> This Rust rewrite would've taken a team of engineers with full-context on the codebase a year of work. With 1 engineer using Fable & closely monitoring Claude Code, we went from start to 100% of the test suite passing on all platforms in 11 days.
This is impressive from a technological standpoint, but it does gloss over the fact that it would have cost $165k in tokens were Bun not part of Anthropic.
The comparison here isn’t completely fair - it would take a small team a year to port it if they spent $0 extra on it.
I’d be interested to see a comparison between spending $165k in 11 days on Claude vs splitting that between 50 people over 11 days for a line-by-line rewrite of the Zig code. I suspect Claude might be faster and therefore cheaper, but maybe not by a lot.
>Combined with the Rust rewrite, ICU changes, and identical code folding, Bun's binary size shrinks by ~20% on Linux & Windows.
People who are surprised by this probably has not seen what Zig code actually looks like. Zig's explicitness and lack of abstraction have a real cost that it is basically one of the most verbose programming languages I've ever seen, it's somehow even more verbose than Go. Basic features of modern languages like pattern matching and generics, and as you can see, having to manually clean up everything means that if you forget once, it's a memory leak. Having SOME abstraction is actually good if it prevents you from making mistakes.
Ironically, Zig is a programming language that's probably best written by LLMs, since they can actually tolerate the verbosity.
Fair point, I phrased that too broadly, and you are right about the loose correlation.
What I was gesturing at, badly, was more that Zig’s low-abstraction / explicit-by-default syntax tends to have you write more boilerplate-y code in general that are more annoying to write and maintain, while not buying you enough over a language with better tooling and ecosystem and compiler optimization like Rust.
Rust's enums can carry data. You can write the same thing in C, but because it does not have the enum feature, you have to do it yourself. They're sometimes called "tagged unions" for a reason, you use a union + a tag when doing it by hand:
I haven't actually compiled this, but it should compile to almost the exact same, if not literally the exact same, machine code. Yet one is way more verbose than the other.
I read my parent ask asking a question: is there a correlation, or not?
I am saying that I do not believe there is a correlation between source code length and binary length. If that's what benced meant by their question, then yes, I agree :)
I’m quite sure there is a certain amount of correlation unfortunately, mainly because there are micro patterns (e.g. IO, allocator) that can’t be modularized into functions. Lots of manual copy-pasta.
I don’t feel the verbosity with Rust. Haven’t written it in a while but now in the LLM era I’m looking forward to saying “sort out the lifetime errors for me”.
Also note that the larger percentages were against already smaller binaries. That smells like there was a single large constant number that got saved somewhere rather than general improvements.
> After that initial shrinkage, the team explored more opportunities for binary size reduction using linker optimizations like Identical Code Folding, removing unused data from ICU, and lazily decompressing small parts of libicu with a zstd dictionary on-demand.
I'd be VERY interested in seeing what the individual effects of those parts were.
Article did a decent job of showing discipline and care and human involvement to assert the automated rewrite was done diligently, as best as it can be when using AI for it. I does make me feel a bit more comfortable about it.
As an aside, I don't know why anyone would not want to use a memory-safe (and possibly race-safe) language in 2026. Rust gives you that in a performant package, so if you are turned off by GCs and immutability for performance reasons, you still have the option to use Rust.
I can understand when you need the absolute best performance and you decide to drop to down to C++, and I also relate with just personal preference, but beyond those it seems a no brainer to me.
I was fairly skeptical about the rewrite when news about it first started going around, and I still don't plan on switching anything to use the Bun rewrite anytime soon, but I appreciate how detailed and well-written the blog post is; it also seems to be primarily human-authored, in my opinion, which is refreshing.
The most significant revelation for me was that Claude Code has been using the rewrite without much fanfare since June 17th.
Without commenting on Bun itself as a project, or the nature of the rewrite, it can't be good for Zig that a naive rewrite away from it fixed memory leaks, improved stability, shrunk binary size by 20%, and improved performance by 5%.
I don't think it's care to categorize this as "a naive rewrite away from [Zig]" - Jarred has been immersed in this project for five years, got to benefit from everything he learned along the way and spent $165,000 of tokens on the most advanced coding LLM anyone has access to.
I expect if he'd spent $165,000 running Fable against the Zig version he could have got a 5% performance improvement, too.
I pay attention when someone makes a hard decision based on a hard-learned lesson. It's like, most who choose to use an ORM just heard of it or want to avoid learning SQL, everyone who removes an ORM learned firsthand horrors.
Did you compare the code before/after? It's a mechanical line-by-line port, and most of the code is identical to the old version, just with Rust syntax. They have an example in the blog post.
People say a lot of things, especially when they have a vested interest in a positive outcome. Bun has been fully vibe coded into another language. There’s no way in hell it’s maintainable. Go read any analysis of the Claude Code leak for proof.
Sure, reasonably well at first glance, but to quote the article:
> I rewrote Bun in Rust using about 50 dynamic workflows in Claude Code run continuously over the course of 11 days.
> Excluding comments, Bun is 535,496 lines of Zig.
