Many repeat the point of “$165k is cheaper than team of multiple engineers working on the rewrite for a year”, which I think is flawed — the team of engineers would have produced idiomatic rust, and it would take probably 100k+ of tokens more to make the bun in rust idiomatic rust.
They also would have produced a team of engineers that knows the Rust codebase.
This is assuming your engineers don't leave for higher-paying roles elsewhere. The market might be cold generally but for engineers working at these frontier AI companies it's red-hot.
And most executives are figuring this into their calculus right now because they were burned badly during COVID. Meta, Google, etc were loose with hiring and engineers flocked from their lower-paying companies in droves. The brain drain was real.
One public company I was at lost nearly 2/3rds of their engineers and mostly to Meta (granted, they had other problems but it was mostly about money -- the offers were excessive). Then market conditions forced them to freeze hiring and they've had a slow exodus of senior talent since as the firefighting has become constant.
AI adoption has only accelerated problems for them.
We've taken this "only two years and then leave" philosophy to an extreme and now companies are totally justified in not investing in their engineers anymore.
The job-hopping followed from companies not investing in their engineers, not the other way around. It was billed as the only way to get a promotion (which usually would come every 1-2 years).
I've been in this industry for a few decades at this point and I was around when this meme was started.
It was purely about maxing your compensation because changing jobs nets you more than promotions & raises.
I've never met an engineer in my life who truly earned a promotion every year and very few every two. Very few companies have org charts that even support that or have that many levels. This logic/advice only really applies at a few companies and people have adopted it no matter where they work. There's not enough growth/hiring at 95% of companies to even come close. The Peter Principle is also a real thing. Everyone's competence has a ceiling.
There's even an implicit understanding of this that people at smaller companies have inflated titles and you typically rank them down 1-2 levels when hiring/acquiring at larger companies.
Now I do agree that companies haven't been investing in their engineers, but that doesn't also mean this isn't a vicious cycle. Employers and employees are in a mexican standoff and things are only going to get worse until one side comes to its senses. Employers have all the leverage for it to not be them.
For the vast majority of companies the average tenure of an engineer sits between 18 and 30 months. They're also mostly hiring young engineers in their 20s. As an employer what is your upside to making such investments before they're at least mid-career? A lot of people seem to want the world and offer nothing in return for it.
> It was purely about maxing your compensation because changing jobs nets you more than promotions & raises.
You don't view it as a problem that companies consistently compensate new hires higher than they are their experienced employees?
Hell, I would have been perfectly happy to never change employers in my entire career to date, if my salary had anywhere near kept pace with my peers who were job-hopping.
> You don't view it as a problem that companies consistently compensate new hires higher than they are their experienced employees?
It depends on the current employee and the incoming employee. They are not interchangeable cogs. Also not everyone consistently provides good value over their tenure -- a lot tend to work hard early and then for various reasons taper off. It's not necessarily their fault, but companies definitely are aware of this.
Personally I tend to value growth of my skillset over growth of income and that has largely informed my movement. I move when there's no longer interesting work if the compensation is at least fair. I've never been about comparing myself to others -- especially when most of my peers' left the industry after 5-10 years.
> a lot tend to work hard early and then for various reasons taper off
And you don’t think that has anything to do with incentives (or lack thereof)?
> I've never been about comparing myself to others
Doesn’t have to be about comparing oneself to others. More about the cost of living increasing because software engineering salaries are rising around you…
> And you don’t think that has anything to do with incentives (or lack thereof)?
No, that's almost never it. It's mostly life circumstances, burnout, frustration with management, etc.
> More about the cost of living increasing because software engineering salaries are rising around you…
We're already talking about very well-paid software engineers here. Appealing to a sense of sympathy over cost of living concerns of 1% earners and them trying to become 0.5% earners isn't exactly a strong argument here...
"Knowing the codebase" seems like an antithetical philosophy to LLM driven engineering orgs.
Well the purpose of the tool is to automate more specialized labor so any knowledge that could benefit the worker has to be downplayed or stigmatized.
I feel your estimate of tokens is a few orders of magnitude off, it’s on the low side.
I use more (albeit cached) when centering a div.
I think they missed a $ sign, i.e. they meant $100k of tokens
This is where we've come to where people proudly proclaim using an AI to do what is a single line of CSS.
"centering div" is a classic problem (/trauma/meme) that sounds trivial but had no universal solution (until 2017?).
And has a dead simple one today.
Also worth noting simple horizontal centering of divs was never a problem, margin: auto was defined in CSS Level 1 in 1996 [1].
It was vertical centering that took a very long time to crack, which really became trivial with Flexbox which was first drafted in 2009[2] but became available unprefixed in browsers between 2012 and 2014 [3] about 13 years ago.
[1] https://www.w3.org/TR/REC-CSS1-961217
[2] https://www.w3.org/TR/2009/WD-css3-flexbox-20090723/
[3] https://caniuse.com/flexbox
Subnote: I'm not counting the display: table hacks.
100k tokens is your pre-prompt and your CLAUDE.md as well as a few files from your root.
What a great joke.
I will have to steal it for an upcoming AI tools meeting I have at work.
Also, pretty clever as centering a div w/ CSS has been notoriously difficult to achieve.
Would they have, in a year? The general plan of attack would likely still be the same: rewrite it in rust while keeping the structure as similar as possible, no matter how unidiomatic, then adjusting the design to make it more idiomatic to rust. Doing both at once is much harder.
> Doing both at once is much harder.
I think with sequential approach (translate -> make idiomatic) it is easier. However, I am not sure that monetary difference is going to be as stark as $165k vs 3 engineers/year. Especially, if you consider that no one knows that is what in the code at the end.
