GPT-6 Astra beats NetHack on third try, first recorded LLM agent ascension
emollick · x · 2026-09-25
Ethan Mollick calls it a startling achievement: GPT-6 Astra beat NetHack — apparently the first recorded LLM-agent ascension. Blogger Kenny documents the run: playing as a dwarven Valkyrie on the Hardfought server via terminal, the model ascended on September 21 after 37,140 turns.
- NetHack tests the gap between knowing and doing: models explain mechanics perfectly but fail execution on stale maps and open menus
- The author's best January handcrafted harness only reached dungeon level 10; the winning agent built its own harness with zero human code
- Full server dumps and three runs are available; BALROG benchmark covers LLM NetHack results systematically
Related event: GPT-6 Astra Becomes First LLM Agent to Beat NetHack(3 posts)→
More from coding & agent
- Researcher: Time Series Will Hand Your Ass to You, Even in Agent Orchestration — RexDouglass · 2026-09-25
- Octen Claims 77 Quality Score, Pitching Cheap Fast Search for Research Agents — kimmonismus · 2026-09-25
- AI Search Startup Octen Promotes $1 Per 1,000 Web Search Calls for Agents — kimmonismus · 2026-09-25
- IterSynth: Role-Decoupled Search Agent Beats Prior 8B Peers by 4.2% — ZJU-REAL · 2026-09-25
- PUBG Ally: A Voice AI Teammate Trained on 39k Real Gameplay Sessions — Beomsoo Kim · 2026-09-25
- AgentKernel: A Trust-Native Operating System for AI Agents — Zhenhua Zou · 2026-09-25