> I work on Bun and this is my branch
>
> This whole thread is an overreaction. 302 comments about code that does not work. We haven’t committed to rewriting. There’s a very high chance all this code gets thrown out completely.
>
> I’m curious to see what a working version of this looks, what it feels like, how it performs and if/how hard it’d be to get it to pass Bun’s test suite and be maintainable. I’d like to be able to compare a viable Rust version and a Zig version side by side.
cargo check reported over 16,000 compiler errors when I wrote that message. It could not print a version number or run JavaScript. I didn’t expect it to work this quickly and I also didn’t expect the performance to be as competitive. There’ll be a blog post with more details.
Basically we are seeing now an "inverse Hofstadter's Law" where doing something with an LLM takes less time thanexpected even when you take into account this law.
I am a Rust developper myself but I really love Zig and Bun. I am just overly curious of all this.
Rust is really fun to work with and the compiler is great, just make sure the rewrite takes compile times into account since larger projects often have to be organized in a way that makes compilation reasonably fast.
how long does it take to compile?
@jarredsumner: It's basically the same as in zig using our faster zig compiler. If we were using the upstream zig compiler, rust port would compile faster.
> I am so tired of worrying about & spending lots of time fixing memory leaks and crashes and stability issues. it would be so nice if the language provided more powerful tools for preventing these things.
Zig is a middle ground. It solves some of the common foot-guns in C, Without the costs of affine substructural typing that offers Rust its super powers.
I am of the opinion that it is horses for courses and not a universal better proposition.
Because my needs don’t fit in with Rust’s decisions very well I will use zig for personal projects when needed. I just need linked lists, graphs etc…
While hopefully someone can provide a more comprehensive explanation here are the two huge wins for my use case.
1) In Zig, accessing an array or slice out of bounds is considered detectable illegal behavior.
2) defer[0] allows you to collocate the the freeing of resources with code.
That at least ‘feels’ safer to me than a bunch of ‘unsafe’ rust that is required for my very specific use case.
I was working on some eBPF code in C and did really miss zig.
For me it fits the Pareto principle but zig is also just a sometimes food for me, so take that for what it is worth.
Fwiw you don't need unsafe for graphs or linked lists in Rust. At least not directly - these things can be abstracted. The petgraph crate is the most popular for graphs. I'm not sure about linked lists because linked lists are the wrong choice 99.9% of the time.
I've written hundreds of thousands of lines of Rust and outside of FFI, I've written I think one line of unsafe Rust.
zig is unmanaged memory. But rust also allows memory leaks, and they're not uncommon in large, complex programs. So this rewrite will not necessarily control for that.
Honestly, I fully support the rewrite to Rust, but he should have just owned this from the start. I'm sure he knew in the back of his mind how dedicated he was to that branch as he had already spent the equivalent of thousands of dollars in tokens by that point.
That's why I said "the equivalent of". Additionally, time and cognitive effort are not free. The work spent on this branch was work that was not spent on other branches. Does that make sense?
6 days is also nothing when you're doing R&D on your company's dime. He could have spent a month trying a dozen different things and thrown away all the code at the end. As long as he ends that month with a clear picture of where to steer the company over the next 5 years, it's time well spent.
Not even the company is spending money. It’s their employee working on a rework of the code owned by the company that owns the infrastructure on which the rework is done. And that company is still yet to turn profit. This work is subsidised by everyone who pays for Claude.
Announcing the decision a week earlier wouldn't help anyone. Maybe he expected it to work (though he didn't say that), but there's no reason to make a final call before seeing that it did work.
Fair enough. I didn't say anything about a "final call". It just feels like there is a middle ground between that and telling people they are overreacting.
Yeah, but he obviously had enough confidence in this project to keep the agents working at it, didn't he? Given infinite time and money, if you prompt an LLM about something enough times, it will eventually work.
Insert something about monkeys, typewriters, and Shakespeare here.
"No one has the intention of building a wall" - Walter Ulbricht, chairman of the central committee, a couple of months before the Berlin Wall was built.
The AI companies and their associates are beginning to surpass that level of denials and lies.
Four days ago there was no intention to rewrite, now it's a simple desire to refactor. It's not adversarial conclusion, it's pointing out the clear hypocrisy.
Running an experiment, the experiment being more successful than you thought, and then deciding to put more effort into a bigger experiment is not hypocrisy. It’s engineering. If you think some of the objective facts they’re putting out (like test coverage and performance) are lies, go and prove it instead of appealing to emotion.
I don’t think the Zig project adopting a strict ‘no LLM’ policy affects the LLM vendors at all. How many developers are working on the Zig project itself that will (maybe) now not buy a Claude subscription? I can buy that this is a marketing stunt, but nobody at the top cares if a relatively small open source project doesn’t allow AI contributions.
Very impressive that they could do this so quickly because I have been on a similar project (porting TypeScript to Rust) for 5 months. But I guess I don't have access to Mythos and unlimited tokens. I'm also close to 100% pass rate. 99.6% at the time of writing.
Rust is perfect for writing all of code using LLM. It's strict type system makes is less likely to make very dumb mistakes that other languages might allow.
Also want to note that writing the code using LLM doesn't remove the need to have a vision for the design and tradeoffs you make as you build a project. So Jarred and his team are the right kind of people to be able to leverage LLMs to write huge amounts of code.
> Rust is perfect for writing all of code using LLM. It's strict type system makes is less likely to make very dumb mistakes that other languages might allow.
I question this. Yes, strong enforcement of invariants at compile time helps the LLM generate functional code since it gets rapid feedback and retraces as opposed to generating buggy code that fails at runtime in edge cases.
On the other hand, Rust is a complex language prone to refactoring avalanches, where a small change in a component forces refactoring distant code. If the initial architecture is bad or lacking, growing the code base incrementally as LLMs typically do will tend towards spaghettification. So I fear a program that compiles and even runs ok, but no longer human readable or maintainable.
It's very easy to just instruct the LLM to build using isolated crates, to maintain boundaries, focus on "ports and adapters", etc, and not run into this - in my experience.
I haven't had any issues with this getting out of hand on >10KLOC vibed rust codebases.
From the languages that I know, Rust is the only language that I can look at a multi-threaded code and understand it. This stuff being checked by the compiler is a huge advantage
I don't see any reason why the approach wouldn't hold just fine, if not better, as the codebase scaled. Indeed this appears to be exactly what the author has done, they mention that they made heavy use of crates.
> On the other hand, Rust is a complex language prone to refactoring avalanches, where a small change in a component forces refactoring distant code.
Are you saying this out of personal experience or just hypothesizing? I am working on a large, complex rust project with Claude Code and do not experience this at all.
- write sleek operator-overloading-based code for simple mathematical operations on your custom pet algebra
- decide that you want to turn it into an autograd library [0]
- realise that you now need either `RefCell` for interior mutability, or arenas to save the computation graph and local gradients
- realise that `RefCell` puts borrow checks on the runtime path and can panic if you get aliasing wrong
- realise that plain arenas cannot use your sleek operator-overloaded expressions, since `a + b` has no access to the arena, so you need to rewrite them as `tape.sum(node_a, node_b)`
- cry
This was my introduction to why you kinda need to know what you will end up building with Rust, or suffer the cascade refactors. In Python, for example, this issue mostly wouldn't happen, since objects are already reference-like, so the tape/graph can stay implicit and you just chug along.
I still prefer Rust, just that these refactor cascades will happen. But they are mechanically doable, because you just need to 'break' one type, and let an LLM correct the fallout errors surfaced by the compiler till you reach a consistent new ownership model, and I suppose this is common enough that LLM saw it being done hundreds of times, haha.
When Microsoft rewrote it in go, there was a comment from one of the leads that they chose it over rust because of the similarity in paradigms (garbage collection, etc), and that using rust would've been more difficult, requiring a lot of "hoop jumping". Now that you've done it... Thoughts?
Yes indeed. More than 1 million lines of code (including tests) is jumping lots of hoops but with LLMs it's not as painful so you can just ask it to do the hard things.
