Epoch: Astra beats most humans at card game but shows little continual learning
teortaxesTex · x · 2026-09-08
Epoch AI's experiment found that after a game-breaking card was banned, Astra still scored higher than most human players, improving significantly by exploring and building strong decks. Yet it showed little sign of continual learning in turn-to-turn tactical play, and Epoch increasingly finds Astra/Fable don't learn as much as humans allowed multiple attempts—suggesting future LLMs may learn on the job far better.
More from Models
- Building a Model for Fancy Math and Code Is Easier Than One That Talks Like a Real Person — rickasaurus · 2026-09-08
- Bigger Context Windows Solve Fitting, Not Memory: Chroma's Context Rot Shows Accuracy Drops Early — Efficient_Joke3384 · 2026-09-08
- Codex Pro users report quota burning noticeably faster: 6% gone after one hour of use — yuntiandeng · 2026-09-08
- ProMax sweeps smartphones while AI model names go mythic — why the two industries diverged — APPSO · 2026-09-08
- What can you actually do with a small language model? Three real use cases — CurieuxExplorer · 2026-09-08
- Astra appears to access Sora via its CoT, one-shotting a generation — mallow610 · 2026-09-08