Yoav Goldberg tempers Astra weight-reading hype: it used a simplified 10x10 Breakout
yoavgo · x · 2026-11-04
NLP researcher Yoav Goldberg boosted a thread by Shauli Ravfogel explaining how Astra "reads neural network weights," while adding a key caveat: he assumed the experiment ran a real Breakout game with graphic input through a CNN, but it was actually a simplified 10x10 version that removes the main complexity — "this is indeed much less impressive."
Context: deep network weights are famously opaque except for the first and last layers, which map to known inputs and outputs; the thread examines whether Astra is reading exactly those boundary layers rather than something more mysterious.
More from Research
- Median human turn-taking pause is -120ms, so voice AI's 500ms brag still misses how people talk — alexisgallagher · 2026-11-04
- ParallACT: when every dev has their own coding agent, coordination still breaks down — EliasEskin · 2026-11-04
- 3DV Conference Archive: 1,351 Papers Across 17 Topics from 2013 to 2026 — ducha_aiki · 2026-11-04
- Locus: autonomous discovery system adapts Boltz-2 and searches antibody sequence space with ESM — yusufroohani · 2026-11-04
- TRL adds native support for MOPD, the distillation trick frontier open models use to merge RL experts — pcuenq · 2026-11-04
- Stanford preprint: CD19-CAR T cells clear EBV+ autoreactive B cells in autoimmune disease — ahandvanish · 2026-11-04