Example of a Claude Code session after 2 hours of "Crunching" that came out without results
https://github.com/mohsen1/tsz/pull/4868 (Edit I force pushed to PR to solve the problem, you can see the initial refuse message in the initial version of PR description)
Funny thing is, the last percent of the test have been so hard to work on that Opus 4.7 routinely bails and says "it's too involved or complicated" so I had to add prompts specifically asking it not to bail.
You should try GPT, I’d be really interested to hear if it works better. (Exclusively using GPT for systems work at $DAYJOB, but compare with opus every couple weeks and GPT consistently gives me better results)
I've been comparing Claude vs Codex using GPT and Claude consistently is better than GPT about reasoning, about writing code, and using the tools as appropriate.
GPT for instance had a lot of issues using git worktrees, and didn't understand how to correctly use it to then merge stuff back into a main branch, vs Claude which seems to do this much more naturally.
GPT also left me with broken tests/code that I had to iterate on manually, Claude is much better about reasoning through code. Primarily Python.
Using xhigh reasoning and a prompt that encourages it to dive deep and validate everything? I find they trade off strengths and it's often Codex that has a slight edge especially on difficult + long running tasks (Opus gets significantly worse past 500k tokens with or without compaction).
That might be opus 4.7 behaviour because I also get that all the time in the past few weeks. Also complex code base, but likely an order of magnitude simpler than yours.
They mentioned that they wanted to port their compiler over to retain existing behavior (vs a re-write) and Rust has a hard time with their cyclic data structures.
shouldn't typed code that uses functional style be kinda the perfect end game for llms? You can parallelize generation at any granularity, easily ring fence changes, reproduce everything, types give clues to the llm.
Rust is amazing, but the way I want to build Rust software breaks down on large projects with LLMs. Maintaining clean boundaries or even just establishing them stops being a flow state and turns into painful reviews that push me into procrastination mode.
Zig is much more type aligned to bun than typescript. And there’s a common interface of C ffi so you could imagine porting it modularly and keeping the test suite in zig
How do we know it is true? The person in question works for Anthropic. And Zig was on the black list for some time due to its slop skepticism.
It could be another marketing stunt like Mythos, which is so dangerous to release that Antrhopic must be bailed out by the government.
We don't know if the timeline is true, whether Rust experts had a hand in it or even if the reported test suite compliance is true. We are dealing with a company of habitual liars and promoters.
Zig isn’t so much on the blacklist because of the culture it carries from its maintainers, but because the ecosystem is no longer easily composed with other GitHub projects/GitHub Actions.
> why: I am so tired of worrying about & spending lots of time fixing memory leaks and crashes and stability issues. it would be so nice if the language provided more powerful tools for preventing these things.
Not a hard number obviously but a clear indication those issues exist.
Is your claim that using Zig ends in an "extremely high amount of crashes/memory bugs?" Wouldn't that mean that it isn't even feasible to make high-quality software with such a tool? There is a lot of quality stuff made with C/C++, so what is Zig doing wrong?
> Is your claim that using Zig ends in an "extremely high amount of crashes/memory bugs?" Wouldn't that mean that it isn't even feasible to make high-quality software with such a tool?
What caused you to hallucinate such a broad blanket statement? The point is the memory unsafety issues they ran into would be categorically impossible in safe Rust, which is why they're doing this in the first place.
It's not hallucination, it's a basic extrapolation. "Bun has had an extremely high amount of crashes/memory bugs due to them using Zig" is the same statement as "using Zig resulted in Bun having an extremely high amount of crashes/memory bugs". It is then natural to ask whether their position is "using Zig results in an extremely high amount of crashes/bugs" in general.
You're not extrapolating, you're swapping their claim for one that's trivially easy to dispute. "Bun has had an extremely high amount of crashes/memory bugs due to them using Zig" (which unlike Rust, doesn't prevent you from writing them) is a completely different statement than your "using Zig results in an extremely high amount of crashes/bugs." Implying that such a generalization was even on the table is insulting.
Yes, obviously you can write high-quality software in Zig. But does Zig categorically reject the kind of bugs Bun was suffering from? Rust does.
It's generalizing from Bun (which might be especially tricky code) to other software that might not have the similar issues. There are lots of different kinds of software.
The statement “there exists a project where zig led to an extremely high amount of crashes/memory bugs” does not imply “all zig projects have an extremely high amount of crashes/memory bugs”.
This is a classic logic problem - eg “there is an orange cat” doesn’t imply “all cats are orange”.
It is much harder to write quality stuff in c/c++ that doesn't have memory bugs (use after free, out of bounds access, use of unitialized memory, double free, memory races, etc.). I wouldn't say it isn't feasible to build high quality software in those languages, but even the highest quality software written in those languages has these types of bugs. Zig is better than c, and maybe a little bit better than c++, especially with respect to spatial memory bugs, but it doesn't provide the same garantees as rust.
The answer is that C (and by extension Zig, C++) code goes through a hardening process. New code in these languages tends to be unsafe. But bugs and vulnerabilities get squashed over time. Bun gets updated fast and so has a lot of new unsafe code.
There’s a lot of leaky crap written in those languages too. One of the core promises of Rust is that the compiler will catch memory issues other languages won’t experience until runtime. If Zig doesn’t offer something similar it’ll make Rust very compelling.
True. But rust does make it a lot harder to leak memory by accident. Rust variables are automatically freed when they go out of scope. Ownership semantics mean the compiler knows when to free almost everything.
it's feasible to write good software but anything on the scale of millions of lines of code will have memory and pointer issues. I've worked in large C++ code bases with people much more experienced and skilled than I was and every single one of them would tell you that at that scale, no matter how economic and simple you program you will produce memory bugs, the smartest person in the world makes errors holding that much stuff in their head.
They're difficult to find, difficult to reason about in big software and you'll always create some. Languages that rule that out are a huge improvement in terms of correctness.
This is correct but people with too big of an ego or affected too much by Dunning-Kruger) will try to say otherwise even when presented with ample evidence. Instead of a valid response you'll get "skill issue" from people that produce segfaulting code on a regular basis.
It is basically Modula-2 / Object Pascal with C like syntax.
While bounds checking, improved argument passing, typed pointers, proper strings and arrays are an improvement over C, it still suffers from use after free cases.
C++ already prevents many of those scenarios, at least for those folks that don't use it as a plain Better C, and actually make use of the standard library in hardned mode. When not, naturally is as bad as C.
Also to note that the tools that Zig offers to prevent that, are also available in C and C++, but people have to actually use them, e.g. I was using Purify back in 2000's.
Then there is the whole point that Zig is not yet 1.0, and who knows what will still change until then.
Translating a project that includes a good test suite from one language to another is known to be a great case where LLMs work well.
When you’re starting with a complete codebase to use as an example and a test suite to check everything it’s much easier to iterate toward the desired goal. The LLM can already see what the goals are and how they’ve been implemented once already, which is a much easier problem than starting from a spec.
You could have said the same thing about steam power or electricity. And it’s not just an analogy: The magic of these things is in being universal information engines. You spend capital to build them, using well-understood, scalable techniques, plug them into electricity, and out comes value.
My point is, there’s no chance of a “haves and have nots” emerging, any more than electricity turned out that way in the modern world.
Electricity might be a good analogy - but for the other side of this argument.
In the US, (nearly) full electrification wasn't achieved until the late 1940's/early 1950's - a process of nearly a century. (A moment of personal trivia, my great grandfather worked on crews electrifying rural areas of the midwest.)
45 million lines would get to ~$1.125 mil for the linux kernel.
950k lines for Bun would get to $23,750
use whatever math you like ofc.
Does an Anthropic/employee pay that, no. Even if it's at a loss in terms of company revenue, it's worth burning the private capital for all kinds of other reasons.
It costs several times what it would cost a small team of engineers, even assuming you gave the engineers more time to do it. I'm guessing (wildly) this was around 0.5M USD in compute time. You do get the result quicker, though.
It wouldn't take a single person who knew what they were doing even a year to re-implement Bun in basically anything, by hand and from scratch, i.e. not even looking at source. Writing the code for something you already understand and have built before is incredibly fast.
I'm sure they'll market what you said, but it's so ridiculous that I would hope people would see through this stuff.
