HTML elements) and the style.css side of things. Like it would be working on the Rust code and forget to check style.css, so it would just manually insert styles on the Rust side even though those elements were already styled on the style.css side.
Codex is also _terrible_ at formatting and will frequently muck things up, so it's mandatory to use it with an autoformatter and instructions to use it. Even with that, Codex will often say that it ran it, but didn't actually run it (or ran it somewhere in the middle instead of at the end) so its pull requests fail CI. Sonnet never seemed to have this issue and just used the prevailing style it saw in the files.
Now, when I say "almost 100% AI", it's maybe 99% because I did have to step in and do some edits myself for things that both failed at. In particular neither can see the actual game running, so they'd make weird mistakes with the design. (Yes, Sonnet in VS Code can see attached images, and potentially can see the DOM of vscode's built in browser, but the vision of all SOTA models is ass so it's effectively useless). I also stepped in once to do one major refactor. The AIs had decided on a very strange, messy, and buggy interpreter implementation at first.
I was thinking that if you had a good enough verified mathematical model of your code using TLA+ or similar you could then use an LLM to generate your code in any language and be confident it is correct. This would be Declarative Programming. Instead of putting in a lot of work writing code that MIGHT do what you intend you put more work into creating the verified model and then the LLM generates code that will do what the model intends.
You're talking ahead of the others in this thread, who do not understand how you got to what you're saying. I've been doing research in this area. You are not only correct, but the implications are staggering, and go further than what you have mentioned above. This is no cult, it is the reorganization of the economics of work.
You’re sounding like a religious zealot recruiting for a cult.
No, it is not possible to prompt every feature, and I suspect people who believe LLMs can accurately program anything in any language are frankly not solving any truly novel or interesting problems, because if they were they’d see the obvious cracks.
> I suspect people who believe LLMs can accurately program anything in any language are frankly not solving any truly novel or interesting problems, because if they were they’d see the obvious cracks.
The vast majority of problems in programming aren't novel or interesting.
I asked this in the other thread (no response, but I was a bit late)
How does anyone using AI like this have confidence that they aren't unintentionally plagiarizing code and violating the terms of whatever license it was released under?
For random personal projects I don't see it mattering that much. But if a large corp is releasing code like this, one would hope they've done some due diligence that they have to just stolen the code from some similar repo on GitHub, laundered through a LLM.
The only section in the readme doesn't mention checking similar projects or libraries for common code:
> Every line was thoroughly reviewed and cross-referenced with relevant RFCs, by security experts with previous experience with those RFCs.
I'm fairly confident that it's not just plagiarizing because I asked the LLM to implement a novel interface with unusual semantics. I then prompted for many specific fine-grain changes to implement features the way I wanted. It seems entirely implausible to me that there could exist prior art that happened to be structured exactly the way I requested.
Note that I came into this project believing that LLMs were plagiarism engines -- I was looking for that! I ended up concluding that this view was not consistent with the output I was actually seeing.
Most of the code generated by LLMs, and especially the code you actually keep from an agent, is mid, replacement-level, boring stuff. If you're not already building projects with LLMs, I think you need to start doing that first before you develop a strong take on this. From what I see in my own work, the code being generated is highly unlikely to be distinguishable. There is more of me and my prompts and decisions in the LLM code than there can possibly be defensible IPR from anybody else, unless the very notion of, like, wrapping a SQLite INSERT statement in Golang is defensible.
The best way I can explain the experience of working with an LLM agent right now is that it is like if every API in the world had a magic "examples" generator that always included whatever it was you were trying to do (so long as what you were trying to do was within the obvious remit of the library).
> How does anyone using AI like this have confidence that they aren't unintentionally plagiarizing code and violating the terms of whatever license it was released under?
All of the big LLM vendors have a "copyright shield" indemnity clause for their paying customers - a guarantee that if you get sued over IP for output from their models their legal team will step in to fight on your behalf.
The consensus for right or wrong, is that LLM produced code (unless repeated verbatim) is equivalent to you or I legitimately stating our novel understanding of mixed sources some of which may be copyrighted.
Safety in the shadow of giant tech companies. People were upset when Microsoft released Copilot trained on GitHub data, but nobody who cared doing do anything about it, and nobody who could have done something about it cared, so it just became the new norm.
This is an excellent question that the AI-boosters always seem to dance around. Three replies already are saying “Nobody cares.” Until they do. I’d be willing to bet that some time in the near future, some big company is going to care a lot and that there will be a landmark lawsuit that significantly changes the LLM landscape. Regulation or a judge is going to eventually decide the extent to which someone can use AI to copy someone else’s IP, and it’s not going to be pretty.
It just presumes a level of fixation in copyright law that I don’t think is realistic. There was a landmark lawsuit MAI v. Peak Computer in 1993, where judges determined that repairing a computer without the permission of the operating system’s author is copyright infringement, and it didn’t change the landscape at all because everyone immediately realized it’s not practical for things to work that way. There’s no realistic world where AI tools end up being extremely useful but nobody uses them because of a court ruling.
> Reading through these commits sparked an idea: what if we treated prompts as the actual source code? Imagine version control systems where you commit the prompts used to generate features rather than the resulting implementation.
Please god, no, never do this. For one thing, why would you not commit the generated source code when storage is essentially free? That seems insane for multiple reasons.
> When models inevitably improve, you could connect the latest version and regenerate the entire codebase with enhanced capability.
How would you know if the code was better or worse if it was never committed? How do you audit for security vulnerabilities or debug with no source code?
My work has involved a project that is almost entirely generated code for over a decade. Not AI generated, the actual work of the project is in creating the code generator.
One of the things we learned very quickly was that having generated source code in the same repository as actual source code was not sustainable. The nature of reviewing changes is just too different between them.