> How do you review a PR with +1 million lines added? How do you start to build the confidence needed to responsibly merge large quantities of LLM-authored code? A language-independent test suite with a million assertions, adversarial code review and when something does go wrong, fixing the process that generates the code instead of hand-fixing the code.
That’s vibe coding. This blog post is an ad for Claude, nothing more.
The same concern applies to every GC language, so it's not necessarily bad for Zig. Bun can have been grown too large for Zig to be effective, while moderately sized projects may still greatly benefit from Zig.
Oh, yeah that might be confusing. I meant "you can say the same thing for GC language if that's true, which isn't necessarily true, so that must be false".
More precisely speaking: GC languages are said to delay memory problems far beyond the horizon, which is often unreachable throughout the project's history. Zig can be a similar case.
True, but rewrites often allow for this sort of benefit in themselves. It's possible rewriting it in zig would have yielded some of the same improvements.
Maybe they'd get the same numeric improvements and bug fixes today (or maybe not, or maybe they'd get even more since the LLM isn't spending time rewriting correct code).
But they wouldn't get a change to the structural issues that created the issues in the first place. They'd end up "ke[eping] fixing these kinds of bugs one-off in perpetuity".
But how would you verify that the agents have written memory safe code? Rust's borrowchecker is a lot faster and actually verifiably safe compared to asking an LLM to fix the safety issues that the Zig version had.
The first scenario was joining a company where a software product barely worked. We did the traditional incremental refactoring / rewriting, but eventually learned how rotten the core was that rewriting from first principles was the best path forward.
The lesson learned here is that the conventional wisdom probably only applies to rewriting complex but working systems.
Then multiple scenarios in the agentic coding age. Between day jobs and hobbies I've reproduced major chunks of complicated software like Salesforce, Gmail, Pioneer Rekordbox with very lean teams.
Much like the blog post, the trick is to get an excellent verification loop with a compiler, linter, and test harness / test suite around the core behaviors.
It's feeling more and more that designing and implementing comprehensive test harnesses is the real work, once you have that let the LLM cook.
"I used a pre-release version of Claude Fable 5 for much of the Rust rewrite."
It'd be interesting if Anthropic became a general software company just because they have access to models that aren't yet released, possibly export-banned.
That's the power of a strong test suite. LLMs excel when you have verifiable rewards. I imagine we'll get a lot more rewritten in rust projects in the future. Rust is also an ideal target for such rewrites as it offers a lot of verification (via its type system) and is low overhead with zero-gc. There's less and less reason to use GC'd languages in the agentic coding era.
I think Rust is a locally optimal target for LLM coding, we might see a better language in the future, but I think Rust will dominate for quite some time.
> There's less and less reason to use GC'd languages in the agentic coding era.
Faster iteration, maybe? Rust's safety guarantee isn't exactly free (while still being very excellent) and does affect iteration time. I have a private project (>300K LoC) that has been translated from Python to TypeScript and the reason we couldn't use Rust was definitely the iteration time.
Eh... rust's safety isn't free, but not having it and wasting time on "oh I forgot to change this call site" also isn't free. On the whole I'd say the safety assists in iteration time.
What costs rust in iteration time in my opinion is the low level (by default) nature of it. There's a faster-to-iterate language that has yet to be created which is rust but we sacrifice performance (and memory fiddling ergonomics for the odd person who does that) so we don't have to worry about things like whether a variable is stack or heap allocated. Which is in the direction of a GCed language but retains the mutable-xor-aliasable semantics.
Between rust and current GCed languages though... I guess I agree with "maybe" in both directions.
The project in question needed lots of near-instant human judgements and the iteration loop had to be extremely tight. Maybe Rust should be reconsidered once it gets stabilized enough, but not right now.
Maybe something like Hylo? But personally I don't see anything displacing rust for the next few years, as I think there's enough rust in the training data for it to be the best "serious" language for agentic systems-level development.
It's really the only systems language in its exact niche.
I'm not very familiar with Hylo, but I think it's in the opposite direction from rust than what I'm suggesting.
I'm suggesting a language where there's no difference between Box<u32> and u32. &Vec<u8> and &[u8] are the same thing. I don't need to write Box::new(...) around my closures to pass them to functions that take a function pointer. This comes with overhead, but in exchange we get simpler less verbose code. I.e. a language that isn't systems level, and isn't particularly machine-empathetic. But still has all the lightweight-formal-methods power of rust with lifetimes and mutable vs shared borrows (and thus references to references) and so on.
My impression of Hylo is that it's purpose is to be a similarly low level systems language to rust, just with a less complicated, and as a consequence less expressive, lightweight formal methods system for proving correctness.
I agree I don't expect rust to be displaced anytime soon. It creates a lot of time to create a good compiler, and a lot more to create the ecosystem of code, tools, and community around it.
They didn't mention the cost of this. Assuming mythos was somewhat involved I'd extrapolate this as: 128 x20 max accounts needed which comes at $25.6k or over 75k in api costs. For 75k you can hire a team of engineers that would produce a better result with sematic conversion and other tricks used in porting from language A to language B at the cost of maybe taking 1 month instead of 10 days.