The saving grace here is a rewrite of a project with a good test suite is the sweet spot: LLMs are great at translation and do great with verifiable goals.
I agree it’s still mind blowing compared to before times, though.
I think a lot of people taking this at face value , a lot of this was possible because of the beyond standard extensive and comprehensive test suit previously built.
It's still an impressive achievement that would have taken even the most competent engineers an exponentially longer time to accomplish.
I just hope it's noted when this is eventually marketed how much human effort went into designing and curating the test suite that even enabled this speed in the first place.
A test suite sort of functions exactly like the ideal scenario for current gen llms. A comprehensive enough test suite essentially forms the spec for agents to implement however they see fit - in this case rust.
You could probably throw away the entire actual source code in certain cases and reimplement the whole thing from scratch just giving an agent access to the tests when it's as well crafted as a project like bun.
I am not sure why people sound so astounded, to be honest. This has been my frank experience of the agentic tools both Codex and Claude since about December.
When given the right constraints this kind of thing is entirely conceivable.
However the important question not being answered here is: does anybody working on it have a full understanding of what has been built?
My experience having constructed similar types of projects using these tools is yes, you could do this in a week or two but now you'll have a month or two of digging through what it made, understanding what was built, and undoing critical yolo leaps of faith it made that you didn't want.
Not to mention to even attempt something like this from scratch would take hundreds of hours if spec work. I see it all day everyday in the aerospace sector. Software engineers have absolutely no idea what deriving a design document and all its associated artifacts actually looks like, and they're in for a rude surprise if the industry really does shift hard that direction
At the very least, it's interesting to be a bystander observering as efforts like this progress. The first thing it makes me wonder is how comprehensive/high quality the test suite is to begin with. Not to cast aspersions, but even at 100% on all platforms I wonder how confident the Bun team would be in migrating.
i think theres a different lesson to be taken from those cases - the LLM will build to what you give a feedback loop for.
if you give just the logical tests, it wont consider the speed at all. if you included tests that measure the speed and ask the llm to match the performance, itll do that too.
its the same class of error as everything else with llms. it has no common sense context for things people consider important. if you dont enforce the boundaries, it will ignore them
I think the industry is moving to English as the programming language, and specifications-context-tdd as the framework for building software.
Many find it distasteful, and many finding liberating. I think it's broadly correlates with how they feel about expressing themselves in english vs say C++.
As a side question, is there anyone who's using LLMs primarily in non-english mode to program? I suspect there's quite a few people using mandarin, but can someone share first-hand account.
I wonder how well Mandarin works for LLM-based programming. On one hand, it's very token efficient as Mandarin script is very dense in meaning. On the other, I suppose this can increase ambiguity.
Character-density and token-efficiency are different things. Latter is data and, therefore, tokenizer specific e.g. take GPT-5's tokenizer o200k_base and run mandarin text and its translation through. Some amount of the time en will beat zh. I just tested with news articles and wikipedia.
After all `def func():` is only 3 tokens on o200k_base.
I can speak, read, and write Taiwanese Mandarin (which is likely relatively underrepresented in the training sets and, which is, in my practical experience, materially different in its usage.)
The authoritative answer for this question would best come from the millions (or tens of millions) of Chinese-speakers who are currently using LLMs to write software.
However, it is my suspicion that you would see no advantages using any language other than English. While there is a certain token-level density to written texts, it seems the benefits of this (and the more recent discussion around “caveman talk”) are quite limited.
Furthermore, consider that the vast majority of textbooks, technical documentation, blog posts, StackOverflow answers, &c. are originally in English. Historically, where these have been translated to Chinese, the translations have often been of very poor quality (and the terminology and phraseology is often incomprehensible unless you also understand some English.) I would suspect that this makes up the overwhelming majority of the training sets for these models.
That said, my experience using the most recent models, is that they are surprisingly language-agnostic in a way that surpasses readily-available human capability. For example, I can prompt the LLM to translate English into something that uses German grammar, Chinese vocabulary, and Japanese characters, and I'll get an output that is worse than what a human expert could do… but where am I going to find a multilingual expert?
(Of course, I have so far only ever been impressed that a model could generate an output but never impressed with the output it did generate. Everything—translations, prose, code—seems universally sloppy and bland and muddy.)
So what I would anticipate the biggest benefit for a Chinese-speaker today… is that if they are disinterested in working internationally, they have significantly less dependency on learning English.
I'm using it 50% English (personal projects)/50% Polish (workplace; reasons being agents.md / team is not that english proficient) and honestly I haven't seen much difference in the output/ambiguity.
Polish prompts tend to be shorter due to the language having a lot of verb forms/conjugations, the only "bad" thing for me is that when it's saying "it broke" it tends to use uncanny / blunt words that make me sometimes laugh.
Interesting. Some questions: Would you say polish is more dense or less dense than english? It's interesting to hear that code quality is not suffering but the response text is sillier or blunter. Any other descrepenacies compared to English?
I would say it certainly can be more dense but even if it's more dense, the tokenizers count it as more. Last time I checked in OpenAI tokenizer for my agents.md it ate 30/40%~ more tokens than the English version at roughly 1:1 meaning.
I think it will eventually be its own dialect of English. Telling LLMs what to do is better using not quite normal English and I think this will continue until it isn't recognizable as natural English anymore, but a new fuzzy programming language (probably >1).
Natural language doesn’t have the precision required for building systems. We already have languages for specifying systems precisely. It’s called “code”…
> 99.8% of bun’s pre-existing test suite passes on Linux x64 glibc in the rust rewrite
OK, they've got a working prototype, congrats! Now it needs to be put into shape so that all the unsafe blocks are eliminated (maybe with a few tiny exceptions), and the code is turned into maintainable, readable, reasonably idiomatic Rust.
I harbor some hope that the (sad) fall of human SWEs will at least be accompanied by language defragmentation. We don't need 38 systems languages once human taste is mostly out of the picture.
So much of the fundamental dynamics of the industry and the job have changed in so little time. Basically over night.
Some days I am so excited at how much I can do now. You can build anything you want, in basically no time! 100% of my software dreams can be a reality.
Some days I am terrified at what's going to happen to the job market.
Suddenly you can get so much with so little. The world only needs so much software.
Is every company that sells software as their core business model going to go out of business?
What will happen if only certain companies or governments get access to the best models?
Certainly companies and governments will have access to better models than the public (in fact, that's already the case with Mythos). The public will still be able to help themselves with models that are behind the frontier.
Obviously there is a huge trend of "rewrite X in Rust". I understand why, Rust is a huge improvement in safety and speed.
My question is, to people even older than me (and I'm certainly not young), does anyone remember this much enthusiasm about people rewriting C code into (C++/Java/Whatever was new and hot)? Because I don't, but maybe I missed it.
No doubt on my side porting was "easy". What I’d find interesting is the ability to maintain and properly care for the code over time for the next iterations.
Do we eventually end up with a codebase that nobody truly understands in depth anymore, where everything is generated and modified through GenAI?
I think the issue is that Zig lost their biggest project, which was a posterboy project for real uses of Zig. Worse, the project felt like Zig wasn't meeting their needs, to the point they abandoned Zig and rewrote their entire project in a different language. Really bad signal for anyone thinking of using Zig for a big project. It is still in beta, but has there been any situation like this, where a upcoming programming language was abandoned by its biggest external project and still was able to be considered a successful language after that?
Presumably the biggest loser in all this is Zig, I only know of the language because of Bun.
But the timescale still gives me pause… just because AI lets us convert a codebase in 6 days doesn’t mean it’s wise. There are surely a lot of downstream implications! It’s always felt a little like Bun is making up a plan as it goes along (and maybe that’s unfair), this seems to underline the point.
Zig is a great low-level language. It's much better than C, while not being so much larger as e.g. Rust or C++. AFAICT Zig does well in embedded development, and should continue to do so. Note that Zig is not even 1.0 yet.
It's been repeated many times that the rejection of the Bun PR was unrelated to their AI-policy. It's also not clear they've "fumbled the ball" given how many projects are complaining about slop PRs.
These tools let you get a massive codebase functional in 6 days. But, presumably, there's no better language to target than Rust (in terms of safety/performance), and therefore the rest of time can be spent making the birthed-in-6-days codebase better.
