We're working on a large Rust codebase, heavily assisted development with Claude and Codex, and one critical workflow is after you have written a spec, have the other LLM critique it thoroughly.
This back and forth will take quite a while, but the resulting implementation plan will be 10x better than the original.
You can automate this by giving Codex a goal, and a skill to call Claude to review the implementation spec until they both agree it's done.
Then, for critical code, have them both implement the spec in a worktree, then BOTH critique each other's implementation.
More often than not, Claude will say to take 2 or 3 pieces from it's design over to Codex, but ship the Codex implementation.
I take this idea even further: After the LLMs have critiqued each other, I introduce a third critique and review it myself as a human. This third party review is most effective at highlighting problems that the LLMs miss, in my experience.
Jokes aside, I agree about having LLMs iterate. Bouncing between GPT and Opus is good in my experience, but even having the same LLM review its own output in a new session started fresh without context will surface a lot of problems.
This process takes a lot of tokens and a lot of time, which is find because I’m reviewing and editing everything myself during that time.
> Unless you can somehow provide some arguments against it,
We're year 4 into this discussion and camps have only gotten more bifrucated. There's no 1-1 discussion to have about this as of now, at least not before the crash.
Your only hope in such discourse is not trying to convince the other party how wrong they are, but appealing to an as of yet undecided party. Be it with reason, or simply pointing out how absurd some comments sound to the average person.
> Your only hope in such discourse is not trying to convince the other party how wrong they are
I don't care about convincing anyone, the ones I reply to or others, but if you take the time to leave a comment, at least make it something to read and think about instead of soundbites like "This is astrology for devs", it's plain boring to read and makes HN worse.
This is precisely how I used to use Beads before I made GuardRails (I wanted something slightly simpler, but similar with more 'guard rails'). I braindump everything I want to build, I ask Claude to do market level research. I then ask Claude to ask clarifying questions, when I ask Claude to be critical of its conclusions and provide the top options and to justify it. I also question Claude and say its okay to disagree with me, be critical, I just want to understand.
By the end you have piecemeal "tickets" for your coding agent, if you have multiple developers you can sync them all up into github, and someone could take some locally, or you can just have Claude work on all of them with subagents. The key feature there is because its all piecemeal the context stays per task.
Then I run a /loop 15m If you're currently working ignore this. Start on the next task in gur if you have not. If you finished all work and cannot pass one gate, work on the next available task.
(Note: gur is my shorthand for GuardRails)
I also added a concept called "gates" so a task cannot complete without an attached gate, gates are arbitrary, they can be reused but when assigned to a task those specific assignments are unique per task. A task is basically anything you want it to be: unit test, try building it, or even seek human confirmation. At least when I was using Beads it did not have "gates" but I'm not sure if it has added anything like it since I stopped using Beads.
Claude will ignore the loop if it's currently working, and when its "out of work" it will review all available tasks.
If anyone's curious its MIT Licensed and on GitHub:
I strongly believe you don’t need to call another model for that. The same model can do result fine. Just not as part of the same context.
I mean that if you ask codex on gpt 5.5 to submit to a plan reviewer subagent that uses gpt5.5, this is enough to have a very good reviewing and reassessment of the plan.
My hypothesis is that it’s even better than opus.
The reason why submitting the product of one LLM to another to review is that you need a fresh trajectory. The previous context might have “guided” the planer into some bias. Removing the context is enough to break free from that trajectory and start fresh.
I hate how seriously people take the output of an LLMs or how reliable they think it is.
Have Claude produce that spec 10 times, use the same prompt and same context. Identical requests, but you'll get 10 unique answers that wil contradict each other with each response seeming extermely confident.
Its scary how confident you people are in these outputs.
If you ask 10 different humans to produce the spec with the same information (prompt and context) they will also produce 10 unique answers that will contradict each other and (depending on who you asked) may be just as confident.
There are real decisions to be made when going from a vague prompt to a spec. It's not surprising that an LLM would produce different specs for the same work on different runs. If the prompt already contained answers to all the decision points that come up when writing the spec then the prompt would already be the spec itself.
I didn’t claim that LLMs are people or that they reason.
If the behavior of the llm is the same as the behavior of reasonable people then the behavior of the llm is reasonable, regardless of how black of a box they generate tokens out of.
Reasonable people will generate divergent specs for the same prompt. Thus it is reasonable for an LLM to generate divergent specs out of the same prompt.
Edit: I use “reasonable” here in the legal sense of the “reasonable person” standard, not to imply any reasoning process.
Aren’t people pattern matching neural networks as well? Why does being a token generator mean something is unreliable?
Further, why does that mean “it doesn’t reason”. Logic can be encoded in language, symbols or code. If I say “all apples are red” -> “all fruit in the bowl are apples” -> “therefor all the fruit are red”. It doesn’t really matter if I understand the logic or what red is or fruit/apples are, the logic is contained in the structure of the syntax. If an LLM can output the conclusion reliably from predictive operations it is able to have the effect of reason and we don’t need to know or care about whether it “understands” the reasoning.
They are token predictors that use statistical techniques to emit the randomly weighted next most likely token given the previous token list.
The result is a strange mimic of human reasoning, because the tokens it predicts are trained on strings that were produced by humans that were reasoning, but that's not the same thing.
Human cognition is complex and poorly understood, and the nature of the mind is an area of study almost as old as consciousness itself. We don't know exactly how it works, or what its exact relationship to the brain is, but we do know that it is not a simple token predictor.
LLMs, by their very nature are constrained to the concept of language and the relationship between existing words in a corpus. This is a box they can not escape.
Modern neuroscience suggests that the human brain is much more vast than that, and in many ways looks like it is constrained by language, but certainly not limited to it.
The structure of language encodes logic in many ways. So the models ability to reason may be an emergent property of the reasoning ability humanity has ejected an extracted from our neural networks and abstracted into language a symbols.
> They are token predictors that use statistical techniques to emit the randomly weighted next most likely token given the previous token list.
Sounds like an implementation detail. Now describe how human reasoning works and explain why that process of chemical and electrical signals results in "reasoning" whereas what LLMs do isn't.
The problem with being this reductive is you can do it to anything, including humans. You can’t be reductive about LLMs and refuse to be reductive about humans - that's poor reasoning, and an LLM would out-reason you on this point, further negating your case.
You have moved goalposts from reasoning to "human cognition". I won't tolerate that sort of slippery wordplay.
Reasoning is making analogies between logical patterns found in conceptual space, with a direction of time (statements precede conclusions). For example. A => B and B => C. You may now deduce A => C. For something fuzzier, A~D and B~E, you may now deduce that D~=>E. This is the sort of thing that higher layer attention mechanism is capable of doing.
> This is a box they can not escape.
Would you say that Helen Keller was less capable of abstract reasoning because she had more constrained access to sensory input?
The problem with that is LLMs can output words or symbols that seen like it used "reason" to produce. But for everything the core algorithm does, it's simply nothing like the wetware reasoning to get to the same answer. So he didn't move goalposts. He always meant the reasoning that stems from human cognition.
Technically if it has that, it'd be singularity no? So basically the premise is they are doing nothing of the sort. Prove any LLM enough and it really does show it has no quarrels contradicting itself or being bossed around. Has no belief / no orientation etc. It's truly mindless but tricks our mind and soul (or whatever) probably.
> Technically if it has that, it'd be singularity no?
reasoning is not black and white. It is possible to reason poorly. Most people cannot do basic math proofs, even math majors struggle with the hardest math proofs. Reasoning in humans is also context/token dependent. I just spent one HOUR trying to show my mom (who has mild dementia) how to use amazon fire (push DOWN until your channel shows up, push RIGHT until the channel becomes big) and she could not figure it out. Rewrote the instructions in japanese and she followed the logic relatively smoothly. Ironically, i'm pretty sure her english is better than her japanese, vocabulary wise.
> it's simply nothing like the wetware reasoning to get to the same answer.
but you don't know how wetware reasoning works, so you are incapable of making that proclamation. I'm pretty sure when I do math proofs (I'm not an amazing mathematician) sometimes I have to literally tick my way through each step of the proof, sometimes breaking it down to super-basic substeps, which to me feels awful lot like what an LLM could be doing. For that matter we don't know how LLM reasoning works but my claim is that these LLMs are in principle capable of reasoning due to architecture.
If this doesn't make sense I suggest you look over the architecture of LLMs carefully and try to understand my point.
(BTW I'm not talking about "reasoning models" with "thinking turns", that's just marketing speak, I'm talking about ANY transformer-based model, even the "dumbest UX architecture" completion models)
Reasoning requires cognition, otherwise there's nothing to reason about, no context or value system to use as a basis for reason.
Decision making can be done by trained machines following rules, but that's different that reasoning. A thermostat isn't reasoning when it decides to turn on the air conditioner, to argue otherwise expands the definition of "reason" to be so broad that it becomes useless.
LLMs are trained on human knowledge and reasoning that results from human cognition, and they are excellent at stochastic mimicry - if the argument is that they are actually reasoning, then some sort of equivalent to human cognition must be present for that to be true. Lacking that, they are nothing more than "token extrusion machines" with some potentially useful characteristics.
