I read all of Cloudflare's Claude-generated commits
maxemitchell.com
maxemitchell.com
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?
Just like any other RFC.
You would not do this because: unlike programming languages, natural languages are ambiguous and thus inadequate to fully specify software.
> 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.
Llms can be configured to have deterministic output too
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.
Not to mention it also performs web searches, web fetching etc which would also make it not deterministic
Clearly it struck a chord with a lot of the folks here though, and it's awesome to read the discourse.
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.
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.
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.
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.
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.
> Now, when you send a request to one of the Gemini 2.5 models, if the request shares a common prefix as one of previous requests, then it’s eligible for a cache hit. We will dynamically pass cost savings back to you, providing the same 75% token discount.
> In order to increase the chance that your request contains a cache hit, you should keep the content at the beginning of the request the same and add things like a user's question or other additional context that might change from request to request at the end of the prompt.
From https://news.ycombinator.com/item?id=43939774 re: same:
> Does this make it appear that the LLM's responses converge on one answer when actually it's just caching?
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).
Although I don’t really understand why you’d only want to store prompts…
What if that model is no longer available?
Except when they aren't. See for instance https://gcc.gnu.org/onlinedocs/gcc-15.1.0/gcc/Developer-Opti... or the __DATE__/__TIME__ macros.
> You can use the -frandom-seed option to produce reproducibly identical object files.
Deterministic.
Also, with regard to __DATE__/__TIME__ macros, those are deterministic, because the current date and time are part of the inputs.
Many compilers are not deterministic when only considering the source files or even the current time. For example, any output produced by iterating over a hash table with pointer keys is likely to depend on ASLR and thus be nondetermistic unless you consider the ASLR randomization to be one of the inputs. Any output that depends on directory iteration order is likely to be consistent on a single computer but vary across computers.
LLMs aren’t magic. They’re software running on inputs like anything else, which means they’re deterministic if you constrain all the inputs.
If LLM generation was like a Makefile step, part of your build process, this concern would make a lot of sense. But nobody, anywhere, does that.
Top level comment of this thread, quoting the article:
> 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.
I've been working on a recent project with heavy use of AI (probably around 100 hours of long-running autonomous AI sprints over the last few weeks), and if you tried to re-run all of my prompts in order, even using the exact same models with the exact same tooling, it would almost certainly fall apart pretty quickly. After the first few, a huge portion of the remaining prompts would be referencing code that wouldn't exist and/or responding to things that wouldn't have been said in the AI's responses. Meta-prompting (prompting agents to prepare prompts for other agents) would be an interesting challenge to properly encode. And how would human code changes be represented, as patches against code that also wouldn't exist?
The whole idea also ignores that AI being fast and cheap compared to human developers doesn't make it infinitely fast or free, or put it in the same league of quickness and cheapness as a compiler. Even if this were conceptually feasible, all it would really accomplish is making it so that any new release of a major software project takes weeks (or more) of build time and thousands of dollars (or more) burned on compute.
It's an interesting thought experiment, but the way I would put it into practice would be to use tooling that includes all relevant prompts / chat logs in each commit message. Then maybe in the future an agent with a more advanced model could go through each commit in the history one by one, take notes on how each change could have been better implemented based on the associated commit message and any source prompts contained therein, use those notes to inform a consolidated set of recommended changes to the current code, and then actually apply the recommendations in a series of pull requests.
There's always this vibe that, like, AI code is like an IOCCC puzzle. No. It's extremely boring mid-code. Any competent developer can review it.
(I assume that isn't what you're actually arguing against, in which case at least one of us must have misread something from the parent chain.)
When you are talking to a person and interpreting what they mean, you have an inherent theory of mind whether you are consciously thinking "how does this person think" or not. It's how we communicate with other people efficiently and it's one of the many things missing with LLM roulette. It's not that you generate a new "theory of mind" with every interaction. It's not something you have to consciously do (although you can, like breathing).
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
Jokes aside, I have worked in projects where auto-generating code was the solution that was chosen and it's always been 100% auto-generated, essentially at compilation time. Any hand-coded stuff needed to handle corner cases or glue pieces together was kept outside of the code generator.
