ohh, so this is what to expect today. github models leak - archive.is/2025.08.07-035…

can you see what i am seeing? "a comedy in three acts"

the remaining 50% deception is in the way the charts try to mislead.

find everyone who vouched gpt3 --> 4 is same as gpt4 --> 5 and block.

even oai devs aren't shilling codex cli on the stream.

gpt5 is genuinely better than o3, not too far off from opus, at least for my pytorch and research scenarios. but the amount of user clutter (models x modes) remains constant

Muon + Stiefiel kexue.fm/archives/11221 solved the open problem at docs.modula.systems/algorithms/man… from Jeremy Bernstein . cc leloy!

re: pure RL from scratch writing to make things as intuitive as possible and compress almost everything i thought of over a long time horizon. no yap sessions today, gpt5 expectations already hijacked it yesterday.

somebody create a polymarket on the auto router being rolled back? what do you think?

Results are in from our internal gpt-oss hackathon! A quick roundup of what we found: (1/8)

adding results and "future scope/followup" pending. it is now closer to a paper than an article. touching grass today. would come back, share and taste sleep.