Why does reasoning require cognition? Isn’t a if else block or switch statement reasoning? Or a formal logic proof? If an LLM produces an output using formal logic or a python script why is that not reasoning? A human would offload the reasoning using similar methods. I know when I took the LSAT, I learned ways to diagram arguments and didn’t have to think/reason about it because the formal logic diagram did the “reasoning for me”.
Aren’t humans just “action potential” extrusion machines? What is unique about our neural pattern recognition to make our cognition different in nature rather than merely degree?
It seems clear at this point that the greatest insight that unlocked our current AI acceleration was scaling alone would unlock emergent properties and abilities.
it's an analogy, it didnt fall on its face at all. it's just a comparison to highlight the point being made was nonsensical. example: you're just a next action generator controlled by trillions of cells and subconscious dna-based behavior. a black box.
moral agency and the ability to learn are implicit in the description you quoted. this isn't some special superpower, all animals have the ability to learn, and many have moral agency. these aren't human specific traits
It appears they don't need to reason or be intelligent to be able to produce working solutions for code. But sure let wild and unmonitored? I wrangle my LLMs like the code monkeys they are. They help materialize code and then you need to sculpt it (and test harness of varying sorts)
It really can be useful. It's very different from old world programming.
An LLM should not "generate specs", a human should. The LLM can work from the specs. It can never infer meaning from a vague prompt. If so, it will start guessing. Every human that ever did functional specification or information analysis at some point knows this. Or has learned the hard way, something with assumptions and asses ;)
The guessing of a LLM for a vague prompt is better than the one of your average developer.
A prompt like "write these two files on disk" will very likely make the LLM do some sort of an atomic write/swap operation, unlike the average developer which will just write the two files and maybe later encounter a race condition bug. You can argue the LLM output is overkill, but it will also be more robust on average.
> It's not surprising that an LLM would produce different specs for the same work on different runs
This is what I don't understand: AI is a computer program with its own data. If we give the same input to that computer program every time, why does it produce different outputs every time? Or does the input include LLM data + our prompt + some random data that computer program picks from its Internet search?
LLMs have a temperature parameter. At zero temperature they are deterministic: they always choose the most likely next token at each step based on what came before and the model weights, and they will always generate the same output given the same input.
As you raise the temperature they will start (pseudo)randomly choosing tokens other than the single most likely token (though that one will still be the most likely to be chosen). It turns out this is almost always better than zero temperature, which has a tendency to get caught in repetitive loops. I imagine all the frontier labs have spent thousands (millions?) of CPU hours tuning the temperature parameters on their models for optimal performance.
> LLMs have a temperature parameter. At zero temperature they are deterministic: they always choose the most likely next token at each step based on what came before and the model weights, and they will always generate the same output given the same input.
"A value proportional to the reciprocal of β is sometimes referred to as the temperature: β = 1/kT, where k is typically 1 or the Boltzmann constant and T is the temperature. A higher temperature results in a more uniform output distribution (i.e. with higher entropy; it is "more random"), while a lower temperature results in a sharper output distribution, with one value dominating."
"Temperature" in the context of softmax does not change a "winning" token, it changes how much probable (in the sense of softmax distribution) winning token will be. If the winning token is "New York", it will be a winner with temperature close to 0 and with temperature of 1e9.
The actual selection of the random token is done separately by using inputs outside of the softmax distribution, for example, by using random number generator. I believe most of LLM configs have a seed for the random number generator.
More than that, generation of code in most programming languages is done with the more guardrails such as beam search guided by schema, syntax and semantics.
The issue is Lllms don't learn, despite the name. A human re-implementing a spec would strive to iterate towards what they feel is a better spec. They can take in their own input and self-correct. The work of implementing the spec gives insight into pain points and strengths, even if they never actually test the spec (they 100% should, but this is to emphasize that struggle for humans is in itself iteration, even before external feedback comes in).
An LLM is isn't deterministic but also isn't iterative without an existing human. You give it the same spec 10 times and it produces 10 results that aren't far off itself but vastly different when you go into the weeds. And not different in a way of improvement. |
So what’s most important is knowing those parameters and the ranges of values, not having the final result. A human, after producing a specs, can the provide the mental model of how he created the specs. Where the inflection points are and what the range of valid results.
What has always mattered is how you decide the specs, not the specs in themselves.
> If you ask 10 different humans to produce the spec with the same information (prompt and context) they will also produce 10 unique answers
But they didn't ask humans, they asked a machine. We expect our machines to behave in predictable ways.
> If the prompt already contained answers to all the decision points that come up when writing the spec then the prompt would already be the spec itself.
This is one of the best arguments against using LLMs I've seen.
It reduces to the classic argument- at the point where you've described a problem and solution in sufficient detail to be confident in the results, you've invented a programming language.
> We expect our machines to behave in predictable ways.
I expect LLMs to produce randomly varying output. Maybe it's the thousands of hours I spent doing monte carlo simulations for my PhD.
> This is one of the best arguments against using LLMs I've seen.
> It reduces to the classic argument- at the point where you've described a problem and solution in sufficient detail to be confident in the results, you've invented a programming language.
I'm not an LLM true believer, but I use codex for various small tasks and it often (not always) does a thoroughly decent job. Yesterday I gave it a pretty vague request to set up a new Home Assistant dashboard and it handled it just fine--I told it what I wanted to see but it figured out itself which helper variables it would need to set up to realize that vision and wrote all the config for it.
I probably could have done it in 15 minutes if I was familiar with Home Assistant's yaml configuration schema and all, but I'm not so it probably would have taken me closer to an hour. Asking codex took me 30 seconds and it did just fine.
I am skeptical that LLM's are going to kill all white collar jobs or whatever anytime soon. Not being able to truly learn things is an issue. Reality has a surprising amount of detail[1], and while codex does well at things like writing Home Assistant configs and setting up a Minecraft server, where there are thousands of examples online of how to do it, when I've asked it to do some more esoteric things it has sometimes failed spectacularly. I don't think having the LLM keep notes and then read them back (filling up the context window) is a real solution here.
But those differences fall within a band of generally accepted results don’t they? And the cost to throw the code away and reimplement is low now. So maybe it doesn’t really matter if the implementation is perfect or identical.
That being said I agree people trust AI too much. Especially people with less experience. It’s easy to forget the models are mirrors of we are as the drivers of the input context not mentors that will guide us to best practices reliably.
>We're working on a large Rust codebase, heavily assisted development with Claude and Codex, and one critical workflow is after you have written a spec, have the other LLM critique it thoroughly.
I do this with other languages, too, not just Rust. Thing is, you have to put a hard stop at some point because the models will always find something to nitpick.
>Testing is the first layer of defense. My system now includes 1,300+ tests — from unit tests to minimal integration tests (e.g., proposer + acceptor only), all the way to multi-replica full integration tests with injected failures. See the project status.
I know LOC is a silly metric, but ~1300 tests for 130k lines averages out to a test per 100 lines - isn't this awfully low for a highly complex piece of code, even discounting the fact that it's vibecoded? 100 LOC can carry a lot of logic for a single test, even for just happy paths.
Considering the domain being distributed systems, and aiming to implement "a Rust-based multi-Paxos consensus engine that not only implements all the features of Azure’s Replicated State Library (RSL)", I don't think we even have to look so deep into it, it's severely lacking tests.
If you're building a distributed system and you don't have more tests and testing code than actual code, by an order of magnitude most likely, then you're missing test coverage.
IIUC only 50k LoC are non-test code, which improves the metric. Whether that's enough tests still depends on the code. If most are getters and setters, the coverage might be ok.
I may have missed it but are those tests written by person or generated? Otherwise how do you know they even test anything (like actually test, not appear to test)
It’s all written by AI and you can’t tell for sure if the tests are good. You can eyeball some but eyeballing 50k lines of code takes a lot of time. You just trust AI and YOLO, find errors later
I'm also shifting to an vibe coding workflow, but I have a genuine question: whenever I use AI for Rust, it makes an insane amount of lifetime errors. I have no idea how people are churning out so many lines of code so quickly.
Honestly, despite all the hype around Rust in the community, the fact that AI can't handle lifetimes reliably makes me reluctant to use it. The AI constantly defaults to spamming .clone() or wrapping things in Rc, completely butchering idiomatic Rust and making the output a pain to work with.
On the other hand, it writes higher-level languages better than I do. For those succeeding with it, how exactly are you configuring or prompting the AI to actually write good, idiomatic Rust
> I'm also shifting to an vibe coding workflow, but I have a genuine question: whenever I use AI for Rust, it makes an insane amount of lifetime errors. I have no idea how people are churning out so many lines of code so quickly.
What harness and model you've been using? For the last few months, essentially since I did the whole "One Human + One Agent = One Browser From Scratch" experiment, I've almost exclusively been doing cross-platform native desktop development with Rust, currently with my own homegrown toolkit basically written from scratch, all with LLMs, mostly with codex.
But I can't remember a single time the agent got stuck on lifetime errors, that's probably the least common issue in regards with agents + Rust I come across. Much bigger issue is the ever-expanding design and LLMs being unable to build proper abstractions that are actually used practically and reduces the amount of code instead of just adding to the hairball.
