Jonathan Rahn
jrahn
AI Lab Lead
Teams
Intro
Jonathan Rahn
AI Lab Lead, Drees & Sommer
Research Focus
Exploring transformer-based strategic reasoning through chess as a testbed, demonstrating that language models can develop sophisticated game-playing capabilities without traditional search algorithms. In collaboration with LAION, developing models that challenge fundamental assumptions about how AI systems learn strategic thinking.
Key Projects
🏆 ROOK-CLF-9M - Classification Chess AI
- 49% action accuracy on ChessBench dataset
- 57% checkmate-in-one accuracy (BIG-bench)
- 9M parameter LLaMA-based decoder reproducing Google DeepMind's searchless chess methodology
🧠 RookWorld-LM - Unified Agent+Environment
- 32.1% checkmate-in-one accuracy (beats ChessGPT-Base 26.5%)
- 99.9% environment simulation accuracy
- Single transformer handling both chess policy and world modeling
- Enables closed-loop self-play without external engines
⚡ ROOK-LM - Chain-of-Thought Reasoning
- 22.2% action accuracy with detailed reasoning traces
- 24.4% checkmate-in-one accuracy
- Trained on 40M positions with Stockfish annotations (6B tokens)
Technical Contributions
- Novel Architectures: Unified world modeling in transformers
- Strategic Tokenization: Custom FEN representations for consistent attention patterns
- Dataset Engineering: 40M+ positions with Stockfish annotations on supercomputing infrastructure
- Open Science: All models, datasets, and code publicly available
Research Impact
Published at LAION Research Notes with collaborators from LAION/JSC and Tokyo Tech/Sakana AI. Contributing to democratization of strategic AI research through open models and reproducible methodologies.
---
Exploring how language models can learn strategic thinking through next-token prediction on appropriately structured data.
Reports
Projects
Links
Activity
Mon
Wed
Fri
NovDecJanFebMarAprMayJunJulAugSepOct
Runs
Name
Project
State
Created
Loading...