This sounds like putting assembly in C code. What was the input language? These two bits ("Not AI generated", "a feature flag") suggest that the code generator didn't have a natural language frontend, but rather a real programming language frontend.
Did you or anyone else inform management that a code generator is essentially a compiler with extra characters? [0] If yes, then what was their response?
I am concerned that your current/past work might have been to build a Compiler-as-a-Service (CaaS). [1] No shade, I'm just concerned that other managers might read all this and then try to build their own CaaS.
[0] Yes, I'm implying that LLMs are compilers. Altman has played us for fools; he's taught a billion people the worst part of programming: fighting the compiler to give you the output you want.
[1] Compiler-as-a-Service is the future our forefathers couldn't imagine warning us about. LLMs are CaaS's; time is a flat circle; where's the exit?; I want off this ride.
> Did you or anyone else inform management that a code generator is essentially a compiler with extra characters?
The output of the code generator was itself fed into a compiler that we also built; and about half of the codegen team (myself included) were themselves developers for the compiler.
I think management is still scared from the 20 year old M4 monstrosity we are still maintaining because writing a compiler would be "too complex".
Nobody commits the compiled code; this is the direction we are moving in, high level source code is the new assembly.
Regenerated code might behave differently, have different bugs(worst case), or not work at all(best case).
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.
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.
Let's assume that a hypothetical future AI is perfect. It will produce correct output 100% of the time, with no bugs, errors, omissions, security flaws, or other failings. It will also generate output instantly and cost nothing to run.
Even with such perfection this idea is doomed to failure because it can only write code based on information in the prompt, which is written by a human. Any ambiguity, unstated assumption, or omission would result in a program that didn't work quite right. Even a perfect AI is not telepathic. So you'd need to explain and describe your intended solution extremely precisely without ambiguity. Especially considering in this "offline generation" case there is no opportunity for our presumed perfect AI to ask clarifying questions.
But, by definition, any language which is precise and clear enough to not produce ambiguity is effectively a programming language, so you've not gained anything over just writing code.
It could also analyze the company website, marketing materials, and so forth, and use that to infer the missing pieces. (Again, something that exists today)
If the company website, marketing materials, and so forth become part of the input, you’ll have to put those in version control as well, as any change is likely to result in a different application being generated (which may or may not be what you want).
There is definitely an argument, for also committing prompts, but it makes no sense to only commit prompts.
Idk kinda different tho.
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.
later
updated to clarify kentonv didn't write this article
> 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'.
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.
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.
How, exactly? Do you think that you're "deciding what gets built and how it's designed" by iterating on the prompts that you feed to the LLM that generates the code?
Or are you saying that you're somehow able to write the "interesting" code, and can instruct the LLM to generate the "boring and obvious" code that needs to be filled-in to make your interesting code work? (This is certainly not what's indicated by your commit history, but, who knows?)
My prompts specify very precisely what should be implemented. I specified the public API and high-level design upfront. I let the AI come up with its own storage schema initially but then I prompted it very specifically through several improvements (e.g. "denormalize this table into this other table to eliminate a lookup"). I designed the end-to-end encryption scheme and told it in detail how to implement it. I pointed out bugs and explained how to fix them. And so on.
All the thinking happened in those prompts. With the details I provided, combined with the OAuth spec, there was really very little room left for any creativity in the code. It was basically connect-the-dots at that point.
> My prompts specify very precisely what should be implemented.
And the precision of your prompt's specifications, has no reliable impact on exactly what code the LLM returns as output.
> With the details I provided, combined with the OAuth spec, there was really very little room left for any creativity in the code. It was basically connect-the-dots at that point.
I truly don't know how you can come to this conclusion, if you have any amount of observed experience with any of the current-gen LLM tools. No amount of prompt engineering gets you a reliable mapping from input query to output code.
> I designed the end-to-end encryption scheme and told it in detail how to implement it. I pointed out bugs and explained how to fix them. And so on.
I guess my response here is that, if you think that this approach to prompt engineering gets you a generated code result that is in any sense equivalent, or even comparable, in terms of quality, to the work that you could produce yourself, as a professional and senior-level software engineer, then, man, we're on different planets. Pointing out bugs and explaining how to fix them in your prompts in no way gets you deterministic, reliable, accurate, high-quality code as output. And actually forget about high-quality, I mean even just bare minimum table-stakes requirements-satisfying stuff.. !
