15M-param RL agent beats Craftax boss 44% of the time for a cent; GPT manages 1% at $83
neuroecology · x · 2026-10-09
The hyperparticle team used an agent meta-harness to train a 15M-parameter RL agent that solves Craftax, a 2D Minecraft-inspired game, without any LLM prior knowledge — something experts doubted a small net could do.
- The 15M agent beats the final boss 44% of the time, in 10 seconds and for about a cent per run
- An LLM-driven approach (GPT) won only 1% of the time, taking 24 hours and $83
Emergent strategies are notable: the agent jails monsters in tiny stone cells and leaves them there. The model trainer is open-sourced on GitHub.
Related event: 15M-Parameter RL Agent Beats Craftax for One Cent, Outpacing GPT(3 posts)→
More from Fun
- Dashboard tracks community progress on OpenAI problem 109 (integer multiplication) — RexDouglass · 2026-10-09
- Kevin Roose's next move gets the p(doom) jokes going — ashleymayer · 2026-10-09
- While waiting for his AI coding limit to reset, a dev went back to hand-writing Go — haydendevs · 2026-10-09
- AI founders dunk on ARR: like the Holy Roman Empire, it may be neither annual, recurring, nor revenue — sarahdrinkwater · 2026-10-09
- tszzl rebuts "future intelligences will love humans" with the megafauna extinction analogy — danfaggella · 2026-10-09
- Satirical timeline: humans get pushed out of code review, sales, then everything — danfaggella · 2026-10-09