The issue I'm trying to overcome now is that each change takes longer and longer to make, unless you're really hardcore about pulling back the design/architecture when the LLM goes overboard. I've only succeeded in having ~10 minute edits in +100K LOC codebases in two of the projects I've done so far, probably because I spent most of the time actually defining and thinking of the design myself instead of outsourcing it to the LLM. But this is the biggest issue I'm hitting over and over with agents right now.
Have you split your 100k loc codebases into smaller crates? If you take a look at eg gitoxide's repo, they've split it in many smaller crates. I think that might help with keeping the scope for the ai small and maybe help with keeping contracts tight and well-defined.
Yes, that absolutely helps (and yes, doing that :) ), I'm going even further and basically hard-enforcing a LOC limit per file too, which helps a lot as well.
The complexities LLMs end up putting themselves in is more about the bigger architecture/design of the program, rather than concrete lines, where things end up so tangled that every change requires 10s of changes across the repository, you know, typical "avoid the hairball" stuff you come across in larger applications...
The feedback loop is the interesting part, if you use standard software engineering practices (modularise, test/document your interfaces, etc) then I find things like Claude Code do an exceptional job: since they can actually run cargo check/test themselves and can validate the tests too.
I'm actually vibe coding a game engine right now using a Hexagonal Architecture, and I ran into this exact same issue when trying to synchronize the feedback loop between the viewport and the editor. To be fair, I probably messed up the domain boundaries myself in the first place, but honestly, the AI-generated code wasn't very effective at solving it either
I'm surprised to hear this. I have not had any issues here at all. The AI might clone things but I don't really care/ mind, I can ask it to refactor to make things zero-copy after, which is how I've often written Rust myself. I've never seen it overly wrap things in Rc.
I see the complete opposite. The lower level the language, the less babysit the agent. Pure asm is the best, only with very advanced SIMD flags it has problems. C is excellent.
But python or typescript are full of errors all the time. I rather fallback to perl than python. Perl has been excellent all along.
> Honestly, despite all the hype around Rust in the community, the fact that AI can't handle lifetimes reliably makes me reluctant to use it. The AI constantly defaults to spamming .clone() or wrapping things in Rc, completely butchering idiomatic Rust and making the output a pain to work with.
This hasn't been true since around gpt-4.5 on the OpenAI side of things. The 5.x models have been pretty much solid on Rust for a while now.
> whenever I use AI for Rust, it makes an insane amount of lifetime errors.
What model are you using, and what frameworks are you using?
This is not a hard problem for LLMs to solve.
Rust is nearly the perfect language for LLMs.
It's exceptionally expressive, and it forbids entirely the most common globally complex bugs that LLMs simply do not (and won't for some time) have the context window size to properly reason about.
Dynamically typed languages are a disaster for LLMs because they allow global complexity WRT to implicit type contracts (that they do not and cannot be relied on to withhold).
If you're going to add types, as someone pointed out earlier, why are you even telling an LLM to write Python anyways?
Rust is barely harder to read than Python with types. It's highly expressive.
You have the `&mut` which seems alien, verbose (safe) concurrency, and lifetimes - which - if you're vibe coding... you don't really need to understand that thoroughly.
You want an LLM to write code in a language where "if it complies, it works" - because... let me tell you, if you vibe code in a language where errors are caught at runtime instead of compile time... It will definitely NOT work.
It's not nearly the perfect language for LLMs and Rust is dramatically harder to read and reason about than Python with types. Other options work better for nearly all apps. I found Kotlin works well:
- Garbage collected so no reasoning tokens or dev cycles are wasted on manual memory management. You say if you're vibe coding you can ignore lifetimes, but in response to a post that says AI can't do a good job and constantly uses escape hatches that lose the benefits of Rust (and can easily make it worse, copying data all over the place is terrible for performance).
- Very fast iteration speed due to JIT, a fast compiler and ability to use precompiled libraries. Rust is slow to compile.
- High level code that reads nearly like English.
- Semantically compatible with Java and Java libs, so lots of code in the training set.
- Unit tests are in separate files from sources. Rust intermixes them, bloating the context window with tests that may not be relevant to the current task.
Unless you're writing kernels, or high integrity systems with memory allocation constrains there is very little reason to use a language like Rust.
Everything that people find great on Rust with exception of the borrow checker, can be found in any compiled language from ML linage. And even that is fading away as they introduce a mix of linear types, dependent types, effects and formal logic.
Write a 250k LOC compiler in Python and then get back to me how well LLMs write in Python...
Sure if you want to vibe code a TODO app where it's literally just copying and pasting one it's already seen 10,000 times before, it can do it in Python.
I wrote and maintain this library of skills and workflows called Rust Bucket[0]
It sets up your repo to ensure agents use a workflow which breaks your user requests down into separate beads, works on them serially, runs a judge agent after every bead is complete to apply code quality rules, and also strict static checks of your code. It's really helpful in extracting long, high-quality turns from the agent. It's what we used to build Offload[1].
Yeah, LLMs suck at named lifetimes. The number of times I have seen Claude reach for indices and clones instead of just using proper named lifetimes is too many to count at this point. Not great for high-performance code!
I think it's due to the lack of quality instructions on what is good Rust code; AI often literally doesn't know what idiomatic Rust is. It can be good to have a reference where you write the basic rules that you want it to follow (ideal to assume it has no idea why spamming clone is bad and you're speaking to someone who has just watched one of those youtube videos with a dude in black t-shirt speaking very slowly and going over basic programming concepts as if they're breaking you out of the matrix).
> have hard gates that disallow doing work in the wrong crate
Maybe I'm using agents wrong, but I'm not sure how you'd end up in that situation in the first place? When I start codex, codex literally only has access to the directory I'm launching it, with no way to navigate, read or edit stuff elsewhere on my disk, as it's wrapped in isolation with copied files into it, with no sync between the host.
Hearing that others seemingly let agents have access to their full computer, I feel like I'm vastly out of date about how development happens nowadays, especially when malware and virus lurks around all the package registries.
tsz is an experiment in giving coding agents full control. On my day job I am a lot more careful. But I've moved on from manually approving every change and instead review the final diff. I noticed manually approving was counterproductive.
Well, some of us rather use languages with automatic resource management in the type system, in whatever form, and thus avoid typing .clone() all over the place.
> So .clone() significantly reduces the mental overhead of using rust with a small performance impact? I'm intrigued :)
No, the performance impact will depend on `impl Clone` for the underlying type, the hotness of the code path, and how sensitive to those two variables your code's domain is. It may be extremely expensive.
> Maybe it's harder to reason about the lifetime semantics while also writing code, and works better as a second phase (the de-cloning).
There are cases where assuming `clone` is possible allows for significant architectural and API simplifications at the expense of performance. In those cases, de-cloning will be involved and may produce significant changes.
If I was forced to write it myself, then I'd agree, I'd use Clojure all day before Rust, because it's such a chore to write, edit and read.
The whole "with AI" kind of reduces my hate for Rust though, and increases the appreciation for how strict the language is, especially when the agents themselves does the whole "do change > see error/warning > adjust code > re-check > repeat" loop themselves, which seems to work better the more strict the language is, as far as I can tell.
The "helpful" error messages from Rust can be a bit deceiving though, as the agents first instinct seems to be to always try what the error message recommends, but sometimes the error is just a symptom of a deeper issue, not the actual root issue.
If I was forced to write it myself, i would love to keep writing ruby. What a wonderful language. I dont write ruby anymore, mostly using golang and python.. but ruby still a joy.
Hardly surprising, you give a strong opinion but you don't actually back that up by any arguments, that stuff tends to be downvoted here. Add some proper reasoning and making it clear why you think as you think, and people will stop downvoting :) Also, stop caring about magic internet points, they don't matter and people downvote random stuff sometimes.
However, if I link to gestalt theory of psychology; The Elements of Typographical Style by Robert Bringhurst; and The Primer of Visual literacy by Donis Dondis, folks will undoubtedly NOT read it and still downvote because they have been in Rust code and so have naturally become accustomed to its monstrous appearance. :)
Perhaps I should design a language that is typographically sound—something like brainf*ck haha
Why would the language being typographically ugly matter? Python's pretty, but it hides a lot of functional nuance behind that. Rust is terse, but it's also expressive in its terseness.
If you want to give it a fair shot, it does take some time to get used to, coming from something like Python or Ruby. I won't deny that. I've found that using LSP-assissted semantic syntax highlighting helps, for me, on the typographic front.
I don't think typographic design is a key consideration in most languages' designs, though, and I don't think it should be. The main thing I look for is consistent, relatively predictable rules around the syntax, as far as that layer of language choice goes.
I really don't get the complain about Rust's syntax, it's almost identical to TypeScript's and nobody complains about TypeScript Synthax being ugly …
(Yes, I know the 'a lifetimes are a bit weird, and that's not something that exist in typescript, but that's also not something you use everyday in Rust either.
We MUST get programming languages and LLMs that do not ever change or break comments.
You can’t have contracts defined in comments in code because there’s no guarantee they won’t be deleted or changed.
Even better, we need the ability to embed directives to LLMs which are NOT comments, but a type of programming construct specifically for this purpose.