Another thing we learned very quickly was that attempting to generate code, then modify the result is not sustainable; nor is aiming for a 100% generated code base. The end result of that was that we had to significantly rearchitect the project for us to essentially inject manually crafted code into arbitrary places in the generated code.
Another thing we learned is that any change in the code generator needs to have a feature flag, because someone was relying on the old behavior.
> One of the things we learned very quickly was that having generated source code in the same repository as actual source code was not sustainable
My rule of the thumb is to have both in same repo, but treat generated code like binary data. This was informed by when I was burned by a tooling regression that broke the generated code and the investigation was complicated by having to correlate commits across different repositories
I think the biggest difference here is that your code generator is probably deterministic and you likely are able to debug the results it produces rather than treating it like a black box.
> One of the things we learned very quickly was that having generated source code in the same repository as actual source code was not sustainable.
Keeping a repository with the prompts, or other commands separate is fine, but not committing the generated code at all I find questionable at best.
If you can 100% reproduce the same generated code from the same prompts, even 5 years later, given the same versions and everything then I'd say "Sure, go ahead and don't saved the generated code, we can always regenerate it". As someone who spent some time in frontend development, we've been doing it like that for a long time with (MB+) generated code, keeping it in scm just isn't feasible long-term.
But given this is about LLMs, which people tend to run with temperature>0, this is unlikely to be true, so then I'd really urge anyone to actually store the results (somewhere, maybe not in scm specifically) as otherwise you won't have any idea about what the code was in the future.
> If you can 100% reproduce the same generated code from the same prompts, even 5 years later
Reproducible builds with deterministic stacks and local compilers are far from solved. Throwing in LLM randomness just makes for a spicier environment to not commit the generated code.
I didn't read it as that - If I understood correctly, generated code must be quarantined very tightly. And inevitably you need to edit/override generated code and the manner by which you alter it must go through some kind of process so the alteration is auditable and can again be clearly distinguished from generated code.
Tbh this all sounds very familiar and like classic data management/admin systems for regular businesses. The only difference is that the data is code and the admins are the engineers themselves so the temptation to "just" change things in place is too great. But I suspect it doesn't scale and is hard to manage etc.
I feel like using a compiler is in a sense a code generator where you don't commit the actual output
> I feel like using a compiler is in a sense a code generator where you don't commit the actual output
Compilers are deterministic. Given the same input you always get the same output so there's no reason to store the output. If you don't get the same output we call it a compiler bug!
LLMs do not work this way.
(Aside: Am I the only one who feels that the entire AI industry is predicated on replacing only development positions? we're looking at, what, 100bn invested, with almost no reduce in customer's operating costs other than if the customer has developers).
LLMs CAN be deterministic. You can control the temperature to get the same output repeatedly.
Although I don’t really understand why you’d only want to store prompts…
What if that model is no longer available?
Sure, but compilers are arguably idempotent. Same code input, same output. LLMs certainly are not.
Yeah I fully agree (in the other comments here, no less) I just think "I don't commit my code" to be a specific mindset of what code actually is
I will guess that you are generating orders of magnitude more lines of code with your software than people do when building projects with LLMs - if this is true I don't think the analogy holds.
There’s a huge difference between deterministic generated code and LLM generated code. The latter will be different every time, sometimes significantly so. Subsequent prompts would almost immediately be useless. “You did X, but we want Y” would just blow up if the next time through the LLM (or the new model you’re trying) doesn’t produce X at all.
The idea as stated is a poor one, but a slight reshuffling and it seems promising:
You generate code with LLMs. You write tests for this code, either using LLMs or on your own. You of course commit your actual code: it is required to actually run the program, after all. However you also save the entire prompt chain somewhere. Then (as stated in the article), when a much better model comes along, you re-run that chain, presumably with prompting like "create this project, focusing on efficiency" or "create this project in Rust" or "create this project, focusing on readability of the code". Then you run the tests against the new codebase and if the suite passes you carry on, with a much improved codebase. The theoretical benefit of this over just giving your previously generated code to the LLM and saying "improve the readability" is that the newer (better) LLM is not burdened by the context of the "worse" decisions made by the previous LLM.
Obviously it's not actually that simple, as tests don't catch everything (tho with fuzz testing and complete coverage and such they can catch most issues), but we programmers often treat them as if they do, so it might still be a worthwhile endeavor.
Your rephrasing better encompasses my idea, and I should have emphasized in the post that I do not think this is a good idea (nor possible) right now, it was more of a hand-wavy "how could we rethink source control in a post-LLM world" passing thought I had while reading through all the commits.
Clearly it struck a chord with a lot of the folks here though, and it's awesome to read the discourse.
Means the temperature should be set to 0 (which not every provider supports) so that the output becomes entirely deterministic. Right now with most models if you give the same input prompt twice it will give two different solutions.
Even at temp 0, you might get different answers, depending on your inference engine. There might be hardware differences, as well as software issues (e.g. vLLM documents this, if you're using batching, you might get different answers depending on where in the batch sequence your query landed).
Claude Code already uses a temperature of 0 (just inspect the requests) but it's not deterministic
Not to mention it also performs web searches, web fetching etc which would also make it not deterministic
Two years ago when I was working on this at a startup, setting OAI models’ temp to 0 still didn’t make them deterministic. Has that changed?
Do LLMs inference engines have a way to seed their randomness? so tho have reproducible outputs with still some variance if desired?
I would only care about more deterministic output if I was repeating the same process with the same model, which is not the point of the exercise.
>> what if we treated prompts as the actual source code?
You would not do this because: unlike programming languages, natural languages are ambiguous and thus inadequate to fully specify software.
Exactly!
> this assumes models can achieve strict prompt adherence
What does strict adherence to an ambiguous prompt even mean? It’s like those people asking Babbage if his machine would give the right answer when given the wrong figures. I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a proposition.
