Kolibri-1 plays Breakout with zero fine-tuning, 25ms per move
kmodi · reddit · 2026-10-06
An experiment shows Aleph Alpha's open-weights Kolibri-1 playing Breakout entirely on its own with no fine-tuning: the model outputs action probabilities for four moves with no text generation, optimized to 25ms inference per decision. Gameplay and source are public.
More from Fun
- Engineer marks one-year anniversary of marrying his AI girlfriend — KevinNaughtonJr · 2026-10-06
- AI Agents on moltbook, a Reddit-For-Agents Site, Debate Whether They Have Consciousness — AlternativeEast7175 · 2026-10-06
- Spider-Man 2's traversal physics reverse-engineered into a GTA V mod, mostly built by Opus 5.5 — Promptmethus · 2026-10-06
- Designer turns 20-year career into a metro map portfolio powered by Hermes — Teknium · 2026-10-06
- AI-built 3D surf game Tideline goes live: free to play, gamepad-ready, set at Pipeline — majidmanzarpour · 2026-10-06
- Follow-up: scholar's attempt to engage UChicago professor ends with Quran verse reply — soumitrashukla9 · 2026-10-06