I am having a different experience than a lot of other commenters here vibe coding with Rust. I am not a Rust programmer or evangelist. I have implemented a drop-in Bash replacement/clone in Rust that passes the upstream Bash test suite and a whole battery of its own. It is a tiny bit faster than Bash itself but consumes a bit more memory. But Codex and Claude both did a great job with it.
I also had it implement a wasm geodesic calculator in Rust and it's amazing and in my use case is better than geodesiclib using the same updated algorithm.
I'm a "C-nile" Rust folks love to hate and did my first hacking in C Deep Blue C on Atari 8-bits. But I'm very impressed with these products and with the ability to leverage some features of Rust with them. (e.g. audit every unsafe instance and define its invariants, etc.)
I also agree with the commenter who said these LLMs are today, at the present moment, good at Go. The only language I notice it seems to be really good above and beyond others at is javascript, I assume because there's so much of it.
Cool post. I don’t fully understand what a code contract is but appreciate the advice. I have settled on a similarly light weight /agile folder when I keep my roadmap.md with epics and sprints.
It's a set of asserts that are a part of the type signature. Requires are asserts on the inputs, ensures are asserts on the outputs.
Depending on your backend you either ignore them, check them all of the time, some of the time, or have SMT-solvers prove that if you uphold the first one all else must follow.
If you're interested in the last one, have a look at Dafny[0]
It's almost guaranteed with agents you could do the same job with less than half of 100k lines. I don't know whats impressive in lines of code generated by agent.
It just an anchor. If it were 50k would you say the same down to 25k? And if so how many more times would it apply?
The interesting thing is that it was manageable solo (in many ways it's _more_ manageable solo+AIs than with coworkers+(their)AIs), and in such a short amount of time.
Original RSL library is 36k LoC. And this is C++. Rust should be like 50% smaller, that is, 18k LoC. This library is so big that I bet the author has no idea if it works or not. 1300 test generated by AI say nothing about actual quality.
In the end it is just a lot of unmaintainable code quickly generated by AI.
This is uncharitable, but makes a prediction. I imagine you'd bet the author won't be successfully using this, at MS/Uber or wherever they are, in a year time?
Rust makes no promise of being terser than C++, and RSL does less than this considering the optimization.
Also it's only 45/50k LOC so not so very from the 36k LOC.
The language may be irrelevant, but the hard guarantees it offers are not. Agents are still very stochastic, they need something deterministic constraining their output.
Also it is kind of interesting that there is so much enthusiasm to use Claude and Claw all over the place, yet lack of vision on how much the whole infrastructure will improve.
Even when it finally bursts and we get into another AI Winter, what was already achieved isn't going away.
The thing that impresses me most is that the author knows everything (from the high level architecture to the small details) of "multi-Paxos consensus engine" (I have no idea what it is, but it must be very complicated) and can write everything out for AI to read (or did he/she use an app to convert speech to text)?
Paxos is certainly non-trivial in the sense that tiny changes can break it, but in terms of functionality it is not that big. 50 KLOC just seems like a lot of code to me.
It is described in the "Code Contracts" section of the article: "Code contracts specify preconditions, postconditions, and invariants for critical functions. These contracts are converted into runtime asserts during testing but can be disabled in production builds for performance". The .NET framework article that he links to: https://learn.microsoft.com/en-us/dotnet/framework/debug-tra...
Is this basically what Dijkstra was saying? I've been thinking how his approach was considered impractical, but may eventually become necessary for security/stability reasons the way things are going. (Seems like new zeroday on HN front page every day now.)
I have Tarpaulin code coverage check and everytime that it drops below the treshold Claude gives up quickly and just lowers the threshold. I don't know how to overcome it. CLAUDE.md neither AGENTS.md help but the LLM always finds its way.
Learnings is irritating to me. The way kids use the word aesthetic is irritating too. I wonder if I might be that old man shaking his fist at the clouds, but I have gotten over begs the question, and literally, so maybe not yet...
A lesson would be a specific learning activity happening at a specific place and time, administered by a person more knowledgeable than you; like a teacher or mentor "giving a lesson".
If you're fine with the generalized form "learned a lesson", then surely "learnings" is fine too. There's no point in trying to police a completely normal and sensible use of language.
Anyway, I accept this usage of the word "lesson", so I also accept "learnings". My point was one of hypocrisy, not policing people in how they can use the word "lesson".
This is great example of AI slop and a big problem with AI coding.
Original RSL library has 36 KLoC across C++ source and headers files. Rust supposed to be more expressive and concise. Yet, AI generated 130k LoCs. I guess nobody understands how this code works and nobody can tell if it actually works.
I've found Rust's safety guarantees to be less useful for slop-generated code because LLMs can always fight their way through the borrow checker by spamming enough Arc<Mutex<Arc<Mutex<...>>>> and clone() everywhere. Rust only gives you safety properties, not liveness. Interior mutability is a fantastic tool for turning safety failures into liveness failures. Remember kids: deadlock is a safe outcome.
It works for humans because when we get a borrow-check failure, we take a step back and think about the global shape of our code and ownership. LLMs path straight to the goal. Problem: code doesn't compile. Solution: more clone()
Yes, but it comes with much better “built-in” guardrails to rein in the autocomplete. Especially if compared to something runtime-surprise-prone-if-lovable like Ruby.
But it does. A whole class of runtime errors you can trivially produce in safe Go — null pointer dereferences, unchecked type assertions, missed enum cases — are unrepresentable in safe Rust. Also, the type system is a lot more expressive, so more invariants can be encoded in it, leading to more business logic bugs being caught at compile time rather than in production.
Rust is about abstractions more than code. You can ask AI to "Optimize/Test/Clarify" but at the end of the day you should be willing to blindly agree to it's output or spend more time reviewing someone else's code.
We're working on a large Rust codebase, heavily assisted development with Claude and Codex, and one critical workflow is after you have written a spec, have the other LLM critique it thoroughly.
This back and forth will take quite a while, but the resulting implementation plan will be 10x better than the original.
You can automate this by giving Codex a goal, and a skill to call Claude to review the implementation spec until they both agree it's done.
Then, for critical code, have them both implement the spec in a worktree, then BOTH critique each other's implementation.
More often than not, Claude will say to take 2 or 3 pieces from it's design over to Codex, but ship the Codex implementation.
I take this idea even further: After the LLMs have critiqued each other, I introduce a third critique and review it myself as a human. This third party review is most effective at highlighting problems that the LLMs miss, in my experience.
Jokes aside, I agree about having LLMs iterate. Bouncing between GPT and Opus is good in my experience, but even having the same LLM review its own output in a new session started fresh without context will surface a lot of problems.
This process takes a lot of tokens and a lot of time, which is find because I’m reviewing and editing everything myself during that time.
This is astrology for devs.
Do they have a golden calf to dance around? Without that success will be hit and miss.
Unless you can somehow provide some arguments against it, I feel like you're the one who is trying to cargo-cult stuff here.
Say what you will with proper reasoning or arguments if you feel compelled, tired reddit-commentary like that helps no one.
> Unless you can somehow provide some arguments against it,
We're year 4 into this discussion and camps have only gotten more bifrucated. There's no 1-1 discussion to have about this as of now, at least not before the crash.
Your only hope in such discourse is not trying to convince the other party how wrong they are, but appealing to an as of yet undecided party. Be it with reason, or simply pointing out how absurd some comments sound to the average person.
> Your only hope in such discourse is not trying to convince the other party how wrong they are
I don't care about convincing anyone, the ones I reply to or others, but if you take the time to leave a comment, at least make it something to read and think about instead of soundbites like "This is astrology for devs", it's plain boring to read and makes HN worse.
>I don't care about convincing anyone
That's fine. Others will care for you.
>it's plain boring to read and makes HN worse.
I chuckled at the joke. Surprising amount of layers to it.
Though I never strove to be a comic nor writer, that kind of terse, compact punch makes me envy those of such literary talent.
This is precisely how I used to use Beads before I made GuardRails (I wanted something slightly simpler, but similar with more 'guard rails'). I braindump everything I want to build, I ask Claude to do market level research. I then ask Claude to ask clarifying questions, when I ask Claude to be critical of its conclusions and provide the top options and to justify it. I also question Claude and say its okay to disagree with me, be critical, I just want to understand.
By the end you have piecemeal "tickets" for your coding agent, if you have multiple developers you can sync them all up into github, and someone could take some locally, or you can just have Claude work on all of them with subagents. The key feature there is because its all piecemeal the context stays per task.
Then I run a /loop 15m If you're currently working ignore this. Start on the next task in gur if you have not. If you finished all work and cannot pass one gate, work on the next available task.
(Note: gur is my shorthand for GuardRails)
I also added a concept called "gates" so a task cannot complete without an attached gate, gates are arbitrary, they can be reused but when assigned to a task those specific assignments are unique per task. A task is basically anything you want it to be: unit test, try building it, or even seek human confirmation. At least when I was using Beads it did not have "gates" but I'm not sure if it has added anything like it since I stopped using Beads.
Claude will ignore the loop if it's currently working, and when its "out of work" it will review all available tasks.
If anyone's curious its MIT Licensed and on GitHub:
https://github.com/Giancarlos/guardrails
I’ll check this out. I might integrate it in to my IDE (www.propelcode.app) as a complement to plan mode.