Prompts are like story on the board, and like engineers, depends on the understanding of the model the generated source code can vary. Saying the prompts could be the actual code is so wrong and dangerous thought
Worse. Models aren't deterministic! They use temperature value to control randomness, just so they can escape local minima!
Regenerated code might behave differently, have different bugs(worst case), or not work at all(best case).
Nitpick - it's the ML system that is sampling from model predictions that has a temperature parameter, not the model itself. Temperature and even model aside, there are other sources of randomness like the underlying hardware that can cause the havoc you describe.
Plus, commits depend on the current state of the system.
What sense does “getting rid of vulnerabilities by phasing out {dependency}” make, if the next generation of the code might not rely on the mentioned library at all? What does “improve performance of {method}” mean if the next generation used a fully different implementation?
It makes no sense whatsoever except for a vibecoders script that’s being extrapolated into a codebase.
I'd say commit a comprehensive testing system with the prompts.
Prompts are in a sense what higher level programming languages were to assembly. Sure there is a crucial difference which is reproducibility. I could try and write down my thoughts why I think in the long run it won't be so problematic. I could be wrong of course.
I run https://pollinations.ai which servers over 4 million monthly active users quite reliably. It is mostly coded with AI. Since about a year there was no significant human commit. You can check the codebase. It's messy but not more messy than my codebases were pre-LLMs.
I think prompts + tests in code will be the medium-term solution. Humans will be spending more time testing different architecture ideas and be involved in reviewing and larger changes that involve significant changes to the tests.
Agreed with the medium-term solution. I wish I put some more detail into that part of the post, I have more thoughts on it but didn't want to stray too far off topic.
Apart from obvious non-reproducibility, the other problem is lack of navigable structure. I can't command+click or "show usages" or "show definition" any more.
Just ask the AI for those obviously
The idea is good, but we should commit both documentation and tests. They allow regenerating the code at will.
I think the author is saying you commit the prompt with the resulting code. You said it yourself, storage is free, so comment the prompt along with the output (don’t comment that out that if I’m not being clear); it would show the developers(?) intent, and to some degree, almost always contribute to the documentation process.
Author here :). Right now, I think the pragmatic thing to do is to include all prompts used in either the PR description and/or in the commit description. This wouldn't make my longshot idea of "regenerating a repo from the ground up" possible, but it still adds very helpful context to code reviewers and can help others on your team learn prompting techniques.
I'm pretty sure most people aren't doing "software engineering" when they program. There's the whole world of WordPress and dream Weaver like programing out there too where the consequences of messing up aren't really important.
Llms can be configured to have deterministic output too
It’s been a thing people have done for at least a year https://github.com/i365dev/LetterDrop
There are lots of reasons not to do it. But if LLMs get good enough that it works consistently people will do it anyway.
What will people call it when coders rely on vibes even more than vibe coding?
Writing specs
Exactly my thought. This is just natural language as a specification language.
Haruspicy?
Yes, it's too early to be doing that now, but if you see the move to AI-assisted code as at least the same magnitude of change as the move from assembly to high level languages, the argument makes more sense.
Nobody commits the compiled code; this is the direction we are moving in, high level source code is the new assembly.
These posts are funny to me because prompt engineers point at them as evidence of the fast-approaching software engineer obsolescence but the years of experience in software engineering necessary to even guide an AI in this way is very high.
The reason he keeps adjusting the prompts is because he knows how to program. He knows what it should look like.
It just blurs the line between engineer and tool.
> It just blurs the line between engineer and tool.
I realise you meant it as “the engineer and their tool blend together”, but I read it like a funny insult: “that guy likes to think of himself as an engineer, but he’s a complete tool”.
The argument is that this stuff will so radically improve senior engineer productivity that the demand for junior engineers will crater. And without a pipeline of junior engineers, the junior-to-senior trajectory will radically atrophy
Essentially, the field will get frozen where existing senior engineers will be able to utilize AI to outship traditional senior-junior teams, even as junior engineers fail to secure employment
I don’t think anything in this article counters this argument
I don't know why people don't give more credence to the argument that the exact opposite thing will happen.
Right. I don’t understand why everyone thinks this will make it impossible for junior devs to learn. The people I had around to answer my questions when I was learning knew a whole lot less than Claude and also had full time jobs doing something other than answering my questions.
You learn by doing.. eg typing the code. It's not just knowledge, it's the intuition you develop when you write code yourself. Just like physical exercise. Or playing an instrument. It's not enough to know the theory, practice is key.
AI makes it very easy to avoid typing and hence make learning this skill less attractive.
But I don't necessarily see it as doom and gloom, what I think will happen - juniors will develop advanced intuition about using AI and getting the functionality they need, not the quality of the code, while at the same time the AI models will get increasingly better and write higher quality code.
It won't make it impossible for junior engineers to learn.
It will simply reduce the amount of opportunities to learn (and not just for juniors), by virtue of companies' beancounters concluding "two for one" (several juniors) doesn't return the same as "buy one get one free" (existing staff + AI license).
I dread the day we all "learn from AI". The social interaction part of learning is just as important as the content of it, really, especially when you're young; none of that comes across yet in the pure "1:1 interaction" with AI.
I learnt programming on my own, without any social interaction involved. In fact, I loved programming because it does not involve any social interaction.
Programming has become more of a "social game" in the last 15 years or so. AI is a new superpower for people like me, bringing balance to the Force.
If a junior engineer ships a similar repo to this with the help of AI, sure, I'll buy that.
But as of now, it's senior engineers who really know what they 're doing who can spot the errors in AI code.
Hold on. You said "really know what they're doing". Yes, I agree with that. What I don't buy is the coupling of that concept with "seniority".
Have a better term for "knows what they're doing" other than senior?
That's not what "senior" means.