I will be a lot more excited when this is possible with <10k of api costs.
> Pre-merge, this took 5.9 billion uncached input tokens, 690 million output tokens, and 72 billion cached input token reads — around $165,000 at API pricing. By hand, I think this would've taken 3 engineers with full context on the codebase about a year, during which time we wouldn't be able to improve Node.js compatibility, fix bugs, fix security issues or implement new features. We never would've done that. The realistic alternative was to do nothing and keep fixing the bugs at the top of this post forever.
I've always felt [0] the people who created Bun had, as their first and foremost goal, a desire to use Zig--and that's great, I like Zig, I like when people build things their own way.
However, I've been skeptical of using Bun, because I want a project whose first and foremost goal is to build good tools that achieve the objectives of the project.
It reminds me of asking game developers: Do you want to build a game, or do you want to build a game engine? Building a game engine is fine, but if you're goal is to make a game, then building an engine is a poor way of achieving your goals.
Likewise, I've wondered if the creators of Bun wanted to build better JavaScript tools, or if they wanted to use Zig.
One thing that I found interesting is that most of the discourse surrounding the topic happened with the assumption that the rewrite was happening with an Opus-like model, and not with Fable. Those assumptions, at least partially, were used as arguments against the fact that the rewrite was feasible and/or a good idea.
Clearly the model itself doesn't completely change the narrative, but at least as a note to myself, I would like to be more careful with assuming the capabilities of the models used internally by Anthropic and affiliated orgs.
> to exhaustively come up with reasons why the changes create bugs or do not work
My biggest issue currently, is I can't seem to get a code review that's about the simplicity of the code, and no /simplify ain't it. Removing certain bugs and generally working seems to be doing alright, especially if it's following either an example code (like in the Bun rewrite case) or a well defined "spec" of how to proceed.
Every time I've rewritten a major project I've made it smaller and faster while fixing all the major bugs and most of the minor ones. My current team has had similar experiences. I'd be curious to see what a Zig -> Zig rewrite of the same magnitude would have done for quality.
Where is the cost breakdown? I feel like this would be the easiest number to determine and write in this post. It's hard to believe that there have been no problems/downsides since the port.
> Pre-merge, this took 5.9 billion uncached input tokens, 690 million output tokens, and 72 billion cached input token reads — around $165,000 at API pricing
> It's hard to believe that there have been no problems/downsides since the port.
I posted on an older article that I thought it probably cost half a million in API pricing. 165k USD is a lot lower. I wonder what the actual compute cost was. When this first hit the news, Opus 4.7 was brand new and required 6x the compute power per user token vs 4.6. The article says they were using Fable, which is way more expensive.
Thanks!! Those are solid numbers but confusing. He reported input, output, and cached input token reads but not cache writes/cached creation input tokens? Maybe cache writes aren't a thing internally?
Adding bespoke animations via Claude Code to the blog post is definitely thematic. It's unclear if they're useful data visualizations as they take a bit of time to parse, but they're neat.
Inspired by this project I ported most of Valkey to Rust here valdr.dev .
The coolest outcome was being able to run a redis comparible store on an a cloudflare durable object so you do I.e. rate limiting for free with little infra.
So I kept hearing that the author did this purely because Anthropic wanted a PR story, but reading this entire very well written post, with meticulous detail, what say you now? I never thought it made any sense for him to do this just because Anthropic asked him to. Sometimes you find yourself fighting the stack you're currently using, and another stack (or programming language) looks like it would alleviate a lot. LLM was just another tool in his toolbelt. I had already ported projects that were old and abandoned before using Claude Code, so I knew it was possible.
I've done rewrites like this, maybe it wasn't Zig to Rust, but I have been able to rewrite sizable projects, from C# to Rust before. I incorporated a similar strategy, have Claude Opus review the codebase, write a spec, then have Claude implement it, while reviewing the spec, and using the codebase as fallback and gospel over the spec. That said, it's not the entire story here as I said, there was a lot of thought put into it, it it had not been done with Claude, I have a feeling he might have started an "experimental" version of Bun in Rust instead, as many developers have done in the past before LLMs.
I would guess the cost to do this with humans would be _at least_ $1.5M in compensation alone (I'm thinking three 500k/year Bay Area engineers) so this is already an order of magnitude cheaper.
Is it worth $165K? I'm less sure of that but it's honestly a moot point - this will get to 5 then 4 digits of cost pretty fast.
I think putting it in terms of API pricing is oversimplifying disingenuously. Anthropic still hasn't pulled the rug out from under us, so I'm sure it cost a great deal of money once everything comes together, likely surpassing 1.5M. Summarily, they got the result faster, which a group of engineers couldn't do, but at a greater expense.
So much of the discourse around this on HN is nonsensical, and I fully agree with you. It's patently absurd that Anthropic would demand him to rewrite Bun into Rust; it's equally absurd that they would demand any sort of stunt at all when Anthropic already pulled off the biggest stunt with Bun: running Claude Code on it. And why on earth would you cannibalize the runtime of your golden goose?