OK. Replace "[expected] deterministic output" with whatever term best fits what this block of text is describing, as that's what I'm talking about. The claim is that a sufficiently-precisely-specified prompt can produce reliably-correct code. Which is just clearly not the case, as of today.
You are just making assertions here with no evidence.
If you prompt the LLM for code, and then you review the code, identify specific problems, and direct the LLM to fix those problems, and repeat, you can, in fact, end up with production-ready code -- in less time than it would take to write by hand.
Proof: My project. I did this. It worked. It's in production.
It seems like you believe this code is not production-ready because it was produced using an LLM which, you believe, cannot produce production-ready code. This is a cyclic argument.
Out of interest: How much did the entire process take and how much would you estimate it to take without the LLM in the loop?
See again here, you're implying that I or my code is disappointing somehow, but with no explanation for how except that it was LLM-assisted. I assert that the code is basically as good as if I'd written it by hand, and if you think I'm just not a competent engineer, like, feel free to Google me.
It's not the typing itself that constrains, it's the detailed but non-essential decision-making. Every line of code requires making several decisions, like naming variables, deciding basic structure, etc. Many of these fine-grained decisions are obvious or don't matter, but it's still mentally taxing, which is why nobody can write code as fast as they can type even when the code is straightforward. LLMs can basically fill in a bunch of those details for you, and reviewing the decisions -- especially the fine-grained ones that don't matter -- is a lot faster than making them.
> How much did the entire process take and how much would you estimate it to take without the LLM in the loop?
I spent about five days mostly focused on prompting the LLM (although I always have many things interrupting me throughout the day, so I wasn't 100% focused). I estimate it would have taken me 2x-5x as long to do by hand, but it's of course hard to say for sure.
I think you're reading a bit too deeply into what I wrote; I explained what I interpreted kiitos' posts as essentially saying. I realize that you've probably had to deal with a lot of people being skeptical to the point of "throwing shade", as it were, so I understand the defensive posture. I am skeptical, but the reason I'm asking questions (alongside the previous bit) is because I'm actually curious about your experiment.
> It's not the typing itself that constrains, it's the detailed but non-essential decision-making. Every line of code requires making several decisions, like naming variables, deciding basic structure, etc. Many of these fine-grained decisions are obvious or don't matter, but it's still mentally taxing, which is why nobody can write code as fast as they can type even when the code is straightforward. LLMs can basically fill in a bunch of those details for you, and reviewing the decisions -- especially the fine-grained ones that don't matter -- is a lot faster than making them.
In your estimation, what is your mental code coverage of the code you ended up with? Do you feel like you have a complete mapping of it, i.e. you could get an external request for change and map it quickly to where it needs to be made and why exactly there?
I know the code structure about as well as if I had written it.
Honestly the code structure is not very complicated. It flows pretty naturally from the interface spec in the readme, and I'd expect anyone who knows OAuth could find their way around pretty easily.
But yes, as part of prompting improvements to the code, I had to fully understand the implementation. My prompts are entirely based on reading the code and deciding what needed to be changed -- not based on any sort of black-box testing of the code (which would be "vibe coding").
I guess I will concede that this is possible, yes. I've never seen it happen, myself, but it could be the case, at some point, in the future.
> in less time than it would take to write by hand.
This is my point of contention. The process you've described takes ages longer than however much time it would take a competent senior-level engineer to just type the code from first principles. No meaningful project has ever been bottle-necked on how long it takes to type characters into editors.
All of that aside, the claim you're making here is that, speaking as a senior IC, the code that an LLM produces, guided by your prompt inputs, is more or less equivalent to any code that you could produce yourself, even controlling for time spent. Which just doesn't match any of my experiences with any current-gen LLM or agent or workflow or whatever. If your universe is all about glue code, where typing is enemy no. 1, and details don't matter, then fair enough, but please understand that this is not usually the domain of senior-level engineers.
That's been my personal experience over the past 1.5 years. LLMs, prompted and guided by me, write code that I would be proud to produce without them.