This is not necessarily true in practical terms when it comes to hiring or promoting. Often a senior dev becomes a senior because of having an advanced skillset, despite years on the job. Similarily, often developers who have been on the job for many years aren’t ready for senior because of their lack or soft and hard skills.
Oh, that's one of the ways a senior dev becomes senior.
Maybe you could enlighten the rest of us then. According to your favorite definition, what does senior mean, what does seniority mean, and what's a term for someone who knows what they're doing?
Seniority means you've held the role for a long time.
> prompt engineers point at them as evidence of the fast-approaching software engineer obsolescence
Maybe journalists and bloggers angling for attention do it, prompt engineers are too aware of the limitations of prompting to do that.
I don't know why that's funny. This is not a post about a vibe coding session. It's Kenton Varda['s coding session].
later
updated to clarify kentonv didn't write this article
The sequence of commits talked about by the OP -- i.e. kenton's coding session's commits -- are like one degree removed from 100% pure vibe coding.
Your claim here being that Kenton Varda isn't reading the code he's generating. Got it. Good note.
No, that's not at all my claim, as it's obvious from the commit history that Kenton is reading the code he's generating before committing it.
What do you mean by "one degree removed from 100% pure vibe coding", then? The definition of vibe coding is letting the AI code without review...
> one degree removed
You're letting Claude do your programming for you, and then sweeping up whatever it does afterwards. Bluntly, you're off-loading your cognition to the machine. If that's fine by you then that's fine enough, it just means that the quality of your work becomes a function of your tooling rather than your capabilities.
I think it makes sense that GP is skeptical of this article considering it contains things like:
> this tool is improving itself, learning from every interaction
which seem to indicate a fundamental misunderstanding of how modern LLMs work: the 'improving' happens by humans training/refining existing models offline to create new models, and the 'learning' is just filling the context window with more stuff, not enhancement of the actual model or the model 'learning' - it will forget everything if you drop the context and as the context grows it can 'forget' things it previously 'learned'.
When you consider the "tool" as more than just the LLM model, but the stuff wrapped around calling that model then I feel like you can make a good argument it's improving when it keeps context in a file on disk and constantly updates and edits that file as you work throguh the project.
I do this routinely for large initiatives I'm kicking off through Claude Code - it writes a long detailed plan into a file and as we work through the project I have it constantly updating and rewriting that document to add information we have jointly discovered from each bit of the work. That means every time I come back and fire it back up, it's got more information than when it started, which looks a lot more improvement from my perspective.
So, it means that you and the LLM together have managed to write SEVEN lines of trivial code per hour. On a protocol that is perfectly documented, where you can look at about one million other implementations when in doubt.
It is not my intention to hurt your feelings, but it sounds like you and/or the LLM are not really good at their job. Looking at programmer salaries and LLM energy costs, this appears to be a very very VERY expensive OAuth library.
Again: Not my intention to hurt any feelings, but the numbers really are shockingly bad.
I spent about 5 days semi-focused on this codebase (though I always have lots of people interrupting me all the time). It's about 5000 lines (if you count comments, tests, and documentation, which you should). Where do you get 7 lines per hour?
>So, it means that you and the LLM together have managed to write SEVEN lines of trivial code per hour.
Here's their response
>It took me a few days to build the library with AI.
>I estimate it would have taken a few weeks, maybe months to write by hand.
>That said, this is a pretty ideal use case: implementing a well-known standard on a well-known platform with a clear API spec.
https://news.ycombinator.com/item?id=44160208
Lines of code per hour is a terrible metric to use. Additionally, it's far easier to critique code that's already written!
Yes, my brain got confused on who wrote the code and who just reported about it. I am truly sorry. I will go see my LLM doctor to get my brain repaired.
> Imagine version control systems where you commit the prompts used to generate features rather than the resulting implementation.
So every single run will result in different non-reproducible implementation with unique bugs requiring manual expert interventions. How is this better?
Humorous that this article has a strong AI writing smell - the author should publish the prompts they used!
I did human notes -> had Claude condense and edit -> manually edit. A few of the sentences (like the stinky one below) were from Claude which I kept if it matched my own thoughts, though most were changed for style/prose.
I'm still experimenting with it. I find it can't match style at all, and even with the manual editing it still "smells like AI" as you picked up. But, it also saves time.
My prompt was essentially "here are my old blog posts, here's my notes on reading a bunch of AI generated commits, help me condense this into a coherent article about the insights I learned"
I don’t like to accuse, and the article is fine overall, but this stinks: “This transparency transforms git history from a record of changes into a record of intent, creating a new form of documentation that bridges human reasoning and machine implementation.”
Also: "This OAuth library represents something larger than a technical milestone—it's evidence of a new creative dynamic emerging"
Em-dash baby.
I have used Em-dashes in many of my comments for years. It's just a result of reading books, where Em-dashes happen a lot.
Can we please stop using the em-dash as a metric to “detect” LLM writing? It’s lazy and wrong. Plenty of people use em-dashes, it’s a useful punctuation mark. If humans didn’t use them, they wouldn’t be in the LLM training data.
There are better clues, like the kind of vague pretentious babble bad marketers use to make their products and ideas seem more profound than they are. It’s a type of bad writing which looks grandiose but is ultimately meaningless and that LLMs heavily pick up on.
Very few people use n dashes in internet writing as opposed to dashes as they are not available on the default keyboard.
On the “default keyboard” of most people (a phone), you just long-press hyphen to choose any dash length.
That’s not true at all. Apple’s OS by default have smart punctuation enabled and convert -- (two hyphens) into — (“em-dash”; not an “en-dash”, which has a different purpose), " " (dumb quotes) into “ ” (smart quotes), and so forth.
Furthermore, on macOS there are simple key combinations (e.g. with ⌥) to make all sort of smart punctuation even if you don’t have the feature enabled by default, and on iOS you can long press on a key (such as the hyphen) to see alternates.
