Erwann Millon (@erwannmillon) 's Twitter Profile
Erwann Millon

@erwannmillon

making gpus go brrrr @krea_ai

ID: 218061659

linkhttps://soundcloud.com/cieux-ofc/rift-dark-melodic-techno-set calendar_today21-11-2010 09:23:08

290 Tweet

1,1K Takipçi

425 Takip Edilen

Erwann Millon (@erwannmillon) 's Twitter Profile Photo

used to wonder why americans don’t eat fruit and vegetables then realized it’s bc american fruit and vegetables taste like shit

Erwann Millon (@erwannmillon) 's Twitter Profile Photo

don't get how people train models w/ NoPE pretraining a small t2i dit and RoPE converges way faster, giving meaningful results in just a few K steps (samples shown @ 5k)

don't get how people train models w/ NoPE
pretraining a small t2i dit and RoPE converges way faster, giving meaningful results in just a few K steps (samples shown @ 5k)
Erwann Millon (@erwannmillon) 's Twitter Profile Photo

Speed up dit pretraining by changing your timestep sampling over time - start biased toward high noise for quick early convergence - gradually shift to uniform to ensure the model performs well across all steps In the first animation, you can see the train timesteps shift over

Erwann Millon (@erwannmillon) 's Twitter Profile Photo

gpt 5 is ass at coding, but they did a good job post-training for multi-turn conversations for non-technical tasks, it consistently anticipates smart follow-up questions and useful next steps.