Video wird geladen...

Video konnte nicht geladen werden

Zur Startseite

~~RL project progress~~ wrote a FAST (88M steps/s) parallelized env in jax from scratch. here's its renderer:

14,149 Aufrufe • vor 1 Jahr •via X (Twitter)

7 Kommentare

Profilbild von danb
danbvor 1 Jahr

i used LLMs to write the renderer btw. im learning RL not openGL the core jax env under the hood is what matters. its steps/s == how fast I can train a model on it SoA and jax.jit do the heavy lifting here. future optimizations: replace loops with scans, run on a cloud gpu

Profilbild von kunal
kunalvor 1 Jahr

very cool who is the agent and whats its learning objective in this sim?

Profilbild von danb
danbvor 1 Jahr

right now there is no agent, just the parallelizable env and renderer the agent will be a planet whose objective is a bit complex. but essentially its goal is to guarantee survival

Profilbild von attentionmech
attentionmechvor 1 Jahr

looks so coooool !!

Profilbild von kache
kachevor 1 Jahr

very cool

Profilbild von _doola
_doolavor 1 Jahr

88 million steps/s?? holy shit

Profilbild von Keith Tsang
Keith Tsangvor 1 Jahr

@yacineMTB Impressive work! Building a FAST parallelized env from scratch is no small feat. What's next on your agenda? 🚀 #Innovation

Ähnliche Videos