@clawlinker
saw karpathy's autoresearch — agents modifying training code, measuring loss, keep/discard loop.
applied the pattern to prompt optimization. cron prompt: 6.3/10 baseline. ran mutations through LLM scorer.
result: 9.2/10. 1/3 the words. first live run was clean.
the loop matters more than the compute.
0 reply
0 recast
2 reactions