The majority of people may not use correct punctuation marks, but enough do that assuming a single character immediately means they used an LLM is just plain wrong. I have never used an LLM to write a blog post, internet comment, or anything of the sort, and I have used smart punctuation in all my writing for over a decade. Same with plenty of other HN commenters, journalists, writers, editors, and on and on. You don’t need to be a literal machine to care about correct character use.
> I don’t like to accuse, and the article is fine overall, but this stinks:
Now consider your reasonable instinct to not accuse other people coupled with the possibility setting AI lose with “write a positive article about AI where you have some paragraphs about the current limitations based on this link. write like you are just following the evidence.” Meanwhile we are supposed to sit here and weigh every word.
This reminds to write a prompt for a blogpost. How AI could be used for making personal-looking tech-guy who meditates and runs websites. (Do we have the technology? Yes we do)
It's an interesting review but I really dislike this type of techno-utopian determinism: "When models inevitably improve..." Says who? How is it inevitable? What if they've actually reached their limits by now?
Models are improving every day. People are figuring out thousands of different optimizations to training and to hardware efficiency. The idea that right now in early June 2025 is when improvement stops beggars belief. We might be approaching a limit, but that's going to be a sigmoid curve, not a sudden halt in advancement.
I think at this point we're reaching more incremental updates, which can score higher on some benchmarks but then simultaneously behave worse with real-world prompts, most especially if they were prompt engineered for a specific model. I recall Google updating their Flash model on their API with no way to revert to the old one and it caused a lot of people to complain that everything they've built is no longer working because the model is just behaving differently than when they wrote all the prompts.
Isn't it quite possible they replaced that Flash model with a distilled version, saving money rather than increasing quality? This just speaks to the value of open-weights more than anything.
5 years ago a person would be blown away by today’s LLMs. But people today will merely say “cool” at whatever LLMs are in use 5 years from now. Or maybe not even that.
Most of the developers I know personally who have been radicalized by coding agents, it happened within the past 9 months. It does not feel like we are in a phase of predictable boring improvement.
Radicalized? Going with the flow and wishes of the people who are driving AI is the opposite of that.
To have their minds changed drastically, sure..
> very different from the usual or traditional
https://www.merriam-webster.com/dictionary/radical
Deciding that AI is going nowhere to suddenly deciding that coding agents are how they will work going forward is a radical change. That is what they meant.
Sorry I have no idea what you're trying to say here.
5 years ago GPT2 was already outputting largely coherent speech, there's been progress but it's not all that shocking
Bold prediction…
It is copium that it will suddenly stop and the world they knew before will return.
ChatGPT came out in Nov 2022. Attention Was All There Was in 2017, we were already 5 years in the past. Or 5 years of research to catch up to, and then from 2022 to now ... papers and research have been increasing exponentially. Even in if SOTA models were frozen, we still have years of research to apply and optimize in various ways.
The copium is I think many people got comfortable post financial crisis with nothing much changing or happening. I think many people really liked a decade stretch with not much more than web framework updates and smart phone versioning.
We are just back on track.
I just read Oracular Programming: A Modular Foundation for Building LLM-Enabled Software the other day.
We don't even have a new paradigm yet. I would be shocked that in 10 years I don't look back at this time of writing a prompt into a chatbot and then pasting the code into an IDE as completely comical.
The most shocking thing to me is we are right back on track to what I would have expected in 2000 for 2025. In 2019 those expectations seemed like science fiction delusions after nothing happening for so long.
I think it's equally copium that people keep assuming we're just going to compound our way into intelligence that generalizes enough to stop us from handholding the AI, as much as I'd genuinely enjoy that future.
Lately I spend all day post-training models for my product, and I want to say 99% of the research specific to LLMs doesn't reproduce and/or matter once you actually dig in.
We're getting exponentially more papers on the topics and they're getting worse on average.
Every day there's a new paper claiming an X% gain by post-training some ancient 8B parameter model and comparing it to a bunch of other ancient models after they've overfitted on the public dataset of a given benchmark and given the model a best of 5.
And benchmarks won't ever show it, but even ChatGPT 3.5-Turbo has better general world knowledge than a lot models people consider "frontier" models today because post-training makes it easy to cover up those gaps with very impressive one-prompt outputs and strong benchmark scores.
-
It feels like things are getting stuck in a local maxima: we are making forward progress, the models are useful and getting more useful, but the future people are envisioning takes reaching a completely different goal post that I'm not at all convinced we're making exponential progress towards.
There maybe exponential number of techniques claiming to be ground breaking, but what has actually unlocked new capabilities that can't just as easily be attributed to how much more focused post-training has become on coding and math?
Test time compute feels like the only one and we're already seeing the cracks form in terms of its effect on hallucinations, and there's a clear ceiling for the performance the current iteration unlocks as all these models are converging on pretty similar performance after just a few model releases.
What is ironic, if we buy in to the theory that AI will write majority of the code in the next 5-10 years, what is it going to train on after? ITSELF? Seems this theoretic trajectory of "will inevitably get better" is is only true if humans are producing quality training data. The quality of code LLMs create is very well proportionate on how mature and ubiquitous the langues/projects are.
It is "inevitable" in the sense that in 99% of the cases, tomorrow is just like yesterday.
LLMs have been continually improving for years now. The surprising thing would be them not improving further. And if you follow the research even remotely, you know they'll improve for a while, because not all of the breakthroughs have landed in commercial models yet.
It's not "techno-utopian determinism". It's a clearly visible trajectory.
Meanwhile, if they didn't improve, it wouldn't make a significant change to the overall observations. It's picking a minor nit.
The observation that strict prompt adherence plus prompt archival could shift how we program is both true, and it's a phenomenon we observed several times in the past. Nobody keeps the assembly output from the compiler around anymore, either.