I still think that generating a Zig-Rust transpiler would be a better approach, given all the LLM quirks, including the ability to just /goal the model with binary-identical LLVM bytecode.
However, an open-sourced tool like that would've greatly harmed the Zig ecosystem and community.
As expected [0] [1], this was a clear advertisement / marketing opportunity of Anthropic's Fable model on rewriting Bun (which powers Claude Code) from Zig into Rust.
Something that would have taken hundreds of developers now took 1 developer with Fable.
Now Claude, rewrite Claude Code from TypeScript to Rust. Make absolutely zero mistakes.
EDIT: the parent has effectively deleted their original comment
> There are a lot of ways to do a terrible job of this. For example, prompting Claude "Rewrite Bun in Rust. Don't make any mistakes." and then praying it would work is not what I did.
This blog post further undermines my trust in Jarred.
He makes it sound like Claude did a fantastic Rust rewrite, and "the work continues."
But when the Rust port merged to main, the state of the code was very, very bad. There were 13,000 instances of `unsafe`, no Miri tests at all, and, sure enough, it exposed UB in safe Rust. https://github.com/oven-sh/bun/issues/30719
Observers could see this coming from a mile away, objected strongly to using AI to RIIR before the code merged. Rather than incorporate feedback and get the code ready for production, Jarred gaslit us all, right here on HN. https://news.ycombinator.com/item?id=48019226
Just 9 days before he merged the Rust rewrite to the main branch, Jarred wrote:
> 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.
It's plausible that Bun's Rust rewrite is now in much better shape than it was in May. But a blog post like this would have been a place to apologize, to accept that it was a very bumpy rollout, to acknowledge that public messaging was extremely poor, and to earn back our trust.
As it stands, I guess I'll have to run my own tests to try to evaluate whether Bun 1.4 is ready for prime time, because I just can't trust Jarred to give us a straight answer.
> But when the Rust port merged to main, the state of the code was very, very bad. There were 13,000 instances of `unsafe`, no Miri tests at all, and, sure enough, it exposed UB in safe Rust.
I mean yeah, that's what this whole post is about. It's about the process of going from that original state to something that's now shipping in production.
That was their estimate at the time, based off the information they had. You can't ask more of someone than that.
Either they estimated poorly, or it ended up the lesser portion of their estimate after all. After all, unless the estimate is 100%, there's always a chance it'll fall into the other portion.
To understand your error, consider that in the month leading up to the 2016 US presidential election, the widely-accepted probabilities were between 70% (Five-Thirty-Eight) and 90% (Reuters) in favour of Clinton.
Should we brace for another front page Zig donation announcement? A fast follow with a “Why Zig?” penance piece, replete with anecdotes about how it is the only true way to express oneself?
> This Rust rewrite would've taken a team of engineers with full-context on the codebase a year of work. With 1 engineer using Fable & closely monitoring Claude Code, we went from start to 100% of the test suite passing on all platforms in 11 days.
This is impressive from a technological standpoint, but it does gloss over the fact that it would have cost $165k in tokens were Bun not part of Anthropic.
The comparison here isn’t completely fair - it would take a small team a year to port it if they spent $0 extra on it.
I’d be interested to see a comparison between spending $165k in 11 days on Claude vs splitting that between 50 people over 11 days for a line-by-line rewrite of the Zig code. I suspect Claude might be faster and therefore cheaper, but maybe not by a lot.
>Combined with the Rust rewrite, ICU changes, and identical code folding, Bun's binary size shrinks by ~20% on Linux & Windows.
People who are surprised by this probably has not seen what Zig code actually looks like. Zig's explicitness and lack of abstraction have a real cost that it is basically one of the most verbose programming languages I've ever seen, it's somehow even more verbose than Go. Basic features of modern languages like pattern matching and generics, and as you can see, having to manually clean up everything means that if you forget once, it's a memory leak. Having SOME abstraction is actually good if it prevents you from making mistakes.
Ironically, Zig is a programming language that's probably best written by LLMs, since they can actually tolerate the verbosity.
Not a compiler expert - shouldn't language verbosity and binary size be, at best, very loosely related?
Fair point, I phrased that too broadly, and you are right about the loose correlation.
What I was gesturing at, badly, was more that Zig’s low-abstraction / explicit-by-default syntax tends to have you write more boilerplate-y code in general that are more annoying to write and maintain, while not buying you enough over a language with better tooling and ecosystem and compiler optimization like Rust.
I don't think you can draw the conclusion that source length and binary size are correlated. For example, in Rust:
Rust's enums can carry data. You can write the same thing in C, but because it does not have the enum feature, you have to do it yourself. They're sometimes called "tagged unions" for a reason, you use a union + a tag when doing it by hand: I haven't actually compiled this, but it should compile to almost the exact same, if not literally the exact same, machine code. Yet one is way more verbose than the other.I think you are saying the same thing as benced - just because Zig source code is verbose is no reason to assume the binary should be larger.
I read my parent ask asking a question: is there a correlation, or not?
