4nshh
28 posts

4nshh
@4n5hh_
Making clankers smart at https://t.co/wRRMAWiDYZ home @SRAVJTI
Mumbai, India Katılım Haziran 2025
52 Takip Edilen13 Takipçiler

Reproduced ABC end-to-end on a single A100-40GB the released code targets 8×H200; I adapted it (gradient accumulation, single-GPU path, configurable checkpoint init) and ran 6 trainings + 7 evals cleanly.
GIF
sujal@sujalDweD
hmm so baseline, 75k checkpoint > 10% success rate for 6/6 grasp-and-place > works best for 5/6 grasp-and-place currently of the shelf, no fine tuning. next we'll fine tuning with correct learning rate, lets see
English
4nshh retweetledi

@kunalswami189 @HaiyuWu1 @lucyjcai General intelligence imo will be a mixture of various components
But things like : long horizon prediction, collapsing multiple possible futures into a single Gaussian prediction will defo make it so tht isn't the end all answer to everything
English

@kunalswami189 @HaiyuWu1 @lucyjcai > We get a shiny new toy (VLAs few months ago) play with it
> Think of it as the answer to everything
>Discover it's limitations
>Jump onto the new thing
English

@Biggest @NotRealMusic Yep was thinking the same, miseducation has more width I think but then again illmatic was created by a 19 year old Nas
Can't pick both are legendary works of art
English

@NotRealMusic Illmatic or Miseducation. Hard to pick between those two for me.
English

@SaiMandhan Either this kid jumps to the new hott buzzword or was born to rule the world, we shall know once the bubble pops
English

Whenever I've faced rejections in life, i remember the fact that PPO was rejected @NeurIPSConf cuz apparently it wasn't that much of an upgrade from baselines.
Funny how PPO has been the baseline to beat for almost half a decade now
Absolute chads @fjwolski @johnschulman2
English

