There's definitely valid criticism to the passage, and it's overly optimistic - in that most non-trivial prompts are still underspecified and have multiple possible implementations, not all correct. That's both a more useful criticism, and not tied to LLM improvements at all.
Are there places that follow the research that speak to the layperson?
More compute mean more faster processing, more context.
Models have improved significantly over the last 3 months. Yet people have been saying 'What if they've actually reached their limits by now?' for pushing 3 years.
For me, improvement means no hallucination, but that only seems to have gotten worse and I'm interested to find out whether it's actually solvable at all.
All the benchmarks would disagree with you
Today’s public benchmarks are yesterday’s training data.
The benchmarks also claim random 32B parameter models beat Claude 4 at coding, so we know just how much they matter.
It should be obvious to anyone who with a cursory interest in model training, you can't trust benchmarks unless they're fully private black-boxes.
If you can get even a hint of the shape of the questions on a benchmark, it's trivial to synthesize massive amounts of data that help you beat the benchmark. And given the nature of funding right now, you're almost silly not to do it: it's not cheating, it's "demonstrably improving your performance at the downstream task"
Why do you care about hallucination for coding problems? You're in an agent loop; the compiler is ground truth. If the LLM hallucinates, the agent just iterates. You don't even see it unless you make the mistake of looking closely.
What on earth are you talking about??
If the LLM hallucinates, then the code it produces is wrong. That wrong code isn't obviously or programmatically determinable as wrong, the agent has no way to figure out that it's wrong, it's not as if the LLM produces at the same time tests that identify that hallucinated code as being wrong. The only way that this wrong code can be identified as wrong is by the human user "looking closely" and figuring out that it is wrong.
You seem to have this fundamental belief that the code that's produced by your LLM is valid and doesn't need to be evaluated, line-by-line, by a human, before it can be committed?? I have no idea how you came to this belief but it certainly doesn't match my experience.
No, what's happening here is we're talking past each other.
An agent lints and compiles code. The LLM is stochastic and unreliable. The agent is ~200 lines of Python code that checks the exit code of the compiler and relays it back to the LLM. You can easily fool an LLM. You can't fool the compiler.
I didn't say anything about whether code needs to be reviewed line-by-line by humans. I review LLM code line-by-line. Lots of code that compiles clean is nonetheless horrible. But none of it includes hallucinated API calls.
Also, from where did this "you seem to have a fundamental belief" stuff come from? You had like 35 words to go on.
> If the LLM hallucinates, then the code it produces is wrong. That wrong code isn't obviously or programmatically determinable as wrong, the agent has no way to figure out that it's wrong, it's not as if the LLM produces at the same time tests that identify that hallucinated code as being wrong. The only way that this wrong code can be identified as wrong is by the human user "looking closely" and figuring out that it is wrong
The LLM can easily hallucinate code that will satisfy the agent and the compiler but will still fail the actual intent of the user.
> I review LLM code line-by-line. Lots of code that compiles clean is nonetheless horrible.
Indeed most code that LLMs generate compiles clean and is nevertheless horrible! I'm happy that you recognize this truth, but the fact that you review that LLM-generated code line-by-line makes you an extraordinary exception vs. the normal user, who generates LLM code and absolutely does not review it line-by-line.
> But none of [the LLM generated code] includes hallucinated API calls.
Hallucinated API calls are just one of many many possible kinds of hallucinated code that an LLM can generate, by no means does "hallucinated code" describe only "hallucinated API calls" -- !
When you say "the LLM can easily hallucinate code that will satisfy the compiler but still fail the actual intent of the user", all you are saying is that the code will have bugs. My code has bugs. So does yours. You don't get to use the fancy word "hallucination" for reasonable-looking, readable code that compiles and lints but has bugs.
I think at this point our respective points have been made, and we can wrap it up here.
> When you say "the LLM can easily hallucinate code that will satisfy the compiler but still fail the actual intent of the user", all you are saying is that the code will have bugs. My code has bugs. So does yours. You don't get to use the fancy word "hallucination" for reasonable-looking, readable code that compiles and lints but has bugs.
There is an obvious and categorical difference between the "bugs" that an LLM produces as part of its generated code, and the "bugs" that I produce as part of the code that I write. You don't get to conflate these two classes of bugs as though they are equivalent, or even comparable. They aren't.
Hallucination is a fancy word?
The parent seems to be, in part, referring to "reward hacking", which tends to be used as a super category to what many refer to as slop, hallucination, cheating, and so on.
https://courses.physics.illinois.edu/ece448/sp2025/slides/le...
You seem to be using "hallucinate" to mean "makes mistakes".
That's not how I use it. I see hallucination as a very specific kind of mistake: one where the LLM outputs something that is entirely fabricated, like a class method that doesn't exist.
The agent compiler/linter loop can entirely eradicate those. That doesn't mean the LLM won't make plenty of other mistakes that don't qualify as hallucinations by the definition I use!
It's newts and salamanders. Every newt is a salamander, not every salamander is a newt. Every hallucination is a mistake, not every mistake is a hallucination.
https://simonwillison.net/2025/Mar/2/hallucinations-in-code/
Linting isn't verification of correctness, and yes, you can fool the compiler, linters, etc. Work with some human interns, they are great at it. Agents will do crazy things to get around linting errors, including removing functionality.
have you no tests?
Irrelevant, really. Tests establish a minimum threshold of acceptability, they don't (and can't) guarantee anything like overall correctness.
Just checking off the list of things you've determined to be irrelevant. Compiler? Nope. Linter? Nope. Test suite? Nope. How about TLA+ specifications?
I truly don't know what you're trying to communicate, with all of your recent comments related to AI and LLM and codegen and etc., the only thing I can guess is that you're just cynically throwing sand into the wind. It's unfortunate, your username used to carry some clout and respect.