I strongly believe you don’t need to call another model for that. The same model can do result fine. Just not as part of the same context.
I mean that if you ask codex on gpt 5.5 to submit to a plan reviewer subagent that uses gpt5.5, this is enough to have a very good reviewing and reassessment of the plan.
My hypothesis is that it’s even better than opus.
The reason why submitting the product of one LLM to another to review is that you need a fresh trajectory. The previous context might have “guided” the planer into some bias. Removing the context is enough to break free from that trajectory and start fresh.
I hate how seriously people take the output of an LLMs or how reliable they think it is.
Have Claude produce that spec 10 times, use the same prompt and same context. Identical requests, but you'll get 10 unique answers that wil contradict each other with each response seeming extermely confident.
Its scary how confident you people are in these outputs.
If you ask 10 different humans to produce the spec with the same information (prompt and context) they will also produce 10 unique answers that will contradict each other and (depending on who you asked) may be just as confident.
There are real decisions to be made when going from a vague prompt to a spec. It's not surprising that an LLM would produce different specs for the same work on different runs. If the prompt already contained answers to all the decision points that come up when writing the spec then the prompt would already be the spec itself.
LLMs aren't people. They don't reason. They're token generators, a black box. Your analogy falls on its face with any scrutiny.
I didn’t claim that LLMs are people or that they reason.
If the behavior of the llm is the same as the behavior of reasonable people then the behavior of the llm is reasonable, regardless of how black of a box they generate tokens out of.
Reasonable people will generate divergent specs for the same prompt. Thus it is reasonable for an LLM to generate divergent specs out of the same prompt.
Edit: I use “reasonable” here in the legal sense of the “reasonable person” standard, not to imply any reasoning process.
Aren’t people pattern matching neural networks as well? Why does being a token generator mean something is unreliable?
Further, why does that mean “it doesn’t reason”. Logic can be encoded in language, symbols or code. If I say “all apples are red” -> “all fruit in the bowl are apples” -> “therefor all the fruit are red”. It doesn’t really matter if I understand the logic or what red is or fruit/apples are, the logic is contained in the structure of the syntax. If an LLM can output the conclusion reliably from predictive operations it is able to have the effect of reason and we don’t need to know or care about whether it “understands” the reasoning.
LLMs do reason (they just sometimes don't reason well).
I assure you I've met many devs and "engineers" that reason less than LLMs, and are black boxes, especially in terms of the code they write.
> LLMs do reason
No, they don't.
They are token predictors that use statistical techniques to emit the randomly weighted next most likely token given the previous token list.
The result is a strange mimic of human reasoning, because the tokens it predicts are trained on strings that were produced by humans that were reasoning, but that's not the same thing.
Human cognition is complex and poorly understood, and the nature of the mind is an area of study almost as old as consciousness itself. We don't know exactly how it works, or what its exact relationship to the brain is, but we do know that it is not a simple token predictor.
LLMs, by their very nature are constrained to the concept of language and the relationship between existing words in a corpus. This is a box they can not escape.
Modern neuroscience suggests that the human brain is much more vast than that, and in many ways looks like it is constrained by language, but certainly not limited to it.
The structure of language encodes logic in many ways. So the models ability to reason may be an emergent property of the reasoning ability humanity has ejected an extracted from our neural networks and abstracted into language a symbols.
> They are token predictors that use statistical techniques to emit the randomly weighted next most likely token given the previous token list.
Sounds like an implementation detail. Now describe how human reasoning works and explain why that process of chemical and electrical signals results in "reasoning" whereas what LLMs do isn't.
The problem with being this reductive is you can do it to anything, including humans. You can’t be reductive about LLMs and refuse to be reductive about humans - that's poor reasoning, and an LLM would out-reason you on this point, further negating your case.
You have moved goalposts from reasoning to "human cognition". I won't tolerate that sort of slippery wordplay.
Reasoning is making analogies between logical patterns found in conceptual space, with a direction of time (statements precede conclusions). For example. A => B and B => C. You may now deduce A => C. For something fuzzier, A~D and B~E, you may now deduce that D~=>E. This is the sort of thing that higher layer attention mechanism is capable of doing.
> This is a box they can not escape.
Would you say that Helen Keller was less capable of abstract reasoning because she had more constrained access to sensory input?
The problem with that is LLMs can output words or symbols that seen like it used "reason" to produce. But for everything the core algorithm does, it's simply nothing like the wetware reasoning to get to the same answer. So he didn't move goalposts. He always meant the reasoning that stems from human cognition.
Technically if it has that, it'd be singularity no? So basically the premise is they are doing nothing of the sort. Prove any LLM enough and it really does show it has no quarrels contradicting itself or being bossed around. Has no belief / no orientation etc. It's truly mindless but tricks our mind and soul (or whatever) probably.
> Technically if it has that, it'd be singularity no?
reasoning is not black and white. It is possible to reason poorly. Most people cannot do basic math proofs, even math majors struggle with the hardest math proofs. Reasoning in humans is also context/token dependent. I just spent one HOUR trying to show my mom (who has mild dementia) how to use amazon fire (push DOWN until your channel shows up, push RIGHT until the channel becomes big) and she could not figure it out. Rewrote the instructions in japanese and she followed the logic relatively smoothly. Ironically, i'm pretty sure her english is better than her japanese, vocabulary wise.
> it's simply nothing like the wetware reasoning to get to the same answer.
but you don't know how wetware reasoning works, so you are incapable of making that proclamation. I'm pretty sure when I do math proofs (I'm not an amazing mathematician) sometimes I have to literally tick my way through each step of the proof, sometimes breaking it down to super-basic substeps, which to me feels awful lot like what an LLM could be doing. For that matter we don't know how LLM reasoning works but my claim is that these LLMs are in principle capable of reasoning due to architecture.
If this doesn't make sense I suggest you look over the architecture of LLMs carefully and try to understand my point.
(BTW I'm not talking about "reasoning models" with "thinking turns", that's just marketing speak, I'm talking about ANY transformer-based model, even the "dumbest UX architecture" completion models)
Humans off load reasoning into language and syntax. Chinese encodes arithmetic into the grammar/syntax patterns better than French for example.
Your posts are generally insightful. Thanks for the contribution. Even if it’s a bit cranky and gruff :)
Reasoning requires cognition, otherwise there's nothing to reason about, no context or value system to use as a basis for reason.
Decision making can be done by trained machines following rules, but that's different that reasoning. A thermostat isn't reasoning when it decides to turn on the air conditioner, to argue otherwise expands the definition of "reason" to be so broad that it becomes useless.
LLMs are trained on human knowledge and reasoning that results from human cognition, and they are excellent at stochastic mimicry - if the argument is that they are actually reasoning, then some sort of equivalent to human cognition must be present for that to be true. Lacking that, they are nothing more than "token extrusion machines" with some potentially useful characteristics.
Why does reasoning require cognition? Isn’t a if else block or switch statement reasoning? Or a formal logic proof? If an LLM produces an output using formal logic or a python script why is that not reasoning? A human would offload the reasoning using similar methods. I know when I took the LSAT, I learned ways to diagram arguments and didn’t have to think/reason about it because the formal logic diagram did the “reasoning for me”.
Aren’t humans just “action potential” extrusion machines? What is unique about our neural pattern recognition to make our cognition different in nature rather than merely degree?
It seems clear at this point that the greatest insight that unlocked our current AI acceleration was scaling alone would unlock emergent properties and abilities.
it's an analogy, it didnt fall on its face at all. it's just a comparison to highlight the point being made was nonsensical. example: you're just a next action generator controlled by trillions of cells and subconscious dna-based behavior. a black box.
> you're just a next action generator controlled by trillions of cells and subconscious dna-based behavior.
With moral agency and the ability to learn (even if we presume you are correct, which I don't think you are).
moral agency and the ability to learn are implicit in the description you quoted. this isn't some special superpower, all animals have the ability to learn, and many have moral agency. these aren't human specific traits
Reductio ad absurdum.
exactly my point lol
It appears they don't need to reason or be intelligent to be able to produce working solutions for code. But sure let wild and unmonitored? I wrangle my LLMs like the code monkeys they are. They help materialize code and then you need to sculpt it (and test harness of varying sorts)
It really can be useful. It's very different from old world programming.
They very obviously reason.
it's kind of crazy to think that the transformer architecture can't encode some primitive form of reasoning.
Wow, there are still people trying to claim they don't reason. What will they have to do before you'll admit that they can?
An LLM should not "generate specs", a human should. The LLM can work from the specs. It can never infer meaning from a vague prompt. If so, it will start guessing. Every human that ever did functional specification or information analysis at some point knows this. Or has learned the hard way, something with assumptions and asses ;)
The guessing of a LLM for a vague prompt is better than the one of your average developer.
A prompt like "write these two files on disk" will very likely make the LLM do some sort of an atomic write/swap operation, unlike the average developer which will just write the two files and maybe later encounter a race condition bug. You can argue the LLM output is overkill, but it will also be more robust on average.
What kind of race condition do you have in mind?
