Hacker News Reader

Comment by geertj | Hacker News Reader

I wonder how much of the rewrite was done using codex itself. It does seem like a perfect use case for it.

I imagine a reasonable amount. The maintainer who is doing most of the Rust rewrite submitted a PR to one of the Ratatui widget libraries I maintain that seemed to be Codex produced[1].

[1]: https://github.com/joshka/tui-markdown/pull/80

This is an interesting proof of what I keep reading about. Where people are quick at making something, like a PR, with AI, but then doing the last 10% is left in the air.

If codex was half as good as they say it is in the presentation video, surely they could’ve sent a request to the one in chatgpt from their phone while waiting for the bus, and it would’ve addressed your comments…

The last 10% you're referring to are nits. That's like the last 0.000001%. also, it could have fixed all these in like a minute by itself.

Why didn't it?

I'm guessing that to do this would have required running codex and ensuring that the specific repo / PR context was in scope. Perhaps the dev just hasn't done this yet?

Some of my response comments were nits where the tooling didn't respect conventions in the code / brought in conventions that weren't in use. I'd expect that minding existing conventions would be something that LLM based code tooling would eventually incorporate explicitly into its context and guardrails. Intuitively this seems like it would be a difficult thing to push down a level into a model's training for various practical reasons.

Thats an interesting question. how much of code at these companies is written by tools they are selling.

I suspect its not much because I never see any stats published by any of these companies.

Aider does publish this, and it's fascinating - https://aider.chat/HISTORY.html

> Aider writes most of its own code, usually about 70-80% of the new code in each release. These statistics are based on the git commit history of the aider repo.

> Whenever aider edits a file, it commits those changes with a descriptive commit message.

interesting model. every prompt response acceptance gets a git commit without human modifications .

Reply on news.ycombinator.com