
Ooof. /provably/ correct code, not probably. Though...both are good.

Occasional OCaml programmer. Host of @signalsthreads. @yminsky@discuss.systems @yminsky.bsky.social threads.net/@yaronminsky

Ooof. /provably/ correct code, not probably. Though...both are good.


Exciting news: we're looking for interns for our formal methods team! Help us turn OCaml into the best language on the planet for writing probably correct code. Hiring in both London and New York. janestreet.com/join-jane-str… janestreet.com/join-jane-str…

I wonder if David is seeing the same thing I am: LLMs of today are tactical geniuses but still learners at strategy. A classic problem is putting fixes at the wrong level in the system: a short-term tactical win but a strategic loss. Those strategic issues compound in weird ways.

I'm struggling with this too. Most code is fundamentally utilitarian, and agents getting it done is pure good. But agents will build on whatever foundations you give them -- and I think those foundations matter more than ever.

After a ~year of heavy LLM use in library development I'm currently trying to ship my first production system with 100% LLM-written code, and... I'm definitely finding that they're bad at this in ways that correspond pretty well to being bad at software design.

Either way, I don't think it's just small mindedness. The inability of current agents to generate clean systems with beautiful abstractions is a real cognitive gap! Maybe it will be closed soon, and maybe it doesn't matter, but the gap is real nonetheless.

But...can they? We don't really know how effectively they can operate in the face of what we think of as bad design. And we don't have large scale, complex projects that have been entirely vibe coded and have shown a lot of impact and success.

One perspective is: it's small minded of us to care too much about traditional notions of taste. We should just care about the results, and if agents can operate effectively without clean abstractions and well factor code, it's just prejudice to care.

One interesting bit of AI discourse: it's pretty broadly agreed that agents aren't great at software design. But...they're still good at getting tasks done. So, maybe software design doesn't matter? I think there's a lot of complexity baked into this question!

+1 for having a kid in NYC: when your baby wakes up at 6 am on a Sunday you can take them to the marathon reading of The Power Broker

I just want to state what it means to be deadlock free HOW HARD COULD IT BE

This was a lot of fun to write! Compilers are so cool and it’s pretty amazing that interns can do major work on them. blog.janestreet.com/writing-…

Here's the actual link: signalsandthreads.com/tradin…

A new Signals & Threads! Come for the discussion of integrating LLMs into chat trading applications, stay for the harrowing story of Chloe's participation in a harrowing 10-day horse race in Mongolia.

This is a really balanced conversation, and bonus appearance of Eliza, the world's stupidest AI, if you want to feel old
I got to host a great panel on AI timelines: @DKokotajlo, @EgeErdil2, @Jsevillamol, and @RyanGreenblatt. It was an interesting event, a kind of follow-up to the old FOOM debate we hosted 15 years ago!