Skip to content

Comment on Day 1 of ARC-AGI-3parent

Comments

Also think about the program-synthesis approach proposed by Poetiq.ai. python programs are being generated and evaluated against previous examples. Then in-context learning is done programmatically via prompt concatenation. If you can "score" online the working and non working examples, then you have a very strong reward signal.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.