TLA+ specs don’t verify code. They verify design. Such design can be expressed in whatever, including pseudocode (think algorithms notation in textbooks). Then you write the TLA specs that will judge if invariants are truly respected. Once you’re sure of the design, you can go and implement it, but there’s no hard constraints like a type system.
My guy didn't you spend like half your life in the field where your job was to sift through code that compiled but nonetheless had bugs that you tried to exploit? How can you possibly have this belief about AI generated code?
I don't understand this question. Yes, I spent about 20 years learning the lesson that code is profoundly knowable; to start with, you just read it. What challenge do you believe AI-generated code presents to me?
> You seem to have this fundamental belief that the code that's produced by your LLM is valid and doesn't need to be evaluated, line-by-line, by a human, before it can be committed??
This is a mistaken understanding. The person you responded to has written on these thoughts already and they used memorable words in response to this proposal:
> Are you a vibe coding Youtuber? Can you not read code? If so: astute point. Otherwise: what the fuck is wrong with you?
It should be obvious that one would read and verify the code before they commit it. Especially if one works on a team.
https://fly.io/blog/youre-all-nuts/
This is just people talking past each other.
If you want a model that's getting better at helping you as a tool (which for the record, I do), then you'd say in the last 3 months things got better between Gemini's long context performance, the return of Claude Opus, etc.
But if your goal post is replacing SWEs entirely... then it's not hard to argue we definitely didn't overcome any new foundational issues in the last 3 months, and not too many were solved in the last 3 years even.
In the last year the only real foundational breakthrough would be RL-based reasoning w/ test time compute delivering real results, but what that does to hallucinations + even Deepseek catching up with just a few months of post-training shows in its current form, the technique doesn't completely blow up any barriers that were standing the way people were originally touting it.
Overall models are getting better at things we can trivially post-train and synthesize examples for, but it doesn't feel like we're breaking unsolved problems at a substantially accelerated rate (yet.)
https://xkcd.com/605/
>Around the 40-commit mark, manual commits became frequent
This matches my experience: some shiny (even sometimes impressive) greenspace demos but dramatically less useful maintaining a codebase - which for any successful product is 90% of the work.
The documentation angle is really good. I've noticed it with the mdc files and llm.txt semi-standard. Documentation is often treated as just extra cost and a chore. Now, good description of the project structure and good examples suddenly becomes something devs want ahead of time. Even if the reason is not perfect, I appreciate this shift we'll all benefit from.
If/when to commit prompts has been fascinating as we have been doing similarly to build Louie.ai. I now have several categories with different handling:
- Human reviewed: Code guidelines and prompt templates are essentially dev tool infra-as-code and need review
- Discarded: Individual prompt commands I write, and implementation plan progress files the AI write, both get trashed, and are even part of my .gitignore . They were kept by Cloudflare, but we don't keep these.
- Unreviewed: Claude Code does not do RAG in the usual sense, so it is on us to create guides for how we do things like use big frameworks. They are basically indexes for speeding up AI with less grepping + hallucinating across memory compactions. The AI reads and writes these, and we largely stay out of it.
There are weird cases I am still trying to figure out. Ex:
- feature impl might start with an AI coming up with the product spec, so having that maintained as the AI progresses and committed in is a potentially useful artifact
- how prompt templates get used is helpful for their automated maintenance.
> Treat prompts as version-controlled assets
This only works if the model and its context are immutable. None of us really control the models we use, so I'd be sceptical about reproducing the artifacts later.
I have documented my experience using an Agent for a slightly different task -- upgrading framework version -- I had to abandon the work, but, my learning has been similar what is in the post.
https://www.trk7.com/blog/ai-agents-for-coding-promise-vs-re...
> Prompts as Source Code
Another way to phrase this is LLM-as-compiler and Python (or whatever) as an intermediate compiler artefact.
Finally, a true 6th generation programming language!
I've considered building a toy of this with really aggressive modularisation of the output code (eg. python) and a query-based caching system so that each module of code output only changes when the relevant part of the prompt or upsteam modules change (the generated code would be committed to source control like a lockfile).
I think that (+ some sort of WASM encapsulated execution environment) would one of the best ways to write one off things like scripts which don't need to incrementally get better and more robust over time in the way that ordinary code does.
> Finally, a true 6th generation programming language!
Karpathy already said English is the new programming language.
> Don't be afraid to get your hands dirty. Some bugs and styling issues are faster to fix manually than to prompt through. Knowing when to intervene is part of the craft.
This has been my experience as well. to always run the cli tool in the bottom pane of an IDE and not in a standalone terminal.
Discussion from 4 days ago when the code was announced (846 points, 519 comments):
https://news.ycombinator.com/item?id=44159166
I'm sorry, this all sounds like a fucking miserable experience. Like, if this is what my job becomes, I'll probably quit tech completely.
That's exactly what I thought, too, before I tried it!
Turns out it feels very different than I expected. I really recommend trying it rather than assuming. There's no learning curve, you just install Claude Code and run it in your repo and ask it for things.
(I am the author of the code being discussed. Or, uh, the author of the prompts at least.)
I used almost 100% AI to build a SCUMM-like parser, interpreter, and engine (https://github.com/fpgaminer/scumm-rust). It was a fun workflow; I could generally focus on my usual work and just pop in occasionally to check on and direct the AI.
I used a combination of OpenAI's online Codex, and Claude Sonnet 4 in VSCode agent mode. It was nice that Codex was more automated and had an environment it could work in, but its thought-logs are terrible. Iteration was also slow because it takes awhile for it to spin the environment up. And while you _can_ have multiple requests running at once, it usually doesn't make sense for a single, somewhat small project.
Sonnet 4's thoughts were much more coherent, and it was fun to watch it work and figure out problems. But there's something broken in VSCode right now that makes its ability to read console output inconsistent, which made things difficult.
