I showed these guys my code. And they made it unfathomably fast The entire benchmark now takes the same time as 1 task before shikhar @vitr

Damn I did not expect this this year. Noam Brown is an ablation study possible that 100% ensures these problems aren't in training distribution? Would help color expectations for scientific discovery I understand its hard with the vast training data but extraordinary claims

🎯🎯🎯 It would be great to see AI papers adopt such rigor in benchmarking (and clearly explain the experiment setup!). Details are conveniently omitted for hype poasting, but **nuances really matter** for a clear understanding + honest discussion of capabilities.

looking for a student researcher September-December. working w/ me+Mayalen Etcheverry, and alongside Alex Mordvintsev+Ettore Randazzo+others in Pi, on either NCAs or Computational Life. must be enrolled in a degree granting program. fill out the interest form by EOW: docs.google.com/forms/d/e/1FAI…

Anyone in Chennai in the next 4 days? Let's meet!

I think working on great things is the only thing worth doing in life. Here be great things.

we were given universal function approximators, it is our god-given right to fit the models to anything and everything

Sam really knows how to hype up product releases. Now I'm crazy curious about gpt-5+

The 3 Body trilogy is so good. Especially the dark forest

I'm tired of seeing Hierarchical Reasoning Model on every X thread. I don't think its great on ARC Nobody else seems to have actually gone into the weeds so I'll do that over the next few days

Called it x.com/evilmathkid/st…

wow maybe we do need to find alternatives to backprop

great things pt 2

In general, I don't think there are any big secrets in AI research today. The only one I'm absolutely clueless about is how OpenAI trained the IMO model if they didn't use any formal verification at all. Any guesses?

If all science/math problems become solvable by AI, what differences would remain between humans and AI? Building AI that can set its own goals seems to be the next obvious step. But how to do this? How do humans do this?