(trl trainer code below) how do you scale language model search? can a language model be guided towards solving new problems that it usually cannot solve purely through guided resampling without training the model whatsoever? turns out the answer is sorta yes. this has important

Freeman Dyson on the Greening of the Galaxy: "the galaxy is very boring as long as it's dead, you look at all these planets and stars we've discovered, they're beautiful, but still they're sort of boring. So I'd like to bring life to the universe and bring the universe to life"

It has been an honor and a privilege seeing this team work. So excited about Gemini 3 and Antigravity and to see what people build with this technology.

this is so cool i think we’re not far from factorio style chip fabs

the tech my company built and was brought over to Google through an acquisition late last year is now live ! https://t.co/T9BrvZv7Db 🚀 very proud of my old team and will be working doubly hard to make Gemini the best it can be so it can power this tech for devs worldwide.

Staying at a hotel in the Bay Area. Doing an intense gym sesh. A guy walks in and it’s my good friend @hsu_steve.

i’m sad to inform you that being extremely methodical about your sleep and treating it like an RPG with bonuses and modifiers does, in fact, work very well. not sure how cavemen just naturally got good sleep tbh.

I’m excited to announce that I’ve joined this amazing team of some of the most brilliant people I’ve had the fortune to meet, working on Gemini Data, LLM Reasoning, and AI for competitive math and code. Let’s go!

maybe conical breathing

box breathing is a bad visual metaphor that never clicked for me allow me to propose "circle breathing" breathe in –– circle expands hold breath in –– circle radius constant (large) breathe out –– circle contracts hold out –– circle constant (small)

there are approximately fifty thousand papers that can be distilled to E [ \nabla . ] \neq \nabla E[.] please check your order of operations kids and enjoy the weekend!

one simple way to cut down on hallucinations for responses from an agent you use (eg for coding) is to ask it to construct a proof of its claim easier to falsify something in the proof than the bigger claim for which you might not have context as a user

newton was on another level, it took a nobel prize winner in physics several years to digest his principia into something readable by modern physicists. before newton there was only darkness.

@willccbb @johnschulman2 https://t.co/rRlB2kLSaT discrete feynman path integral from colin morningstar's review paper who i worked with on this stuff as an undergrad!

@willccbb @johnschulman2 hopefully more on this soon

@willccbb @johnschulman2

@willccbb @johnschulman2