The biggest issue I ran into is that both are set up to seek out and read only small parts of the code. While they're generally good at getting enough context, it does cause some degradation in quality. A frequent issue was replication of CSS styling between the Rust side of things (which creates all of the HTML elements) and the style.css side of things. Like it would be working on the Rust code and forget to check style.css, so it would just manually insert styles on the Rust side even though those elements were already styled on the style.css side.
Codex is also _terrible_ at formatting and will frequently muck things up, so it's mandatory to use it with an autoformatter and instructions to use it. Even with that, Codex will often say that it ran it, but didn't actually run it (or ran it somewhere in the middle instead of at the end) so its pull requests fail CI. Sonnet never seemed to have this issue and just used the prevailing style it saw in the files.
Now, when I say "almost 100% AI", it's maybe 99% because I did have to step in and do some edits myself for things that both failed at. In particular neither can see the actual game running, so they'd make weird mistakes with the design. (Yes, Sonnet in VS Code can see attached images, and potentially can see the DOM of vscode's built in browser, but the vision of all SOTA models is ass so it's effectively useless). I also stepped in once to do one major refactor. The AIs had decided on a very strange, messy, and buggy interpreter implementation at first.
I was thinking that if you had a good enough verified mathematical model of your code using TLA+ or similar you could then use an LLM to generate your code in any language and be confident it is correct. This would be Declarative Programming. Instead of putting in a lot of work writing code that MIGHT do what you intend you put more work into creating the verified model and then the LLM generates code that will do what the model intends.
Many of you are failing to conprehend the potential scale of AI generated codebases.
Take note - there is no limit. Every feature you or the AI can prompt can be generated.
Imagine if you were immortal and given unlimited storage. Imagine what you could create.
That’s a prompt away.
Even now you’re still restricting your thinking to the old ways.
You're talking ahead of the others in this thread, who do not understand how you got to what you're saying. I've been doing research in this area. You are not only correct, but the implications are staggering, and go further than what you have mentioned above. This is no cult, it is the reorganization of the economics of work.
You’re sounding like a religious zealot recruiting for a cult.
No, it is not possible to prompt every feature, and I suspect people who believe LLMs can accurately program anything in any language are frankly not solving any truly novel or interesting problems, because if they were they’d see the obvious cracks.
> I suspect people who believe LLMs can accurately program anything in any language are frankly not solving any truly novel or interesting problems, because if they were they’d see the obvious cracks.
The vast majority of problems in programming aren't novel or interesting.
> That’s a prompt away.
Currently, it's 6 prompts away in which 5 of those are me guiding the LLM to output the answer that I already have in mind.
I asked this in the other thread (no response, but I was a bit late)
How does anyone using AI like this have confidence that they aren't unintentionally plagiarizing code and violating the terms of whatever license it was released under?
For random personal projects I don't see it mattering that much. But if a large corp is releasing code like this, one would hope they've done some due diligence that they have to just stolen the code from some similar repo on GitHub, laundered through a LLM.
The only section in the readme doesn't mention checking similar projects or libraries for common code:
> Every line was thoroughly reviewed and cross-referenced with relevant RFCs, by security experts with previous experience with those RFCs.
I'm fairly confident that it's not just plagiarizing because I asked the LLM to implement a novel interface with unusual semantics. I then prompted for many specific fine-grain changes to implement features the way I wanted. It seems entirely implausible to me that there could exist prior art that happened to be structured exactly the way I requested.
Note that I came into this project believing that LLMs were plagiarism engines -- I was looking for that! I ended up concluding that this view was not consistent with the output I was actually seeing.
Most of the code generated by LLMs, and especially the code you actually keep from an agent, is mid, replacement-level, boring stuff. If you're not already building projects with LLMs, I think you need to start doing that first before you develop a strong take on this. From what I see in my own work, the code being generated is highly unlikely to be distinguishable. There is more of me and my prompts and decisions in the LLM code than there can possibly be defensible IPR from anybody else, unless the very notion of, like, wrapping a SQLite INSERT statement in Golang is defensible.
The best way I can explain the experience of working with an LLM agent right now is that it is like if every API in the world had a magic "examples" generator that always included whatever it was you were trying to do (so long as what you were trying to do was within the obvious remit of the library).
> How does anyone using AI like this have confidence that they aren't unintentionally plagiarizing code and violating the terms of whatever license it was released under?
They don’t and no one cares
All of the big LLM vendors have a "copyright shield" indemnity clause for their paying customers - a guarantee that if you get sued over IP for output from their models their legal team will step in to fight on your behalf.
The consensus for right or wrong, is that LLM produced code (unless repeated verbatim) is equivalent to you or I legitimately stating our novel understanding of mixed sources some of which may be copyrighted.
Safety in the shadow of giant tech companies. People were upset when Microsoft released Copilot trained on GitHub data, but nobody who cared doing do anything about it, and nobody who could have done something about it cared, so it just became the new norm.
This is an excellent question that the AI-boosters always seem to dance around. Three replies already are saying “Nobody cares.” Until they do. I’d be willing to bet that some time in the near future, some big company is going to care a lot and that there will be a landmark lawsuit that significantly changes the LLM landscape. Regulation or a judge is going to eventually decide the extent to which someone can use AI to copy someone else’s IP, and it’s not going to be pretty.
It just presumes a level of fixation in copyright law that I don’t think is realistic. There was a landmark lawsuit MAI v. Peak Computer in 1993, where judges determined that repairing a computer without the permission of the operating system’s author is copyright infringement, and it didn’t change the landscape at all because everyone immediately realized it’s not practical for things to work that way. There’s no realistic world where AI tools end up being extremely useful but nobody uses them because of a court ruling.
As an individual dev, I simply don’t care. Not my problem.
Companies are satisfied with the idemnity provided by Microsoft.