In order to support dynamically on and off-boarding compute to the run, we introduced a new distributed abstraction `ElasticDeviceMesh` which manages the resizing of process groups without requiring a cold restart. More info on this, along with some other interesting innovations we did to pull this off in our blog post:

This seems like a good idea more people should be aware of. The choice of hyperparameters for the relative heights of the correct and wrong piece are interesting though. Is there an accuracy difference between Reward B and C? If not, does that mean you should always pick larger correct reward > wrong?

@nooriefyi you dont differentiate through the reward?

@tenderizzation sacril3ge

@Ar_Douillard has schmidhuber tried to claim your work yet?

@cloneofsimo @main_horse might be interesting for @wanchao_ to comment abit on this

@Ar_Douillard @kalomaze do they scale outside of google?

@tenderizzation this ham is pretty cunning

@sachoslks Yup! I think open source should allow for derivatives that are closed.

@afurgs bigger and better than before!

@bronzeagepapi We will be sure to inform you guys if it discovers any quantum mechanic glitches during the run 🫡

RT @MatternJustus: The path to our first reasoning model consists of three steps: 1. Generating cold-start reasoning data 2. SFT on cold-s…

@shxf0072 has there been any papers using GRPO for non LLM tasks? why should it only be great at optimizing LLM?

@teortaxesTex We need more memes! Memes allow people to say what they actually want to say without the full commitment of having said it explicitly

@manveerxyz lfg 🫡

RT @srush_nlp: Got talked into giving a DeepSeek talk this afternoon Not sure I have anything new to say here! But…

@CamutoDante i dont quite get what u mean here. intermediate activations are deterministic given the same input sequence and model. there isnt much a malicious or a benevolent actor can do to alter them

RT @huseinzol05: @mesolitica released Malaysian TTS models including dataset! Special thanks to @jackminong and @PrimeIntellect for the com…

@gabrielchua_ @PrimeIntellect @aiDotEngineer @swyx sent u a dm :)

@huseinzol05 tldr; hash activations good