Astra IDs Breakout, Pong, Pac-Man and Freeway from raw model weights alone
burny_tech · x · 2026-11-03
rikiyatakehi shares an interpretability demo: feeding Astra the raw weights of four game policies one at a time. All four share the exact same architecture, inputs, and actions, yet with no tools or options provided, Astra correctly named all four games: Breakout, Pong, Pac-Man, and Freeway.
More from Fun
- AI audit finds physical therapist who billed Medicare for 9,006 hours in an 8,760-hour year — SumitGup · 2026-11-03
- 'Is it the homage?' legally-distinct game character chart cracks up Zed's CEO — zeeg · 2026-11-03
- Jack Clark's weekly Tech Tales: children play god in 2040s city-building game — jackclarkSF · 2026-11-03
- Brave, the 'lightweight' browser, roasted over massive RAM usage — gnukeith · 2026-11-03
- Joke test: nudging Qwen's persona takes 0.69% weights, others' safety filters flagged it — dejavucoder · 2026-11-03
- AI Video Fails Double Feature: Sleepwalking Kid and the Runaway Coffee Machine — nastyalarm · 2026-11-03