Skip to content

Comment on Show HN: AI agents play SimCity through a REST APIparent

Comments

I think the way to see this as the organic process of discovering hard-to-game benchmarks. The loop is:

1. People discover things LLMs can kind of do, but very poorly.

2. Frontier labs sample these discoveries and incorporate them into benchmarks to monitor internally.

3. Next generation model improves on said benchmarks, and the improvements generalize to improvements on loosely correlated real world tasks.

AboutSource Built by g1lg1l

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