I am saying that I do not believe there is a correlation between source code length and binary length. If that's what benced meant by their question, then yes, I agree :)
I’m quite sure there is a certain amount of correlation unfortunately, mainly because there are micro patterns (e.g. IO, allocator) that can’t be modularized into functions. Lots of manual copy-pasta.
It required a little bit of messing with optimisation settings and library generation in Rust, but they emit very very similar x86-64 assembly:
https://godbolt.org/z/89W4srz4d
Nice, thank you for picking up after my laziness. Surely only a few bytes different in the binary, and much, much smaller of a delta than the source.
> Ironically, Zig is a programming language that's probably best written by LLMs, since they can tolerate actually tolerate the verbosity.
Rust in my opinion feels the same.
I don’t feel the verbosity with Rust. Haven’t written it in a while but now in the LLM era I’m looking forward to saying “sort out the lifetime errors for me”.
I trust a lot more Rust code generated by an LLM than anything else ngl.
Agree. But just because it feels the same doesn’t mean it compiles the same.
That can often depend on how you write it.
Naive was 4-9% on the initial pass.
Also note that the larger percentages were against already smaller binaries. That smells like there was a single large constant number that got saved somewhere rather than general improvements.
> After that initial shrinkage, the team explored more opportunities for binary size reduction using linker optimizations like Identical Code Folding, removing unused data from ICU, and lazily decompressing small parts of libicu with a zstd dictionary on-demand.
I'd be VERY interested in seeing what the individual effects of those parts were.
Article did a decent job of showing discipline and care and human involvement to assert the automated rewrite was done diligently, as best as it can be when using AI for it. I does make me feel a bit more comfortable about it.
As an aside, I don't know why anyone would not want to use a memory-safe (and possibly race-safe) language in 2026. Rust gives you that in a performant package, so if you are turned off by GCs and immutability for performance reasons, you still have the option to use Rust.
I can understand when you need the absolute best performance and you decide to drop to down to C++, and I also relate with just personal preference, but beyond those it seems a no brainer to me.
I was fairly skeptical about the rewrite when news about it first started going around, and I still don't plan on switching anything to use the Bun rewrite anytime soon, but I appreciate how detailed and well-written the blog post is; it also seems to be primarily human-authored, in my opinion, which is refreshing.
The most significant revelation for me was that Claude Code has been using the rewrite without much fanfare since June 17th.
Without commenting on Bun itself as a project, or the nature of the rewrite, it can't be good for Zig that a naive rewrite away from it fixed memory leaks, improved stability, shrunk binary size by 20%, and improved performance by 5%.
I don't think it's care to categorize this as "a naive rewrite away from [Zig]" - Jarred has been immersed in this project for five years, got to benefit from everything he learned along the way and spent $165,000 of tokens on the most advanced coding LLM anyone has access to.
I expect if he'd spent $165,000 running Fable against the Zig version he could have got a 5% performance improvement, too.
I pay attention when someone makes a hard decision based on a hard-learned lesson. It's like, most who choose to use an ORM just heard of it or want to avoid learning SQL, everyone who removes an ORM learned firsthand horrors.
I would guess that people looking to use Zig understand that those are project concerns and not language concerns.
The stability gains are a direct language concern as mentioned throughout the article.
Yeah but they turned it into something unreadable. Call it a skill issue if you wish.
I just haven’t found another language that just makes sense. Zig doesn’t hide anything from you
>they turned it into something unreadable
Did you compare the code before/after? It's a mechanical line-by-line port, and most of the code is identical to the old version, just with Rust syntax. They have an example in the blog post.
The article explicitly mentions the maintainability as a foremost concern.
People say a lot of things, especially when they have a vested interest in a positive outcome. Bun has been fully vibe coded into another language. There’s no way in hell it’s maintainable. Go read any analysis of the Claude Code leak for proof.
Claude Code is entirely vibe-coded for a long time. Bun isn't. You go read and compare the actual Bun code; it reads reasonably well [1].
[1] For example, as a random sample, https://github.com/oven-sh/bun/blob/bun-v1.3.14/src/css/medi... -> https://github.com/oven-sh/bun/blob/4924862cffbf671792d47c92...
Sure, reasonably well at first glance, but to quote the article:
> I rewrote Bun in Rust using about 50 dynamic workflows in Claude Code run continuously over the course of 11 days.
> Excluding comments, Bun is 535,496 lines of Zig.
> How do you review a PR with +1 million lines added? How do you start to build the confidence needed to responsibly merge large quantities of LLM-authored code? A language-independent test suite with a million assertions, adversarial code review and when something does go wrong, fixing the process that generates the code instead of hand-fixing the code.
That’s vibe coding. This blog post is an ad for Claude, nothing more.
Doesn't look like vibecoding to me. It does look like a Claude ad, but they do have a vested interest in not screwing up Bun now that they own it.
You're entitled to call things as you wish, of course, but your definition of "vibe-coding" differs quite a bit from mine.
