LittlePly open chess agent: 87M from-scratch model, searchable harness, free to play
andreisavu · x · 2026-11-11
Andrei Savu publishes the LittlePly project site and previews act two: exploring learning to search through program generation.
Project highlights:
- CM-14: an 87M-parameter model trained from scratch on human games, outputting move policies and win/draw/loss estimates
- The harness executes: legality checks, uncertainty-guided search, caches, bounded tools, clock-aware execution
- Play the full deployed system on Lichess — not the unassisted browser demo
- Three paths: a five-chapter learning course (policy to agent), an open research notebook (what improved, what failed), and the architecture source
- Designed as a worthwhile opponent: surprising enough to learn from, fallible enough to beat
Related event: LittlePly: Open-Source Chess Agent Combines 87M Model with Search Harness(4 posts)→
More from coding & agent
- 'Agentic AI for Low-Level Kernel Fusion Is a Dead End': Devs Clash Over Megakernels — idanbeck · 2026-11-11
- testingcatalog daily: Gemini 4 Flash due today, GPT-6.2 Sol/Astra this week, Gemini Robotics 2.5 hits 96% — testingcatalog · 2026-11-11
- Cloudflare announces OpenAI Codex can now run on Durable Objects like it runs on your laptop — tekbog · 2026-11-11
- What a project home should hold: decision log, meeting notes, progress and lessons learned — dSebastien · 2026-11-11
- Engineer hid a pirate-talk prompt injection in his LinkedIn — latest SOTA models now spot the trap — smlpth · 2026-11-11
- Fable 5.2 missing from Grok Bot: Squad founder warns against single-provider AI stacks — tibo_maker · 2026-11-11