> It's not surprising that an LLM would produce different specs for the same work on different runs This is what I don't understand: AI is a computer program with its own data. If we give the same input to that computer program every time, why does it produce different outputs every time? Or does the input include LLM data + our prompt + some random data that computer program picks from its Internet search?
LLMs have a temperature parameter. At zero temperature they are deterministic: they always choose the most likely next token at each step based on what came before and the model weights, and they will always generate the same output given the same input.
As you raise the temperature they will start (pseudo)randomly choosing tokens other than the single most likely token (though that one will still be the most likely to be chosen). It turns out this is almost always better than zero temperature, which has a tendency to get caught in repetitive loops. I imagine all the frontier labs have spent thousands (millions?) of CPU hours tuning the temperature parameters on their models for optimal performance.
"A value proportional to the reciprocal of β is sometimes referred to as the temperature: β = 1/kT, where k is typically 1 or the Boltzmann constant and T is the temperature. A higher temperature results in a more uniform output distribution (i.e. with higher entropy; it is "more random"), while a lower temperature results in a sharper output distribution, with one value dominating."
"Temperature" in the context of softmax does not change a "winning" token, it changes how much probable (in the sense of softmax distribution) winning token will be. If the winning token is "New York", it will be a winner with temperature close to 0 and with temperature of 1e9.
The actual selection of the random token is done separately by using inputs outside of the softmax distribution, for example, by using random number generator. I believe most of LLM configs have a seed for the random number generator.
More than that, generation of code in most programming languages is done with the more guardrails such as beam search guided by schema, syntax and semantics.
Nah. Even with zero temperature this is still variation.
The issue is Lllms don't learn, despite the name. A human re-implementing a spec would strive to iterate towards what they feel is a better spec. They can take in their own input and self-correct. The work of implementing the spec gives insight into pain points and strengths, even if they never actually test the spec (they 100% should, but this is to emphasize that struggle for humans is in itself iteration, even before external feedback comes in).
An LLM is isn't deterministic but also isn't iterative without an existing human. You give it the same spec 10 times and it produces 10 results that aren't far off itself but vastly different when you go into the weeds. And not different in a way of improvement. |
So what’s most important is knowing those parameters and the ranges of values, not having the final result. A human, after producing a specs, can the provide the mental model of how he created the specs. Where the inflection points are and what the range of valid results.
What has always mattered is how you decide the specs, not the specs in themselves.
> If you ask 10 different humans to produce the spec with the same information (prompt and context) they will also produce 10 unique answers
But they didn't ask humans, they asked a machine. We expect our machines to behave in predictable ways.
> If the prompt already contained answers to all the decision points that come up when writing the spec then the prompt would already be the spec itself.
This is one of the best arguments against using LLMs I've seen.
It reduces to the classic argument- at the point where you've described a problem and solution in sufficient detail to be confident in the results, you've invented a programming language.
> We expect our machines to behave in predictable ways.
I expect LLMs to produce randomly varying output. Maybe it's the thousands of hours I spent doing monte carlo simulations for my PhD.
> This is one of the best arguments against using LLMs I've seen.
> It reduces to the classic argument- at the point where you've described a problem and solution in sufficient detail to be confident in the results, you've invented a programming language.
I'm not an LLM true believer, but I use codex for various small tasks and it often (not always) does a thoroughly decent job. Yesterday I gave it a pretty vague request to set up a new Home Assistant dashboard and it handled it just fine--I told it what I wanted to see but it figured out itself which helper variables it would need to set up to realize that vision and wrote all the config for it.
I probably could have done it in 15 minutes if I was familiar with Home Assistant's yaml configuration schema and all, but I'm not so it probably would have taken me closer to an hour. Asking codex took me 30 seconds and it did just fine.
I am skeptical that LLM's are going to kill all white collar jobs or whatever anytime soon. Not being able to truly learn things is an issue. Reality has a surprising amount of detail[1], and while codex does well at things like writing Home Assistant configs and setting up a Minecraft server, where there are thousands of examples online of how to do it, when I've asked it to do some more esoteric things it has sometimes failed spectacularly. I don't think having the LLM keep notes and then read them back (filling up the context window) is a real solution here.
[1] http://johnsalvatier.org/blog/2017/reality-has-a-surprising-...
I haven't made the argument that LLMs aren't useful, I can see cases where they are.
I don't think they include areas where correctness, determinism or human reasoning are important.
At least, not in isolation.
But those differences fall within a band of generally accepted results don’t they? And the cost to throw the code away and reimplement is low now. So maybe it doesn’t really matter if the implementation is perfect or identical.
That being said I agree people trust AI too much. Especially people with less experience. It’s easy to forget the models are mirrors of we are as the drivers of the input context not mentors that will guide us to best practices reliably.
Imagine making this your entire identity
>We're working on a large Rust codebase, heavily assisted development with Claude and Codex, and one critical workflow is after you have written a spec, have the other LLM critique it thoroughly.
I do this with other languages, too, not just Rust. Thing is, you have to put a hard stop at some point because the models will always find something to nitpick.
The return of pair programming.
It's incredible how much developers will do to avoid having to look at or think about code.
What is incredible is that these people have the gall to call themselves developers.
>Testing is the first layer of defense. My system now includes 1,300+ tests — from unit tests to minimal integration tests (e.g., proposer + acceptor only), all the way to multi-replica full integration tests with injected failures. See the project status.
I know LOC is a silly metric, but ~1300 tests for 130k lines averages out to a test per 100 lines - isn't this awfully low for a highly complex piece of code, even discounting the fact that it's vibecoded? 100 LOC can carry a lot of logic for a single test, even for just happy paths.
Considering the domain being distributed systems, and aiming to implement "a Rust-based multi-Paxos consensus engine that not only implements all the features of Azure’s Replicated State Library (RSL)", I don't think we even have to look so deep into it, it's severely lacking tests.
If you're building a distributed system and you don't have more tests and testing code than actual code, by an order of magnitude most likely, then you're missing test coverage.
IIUC only 50k LoC are non-test code, which improves the metric. Whether that's enough tests still depends on the code. If most are getters and setters, the coverage might be ok.
I may have missed it but are those tests written by person or generated? Otherwise how do you know they even test anything (like actually test, not appear to test)
It’s all written by AI and you can’t tell for sure if the tests are good. You can eyeball some but eyeballing 50k lines of code takes a lot of time. You just trust AI and YOLO, find errors later
I'm also shifting to an vibe coding workflow, but I have a genuine question: whenever I use AI for Rust, it makes an insane amount of lifetime errors. I have no idea how people are churning out so many lines of code so quickly.
Honestly, despite all the hype around Rust in the community, the fact that AI can't handle lifetimes reliably makes me reluctant to use it. The AI constantly defaults to spamming .clone() or wrapping things in Rc, completely butchering idiomatic Rust and making the output a pain to work with.
On the other hand, it writes higher-level languages better than I do. For those succeeding with it, how exactly are you configuring or prompting the AI to actually write good, idiomatic Rust
> I'm also shifting to an vibe coding workflow, but I have a genuine question: whenever I use AI for Rust, it makes an insane amount of lifetime errors. I have no idea how people are churning out so many lines of code so quickly.
What harness and model you've been using? For the last few months, essentially since I did the whole "One Human + One Agent = One Browser From Scratch" experiment, I've almost exclusively been doing cross-platform native desktop development with Rust, currently with my own homegrown toolkit basically written from scratch, all with LLMs, mostly with codex.
But I can't remember a single time the agent got stuck on lifetime errors, that's probably the least common issue in regards with agents + Rust I come across. Much bigger issue is the ever-expanding design and LLMs being unable to build proper abstractions that are actually used practically and reduces the amount of code instead of just adding to the hairball.
The issue I'm trying to overcome now is that each change takes longer and longer to make, unless you're really hardcore about pulling back the design/architecture when the LLM goes overboard. I've only succeeded in having ~10 minute edits in +100K LOC codebases in two of the projects I've done so far, probably because I spent most of the time actually defining and thinking of the design myself instead of outsourcing it to the LLM. But this is the biggest issue I'm hitting over and over with agents right now.
Have you split your 100k loc codebases into smaller crates? If you take a look at eg gitoxide's repo, they've split it in many smaller crates. I think that might help with keeping the scope for the ai small and maybe help with keeping contracts tight and well-defined.
Yes, that absolutely helps (and yes, doing that :) ), I'm going even further and basically hard-enforcing a LOC limit per file too, which helps a lot as well.
The complexities LLMs end up putting themselves in is more about the bigger architecture/design of the program, rather than concrete lines, where things end up so tangled that every change requires 10s of changes across the repository, you know, typical "avoid the hairball" stuff you come across in larger applications...
> basically hard-enforcing a LOC limit per file too, which helps a lot as well
this. create pre-commit hooks that enforce project conventions, code quality checks, and regression testing. it saves you so much headache
A lefthook:
format: glob: ".rs" run: cargo fmt -- --check
lint: glob: ".rs" run: cargo clippy -- -D warnings
tests: run: cargo test
audit: run: cargo audit
+ hooks that shove the lefthook automatically in the ai's face
---
rustfmt.toml:
edition = "2021" newline_style = "Unix" use_small_heuristics = "Max" max_width = 100
use "stage_fixed" to automatically persist the formatting :)
Thank you!