I have tried using AI to make changes to the Cloudflare Workers Runtime -- my usual main project, which I started, and know like the back of my hand, and which incidentally handles over a trillion web requests every day -- and in general in that case I haven't found it saved me much time. (Though I've been a bit surprised by the fact that it can find its way around the code at all, it's a pretty complicated C++ codebase.)
It really depends on the use case.
Sometimes I give the agent a vague design ("make XYZ configurable") and it implements it the wrong way, so I'll tell it to do it again with more precise instructions ("use a config file instead of a CLI argument"). The best thing is you can tell it after it wrote 500 lines of code and updated all the tests, and its feelings won't be hurt one bit :)
It can be useful as a research tool too, for instance I was porting a library to a new language, and I told the agent to 1) find all the core types and 2) for each type, run a subtask to compare the implementation in each language and write a markdown file that summarizes the differences with some code samples. 20 min later I had a neat collection of reports that I could refer to while designing the API in the new language.
I guarantee you that Kenton Varda's generators generate more code than any other code generators that aren't compilers. ;)
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
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.
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.
I think there is also a big difference between being forced to use an LLM in a certain way, and being able to completely design your interaction with the LLM yourself. The former I imagine can be indeed tedious and frustrating, the latter is just miraculous.
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.
I think asking questions to the AI for your own understanding totally makes sense, but there is a benefit when you actually create the code versus asking the AI to do it.
"Most of the code one writes in a given environment is rote" is true in the same sense that most of the words one writes in any given bit of text are rote e.g. conjunctions, articles, prepositions, etc.
Straight-line replacement-grade mid code that just does the things a prompt tells it to in the least clever most straightforward way possible is usually a good thing; that long clunky string of modifiers goes by the name "maintainability".
But as of now, it's senior engineers who really know what they 're doing who can spot the errors in AI code.
You need time to accumulate experience. You need experience, time in the proverbial trenches, to be a senior engineer.
You need to be doing different things too, not just implementing the same cookie cutter code repeatedly. If you are doing that, and havent automated it, you are not a senior engineer.
What makes people think that an increase in senior engineer productivity causes demand for junior engineers to decrease?
I think it will have the opposite effect: an increase in senior engineer productivity enables the company to add more features to its products, making it more valuable to its customers, who can therefore afford to pay more for the software. With this increase in revenue, the company is able to hire more junior engineers.
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”.
Maybe journalists and bloggers angling for attention do it, prompt engineers are too aware of the limitations of prompting to do that.
I guess that's where a big miss in understanding so much of the messaging about generative AI in coding happens for me, and why the Fly.io skepticism blog post irritated me so much as well.
It _is_ how collaboration with a person works, but the when you have to fix the issues that the tool created, you aren't collaborating with a person, you're making up for a broken tool.
I can't think of any field where I'd be expected to not only put up with, but also celebrate, a tool that screwed up and required manual intervention so often.
The level of anthropomorphism that occurs in order to advocate on behalf of generative AI use leads to saying things like "it's how collaboration works" here, when I'd never say the same thing about the table saw in my woodshop, or even the relatively smart cruise control on my car.
Generative AI is still just a tool built by people following a design, and which purportedly makes work easier. But when my saw tears out cuts that I have to then sand or recut, or when my car slams on the brakes because it can't understand a bend in the road around a parking lane, I don't shrug and ascribe them human traits and blame myself for being frustrated over how they collaborate with me.
I love the paradigm shift but hate when the hype is uninformed or dishonest or not treating it with an eye for quality.
So every single run will result in different non-reproducible implementation with unique bugs requiring manual expert interventions. How is this better?
To have their minds changed drastically, sure..
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.
If your second paragraph makes that reply irrelevant, are you saying the meaning was "Your use of 'radicalized' is technically correct but I still think you shouldn't have used it here"?
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.
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.
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.
It feels a bit like Halide, where the goal and the strategy are separated so that each can be optimized independently.
Those new paradigms are being discovered by hordes of vibecoders, myself included. I am having wonderful results with TDD and AI assisted design.
IDEs are now mostly browsers for code, and I no longer copy and paste with a chatbot.
Curious what you think about the Oracular paper. One area that I have been working on for the last couple weeks is extracting ToT for the domain and then using the LLM to generate an ensemble of exploration strategies over that tree.