This is remarkable. Man, there are all those ancient things that "we've lost the source code for". One time, in a past job 10 years ago we were reimplementing something that was lost to the sands of time, using an out of date spec it had used. It was such a tedious job with verification but we got there. Amazing how easy that would be today.
Why not? I think we are perfectly capable on generating a test and validation environment that we can use for correctness. Most likely llms could do this better than engineers with zero to none domain and language knowledge can do these days. From that point on, rewrites would become feasable (not easy, feasable).
Hmm, that's unfortunate - why does so much Rust stuff seem to default to MIT/BSD ? Just because Mozilla used that for most of the Rust stuff ?
Do developers using Rust even know the difference ? Like how anyone can basically take all you work & base a proprietary fork on it with maybe saying "thanks" (attribution) if they feel like it ? :P
I'm very aware when I pick Apache-2. I want attribution for my work, but I don't care about open source purity. I respect closed source software and I put my open source code up for free because I don't care to profit off of my hobbies.
Kinda crazy to use AI to switch from zig to rust in a tool that runs js. Bin bun and use a real lang to begin with. No reason to have that extra layer anymore.
Still, ~1M LOC ported in a work week (400 LOC/min, wtf?) and almost all of it working is pretty wild. I hope the guy managed to maintain normal function, cause I found that getting into the flow but with AI is even more self-consuming and intoxicating than without it, which was already potentially rather rough.
@simonw explains how hilariously misguided that paper is in one of the top comments, and how it doesn't apply remotely to a real agent harness. Plus it's not even clearly relevant here, because the model isn't trying to regurgitate the original document, but generate a new one, and there are guardrails to put it back on track in the form or a compiler and tests. Also, the test suite is very thorough, and pre-existing, and the vast majority passes already. This is skepticism for the sake of it.
This is entirely possible with Claude as it existed even last year.
The LLMs are quite good at re-writes and even better when provided an 'oracle' like a well rounded test suite or existing implementation to work against.
Its part of the reason we keep seeing "I rewrote <library> in <language>" posts on hackernews and when you look at the repo its more like I prompted claude to rewrite this repo in rust or whatever.
> why: I am so tired of worrying about & spending lots of time fixing memory leaks and crashes and stability issues. it would be so nice if the language provided more powerful tools for preventing these things.
As expected, Modula-2 / Objective Pascal like safety was great during the last century, before automatic resource management, and improved type system became common in this century.
Naturally also have to note, wasn't this supposed to be only an experiment, nothing serious?
From 4 days ago: https://news.ycombinator.com/item?id=48019226
cargo check reported over 16,000 compiler errors when I wrote that message. It could not print a version number or run JavaScript. I didn’t expect it to work this quickly and I also didn’t expect the performance to be as competitive. There’ll be a blog post with more details.
That's a post I am eagerly waiting to read.
Basically we are seeing now an "inverse Hofstadter's Law" where doing something with an LLM takes less time thanexpected even when you take into account this law.
I am a Rust developper myself but I really love Zig and Bun. I am just overly curious of all this.
Rust is really fun to work with and the compiler is great, just make sure the rewrite takes compile times into account since larger projects often have to be organized in a way that makes compilation reasonably fast.
> I am so tired of worrying about & spending lots of time fixing memory leaks and crashes and stability issues. it would be so nice if the language provided more powerful tools for preventing these things.
haven't used zig...(only used rust)
but zig doesn't solve those problems?
Zig is a middle ground. It solves some of the common foot-guns in C, Without the costs of affine substructural typing that offers Rust its super powers.
I am of the opinion that it is horses for courses and not a universal better proposition.
Because my needs don’t fit in with Rust’s decisions very well I will use zig for personal projects when needed. I just need linked lists, graphs etc…
While hopefully someone can provide a more comprehensive explanation here are the two huge wins for my use case.
1) In Zig, accessing an array or slice out of bounds is considered detectable illegal behavior.
2) defer[0] allows you to collocate the the freeing of resources with code.
That at least ‘feels’ safer to me than a bunch of ‘unsafe’ rust that is required for my very specific use case.
I was working on some eBPF code in C and did really miss zig.
For me it fits the Pareto principle but zig is also just a sometimes food for me, so take that for what it is worth.
[0] https://zig.guide/language-basics/defer/
Fwiw you don't need unsafe for graphs or linked lists in Rust. At least not directly - these things can be abstracted. The petgraph crate is the most popular for graphs. I'm not sure about linked lists because linked lists are the wrong choice 99.9% of the time.
I've written hundreds of thousands of lines of Rust and outside of FFI, I've written I think one line of unsafe Rust.
zig is unmanaged memory. But rust also allows memory leaks, and they're not uncommon in large, complex programs. So this rewrite will not necessarily control for that.
Nope! Zig is like C in this regard. There’s no borrow checker. Managing memory is your responsibility.
It gives you a few more tools than C - like a debug allocator, bounds checked array slices and so on. But it’s not a memory safe language like rust.
It's not.. but im pretty sure it could be. could probably even take this (WIP) idea and bolt on a formal verifier pretty easily.
https://github.com/ityonemo/clr
Peter Naur: Programming as Theory Building
Bun: Hold my beer
Looks like he did the maintainability performance and test suite checks and made his decision :)
Honestly, I fully support the rewrite to Rust, but he should have just owned this from the start. I'm sure he knew in the back of his mind how dedicated he was to that branch as he had already spent the equivalent of thousands of dollars in tokens by that point.
Bun was VC funded and acquired by Anthropic. He's spending company money, not his own money.
That's why I said "the equivalent of". Additionally, time and cognitive effort are not free. The work spent on this branch was work that was not spent on other branches. Does that make sense?
6 days is also nothing when you're doing R&D on your company's dime. He could have spent a month trying a dozen different things and thrown away all the code at the end. As long as he ends that month with a clear picture of where to steer the company over the next 5 years, it's time well spent.
Had my former employers been so lenient with how I spend company time, I might still be an office worker instead of self-employed!
Not even the company is spending money. It’s their employee working on a rework of the code owned by the company that owns the infrastructure on which the rework is done. And that company is still yet to turn profit. This work is subsidised by everyone who pays for Claude.
Announcing the decision a week earlier wouldn't help anyone. Maybe he expected it to work (though he didn't say that), but there's no reason to make a final call before seeing that it did work.
Fair enough. I didn't say anything about a "final call". It just feels like there is a middle ground between that and telling people they are overreacting.
Yeah but with no guarantee that it was going to work, why should he have?
Yeah, but he obviously had enough confidence in this project to keep the agents working at it, didn't he? Given infinite time and money, if you prompt an LLM about something enough times, it will eventually work.
Insert something about monkeys, typewriters, and Shakespeare here.
He was 2 days into a project that ended up taking 6. You're being extremely unreasonable.
But you didn’t have to sit and type. Assuming that you look at what it did, why not?
Yeah, that means it's an extremely successful experiment so far.
"No one has the intention of building a wall" - Walter Ulbricht, chairman of the central committee, a couple of months before the Berlin Wall was built.
The AI companies and their associates are beginning to surpass that level of denials and lies.
It’s disrespectful to immediately jump to adversarial conclusions from a simple desire to refactor and poor netiquette.
The right to be suspicious of the motives of powerful people is infinitely more important than protecting their feelings from being hurt by suspicion.
Not to mention invoking a major historical event, appeal to emotion move.
Four days ago there was no intention to rewrite, now it's a simple desire to refactor. It's not adversarial conclusion, it's pointing out the clear hypocrisy.
Running an experiment, the experiment being more successful than you thought, and then deciding to put more effort into a bigger experiment is not hypocrisy. It’s engineering. If you think some of the objective facts they’re putting out (like test coverage and performance) are lies, go and prove it instead of appealing to emotion.
you know this whole exercise is both a marketing exercise and a way to make noise.
would the world come to a standstill tomorrow if every Bun instance out there ran on Node.js ?
they know their A.I can't sell without the noise that it's now on the edge of the frontier. this is hype.
zig adopting a strict 'no LLM' policy affects the LLM vendors.
I don’t think the Zig project adopting a strict ‘no LLM’ policy affects the LLM vendors at all. How many developers are working on the Zig project itself that will (maybe) now not buy a Claude subscription? I can buy that this is a marketing stunt, but nobody at the top cares if a relatively small open source project doesn’t allow AI contributions.
What's your point
Very impressive that they could do this so quickly because I have been on a similar project (porting TypeScript to Rust) for 5 months. But I guess I don't have access to Mythos and unlimited tokens. I'm also close to 100% pass rate. 99.6% at the time of writing.
https://tsz.dev
Rust is perfect for writing all of code using LLM. It's strict type system makes is less likely to make very dumb mistakes that other languages might allow.
Also want to note that writing the code using LLM doesn't remove the need to have a vision for the design and tradeoffs you make as you build a project. So Jarred and his team are the right kind of people to be able to leverage LLMs to write huge amounts of code.
> Rust is perfect for writing all of code using LLM. It's strict type system makes is less likely to make very dumb mistakes that other languages might allow.
I question this. Yes, strong enforcement of invariants at compile time helps the LLM generate functional code since it gets rapid feedback and retraces as opposed to generating buggy code that fails at runtime in edge cases.
On the other hand, Rust is a complex language prone to refactoring avalanches, where a small change in a component forces refactoring distant code. If the initial architecture is bad or lacking, growing the code base incrementally as LLMs typically do will tend towards spaghettification. So I fear a program that compiles and even runs ok, but no longer human readable or maintainable.
It's very easy to just instruct the LLM to build using isolated crates, to maintain boundaries, focus on "ports and adapters", etc, and not run into this - in my experience.
I haven't had any issues with this getting out of hand on >10KLOC vibed rust codebases.
From the languages that I know, Rust is the only language that I can look at a multi-threaded code and understand it. This stuff being checked by the compiler is a huge advantage
> I haven't had any issues with this getting out of hand on >10KLOC vibed rust codebases.
This rewrite is >750k lines of Rust
I don't see any reason why the approach wouldn't hold just fine, if not better, as the codebase scaled. Indeed this appears to be exactly what the author has done, they mention that they made heavy use of crates.
> On the other hand, Rust is a complex language prone to refactoring avalanches, where a small change in a component forces refactoring distant code.
Are you saying this out of personal experience or just hypothesizing? I am working on a large, complex rust project with Claude Code and do not experience this at all.
It can happen like this:
- write sleek operator-overloading-based code for simple mathematical operations on your custom pet algebra
- decide that you want to turn it into an autograd library [0]
- realise that you now need either `RefCell` for interior mutability, or arenas to save the computation graph and local gradients
- realise that `RefCell` puts borrow checks on the runtime path and can panic if you get aliasing wrong
- realise that plain arenas cannot use your sleek operator-overloaded expressions, since `a + b` has no access to the arena, so you need to rewrite them as `tape.sum(node_a, node_b)`
- cry
This was my introduction to why you kinda need to know what you will end up building with Rust, or suffer the cascade refactors. In Python, for example, this issue mostly wouldn't happen, since objects are already reference-like, so the tape/graph can stay implicit and you just chug along.
I still prefer Rust, just that these refactor cascades will happen. But they are mechanically doable, because you just need to 'break' one type, and let an LLM correct the fallout errors surfaced by the compiler till you reach a consistent new ownership model, and I suppose this is common enough that LLM saw it being done hundreds of times, haha.
[0] https://github.com/karpathy/micrograd
When Microsoft rewrote it in go, there was a comment from one of the leads that they chose it over rust because of the similarity in paradigms (garbage collection, etc), and that using rust would've been more difficult, requiring a lot of "hoop jumping". Now that you've done it... Thoughts?
Yes indeed. More than 1 million lines of code (including tests) is jumping lots of hoops but with LLMs it's not as painful so you can just ask it to do the hard things.
Example of a Claude Code session after 2 hours of "Crunching" that came out without results https://github.com/mohsen1/tsz/pull/4868 (Edit I force pushed to PR to solve the problem, you can see the initial refuse message in the initial version of PR description)
Funny thing is, the last percent of the test have been so hard to work on that Opus 4.7 routinely bails and says "it's too involved or complicated" so I had to add prompts specifically asking it not to bail.
You should try GPT, I’d be really interested to hear if it works better. (Exclusively using GPT for systems work at $DAYJOB, but compare with opus every couple weeks and GPT consistently gives me better results)
OpenAI gave me that 10x boost and used it all already for this week. I'm guessing the last 50 tests is only doable by GPT 5.5 xhigh
I've been comparing Claude vs Codex using GPT and Claude consistently is better than GPT about reasoning, about writing code, and using the tools as appropriate.
GPT for instance had a lot of issues using git worktrees, and didn't understand how to correctly use it to then merge stuff back into a main branch, vs Claude which seems to do this much more naturally.
GPT also left me with broken tests/code that I had to iterate on manually, Claude is much better about reasoning through code. Primarily Python.
Using xhigh reasoning and a prompt that encourages it to dive deep and validate everything? I find they trade off strengths and it's often Codex that has a slight edge especially on difficult + long running tasks (Opus gets significantly worse past 500k tokens with or without compaction).
Do you have any write ups on your workflow with Claude and github dev?
That might be opus 4.7 behaviour because I also get that all the time in the past few weeks. Also complex code base, but likely an order of magnitude simpler than yours.
Is GC useful for a static type checker? Or did they make a new runtime?
They mentioned that they wanted to port their compiler over to retain existing behavior (vs a re-write) and Rust has a hard time with their cyclic data structures.
shouldn't typed code that uses functional style be kinda the perfect end game for llms? You can parallelize generation at any granularity, easily ring fence changes, reproduce everything, types give clues to the llm.
Rust is amazing, but the way I want to build Rust software breaks down on large projects with LLMs. Maintaining clean boundaries or even just establishing them stops being a flow state and turns into painful reviews that push me into procrastination mode.
Wow, amazing work.
Pretty impressive that it is faster than the Go version already.
Thank you!
It's much faster in single file benchmarks (3 to 5x)
https://tsz.dev/benchmarks/micro
I have optimizations planned for large projects that I'm still flushing out.
Zig is much more type aligned to bun than typescript. And there’s a common interface of C ffi so you could imagine porting it modularly and keeping the test suite in zig
How do we know it is true? The person in question works for Anthropic. And Zig was on the black list for some time due to its slop skepticism.
It could be another marketing stunt like Mythos, which is so dangerous to release that Antrhopic must be bailed out by the government.
We don't know if the timeline is true, whether Rust experts had a hand in it or even if the reported test suite compliance is true. We are dealing with a company of habitual liars and promoters.
> How do we know it is true?
The branch is open.
You can check it out and run the tests if you don’t believe it.
Zig isn’t so much on the blacklist because of the culture it carries from its maintainers, but because the ecosystem is no longer easily composed with other GitHub projects/GitHub Actions.
> We are dealing with a company of habitual liars and promoters.
Any sources to back this up?
I just want to comment that I think it's a good change if we look past the AI involvement.
Bun has had an extremely high amount of crashes/memory bugs due to them using Zig, unlike Deno which is Rust.
Of course, if Bun's Rust port has tons of `unsafe`, it won't magically solve them all, but it'll still get better
> Bun has had an extremely high amount of crashes/memory bugs
Any stats/source? Not that I think it's false
> and the ugly parts look uglier (unsafe) which encourages refactoring.
Looks like Bun owes that to itself to some extent, not solely because of the language
I believe the author is the creator of Bun.
You want a better source than the actual author of Bun?
Authors can't exaggerate? Maybe some actual numbers can convince people.
The cool thing is the author doesn't actually have to convince anyone
Not that it's a particularly accurate stat, but:
https://github.com/oven-sh/bun/issues?q=is%3Aissue%20state%3...
119 open, 885 closed
https://github.com/denoland/deno/issues?q=is%3Aissue%20state...
10 open, 46 closed
FTA:
> why: I am so tired of worrying about & spending lots of time fixing memory leaks and crashes and stability issues. it would be so nice if the language provided more powerful tools for preventing these things.
Not a hard number obviously but a clear indication those issues exist.
I don’t understand: just use an agent to find all memory leaks and segfaults. I don’t get the argument if you are gonna vibe code anyway.
With unlimited tokens make it a lint rule or auto formatter.
If you look at percent of segfault errors in each repo, Bun had a much larger percent. Although don't quote on me.
Is your claim that using Zig ends in an "extremely high amount of crashes/memory bugs?" Wouldn't that mean that it isn't even feasible to make high-quality software with such a tool? There is a lot of quality stuff made with C/C++, so what is Zig doing wrong?
> Is your claim that using Zig ends in an "extremely high amount of crashes/memory bugs?" Wouldn't that mean that it isn't even feasible to make high-quality software with such a tool?
What caused you to hallucinate such a broad blanket statement? The point is the memory unsafety issues they ran into would be categorically impossible in safe Rust, which is why they're doing this in the first place.
It's not hallucination, it's a basic extrapolation. "Bun has had an extremely high amount of crashes/memory bugs due to them using Zig" is the same statement as "using Zig resulted in Bun having an extremely high amount of crashes/memory bugs". It is then natural to ask whether their position is "using Zig results in an extremely high amount of crashes/bugs" in general.
You're not extrapolating, you're swapping their claim for one that's trivially easy to dispute. "Bun has had an extremely high amount of crashes/memory bugs due to them using Zig" (which unlike Rust, doesn't prevent you from writing them) is a completely different statement than your "using Zig results in an extremely high amount of crashes/bugs." Implying that such a generalization was even on the table is insulting.
Yes, obviously you can write high-quality software in Zig. But does Zig categorically reject the kind of bugs Bun was suffering from? Rust does.
The point is that the "extremely high amount of crashes/bugs" is maybe not the fault of Zig after all, as was implied.
How software behaves is very obviously downstream of the tools (in this case programming language) used to build it.
It's generalizing from Bun (which might be especially tricky code) to other software that might not have the similar issues. There are lots of different kinds of software.
The statement “there exists a project where zig led to an extremely high amount of crashes/memory bugs” does not imply “all zig projects have an extremely high amount of crashes/memory bugs”.
This is a classic logic problem - eg “there is an orange cat” doesn’t imply “all cats are orange”.
It is much harder to write quality stuff in c/c++ that doesn't have memory bugs (use after free, out of bounds access, use of unitialized memory, double free, memory races, etc.). I wouldn't say it isn't feasible to build high quality software in those languages, but even the highest quality software written in those languages has these types of bugs. Zig is better than c, and maybe a little bit better than c++, especially with respect to spatial memory bugs, but it doesn't provide the same garantees as rust.
The answer is that C (and by extension Zig, C++) code goes through a hardening process. New code in these languages tends to be unsafe. But bugs and vulnerabilities get squashed over time. Bun gets updated fast and so has a lot of new unsafe code.
> There is a lot of quality stuff made with C/C++
There’s a lot of leaky crap written in those languages too. One of the core promises of Rust is that the compiler will catch memory issues other languages won’t experience until runtime. If Zig doesn’t offer something similar it’ll make Rust very compelling.
rust does not promise leak safety.
True. But rust does make it a lot harder to leak memory by accident. Rust variables are automatically freed when they go out of scope. Ownership semantics mean the compiler knows when to free almost everything.
> Wouldn't that mean that it isn't even feasible to make high-quality software with such a tool?
plenty of other companies/entities making high quality software in zig? tigerbeetle, zig itself for example.
Bun's entire history has been a kind of haphazard move as fast as you can story, so...
it's feasible to write good software but anything on the scale of millions of lines of code will have memory and pointer issues. I've worked in large C++ code bases with people much more experienced and skilled than I was and every single one of them would tell you that at that scale, no matter how economic and simple you program you will produce memory bugs, the smartest person in the world makes errors holding that much stuff in their head.
They're difficult to find, difficult to reason about in big software and you'll always create some. Languages that rule that out are a huge improvement in terms of correctness.
This is correct but people with too big of an ego or affected too much by Dunning-Kruger) will try to say otherwise even when presented with ample evidence. Instead of a valid response you'll get "skill issue" from people that produce segfaulting code on a regular basis.
It is basically Modula-2 / Object Pascal with C like syntax.
While bounds checking, improved argument passing, typed pointers, proper strings and arrays are an improvement over C, it still suffers from use after free cases.
C++ already prevents many of those scenarios, at least for those folks that don't use it as a plain Better C, and actually make use of the standard library in hardned mode. When not, naturally is as bad as C.
Also to note that the tools that Zig offers to prevent that, are also available in C and C++, but people have to actually use them, e.g. I was using Purify back in 2000's.
Then there is the whole point that Zig is not yet 1.0, and who knows what will still change until then.
And they're clearly marked as `unsafe`, so easy to find, which gives them a nice list of issues to address.
6 days of work to do this. Even if it doesn't end up becoming meaningful, it shows just how tokens and work done will be linked now and in the future.
It's going to be hard to compete with someone or a company that has more compute. They will just be able to do things you can't.
Translating a project that includes a good test suite from one language to another is known to be a great case where LLMs work well.
When you’re starting with a complete codebase to use as an example and a test suite to check everything it’s much easier to iterate toward the desired goal. The LLM can already see what the goals are and how they’ve been implemented once already, which is a much easier problem than starting from a spec.
You could have said the same thing about steam power or electricity. And it’s not just an analogy: The magic of these things is in being universal information engines. You spend capital to build them, using well-understood, scalable techniques, plug them into electricity, and out comes value.
My point is, there’s no chance of a “haves and have nots” emerging, any more than electricity turned out that way in the modern world.
Electricity might be a good analogy - but for the other side of this argument.
In the US, (nearly) full electrification wasn't achieved until the late 1940's/early 1950's - a process of nearly a century. (A moment of personal trivia, my great grandfather worked on crews electrifying rural areas of the midwest.)
>My point is, there’s no chance of a “haves and have nots” emerging, any more than electricity turned out that way in the modern world.
Energy costs vary widely across the world and that has enormous capacity for the economies of different countries and their industrial capacity.
https://worldpopulationreview.com/country-rankings/cost-of-e...
Electricity looks pretty even. Higher in Europe but they can afford that.
Due to purchasing power parity, it is actually much hhigher in poorer countries, in that they are absolutely still asking the have nots.
I can't help but wonder what this cost in USD assuming you paid standard rates from Anthropic. Can someone even ballpark the price?
10k lines ~$250 in OpenAI API calls (no plan)
45 million lines would get to ~$1.125 mil for the linux kernel.
950k lines for Bun would get to $23,750
use whatever math you like ofc.
Does an Anthropic/employee pay that, no. Even if it's at a loss in terms of company revenue, it's worth burning the private capital for all kinds of other reasons.
Much less than what it’d costs for a team of rust engineers.
This is both amazing and scary; has been for a while now.
It costs several times what it would cost a small team of engineers, even assuming you gave the engineers more time to do it. I'm guessing (wildly) this was around 0.5M USD in compute time. You do get the result quicker, though.
With less employees....
Isn’t just one guy?
Exactly
This is exactly how Anthropic will market this rewrite towards companies thinking about doing more layoffs.
1 person did a rust rewrite that took 6 days that would have taken hundreds of engineers more than a year to do.
It wouldn't take a single person who knew what they were doing even a year to re-implement Bun in basically anything, by hand and from scratch, i.e. not even looking at source. Writing the code for something you already understand and have built before is incredibly fast.
I'm sure they'll market what you said, but it's so ridiculous that I would hope people would see through this stuff.
> 1 person did a rust rewrite that took 6 days that would have taken hundreds of engineers more than a year to do.
The entire bun team was only about a dozen people and they wrote it from scratch.
It would not take hundreds of engineers to port the existing codebase to another language.
I think this is a cool experiment, but some of these claims are getting absurd.
The saving grace here is a rewrite of a project with a good test suite is the sweet spot: LLMs are great at translation and do great with verifiable goals.
I agree it’s still mind blowing compared to before times, though.
> would have taken hundreds of engineers more than a year
This is estimating what, 10 lines per day each? No way translating code is anywhere near that slow.