The irony of the machines having no mechanical sympathy is just too good
The feedback loop is the interesting part, if you use standard software engineering practices (modularise, test/document your interfaces, etc) then I find things like Claude Code do an exceptional job: since they can actually run cargo check/test themselves and can validate the tests too.
What kinds of programs are you writing and with what models? I'm curious if the lifetimes your programs require are trickier than most.
I'm actually vibe coding a game engine right now using a Hexagonal Architecture, and I ran into this exact same issue when trying to synchronize the feedback loop between the viewport and the editor. To be fair, I probably messed up the domain boundaries myself in the first place, but honestly, the AI-generated code wasn't very effective at solving it either
I'm surprised to hear this. I have not had any issues here at all. The AI might clone things but I don't really care/ mind, I can ask it to refactor to make things zero-copy after, which is how I've often written Rust myself. I've never seen it overly wrap things in Rc.
I've not done any particular/ special prompting.
I see the complete opposite. The lower level the language, the less babysit the agent. Pure asm is the best, only with very advanced SIMD flags it has problems. C is excellent.
But python or typescript are full of errors all the time. I rather fallback to perl than python. Perl has been excellent all along.
I’ve been writing almost exclusively Rust with LLMs and rarely ever hit this. I guess maybe the kind of work you are doing?
> Honestly, despite all the hype around Rust in the community, the fact that AI can't handle lifetimes reliably makes me reluctant to use it. The AI constantly defaults to spamming .clone() or wrapping things in Rc, completely butchering idiomatic Rust and making the output a pain to work with.
This hasn't been true since around gpt-4.5 on the OpenAI side of things. The 5.x models have been pretty much solid on Rust for a while now.
> whenever I use AI for Rust, it makes an insane amount of lifetime errors.
What model are you using, and what frameworks are you using?
This is not a hard problem for LLMs to solve.
Rust is nearly the perfect language for LLMs.
It's exceptionally expressive, and it forbids entirely the most common globally complex bugs that LLMs simply do not (and won't for some time) have the context window size to properly reason about.
Dynamically typed languages are a disaster for LLMs because they allow global complexity WRT to implicit type contracts (that they do not and cannot be relied on to withhold).
If you're going to add types, as someone pointed out earlier, why are you even telling an LLM to write Python anyways?
Rust is barely harder to read than Python with types. It's highly expressive.
You have the `&mut` which seems alien, verbose (safe) concurrency, and lifetimes - which - if you're vibe coding... you don't really need to understand that thoroughly.
You want an LLM to write code in a language where "if it complies, it works" - because... let me tell you, if you vibe code in a language where errors are caught at runtime instead of compile time... It will definitely NOT work.
It's not nearly the perfect language for LLMs and Rust is dramatically harder to read and reason about than Python with types. Other options work better for nearly all apps. I found Kotlin works well:
- Garbage collected so no reasoning tokens or dev cycles are wasted on manual memory management. You say if you're vibe coding you can ignore lifetimes, but in response to a post that says AI can't do a good job and constantly uses escape hatches that lose the benefits of Rust (and can easily make it worse, copying data all over the place is terrible for performance).
- Very fast iteration speed due to JIT, a fast compiler and ability to use precompiled libraries. Rust is slow to compile.
- High level code that reads nearly like English.
- Semantically compatible with Java and Java libs, so lots of code in the training set.
- Unit tests are in separate files from sources. Rust intermixes them, bloating the context window with tests that may not be relevant to the current task.
Then your domain problem you’re trying to solve doesn’t benefit from Rust.
Sounds like your work doesn’t need Rust and that’s ok.
But don’t generalize.
Unless you're writing kernels, or high integrity systems with memory allocation constrains there is very little reason to use a language like Rust.
Everything that people find great on Rust with exception of the borrow checker, can be found in any compiled language from ML linage. And even that is fading away as they introduce a mix of linear types, dependent types, effects and formal logic.
Write a 250k LOC compiler in Python and then get back to me how well LLMs write in Python...
Sure if you want to vibe code a TODO app where it's literally just copying and pasting one it's already seen 10,000 times before, it can do it in Python.
I didn't say it's better to use Python!
This is likely a fully LLM-generated reply.
I wrote and maintain this library of skills and workflows called Rust Bucket[0]
It sets up your repo to ensure agents use a workflow which breaks your user requests down into separate beads, works on them serially, runs a judge agent after every bead is complete to apply code quality rules, and also strict static checks of your code. It's really helpful in extracting long, high-quality turns from the agent. It's what we used to build Offload[1].
0: https://github.com/imbue-ai/rust-bucket : A rusty bucket to carry your slop ;)
1: https://github.com/imbue-ai/offload
rust-bucket is 404, did you make it private?
Thanks for the flag--I guess I must have never made it public.
Fixed.
Yeah, LLMs suck at named lifetimes. The number of times I have seen Claude reach for indices and clones instead of just using proper named lifetimes is too many to count at this point. Not great for high-performance code!
I think it's due to the lack of quality instructions on what is good Rust code; AI often literally doesn't know what idiomatic Rust is. It can be good to have a reference where you write the basic rules that you want it to follow (ideal to assume it has no idea why spamming clone is bad and you're speaking to someone who has just watched one of those youtube videos with a dude in black t-shirt speaking very slowly and going over basic programming concepts as if they're breaking you out of the matrix).
Lots and lots of guardrails to not allow slop.
In tsz I have hard gates that disallow doing work in the wrong crate etc.
https://github.com/mohsen1/tsz
> have hard gates that disallow doing work in the wrong crate
Maybe I'm using agents wrong, but I'm not sure how you'd end up in that situation in the first place? When I start codex, codex literally only has access to the directory I'm launching it, with no way to navigate, read or edit stuff elsewhere on my disk, as it's wrapped in isolation with copied files into it, with no sync between the host.
Hearing that others seemingly let agents have access to their full computer, I feel like I'm vastly out of date about how development happens nowadays, especially when malware and virus lurks around all the package registries.
tsz is an experiment in giving coding agents full control. On my day job I am a lot more careful. But I've moved on from manually approving every change and instead review the final diff. I noticed manually approving was counterproductive.
Right, I'm giving my agents full control too, but not sure why that'd exclude putting them in a sandbox?
Clone is not "butchering idiomatic Rust", we gotta stop this nonsense
Well, some of us rather use languages with automatic resource management in the type system, in whatever form, and thus avoid typing .clone() all over the place.
Sorry, should clarify. .clone() itself isn't inherently unidiomatic when used .
My issue is specifically with how the AI uses it. In AI code, .clone() is almost always used as a brute-force escape hatch
Just like for me as an amateur Rust enjoyer then
So .clone() significantly reduces the mental overhead of using rust with a small performance impact? I'm intrigued :)
Maybe it's harder to reason about the lifetime semantics while also writing code, and works better as a second phase (the de-cloning).
Honestly Rust is an UGLY language. For whatever powers it possesses in memory safety, its cryptic symbology is reminiscent of assembly.
This is a problem when language designers are mathematicians and don’t understand typographical nuance and visual weights.
If I was forced to write it myself, then I'd agree, I'd use Clojure all day before Rust, because it's such a chore to write, edit and read.
The whole "with AI" kind of reduces my hate for Rust though, and increases the appreciation for how strict the language is, especially when the agents themselves does the whole "do change > see error/warning > adjust code > re-check > repeat" loop themselves, which seems to work better the more strict the language is, as far as I can tell.
The "helpful" error messages from Rust can be a bit deceiving though, as the agents first instinct seems to be to always try what the error message recommends, but sometimes the error is just a symptom of a deeper issue, not the actual root issue.
If I was forced to write it myself, i would love to keep writing ruby. What a wonderful language. I dont write ruby anymore, mostly using golang and python.. but ruby still a joy.
It’s funny I got downvoted immediately as expected.
I mean God help us should a crustacean try to understand the merits of my claim.
“Oh he’s saying something negative about rust…” Downvote!
I think with AI the language should still be readable. Humans need to be able to understand what’s going on!
Hardly surprising, you give a strong opinion but you don't actually back that up by any arguments, that stuff tends to be downvoted here. Add some proper reasoning and making it clear why you think as you think, and people will stop downvoting :) Also, stop caring about magic internet points, they don't matter and people downvote random stuff sometimes.
You’re right on both counts.
However, if I link to gestalt theory of psychology; The Elements of Typographical Style by Robert Bringhurst; and The Primer of Visual literacy by Donis Dondis, folks will undoubtedly NOT read it and still downvote because they have been in Rust code and so have naturally become accustomed to its monstrous appearance. :)
Perhaps I should design a language that is typographically sound—something like brainf*ck haha
I would rather have people running for languages like D, but the automatic resource management stigma is high among certain classes of developers.
Why would the language being typographically ugly matter? Python's pretty, but it hides a lot of functional nuance behind that. Rust is terse, but it's also expressive in its terseness.
If you want to give it a fair shot, it does take some time to get used to, coming from something like Python or Ruby. I won't deny that. I've found that using LSP-assissted semantic syntax highlighting helps, for me, on the typographic front.
I don't think typographic design is a key consideration in most languages' designs, though, and I don't think it should be. The main thing I look for is consistent, relatively predictable rules around the syntax, as far as that layer of language choice goes.
