Laptop LLM identifies training games from raw weights

A 27B parameter model running on a laptop successfully identified the training game for 3 out of 4 Atari-style games using only raw weights, a process that took nearly three hours. The model failed to identify "Freeway," mistaking it for a Pac-Man-like game, which the author notes is informative as it demonstrates the model's tendency to categorize unfamiliar patterns with the closest known concept. This experiment, a reproduction of a prior weight-space programming paper, highlights that mid-size open models can extract semantic labels from weights, though it does not prove true understanding or negate the need for code execution in practical applications. AI

IMPACT Demonstrates that even smaller, locally run models can extract semantic information from raw weights, suggesting potential for new forms of model analysis.

RANK_REASON The item describes an experiment reproducing a prior research paper on weight-space programming, detailing the methodology and results of an LLM identifying training data from raw weights. [lever_c_demoted from research: ic=1 ai=1.0]

Read on dev.to — LLM tag →

• 27B model
• Atari
• Breakout
• exp1_which_game
• Kimi K3.1 Pro
• lm_understand-logs
• OpenRouter
• Pac-Man
• Pong
• Qwen
• RikiyaT
• Simon Willison
• Weight-Space Programming

• model release
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →