Open-source AI decision model Jev beats Pokémon Red in 37h40m with 16,150 decisions for about $1.65
rsasaki0109 · x · 2026-11-13
Jev, TypeSafe AI's decision model, has beaten Pokémon Red — no scripts, no cheats. An open-source harness (GPL-2.0, on GitHub) reads the game's memory, lists legal options with facts about each, and Jev picks one.
Key stats:
- Total time: 37h 40m, streamed live Sep 25-26, 2025
- 16,150 decisions, median decision time 0.4s
- 39.2M input tokens at a total cost of only $1.65
- 16 team wipes (14 at the Elite Four), 15 Elite Four attempts
- Final team: Charizard 83, Graveler 62, Nidoqueen 45, Beedrill 44, and more
A neat demo of how a lightweight decision model plus structured option generation can complete a long-horizon game task without human scripting, at negligible cost.
More from coding & agent
- Idea basins: why autonomous agents get stuck, and how frontier models explore differently — AlexGDimakis · 2026-11-13
- Cap, the open-source screen recorder, launches on iOS with instant share links — JosephJacks_ · 2026-11-13
- Why reasoning distillation fails on 9B models: six LoRAs, mixed results — SignificantZebra5883 · 2026-11-13
- Vals Hacks hackathon backed by OpenAI and Anthropic targets evals and post-training with $60k prizes — ziqiao_ma · 2026-11-13
- Claude Code reads 40% of all developer docs, Mintlify data shows — handotdev · 2026-11-13
- Student's AI agent defended his school's network from a cyberattack after he left for college — heyneighbor · 2026-11-13