> 1 person did a rust rewrite that took 6 days that would have taken hundreds of engineers more than a year to do.
Even cheaper would just be to not do it in the first place. Was there a pressing need to rewrite it?
I think a lot of people taking this at face value , a lot of this was possible because of the beyond standard extensive and comprehensive test suit previously built.
It's still an impressive achievement that would have taken even the most competent engineers an exponentially longer time to accomplish.
I just hope it's noted when this is eventually marketed how much human effort went into designing and curating the test suite that even enabled this speed in the first place.
A test suite sort of functions exactly like the ideal scenario for current gen llms. A comprehensive enough test suite essentially forms the spec for agents to implement however they see fit - in this case rust.
You could probably throw away the entire actual source code in certain cases and reimplement the whole thing from scratch just giving an agent access to the tests when it's as well crafted as a project like bun.
Look what it can do in 6 days!
Ignore the hundreds of thousands of hours put into the original architecture and test suite that made it possible in the first place.
Exactly this.
I am not sure why people sound so astounded, to be honest. This has been my frank experience of the agentic tools both Codex and Claude since about December.
When given the right constraints this kind of thing is entirely conceivable.
However the important question not being answered here is: does anybody working on it have a full understanding of what has been built?
My experience having constructed similar types of projects using these tools is yes, you could do this in a week or two but now you'll have a month or two of digging through what it made, understanding what was built, and undoing critical yolo leaps of faith it made that you didn't want.
Not to mention to even attempt something like this from scratch would take hundreds of hours if spec work. I see it all day everyday in the aerospace sector. Software engineers have absolutely no idea what deriving a design document and all its associated artifacts actually looks like, and they're in for a rude surprise if the industry really does shift hard that direction
At the very least, it's interesting to be a bystander observering as efforts like this progress. The first thing it makes me wonder is how comprehensive/high quality the test suite is to begin with. Not to cast aspersions, but even at 100% on all platforms I wonder how confident the Bun team would be in migrating.
Just a cautionary case of porting to Rust using AI
https://blog.katanaquant.com/p/your-llm-doesnt-write-correct...
i think theres a different lesson to be taken from those cases - the LLM will build to what you give a feedback loop for.
if you give just the logical tests, it wont consider the speed at all. if you included tests that measure the speed and ask the llm to match the performance, itll do that too.
its the same class of error as everything else with llms. it has no common sense context for things people consider important. if you dont enforce the boundaries, it will ignore them
Question is, are our optimization functions well specified enough? (No)
How important is well specified opt function? No one knows. We will find out
Discussed here if anyone's interested:
LLMs work best when the user defines their acceptance criteria first - https://news.ycombinator.com/item?id=47283337 - March 2026 (422 comments)
I think the industry is moving to English as the programming language, and specifications-context-tdd as the framework for building software.
Many find it distasteful, and many finding liberating. I think it's broadly correlates with how they feel about expressing themselves in english vs say C++.
As a side question, is there anyone who's using LLMs primarily in non-english mode to program? I suspect there's quite a few people using mandarin, but can someone share first-hand account.
I use French nearly all the time, it works well. Not that I can't write English prompts, but I find it easier to use my native language.
I wonder how well Mandarin works for LLM-based programming. On one hand, it's very token efficient as Mandarin script is very dense in meaning. On the other, I suppose this can increase ambiguity.
Character-density and token-efficiency are different things. Latter is data and, therefore, tokenizer specific e.g. take GPT-5's tokenizer o200k_base and run mandarin text and its translation through. Some amount of the time en will beat zh. I just tested with news articles and wikipedia.
After all `def func():` is only 3 tokens on o200k_base.
I can speak, read, and write Taiwanese Mandarin (which is likely relatively underrepresented in the training sets and, which is, in my practical experience, materially different in its usage.)
The authoritative answer for this question would best come from the millions (or tens of millions) of Chinese-speakers who are currently using LLMs to write software.
However, it is my suspicion that you would see no advantages using any language other than English. While there is a certain token-level density to written texts, it seems the benefits of this (and the more recent discussion around “caveman talk”) are quite limited.
Furthermore, consider that the vast majority of textbooks, technical documentation, blog posts, StackOverflow answers, &c. are originally in English. Historically, where these have been translated to Chinese, the translations have often been of very poor quality (and the terminology and phraseology is often incomprehensible unless you also understand some English.) I would suspect that this makes up the overwhelming majority of the training sets for these models.
That said, my experience using the most recent models, is that they are surprisingly language-agnostic in a way that surpasses readily-available human capability. For example, I can prompt the LLM to translate English into something that uses German grammar, Chinese vocabulary, and Japanese characters, and I'll get an output that is worse than what a human expert could do… but where am I going to find a multilingual expert?
(Of course, I have so far only ever been impressed that a model could generate an output but never impressed with the output it did generate. Everything—translations, prose, code—seems universally sloppy and bland and muddy.)
So what I would anticipate the biggest benefit for a Chinese-speaker today… is that if they are disinterested in working internationally, they have significantly less dependency on learning English.
I'm using it 50% English (personal projects)/50% Polish (workplace; reasons being agents.md / team is not that english proficient) and honestly I haven't seen much difference in the output/ambiguity.
Polish prompts tend to be shorter due to the language having a lot of verb forms/conjugations, the only "bad" thing for me is that when it's saying "it broke" it tends to use uncanny / blunt words that make me sometimes laugh.
Interesting. Some questions: Would you say polish is more dense or less dense than english? It's interesting to hear that code quality is not suffering but the response text is sillier or blunter. Any other descrepenacies compared to English?
I would say it certainly can be more dense but even if it's more dense, the tokenizers count it as more. Last time I checked in OpenAI tokenizer for my agents.md it ate 30/40%~ more tokens than the English version at roughly 1:1 meaning.
I think it will eventually be its own dialect of English. Telling LLMs what to do is better using not quite normal English and I think this will continue until it isn't recognizable as natural English anymore, but a new fuzzy programming language (probably >1).
>Telling LLMs what to do is better using not quite normal English
What are your prompts like?
I'm teaching my kids to be fluent in tokenese
Natural language doesn’t have the precision required for building systems. We already have languages for specifying systems precisely. It’s called “code”…
I agree, and those are still too focused on code generation for specific languages are fighting the last war.
It is the revenge of UML modeling.
Eventually it will get good enough that what comes out of agent work, is a matter of formal specification.
Assuming that code is actually needed and cannot be achieved as pure agent orchestration workflows.
I'm using it in english / albanian. Not much difference really. Impressive.
> 99.8% of bun’s pre-existing test suite passes on Linux x64 glibc in the rust rewrite
OK, they've got a working prototype, congrats! Now it needs to be put into shape so that all the unsafe blocks are eliminated (maybe with a few tiny exceptions), and the code is turned into maintainable, readable, reasonably idiomatic Rust.
I wonder how long is it going to take.
About 2 months, or 60 days, if we go by the old 90/10 rule.
Not sure that rule is even applicable anymore, but I don't have a better heuristic to make guesses by either.
Bun is going this route because their proposed fix wasn’t great. https://ziggit.dev/t/bun-s-zig-fork-got-4x-faster-compilatio...
Cannot imagine this agent rewrite had anyone review any the code (you can’t at that speed).
I’m positive this will go extremely well :p
I harbor some hope that the (sad) fall of human SWEs will at least be accompanied by language defragmentation. We don't need 38 systems languages once human taste is mostly out of the picture.
What a time to be alive.
So much of the fundamental dynamics of the industry and the job have changed in so little time. Basically over night.
Some days I am so excited at how much I can do now. You can build anything you want, in basically no time! 100% of my software dreams can be a reality.
Some days I am terrified at what's going to happen to the job market.
Suddenly you can get so much with so little. The world only needs so much software.
Is every company that sells software as their core business model going to go out of business?
What will happen if only certain companies or governments get access to the best models?
Certainly companies and governments will have access to better models than the public (in fact, that's already the case with Mythos). The public will still be able to help themselves with models that are behind the frontier.
Obviously there is a huge trend of "rewrite X in Rust". I understand why, Rust is a huge improvement in safety and speed.
