@github.awesome: Andrej Karpathy just open-sourced autoresearch. You give an AI agent a real LLM training setup, point it at an instruction file, and go to sleep. The agent autonomously modifies the architecture, runs a fixed 5-minute training loop on a single GPU, checks validation loss, and decides whether to keep or discard the changes — then repeats the cycle all night. By morning you wake up to a full log of automated experiments and an optimized model. #github #opensource
Going to sleep and leaving an LLM running is a very expensive idea
2026-03-07 16:08:09
18
AaronMuscles :
He wants a Openclaw deal
2026-03-08 00:01:05
4
halkpaw :
Claude optimized my research by just deleting data, using the wrong modeling technique intentionally, or intentionally ignoring the spirit of the work. Would this really work like go to sleep and wake up and it’s done?
2026-03-07 15:08:14
2
roland myerson :
we shipped an agent that does this end to end, latency gains are real once you get the orchestration loop tight. the trick is keeping the reward signal lightweight - don't let the agent overthink eval, just ship iterations fast
2026-05-01 22:30:43
0
Kevin Rehak :
Wake up to $100 ai bill
2026-03-08 02:29:39
9
lutzy :
Auto research?
2026-03-08 04:23:53
0
jdDoubleWide :
huh?
2026-03-08 23:45:46
0
🌸Just D🌸 :
What’s the file called
2026-03-08 19:05:37
0
Thomas - sourcetms | AI / Tech :
Crazy release to be honest
2026-03-08 19:32:16
0
reverb.tv :
When I woke up, it had already eaten up all my tokens
2026-03-07 17:50:28
1
Lisa Deep in the ❤️ of Texas :
😁😁😁
2026-05-24 05:45:50
0
To see more videos from user @github.awesome, please go to the Tikwm
homepage.