I really don't get the complain about Rust's syntax, it's almost identical to TypeScript's and nobody complains about TypeScript Synthax being ugly …
(Yes, I know the 'a lifetimes are a bit weird, and that's not something that exist in typescript, but that's also not something you use everyday in Rust either.
To me it looks clean and concise
I’m curious why? Also I’m curious how long you have programmed in Rust?
How many of those tests have you actually read yourself if all of them are generated by AI (also when you're sleeping) ?
This is from 2025 - I would like to see an update now how that system turned out to be after the vibe hype
I feel like there's very little blogs that actually follow up on their experiment. It's just dopamine city.
To me, the real question after reading this, is: Is your new implementation of Azure’s RSL now being used?
If it is, and it works well, then to me this is far more meaningful than the fact that AI wrote 130K lines of code.
We MUST get programming languages and LLMs that do not ever change or break comments.
You can’t have contracts defined in comments in code because there’s no guarantee they won’t be deleted or changed.
Even better, we need the ability to embed directives to LLMs which are NOT comments, but a type of programming construct specifically for this purpose.
I am having a different experience than a lot of other commenters here vibe coding with Rust. I am not a Rust programmer or evangelist. I have implemented a drop-in Bash replacement/clone in Rust that passes the upstream Bash test suite and a whole battery of its own. It is a tiny bit faster than Bash itself but consumes a bit more memory. But Codex and Claude both did a great job with it.
I also had it implement a wasm geodesic calculator in Rust and it's amazing and in my use case is better than geodesiclib using the same updated algorithm.
I'm a "C-nile" Rust folks love to hate and did my first hacking in C Deep Blue C on Atari 8-bits. But I'm very impressed with these products and with the ability to leverage some features of Rust with them. (e.g. audit every unsafe instance and define its invariants, etc.)
I also agree with the commenter who said these LLMs are today, at the present moment, good at Go. The only language I notice it seems to be really good above and beyond others at is javascript, I assume because there's so much of it.
Cool post. I don’t fully understand what a code contract is but appreciate the advice. I have settled on a similarly light weight /agile folder when I keep my roadmap.md with epics and sprints.
It's a set of asserts that are a part of the type signature. Requires are asserts on the inputs, ensures are asserts on the outputs.
Depending on your backend you either ignore them, check them all of the time, some of the time, or have SMT-solvers prove that if you uphold the first one all else must follow.
If you're interested in the last one, have a look at Dafny[0]
[0] https://dafny.org/
It's almost guaranteed with agents you could do the same job with less than half of 100k lines. I don't know whats impressive in lines of code generated by agent.
It just an anchor. If it were 50k would you say the same down to 25k? And if so how many more times would it apply?
The interesting thing is that it was manageable solo (in many ways it's _more_ manageable solo+AIs than with coworkers+(their)AIs), and in such a short amount of time.
Original RSL library is 36k LoC. And this is C++. Rust should be like 50% smaller, that is, 18k LoC. This library is so big that I bet the author has no idea if it works or not. 1300 test generated by AI say nothing about actual quality.
In the end it is just a lot of unmaintainable code quickly generated by AI.
I see no reason for Rust to be shorter than C++,. when using latest standards.
This is uncharitable, but makes a prediction. I imagine you'd bet the author won't be successfully using this, at MS/Uber or wherever they are, in a year time?
Rust makes no promise of being terser than C++, and RSL does less than this considering the optimization.
Also it's only 45/50k LOC so not so very from the 36k LOC.
Yes, I would bet it won't go anywhere.
The blog post mentioned the project is 130k LoC multiple times. Where 45/50k LoC comes from?
>Rust makes no promise of being terser than C++
True, but Rust has no header files, this alone is a great LoC saver.
50k LOC wouled be the rust code without tests.
But it's not apples to apples because they seem to have done much more performance work though, this is far from code golfing.
Has Rust code generally been found shorter than C++ in practice? I don't see an obvious reason for it.
the interesting thing is how fast it becomes unmanagable.
Also that, I suspect that's correlated to how practical is to have multiple people (with their agents) iterating on it.
> It's almost guaranteed with agents you could do the same job with less than half of 100k lines.
That's great, non-test code is only ~47k lines of code.
For a startup with limited funding, building a product is no more a bottleneck. every one doesn't have the same access to funding!
Contrarian view: Why English will never be a programming language. https://www.slater.dev/2026/05/why-english-will-never-be-a-p...
The moment a language is the output of a natural language compiler, the language itself is kind of irrelevant.
Change the skills, ask the agent to do exactly the same in something else.
I am slowly focusing on agent orchestration tools, which make the actual programming language as relevant as doing SOA with BPEL.
The language may be irrelevant, but the hard guarantees it offers are not. Agents are still very stochastic, they need something deterministic constraining their output.
That is where formalisms come into play.
Also it is kind of interesting that there is so much enthusiasm to use Claude and Claw all over the place, yet lack of vision on how much the whole infrastructure will improve.
Even when it finally bursts and we get into another AI Winter, what was already achieved isn't going away.
The thing that impresses me most is that the author knows everything (from the high level architecture to the small details) of "multi-Paxos consensus engine" (I have no idea what it is, but it must be very complicated) and can write everything out for AI to read (or did he/she use an app to convert speech to text)?
Paxos is certainly non-trivial in the sense that tiny changes can break it, but in terms of functionality it is not that big. 50 KLOC just seems like a lot of code to me.
Is the idea of the runtime contracts similar to the idea of runtime validation? Or are they different in some way?
It is described in the "Code Contracts" section of the article: "Code contracts specify preconditions, postconditions, and invariants for critical functions. These contracts are converted into runtime asserts during testing but can be disabled in production builds for performance". The .NET framework article that he links to: https://learn.microsoft.com/en-us/dotnet/framework/debug-tra...
Is this basically what Dijkstra was saying? I've been thinking how his approach was considered impractical, but may eventually become necessary for security/stability reasons the way things are going. (Seems like new zeroday on HN front page every day now.)
Ah, I missed the reference. Thanks a lot!
I have Tarpaulin code coverage check and everytime that it drops below the treshold Claude gives up quickly and just lowers the threshold. I don't know how to overcome it. CLAUDE.md neither AGENTS.md help but the LLM always finds its way.
Lessons. There's no such thing as learnings.
Learnings is irritating to me. The way kids use the word aesthetic is irritating too. I wonder if I might be that old man shaking his fist at the clouds, but I have gotten over begs the question, and literally, so maybe not yet...
I understand the instinct, but that's a bit too prescriptive for me.
https://en.wiktionary.org/wiki/learnings
A lesson would be a specific learning activity happening at a specific place and time, administered by a person more knowledgeable than you; like a teacher or mentor "giving a lesson".
If you're fine with the generalized form "learned a lesson", then surely "learnings" is fine too. There's no point in trying to police a completely normal and sensible use of language.
So when you cause an incident because you did not pay attention and "learn your lesson" who's the mentor?
The universe.
Anyway, I accept this usage of the word "lesson", so I also accept "learnings". My point was one of hypocrisy, not policing people in how they can use the word "lesson".
This is great example of AI slop and a big problem with AI coding.
Original RSL library has 36 KLoC across C++ source and headers files. Rust supposed to be more expressive and concise. Yet, AI generated 130k LoCs. I guess nobody understands how this code works and nobody can tell if it actually works.
All unit tests can pass if you don't assert anything. Just have to make sure to read through all 130k lines of code to check.
I've found Rust's safety guarantees to be less useful for slop-generated code because LLMs can always fight their way through the borrow checker by spamming enough Arc<Mutex<Arc<Mutex<...>>>> and clone() everywhere. Rust only gives you safety properties, not liveness. Interior mutability is a fantastic tool for turning safety failures into liveness failures. Remember kids: deadlock is a safe outcome.
It works for humans because when we get a borrow-check failure, we take a step back and think about the global shape of our code and ownership. LLMs path straight to the goal. Problem: code doesn't compile. Solution: more clone()
How are you keeping the requirement, design, and tasks docs in sync as the code evolves? I'm curious if anyone's landed on a good workflow for this.
Rust code generation consumes lot of token
Go is much better target, i've observed rails/ruby code is also much easier for AI to spit out.
And Haskell flies with AI
Yes, but it comes with much better “built-in” guardrails to rein in the autocomplete. Especially if compared to something runtime-surprise-prone-if-lovable like Ruby.
This is why I suggested Go.
Rust doesn't add anything over Go for LLM coding.
But it does. A whole class of runtime errors you can trivially produce in safe Go — null pointer dereferences, unchecked type assertions, missed enum cases — are unrepresentable in safe Rust. Also, the type system is a lot more expressive, so more invariants can be encoded in it, leading to more business logic bugs being caught at compile time rather than in production.
For that I would chose neither, rather go with Haskell, OCaml, F#, Lean, Dafny, FStar, Scala, Kotlin,...
Rust is about abstractions more than code. You can ask AI to "Optimize/Test/Clarify" but at the end of the day you should be willing to blindly agree to it's output or spend more time reviewing someone else's code.
Where can we read the code?
Seriously, "Learnings"? Learn you an English.