Congratulations on the raise!
Big fan of your open-endedness work here. We've been getting a lot of inspiration from it at @imbue-ai.bsky.social . Looking forward to see what else Sakana will be developing with the new funding!

It is off by default. You have to enable it from what I've read.

Duh, if she enabled the browser memories feature, of course it will do that. It's not like they're hiding it, it's one of the advertised features and afaik can be turned off.

If you can read this, that means I tried and it worked.

The fact that I can point a coding agent to an arbitrary open-source software and have it implement some missing feature in a few minutes is wild. The one-shot implementation quality might not be sufficient for merging it upstream, but it's quite often enough for my one-off use.

We just released a beta of such a product last week 🤞https://imbue.com/sculptor-announce/
Runs Claude Code in Docker containers with some special collaboration sauce.

We might open source Sculptor in the future, but aren't quite ready for it yet. (between us though, it's written in Python and not obfuscated, so you can also just look at its code)

Ah, interesting. I haven't thought of that.

Unfortunately I don't think they're made uncomfortable by the rituals, but by Leo criticizing Trump's deportation policy. They just have to latch on to random things now to present the pope in a bad light.

Interesting that Sonnet 4.5 is a lot more expensive than Sonnet 4 in this eval. I assume it generates more tokens for the same problem?

Yeah I feel like LLMs are such a cool tool for diving into unfamiliar languages. Both for answering questions, but also for generating example code at various levels of complexity. I've recently been trying to learn some Lisp again, including by vibe coding some. bsky.app/profile/dmew...

Ah yeah. I have never written Cobol myself, but happened to have listened to this podcast just a week ago. youtu.be/Rdm3fgxbLOE?... So I recognized some of the structures when I looked at your webserver.

Oh, I see you used Claude to write it... Either way, maybe you still heve a sense of the above from reviewing its code?

Very cool! Was there anything in the structure of Cobol that you found made it challenging, or just a lack of suitable libraries? (I'd guess it's not a great language for parsing, but then neither is C, is it?)

When I first read your summary, I thought you had changed the log prop sampling so it could never sample the stop token for 10 hours. I wonder if that approach would give similar behavior or something different. In that case, there would be no countdown time or user messages at all.

Cool experiment! I agree that it might reveal something about a model's ability (tendency?) to seek novelty when it can't make progress with one approach.