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"
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.
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.
The subtext is pretty obvious, I think: that standards, on message boards, are being set for LLM-generated code that are ludicrously higher than would be set for people-generated code.
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" -- !
I think at this point our respective points have been made, and we can wrap it up here.
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...
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.
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/
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.)
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.
Why is this such a big deal? This library is not even that interesting. It is very straightforward task I expect most programers will be able to pull off easily. 2/3 of the code is type interfaces and comments. The rest is by book implementation of a protocol that is not even that complex.
Please, there are some React JSX files in your code base with a lot more complexities and intricacies than this.
Has anyone even read the code at all?
As you say, the code is not interesting, it deals with a well known topic
And it required lots of man power to get done
tldr: this is a non-event disguised as incredible success. No doubt cloudflare is making money with that AI crap, somehow.
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)
Em-dash baby.
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.
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.
You can certainly select an em dash but most don’t know what it means and don’t use it.
It’s certainly not infallible proof but multiple uses of it in comments online (vs published material or newspapers) are very unusual, so I think it’s an interesting indicator. I completely agree it is common in some texts, usually ones from publishing houses with style guides but also people who know about writing or typography.
I don't see anyone doing that here. LLM writing was brought up because of the writing style, not the dash. It just reinforces the suspicion.
> If humans didn’t use them, they wouldn’t be in the LLM training data.
Humans weren't using them in every context as they are now. They might've been used in books but blog posts and work documents weren't full of them.
It's not a definite thing but it's absolutely a good indicator.
Plenty of people care about typographic punctuation, and others use software (such as Apple’s OSs, markdown converters, publishing and editing tools) which auto-converts smart punctuation. Heck, tools for doing that are older than Markdown, and that is already two decades old.
https://daringfireball.net/projects/smartypants/
Look, nowhere have I said using an em-dash can’t be an indicator, my objection is people using it as the indicator. It’s become a meme. Too many people act like if the existence of a single em-dash immediately and conclusively proves it was written by an LLM. It does not.
That's where I stopped reading. If they needed "AI" for turning their git history into a record of intent ("transparency"), then they had been doing it all wrong, previously. Git commit messages have always been a "form of documentation that bridges human reasoning" -- namely, with another human's (the reader's) reasoning.
If you don't walk your reviewer through your patch, in your commit message, as if you were teaching them, then you're doing it wrong.
Left a bad taste in my mouth.
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"
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.
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!
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.)
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.
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.
Companies are satisfied with the idemnity provided by Microsoft.
They don’t and no one cares
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).
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.
So for direct copies like what you are talking about that would be picked up.
For copying concepts from other libraries, seems like a problem with or without LLM's.
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.
Karpathy already said English is the new programming language.
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.
https://www.trk7.com/blog/ai-agents-for-coding-promise-vs-re...
- 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.
But things like styling and unused code removal have been automated for a long time already, thanks to non-AI tools; assuming that the AI agent has access to those tools (e.g. assuming the agent can trigger a linter), then the engineer could have just included these steps in the prompts instead of running them manually.
EDIT - I still think there are aspects where AI is obviously lacking, I just think those specific examples are not among them
And they probably will be. Looks like prompts have become the new higher-level coding language, the same way JavaScript is a human-friendly abstraction of an existing programming language (like C), which is already a more accessible way to write assembly itself, and the same goes for the underlying binary code... I guess we eventually reached the final step in the development chain, bridging the gap between hardware instructions and human language.
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.
I think this is valuable data, but it is also out of distribution data. Prior to AI models writing code, this won't be present in the training set. Additional training will probably be needed to correlate better results with the new input stream, and also to learn that some of the records would be of its own unreliability and to develop a healthy scepticism of what it has said in the past.
There's a lot of talk about model collapse with models training purely on their own output, or AI slop infecting training data sets, but ultimately it is all data. Combined with a signal to say which bits were ultimately beneficial, it can all be put to use. Even the failures can provide a good counterfactual signal for constrastive learning.
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.
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.
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.
The vast majority of problems in programming aren't novel or interesting.
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.
and the overwhelming majority of humanity will be worse off for it
E = MC^2 + AI
A new equation for physics?
The potential of AI is unlimited.