today I pushed @NVIDIAAI personaplex 7b to it's limits. for the @huggingface x @Gradio hackathon with @NVIDIAAI , we made an app that uses personaplex to teach (by passing in context from nemotron 3 omni), and uses a finetuned nemotron 3 omni to generate visuals that show in real time as the model speaks, made just for you!
tried to get it on huggingface but they keep 503ing when I tried to upload
was really fun to use @modal, @OpenAI Codex, and figuring out clever ways to use @NVIDIAAI's personaplex and nemotron models in novel ways!
found a store that had 2 left at 10k and rushed to buy
now 3 of them total! (will try and get 1 more for a full 4x node)
if you guys have any suggestions for quants/post-trains I should try and attempt, lmk (will be open sourcing a lot!)
really fun time at the autonomous healthcare hackathon
had to leave a bit earlier than I wanted to but it was really cool!
thanks @xai for the shirt & water
all of it in just a couple days!
additionally: i measured the information a controller needs for this manuever, turns out it's around 250k in any basis (knots, fourier, neural net weights).
this makes yacine's 1m params not seem like bloat (which is what I initially thought when I saw it lol)