Positive users praise Humans& open-source 4-bit RL recipe as very cool and interesting for long-horizon training while negative users express sadness about benchmark compression or respond with insults.
Based on 14 visible X reactions from 25 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@ApplyWiseAi @humansand This is exactly what i was wondering. Glad to see simple minded people joining in on discussions that are bigger than they are.
@eliebakouch sad to see a sota benchmark get compressed down to fit two standard deviations
@niloofar_mire @humansand very cool, awesome work
@humansand so gooood
The methodology resolves training instabilities associated with NVFP4-RL.
@humansand very cool!
@humansand Very interested!
When you read a closed source blog pose, and they say "RL systems at scale take lots of engineer work", this is exactly the sort of thing they mean. Really great work by the sand team to go through the hard science and engineering effort to get low precision RL working AND go through the additional effort to communicate it so clearly. Take some time today to read through the blog, there's lots of good tidbits
We have open-sourced our 4-bit hardware native RL recipe, w/ NO degradation compared to 8 bit!! Checkout our blogpost! https://x.com/niloofar_mire/status/2075709107697525116/photo/1 https://twitter.com/humansand/status/2075618383631167692
Proud of the excellent work coming out of our infra team - excited for continued open source collab opportunities! https://twitter.com/humansand/status/2075618383631167692
i'm sad this is not happening near 6/7 of the range instead https://x.com/eliebakouch/status/2075623167683338685/photo/1 https://twitter.com/humansand/status/2075618383631167692
Congrats @humansand on native NVFP4 training recipe, landing in Miles! https://twitter.com/humansand/status/2075618383631167692
Nice work!!! https://twitter.com/saurabh_shah2/status/2075619746566029428
Positive users praise Humans& open-source 4-bit RL recipe as very cool and interesting for long-horizon training while negative users express sadness about benchmark compression or respond with insults.
Based on 14 visible X reactions from 25 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@humansand Very interested!