Sure, you can argue that now it doesn’t matter — agents and all that, but I am not so sure.
> the team of engineers would have produced idiomatic rust
Not necessarily. Didn’t Microsoft port the TS compiler to Go and they did it by translating the TS? It wasn’t idiomatic Go.
Yes, and curiously many of the tellings of this story don't account for the costs of the human software engineers (Jarred and the other bun team members) who guided all of this work.
IMO a team of engineers (presuming this is pre-AI) would have improved the existing Zig codebase, rather than spent time and money on a Rust port in the first place. In fact, I still think most teams of engineers would choose to improve what they have even now, circa AI.
I doubt file by file rewriting takes as much work as rewriting and refactoring the code - especially at this scale.
Yes this is a typical case of showing results fast. I wonder what the cost for the remaining 10℅ of debugging and fixing all the bloat will be. Not so cheap I am guessing
> produced idiomatic rust
I keep seeing this. What is "un-idiomatic" rust?
Rust contains the ability to do everything that C can, if you use `unsafe`. And a file-by-file rewrite in Rust from another language usually involves keeping the ABI and API between files very C-like (and unsafe). In practice this means that the result of a first pass this way has all the memory safety of C code, but with worse readability because Rust makes unsafe things less ergonomic.
To actually get the safety benefits of Rust in a real way you have to rework those files to not treat eachother as C code. This is the interesting and the difficult part of a rewrite in Rust, and one that an unsupervised LLM rewrite is probably not even going to attempt.
I don't think this latter stage has actually happened with Bun's codebase. The result of the Rust rewrite in Bun's case is actually less safe than the Zig it is replacing, especially because (I am told) a lot of the new Rust code is unsound (introducing UB).
> To actually get the safety benefits of Rust in a real way you have to rework those files to not treat eachother as C code. This is the interesting and the difficult part of a rewrite in Rust, and one that an unsupervised LLM rewrite is probably not even going to attempt.
You can use deterministic linting for `unsafe` usage, have agents target `unsafe`, etc. It's pretty easy. You can even run `miri` against the code and give that as a tool for LLM feedback. I've done this all before and it works fine.
> The result of the Rust rewrite in Bun's case is actually less safe than the Zig it is replacing, especially because (I am told) a lot of the new Rust code is unsound (introducing UB).
I'm unconvinced that this is true. How could you tell? You only know about the rust bugs because rust makes them grep'able/ trivial to verify, there's no way of knowing which bugs existed in zig that didn't translate. Regardless, the problem is now trivial to understand in Rust and start to target.
> one that an unsupervised LLM rewrite is probably not even going to attempt.
Why not? A complete test suite exists, so it just boils down to "reimplement this code to reduce the number of 'unsafe' references, while keeping the tests passing" (the last part isn't even needed since Claude loves to run tests and linters anyway).
An LLM definitely can do that work, but LLMs have a tendency to follow the path of least resistance unless you force them to do things properly and carefully supervise them.
> LLMs have a tendency to follow the path of least resistance
That's a really strong motivation for the Rust port. It's literally impossible to take a path not leading to an unacceptable outcome because the Rust compiler fails the compile if code no longer marked "unsafe" is still actually unsafe, and the tests fail if the implementation is incorrect. The compiler actively helps to force the LLM to do things properly.
…only if you actually force the LLM not to use `unsafe` somehow! And one must remember that `unsafe` can have all sorts of non-local effects. One way to be maliciously compliant (or in the LLM's case, I guess it's more incompetence than malice) is to make a “safe” function that does an `unsafe` operation for you, and then replace every usage of that unsafe operation with your “safe” function. You can remove hundreds of unsafe blocks that way without any improvement in safety.
Presumably lots of stuff wrapped in 'unsafe'. I'm not a rust guy or a rust fan, and the last time I wrote something with Rust was like 7 years ago, but to my memory, because of how Rust manages mutability there are many access patterns that are Rust-specific; idiomatic Rust is going to use this patterns but they would be unlikely to show up in a language with a different type system / borrow checker / etc. etc.
So "unidiomatic" rust means, 'don't use these parts of the language even though they're first-class citizens?'
How odd. Lot of mental gymnastics going on there.
The Bun author himself writes that the port is not idiomatic Rust:
(from: https://bun.com/blog/bun-in-rust)The original version of Bun was already a line-by-line (manual) port of esbuild from Go to Zig (also mentioned in this post), so the Zig code already wasn't "Zig-idiomatic" and apparently riddled with problems. That doesn't inspire much confidence in the LLM-translated Rust version tbh.
> That doesn't inspire much confidence in the LLM-translated Rust version tbh.
And yet Claude Code, which is the primary consumer, continues to chug away without any issue (that I, as a fairly heavy user, have encountered).
> "unidiomatic" rust means, 'don't use these parts of the language even though they're first-class citizens?'
that is indeed a roughly accurate definition of what "unidiomatic" means
It's mostly the same in any general purpose programming language though, parts of the language features are meant (or became seens as) tools to be used in very narrow cases, which people less familiar with the language tend to "overuse".
Eg. goto/jump, bare "except" statements in python, etc... In Rust, unsafe has its uses and is a first class language feature, but it certainly isn't idiomatic to use it to mimic access patterns from other languages (unless absolutely necessary).
It means "don't misuse parts of the language in ways that circumvent best practices" one of which is "keep unsafe regions as small and few as possible" not "never use unsafe"
Meh. The (a major?) goal of Rust is this sort of type safety flowing down to memory bugs. Not every possible access pattern was known at the beginning, and generally they developed a philosophy of keeping the "unsafe" memory access patterns extremely bounded and heavily inspected. I think that's a fair and admirable goal.
[dead]