My question is, to people even older than me (and I'm certainly not young), does anyone remember this much enthusiasm about people rewriting C code into (C++/Java/Whatever was new and hot)? Because I don't, but maybe I missed it.
If this goes through, it feels like it will stoke rust on zig violence
No doubt on my side porting was "easy". What I’d find interesting is the ability to maintain and properly care for the code over time for the next iterations. Do we eventually end up with a codebase that nobody truly understands in depth anymore, where everything is generated and modified through GenAI?
Thanks for the sharing
The Pareto principle is in play here. It might take years to get that last percentage point.
The branch in question: https://github.com/oven-sh/bun/compare/main...claude/phase-a...
What does this mean for Zig?
Few big popular projects use Zig, if they start to move away from it, what Zig's future will look like?
I think the issue is that Zig lost their biggest project, which was a posterboy project for real uses of Zig. Worse, the project felt like Zig wasn't meeting their needs, to the point they abandoned Zig and rewrote their entire project in a different language. Really bad signal for anyone thinking of using Zig for a big project. It is still in beta, but has there been any situation like this, where a upcoming programming language was abandoned by its biggest external project and still was able to be considered a successful language after that?
It means nothing for Zig. Zig isn't even out of beta yet.
Jarred has already said on Twitter that this was only an experiment for comparisons and very, very unlikely that they'd switch to Rust.
https://news.ycombinator.com/item?id=48077663
Obviously bun having been acquired by Anthropic changes the arithmetic a bit, but I'd love to see the token cost/consumption of this initiative.
And here I am trying to get an LLM to add types to a 100k line Ruby repository for 2 days, and it's not going so hot...
A SMT solver may work better.
Presumably the biggest loser in all this is Zig, I only know of the language because of Bun.
But the timescale still gives me pause… just because AI lets us convert a codebase in 6 days doesn’t mean it’s wise. There are surely a lot of downstream implications! It’s always felt a little like Bun is making up a plan as it goes along (and maybe that’s unfair), this seems to underline the point.
Zig is a great low-level language. It's much better than C, while not being so much larger as e.g. Rust or C++. AFAICT Zig does well in embedded development, and should continue to do so. Note that Zig is not even 1.0 yet.
Yeah but now they got the fame of the language that fumbled the ball because of an overly onerous anti-AI stance.
It's been repeated many times that the rejection of the Bun PR was unrelated to their AI-policy. It's also not clear they've "fumbled the ball" given how many projects are complaining about slop PRs.
Lol. What a goofy take.
But the author said "the code truly works, passing the test suite on Linux and soon other platforms" which just sounds really wise.
These tools let you get a massive codebase functional in 6 days. But, presumably, there's no better language to target than Rust (in terms of safety/performance), and therefore the rest of time can be spent making the birthed-in-6-days codebase better.
This is remarkable. Man, there are all those ancient things that "we've lost the source code for". One time, in a past job 10 years ago we were reimplementing something that was lost to the sands of time, using an out of date spec it had used. It was such a tedious job with verification but we got there. Amazing how easy that would be today.
Serious question… Who’s going to want to run a vibe coded runtime in production?
I don’t see how this is a good look for Bun?
>this is a 960,000 LOC rewrite, the code truly works, passing the test suite on Linux and soon other platforms
I wonder how much of this is original size vs rust requiring verbosity vs the LLM being verbose in general.
Not a criticism, I do believe language translation it's the one field that AI is mature enough to near one shot projects.
would be fun to do zig -> rust -> zig and to measure the delta
(in a VAE-ish way, kl div on the embeddings?)
also feels like a good posttraining task
Good enough for a side project, not good enough for transferring banking system from cobol
That is actually what companies like IBM and Unisys are already doing today, LLM assisted porting.
https://research.ibm.com/publications/enterprise-scale-cobol...
Why not? I think we are perfectly capable on generating a test and validation environment that we can use for correctness. Most likely llms could do this better than engineers with zero to none domain and language knowledge can do these days. From that point on, rewrites would become feasable (not easy, feasable).
Meh. I prefer Java, all hours of the day, every day of the week.
Interesting! I wonder how the performance is compared to the Zig version
What license is this ? Let me guess, its is no GPL...
Unlike the GNU coreutils rewrite in Rust, the Bun rewrite in Rust is being undertaken by the owners of the project.
That said, yes, you’re correct that Bun isn’t GPL: https://github.com/oven-sh/bun?tab=License-1-ov-file
Hmm, that's unfortunate - why does so much Rust stuff seem to default to MIT/BSD ? Just because Mozilla used that for most of the Rust stuff ?
Do developers using Rust even know the difference ? Like how anyone can basically take all you work & base a proprietary fork on it with maybe saying "thanks" (attribution) if they feel like it ? :P
I'm very aware when I pick Apache-2. I want attribution for my work, but I don't care about open source purity. I respect closed source software and I put my open source code up for free because I don't care to profit off of my hobbies.
for the same reason most ruby and javascript/typescript stuff is. Heck, even most python.
Most of them never got into the GPL in the first place.
Your guess is correct! Congrats. Bun itself is not GPL either by the way. Oh, rust compiler itself isn't GPL either.
How many tokens did this port consume?
Kinda crazy to use AI to switch from zig to rust in a tool that runs js. Bin bun and use a real lang to begin with. No reason to have that extra layer anymore.
Bunner
Curious how the test suite was applied. Was it ported from Zig to Rust beforehand?
Almost all of Bun's tests are written in JavaScript run in Bun itself.
> and crashes and stability issues
inb4 .unwrap() / slice / etc hell + livelocks & deadlocks + resource leaks & toctou bugs + larger exposure to supply chain attacks
Still, ~1M LOC ported in a work week (400 LOC/min, wtf?) and almost all of it working is pretty wild. I hope the guy managed to maintain normal function, cause I found that getting into the flow but with AI is even more self-consuming and intoxicating than without it, which was already potentially rather rough.
Deleted
@simonw explains how hilariously misguided that paper is in one of the top comments, and how it doesn't apply remotely to a real agent harness. Plus it's not even clearly relevant here, because the model isn't trying to regurgitate the original document, but generate a new one, and there are guardrails to put it back on track in the form or a compiler and tests. Also, the test suite is very thorough, and pre-existing, and the vast majority passes already. This is skepticism for the sake of it.
Perhaps you can elaborate on how your comment is relevant to the Bun's experiment here.
The fastest large-scale rewrite in the history of software engineering, likely
3 years from now: Linux ported to Rust in 6 days.
And on the seventh day Claude ended His work which He had done, and He rested on the seventh day from all His work which He had done
Being anthropic accuired project does he have access to mythos or it’s normal Claude we plebs have access to
This is entirely possible with Claude as it existed even last year.
The LLMs are quite good at re-writes and even better when provided an 'oracle' like a well rounded test suite or existing implementation to work against.
Its part of the reason we keep seeing "I rewrote <library> in <language>" posts on hackernews and when you look at the repo its more like I prompted claude to rewrite this repo in rust or whatever.
As an Anthropic acquihire, not only does he have access to every model and service but he probably has infinite tokens available.
Bun powers Claude.
Also, isn't it a great ad for Anthropic itself? One wonders
Indeed, knowing the amount of tokens spent would be very interesting.
Explain it for dummies. Isn't Zig a programming language? Why are they re-writting a programming language in another programming language?
They're not rewriting zig. They're rewriting bun, which is currently written in zig
jared's post is singlehandedly shitting on Zig's reputation. not good juju for him to post like that.
"I am so tired of worrying about & spending lots of time fixing memory leaks and crashes and stability issues"
bun was zig's poster child. if it moves away, it becomes yet another random language like nim or crystal.
Ew
> why: I am so tired of worrying about & spending lots of time fixing memory leaks and crashes and stability issues. it would be so nice if the language provided more powerful tools for preventing these things.
As expected, Modula-2 / Objective Pascal like safety was great during the last century, before automatic resource management, and improved type system became common in this century.
Naturally also have to note, wasn't this supposed to be only an experiment, nothing serious?
An update on Bun’s experimental migration from Zig to Rust:
The Rust rewrite now passes 99.8% of Bun’s pre-existing Linux x64 glibc test suite.