It's way late and seriously over budget, hard to call that a success.

It's an unconditional success no matter how you slice it.

~3 months start to finish and the price of 1 engineer for a year, to rewrite a critical infra codebase that tens of millions of user software relies on. Truly the goalposts keep moving. This kind of project would've have been 10x the effort and price a few years ago.

Yeah, crucially, a lot of such projects never get greenlit.

Claude building stuff means a lot of things that everybody wanted to do, just couldn't allocate the attention or other resources - now get built.

It's likely much closer to that price if you don't calculate with subsidized token pricing.

Once a model like Fable/Mythos is trained, that was used for this project, the inference costs are profitable. CapEx vs. OpEx. Every new Zig->Rust port of this caliber using Fable/Mythos costs the standard token price, it's not subsidized. The model already exists for the original CapEx, you can do 1,000 ports of this for standard token pricing.

What budget?

What are you talking about? What is the budget? Was there ever a deadline? Who ever decided those?

This couldn't be further from the truth. It was if anything severely under budget of you compare it to what it'd normally cost in engineering time.

What exactly has this update led this to be a "huge win"?

The things stated in the announcement?

Mainly node compatibility though more test coverage and different performance improvements.

I fully AI-driven rewrite of a massive codebase from one language to another within a reasonable release-cycle has never been done before.

There was a lot of sentiment around the lines of:

- Bun is done for, it's all slop now

- there's no way a fully AI-written codebase can work

- Jarred is an idiot for trusting AI to rewrite the whole thing

- they're never going to release this, there's going to be a billion of bugs

This release proved all of this wrong. It also serves as a great advertisement of how capable Anthropic models and harness are to enable this.

I don't think this release definively proved all of this wrong, no.

That's true, it was when they stealthed released it in a application that millions use and no one noticed.

Which one did it not prove wrong? The fully AI written version of Bun has released, fixed a ton of bugs that existed in the zig version, added a lot of new features and performance improvements and has been shipping with every single installation of claude code for months, the most widely used LLM-harness out there.

By every single measurable metric this is better software than the old hand-written version ever was.

It is for sure a big accomplishment, but without having paid too much attention to the drama, I understand that the porting was meant to show Anthropic's models amazing capabilities. From what we know, the process took longer than expected and involved dozens of engineers from Anthropic actively working on it, presumably each one with dozens of SOTA agents helping. It definitely was an expensive process when you sum up the engineers hourly rate and the tokens consumed by frontier LLMs. It's not something that a small team engineers can accomplish with a Claude Max subscription (obviously not, but those are the vibes some people were getting with the project).