5M-parameter action model ANA predicts human Magic picks with 90% accuracy in 12ms on CPU
Scobleizer · x · 2026-08-05
Hugo Thomel's team introduces ANA (A Normal Adversary), an action model for Magic: The Gathering that mimics human play. With only 5M parameters, it runs on CPU, making decisions in 12ms and predicting human picks with 90% accuracy. The team sees this as a step towards people-shaped policies, with future applications in robotics. Try it now.
More from Models
- Nous Research Co-founder: Model Sycophancy is a 'Reward Hack', Not Loyalty — petergyang · 2026-08-05
- Security Expert Mocks AI Eval Labs Over Basic Cybersecurity Flaws — nptacek · 2026-08-05
- Uncensored Qwen3-VL-32B Model with INT8 Quantization Hits Hugging Face — ethanfel · 2026-08-05
- DeepSeek Lags in Vision Productization Despite Strong Research Roots — teortaxesTex · 2026-08-05
- Testing Grok's Topological Identity Theory: Local Correction ≠ Durable Learning — tallmetommy · 2026-08-05
- Kimi K3 ECI 157 Beats V4-Pro; DeepSeek V4-Flash 153 Second Open Model — Jsevillamol · 2026-08-05