Kimi K3.1 reads raw network weights and infers the policy was trained on Breakout
teortaxesTex · x · 2026-11-03
Developer bigeaglexd tested Kimi K3.1 with an exp1whichgame task: feed the model a policy network's raw weights and ask what game the policy was trained on — a task type the model was never trained on.
The published reasoning trace shows K3.1 manually parsing the W1 matrix row by row, mapping 500 rows via the cell×5 index scheme, spotting high-weight input channels, inferring a horizontal-paddle control structure from output unit patterns, then ruling out Snake, Pac-Man, Pong, Space Invaders and others based on the reward layout (fixed 3-row reward bank at top, single bottom agent, horizontal-only actions). It concluded "Breakout" with 97% confidence.
teortaxesTex quipped that "the advantage of open models is that you can read what they think" — and joked "don't show this to Yud." The demo is a striking sample of long-chain reasoning interpretability and rare weight-level comprehension.
More from Fun
- AI swarm of "neural bees" communicates in neuralese, showcasing agent ecologies — Sauers_ · 2026-11-03
- AI detector flags an AI-written alignment essay as 100% human — amplifiedamp · 2026-11-03
- Professor fact-checks Musk: vacuum is the best insulator, space datacenters won't happen — teortaxesTex · 2026-11-03
- Human mops while their paid AI agent renders an After Effects project solo — ZeroStateReflex · 2026-11-03
- Vibe check: 'Nothing beats being two drinks deep in a Waymo blaring music' — generativist · 2026-11-03
- Devs flip on code bloat: from 'code is a liability' to praising AI slop spitting 50k unreviewed lines a day — rickasaurus · 2026-11-03