> There’s no way in hell it’s maintainable
This is not an assertion you are qualified to make
> Go read any analysis of the Claude Code leak for proof
You seem to be implying that Claude code is unmaintainable. Yet they appear to be maintaining it just fine. Did I misunderstand your implication?
The same concern applies to every GC language, so it's not necessarily bad for Zig. Bun can have been grown too large for Zig to be effective, while moderately sized projects may still greatly benefit from Zig.
I thought Zig was supposed to be a C replacement (as in, it doesn't actually provide full safety in the way that Rust or a GC language would)?
Oh, yeah that might be confusing. I meant "you can say the same thing for GC language if that's true, which isn't necessarily true, so that must be false".
More precisely speaking: GC languages are said to delay memory problems far beyond the horizon, which is often unreachable throughout the project's history. Zig can be a similar case.
True, but rewrites often allow for this sort of benefit in themselves. It's possible rewriting it in zig would have yielded some of the same improvements.
A sophisticated rewrite? Sure. This was a naive like-for-like rewrite, though.
Wouldn't the same improvements have been made in zig if they instructed the agents to improve instead of rewrite?
Maybe they'd get the same numeric improvements and bug fixes today (or maybe not, or maybe they'd get even more since the LLM isn't spending time rewriting correct code).
But they wouldn't get a change to the structural issues that created the issues in the first place. They'd end up "ke[eping] fixing these kinds of bugs one-off in perpetuity".
But how would you verify that the agents have written memory safe code? Rust's borrowchecker is a lot faster and actually verifiably safe compared to asking an LLM to fix the safety issues that the Zig version had.
From a PL Theory perspective, Zig is vibe-coded.
Not sure why people use it.
> Historically, rewrites are a terrible idea.
This changed for me over the last 5 years.
The first scenario was joining a company where a software product barely worked. We did the traditional incremental refactoring / rewriting, but eventually learned how rotten the core was that rewriting from first principles was the best path forward.
The lesson learned here is that the conventional wisdom probably only applies to rewriting complex but working systems.
Then multiple scenarios in the agentic coding age. Between day jobs and hobbies I've reproduced major chunks of complicated software like Salesforce, Gmail, Pioneer Rekordbox with very lean teams.
Much like the blog post, the trick is to get an excellent verification loop with a compiler, linter, and test harness / test suite around the core behaviors.
It's feeling more and more that designing and implementing comprehensive test harnesses is the real work, once you have that let the LLM cook.
"I used a pre-release version of Claude Fable 5 for much of the Rust rewrite."
It'd be interesting if Anthropic became a general software company just because they have access to models that aren't yet released, possibly export-banned.
That's the power of a strong test suite. LLMs excel when you have verifiable rewards. I imagine we'll get a lot more rewritten in rust projects in the future. Rust is also an ideal target for such rewrites as it offers a lot of verification (via its type system) and is low overhead with zero-gc. There's less and less reason to use GC'd languages in the agentic coding era.
I think Rust is a locally optimal target for LLM coding, we might see a better language in the future, but I think Rust will dominate for quite some time.
> There's less and less reason to use GC'd languages in the agentic coding era.
Faster iteration, maybe? Rust's safety guarantee isn't exactly free (while still being very excellent) and does affect iteration time. I have a private project (>300K LoC) that has been translated from Python to TypeScript and the reason we couldn't use Rust was definitely the iteration time.
I like using Odin with LLMs for this. it's a simple statically typed language with no GC and very fast compile
Eh... rust's safety isn't free, but not having it and wasting time on "oh I forgot to change this call site" also isn't free. On the whole I'd say the safety assists in iteration time.
What costs rust in iteration time in my opinion is the low level (by default) nature of it. There's a faster-to-iterate language that has yet to be created which is rust but we sacrifice performance (and memory fiddling ergonomics for the odd person who does that) so we don't have to worry about things like whether a variable is stack or heap allocated. Which is in the direction of a GCed language but retains the mutable-xor-aliasable semantics.
Between rust and current GCed languages though... I guess I agree with "maybe" in both directions.
The project in question needed lots of near-instant human judgements and the iteration loop had to be extremely tight. Maybe Rust should be reconsidered once it gets stabilized enough, but not right now.
Maybe something like Hylo? But personally I don't see anything displacing rust for the next few years, as I think there's enough rust in the training data for it to be the best "serious" language for agentic systems-level development.
It's really the only systems language in its exact niche.
I'm not very familiar with Hylo, but I think it's in the opposite direction from rust than what I'm suggesting.
I'm suggesting a language where there's no difference between Box<u32> and u32. &Vec<u8> and &[u8] are the same thing. I don't need to write Box::new(...) around my closures to pass them to functions that take a function pointer. This comes with overhead, but in exchange we get simpler less verbose code. I.e. a language that isn't systems level, and isn't particularly machine-empathetic. But still has all the lightweight-formal-methods power of rust with lifetimes and mutable vs shared borrows (and thus references to references) and so on.
My impression of Hylo is that it's purpose is to be a similarly low level systems language to rust, just with a less complicated, and as a consequence less expressive, lightweight formal methods system for proving correctness.
