🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

There are many algorithms for constructing pre-training data mixtures—which one should we use? Turns out: many of them fall under one framework, have similar issues, and can be improved with a straightforward modification. Introducing Aioli! 🧄 1/9

India just needs more traffic signals and more people who will follow those signals. Unnecessary slow down because people are going any which way at intersections.

Are you doing RL and at #ICML? Let's chat. DM me if possible!

Nice, good job Devin!

This is genius

We are organizing a dinner today for researchers in RL and Data. There are a limited number of slots remaining. Please DM me to join.

Some more drama

Open Thoughts delivers again. Congrats team for a small but powerful reasoning model. Writeup: open-thoughts.ai/blog/ot3_small

I know OpenAI has two employees one called Rhythm and another called Lyric.

This is incredible and congrats to the team, but I don't know why half of twitter is surprised by this result.

So it's time to start doing multi agent RL.

in case you missed it ;)

The Veo videos are so much fun. Someone should create a YouTube for Veo videos..

which model are they going to drop? maximally truth seeking model?

Interesting read

Anyone knows adam?

Cursor thinks my name is jose

🚀 GSPO: Group Sequence Policy Optimization — a breakthrough RL algorithm for scaling LMs! 🔹 Sequence-level optimization — theoretically sound & matching reward 🔹 Rock-solid stability for large MoE models — no collapse 🔹 No hacks like Routing Replay — simpler, cleaner

Everything is from scratch at Google. I mean even the hardware!

TIL that Jensen commissioned a chip specially for Carmack, for Quake. And since that chip worked really well, Carmack told his followers to use Nvidia chip for quake, which helped Nvidia a great deal.