I agree I don't expect rust to be displaced anytime soon. It creates a lot of time to create a good compiler, and a lot more to create the ecosystem of code, tools, and community around it.
They didn't mention the cost of this. Assuming mythos was somewhat involved I'd extrapolate this as: 128 x20 max accounts needed which comes at $25.6k or over 75k in api costs. For 75k you can hire a team of engineers that would produce a better result with sematic conversion and other tricks used in porting from language A to language B at the cost of maybe taking 1 month instead of 10 days.
I will be a lot more excited when this is possible with <10k of api costs.
> Pre-merge, this took 5.9 billion uncached input tokens, 690 million output tokens, and 72 billion cached input token reads — around $165,000 at API pricing. By hand, I think this would've taken 3 engineers with full context on the codebase about a year, during which time we wouldn't be able to improve Node.js compatibility, fix bugs, fix security issues or implement new features. We never would've done that. The realistic alternative was to do nothing and keep fixing the bugs at the top of this post forever.
Even at $165K, it's worth it to have a better base to build on top of—especially since it didn't take a year's worth of time for three programmers.
I've always felt [0] the people who created Bun had, as their first and foremost goal, a desire to use Zig--and that's great, I like Zig, I like when people build things their own way.
However, I've been skeptical of using Bun, because I want a project whose first and foremost goal is to build good tools that achieve the objectives of the project.
It reminds me of asking game developers: Do you want to build a game, or do you want to build a game engine? Building a game engine is fine, but if you're goal is to make a game, then building an engine is a poor way of achieving your goals.
Likewise, I've wondered if the creators of Bun wanted to build better JavaScript tools, or if they wanted to use Zig.
[0]: https://news.ycombinator.com/item?id=35970044
So does that mean the rewrite made you less skeptical?
One thing that I found interesting is that most of the discourse surrounding the topic happened with the assumption that the rewrite was happening with an Opus-like model, and not with Fable. Those assumptions, at least partially, were used as arguments against the fact that the rewrite was feasible and/or a good idea.
Clearly the model itself doesn't completely change the narrative, but at least as a note to myself, I would like to be more careful with assuming the capabilities of the models used internally by Anthropic and affiliated orgs.
> Claude Code v2.1.181 (released June 17th) and later use the Rust port of Bun.
It seems the reports of Bun's death have been greatly exaggerated.
> to exhaustively come up with reasons why the changes create bugs or do not work
My biggest issue currently, is I can't seem to get a code review that's about the simplicity of the code, and no /simplify ain't it. Removing certain bugs and generally working seems to be doing alright, especially if it's following either an example code (like in the Bun rewrite case) or a well defined "spec" of how to proceed.
Every time I've rewritten a major project I've made it smaller and faster while fixing all the major bugs and most of the minor ones. My current team has had similar experiences. I'd be curious to see what a Zig -> Zig rewrite of the same magnitude would have done for quality.
Where is the cost breakdown? I feel like this would be the easiest number to determine and write in this post. It's hard to believe that there have been no problems/downsides since the port.
> Where is the cost breakdown?
From the article
> Pre-merge, this took 5.9 billion uncached input tokens, 690 million output tokens, and 72 billion cached input token reads — around $165,000 at API pricing
> It's hard to believe that there have been no problems/downsides since the port.
A significant portion of the article was dedicated to the 19 regressions they've found. Starting here: https://bun.com/blog/bun-in-rust#porting-mistakes
I posted on an older article that I thought it probably cost half a million in API pricing. 165k USD is a lot lower. I wonder what the actual compute cost was. When this first hit the news, Opus 4.7 was brand new and required 6x the compute power per user token vs 4.6. The article says they were using Fable, which is way more expensive.
Thanks!! Those are solid numbers but confusing. He reported input, output, and cached input token reads but not cache writes/cached creation input tokens? Maybe cache writes aren't a thing internally?
> Compiler errors are a better feedback loop than a style guide
So essentially this whole re-write was about making Bun LLM compatible.
Adding bespoke animations via Claude Code to the blog post is definitely thematic. It's unclear if they're useful data visualizations as they take a bit of time to parse, but they're neat.
Inspired by this project I ported most of Valkey to Rust here valdr.dev .
The coolest outcome was being able to run a redis comparible store on an a cloudflare durable object so you do I.e. rate limiting for free with little infra.
> In Bun v1.3.14, every build leaks about 3 MB, forever
I'm sorry but that is insane, how was this never fixed before the rewrite?
So I kept hearing that the author did this purely because Anthropic wanted a PR story, but reading this entire very well written post, with meticulous detail, what say you now? I never thought it made any sense for him to do this just because Anthropic asked him to. Sometimes you find yourself fighting the stack you're currently using, and another stack (or programming language) looks like it would alleviate a lot. LLM was just another tool in his toolbelt. I had already ported projects that were old and abandoned before using Claude Code, so I knew it was possible.
> what say you now?
I think that when you have a $165,000 hammer, all of your problems begin to look a lot like nails.
I've done rewrites like this, maybe it wasn't Zig to Rust, but I have been able to rewrite sizable projects, from C# to Rust before. I incorporated a similar strategy, have Claude Opus review the codebase, write a spec, then have Claude implement it, while reviewing the spec, and using the codebase as fallback and gospel over the spec. That said, it's not the entire story here as I said, there was a lot of thought put into it, it it had not been done with Claude, I have a feeling he might have started an "experimental" version of Bun in Rust instead, as many developers have done in the past before LLMs.
I would guess the cost to do this with humans would be _at least_ $1.5M in compensation alone (I'm thinking three 500k/year Bay Area engineers) so this is already an order of magnitude cheaper.
Is it worth $165K? I'm less sure of that but it's honestly a moot point - this will get to 5 then 4 digits of cost pretty fast.
I think putting it in terms of API pricing is oversimplifying disingenuously. Anthropic still hasn't pulled the rug out from under us, so I'm sure it cost a great deal of money once everything comes together, likely surpassing 1.5M. Summarily, they got the result faster, which a group of engineers couldn't do, but at a greater expense.
So much of the discourse around this on HN is nonsensical, and I fully agree with you. It's patently absurd that Anthropic would demand him to rewrite Bun into Rust; it's equally absurd that they would demand any sort of stunt at all when Anthropic already pulled off the biggest stunt with Bun: running Claude Code on it. And why on earth would you cannibalize the runtime of your golden goose?
I still think that generating a Zig-Rust transpiler would be a better approach, given all the LLM quirks, including the ability to just /goal the model with binary-identical LLVM bytecode.
However, an open-sourced tool like that would've greatly harmed the Zig ecosystem and community.
> would've greatly harmed the Zig ecosystem and community
People looking to abandon the ship first chance are unlikely to contribute much to the ecosystem and community.
As expected [0] [1], this was a clear advertisement / marketing opportunity of Anthropic's Fable model on rewriting Bun (which powers Claude Code) from Zig into Rust.
Something that would have taken hundreds of developers now took 1 developer with Fable.
Now Claude, rewrite Claude Code from TypeScript to Rust. Make absolutely zero mistakes.
[0] https://news.ycombinator.com/item?id=48073893
[1] https://news.ycombinator.com/item?id=48240829
EDIT: the parent has effectively deleted their original comment
> There are a lot of ways to do a terrible job of this. For example, prompting Claude "Rewrite Bun in Rust. Don't make any mistakes." and then praying it would work is not what I did.
*whoosh* goes the joke.
Hacker News does not have a meme-y culture, and this post takes this topic pretty seriously and is technically interesting.
It's not so much that I missed that it was a joke, I just don't think that it really added to the discussion.
What you've edited it to is a much better comment.
1 Developer with a 200k budget for tools.
This slop rewrite introduced new vulnerabilities and regressions.
Care to elaborate?
This blog post further undermines my trust in Jarred.
He makes it sound like Claude did a fantastic Rust rewrite, and "the work continues."
But when the Rust port merged to main, the state of the code was very, very bad. There were 13,000 instances of `unsafe`, no Miri tests at all, and, sure enough, it exposed UB in safe Rust. https://github.com/oven-sh/bun/issues/30719
Observers could see this coming from a mile away, objected strongly to using AI to RIIR before the code merged. Rather than incorporate feedback and get the code ready for production, Jarred gaslit us all, right here on HN. https://news.ycombinator.com/item?id=48019226
Just 9 days before he merged the Rust rewrite to the main branch, Jarred wrote:
> 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.
It's plausible that Bun's Rust rewrite is now in much better shape than it was in May. But a blog post like this would have been a place to apologize, to accept that it was a very bumpy rollout, to acknowledge that public messaging was extremely poor, and to earn back our trust.
As it stands, I guess I'll have to run my own tests to try to evaluate whether Bun 1.4 is ready for prime time, because I just can't trust Jarred to give us a straight answer.
Pre-release code had bugs that were fixed before the release? Why is that a problem? That's the point of having a testing and release process
what about new bugs introduced after the rewrite?
> But when the Rust port merged to main, the state of the code was very, very bad. There were 13,000 instances of `unsafe`, no Miri tests at all, and, sure enough, it exposed UB in safe Rust.
I mean yeah, that's what this whole post is about. It's about the process of going from that original state to something that's now shipping in production.
> We haven’t committed to rewriting. There’s a very high chance all this code gets thrown out completely.
God forbid an engineer express uncertainty.
Uncertainty is one thing, but a high chance means it’s 51% or higher to me.
Based on that, the bun rewrite messaging was fairly misleading.
That was their estimate at the time, based off the information they had. You can't ask more of someone than that.
Either they estimated poorly, or it ended up the lesser portion of their estimate after all. After all, unless the estimate is 100%, there's always a chance it'll fall into the other portion.
To understand your error, consider that in the month leading up to the 2016 US presidential election, the widely-accepted probabilities were between 70% (Five-Thirty-Eight) and 90% (Reuters) in favour of Clinton.
Should we brace for another front page Zig donation announcement? A fast follow with a “Why Zig?” penance piece, replete with anecdotes about how it is the only true way to express oneself?