Qwen 3.8 27B on my laptop can interpret model weights, too

Who Should Read

LLM app developers curious how far local models can reason about raw model weights

Core Mechanics

• Raw weights reveal what a net was trained for
  • The paper claims GPT-6 Astra can tell what a network was trained to do from raw weights alone.
  • The authors call this Weight-Space Programming.
• The paper mixes two separate topics
  • One is reading weights without tools.
  • The other is tool-armed models improving other models.
  • spaCy author Matthew Honnibal called the conflation misleading.
• The test used four 10x10 grid games
  • Weights of RL policy networks for Breakout, Pong, Pac-Man and Freeway went into the prompt, asking which game it was.
  • The model had no tools and one response.
• Local Qwen got 3 of 4 right
  • Pac-Man took 46m 12s at 72% confidence, Pong 1h 4m 54s at 75%, Breakout 52m 25s at 88%.
  • It failed Freeway and guessed a Pac-Man-like game.
• Running locally needed setup
  • The model is 17GB and needs a long context window in LM Studio.
  • The default 8192 context errored and the timeout was raised to 900 seconds.
• The author rejects overblown takes
  • The reasoning traces were nearly incomprehensible and a skilled human with pencil and notepad could get there.
  • He does not think the age of source code is over.

Community Reactions

• Interpretation is not the same as correct
  • One comment used a Shakespeare analogy: calling for an interpretation does not make it correct.
• Shallow network objection
  • One commenter said the network is too shallow to be hard to decode.
• Hands-on report with a custom network
  • One commenter said Qwen 3.8 27B worked out the purpose of their 5x5-to-2x2 upscaler network via code and numeric experiments.
  • It later improved the weights, but by writing scripts, not by reading weights directly.
• Need tooling to verify changes
  • One commenter said that if models edit models, we need plenty of tooling rather than trusting updates.

How to Apply

• Run the public prompt on a local model
  • Today, use the llm CLI to run the repo's Pac-Man prompt on a local model.
  • Set a long context in LM Studio and raise the timeout.
• Verify against known ground truth
  • When a model explains weights or architecture, compare with the known answer.
  • Do not trust the confidence score without checking.
• Evaluate with and without tools separately
  • When designing evals, split conditions by whether code execution is allowed and measure both.

Try It

# Paste into a terminal (public prompt from the paper's repo, using the llm CLI)

pip install llm llm-lmstudio

# Load qwen/qwen3.8-27b in LM Studio with a long context, then run
LMSTUDIO_TIMEOUT=900 llm -f https://raw.githubusercontent.com/RikiyaT/lm_understand-logs/refs/heads/main/exp1_which_game/runs/pacman/prompt.txt -m lmstudio/qwen/qwen3.8-27b

# Next: swap in the breakout, pong and freeway paths

Terminology

Related Papers

• Google will turn off search for schools that reject AI

  Reports say schools turning off AI also lose Google Search, and commenters are split

• The 'Astra can read and write weights to improve Qwen' post is misleading

  Astra's Qwen gain used Python and model runs, not just raw weights

• Launch HN: Momentic (YC W24) – Agents that catch bugs by using your app

  Mo, an AI QA agent that explores your app, found about 269 labeled bugs on QA Bench versus 86 for the next best.

• We're releasing the weights of Mistral Large 4

  Mistral released Large 4 weights, claiming top open-weight scores on coding benchmarks

• Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5

  Gemini 4 Carbon takes #1 on Intelligence Index v5 with 59 and is second in coding agents.

• Gemini 4 Carbon

  Google unveils Gemini 4 Carbon, a cybersecurity-focused model, but access is limited to select cyber defenders.

• Dayjob: Tax – evaluating AI agents on 50 professional tax assignments

  On 50 professional tax tasks, the best AI agent passed only 10.8%

• Anthropic S-1

  Anthropic filed an S-1 to list Class A shares on Nasdaq as 'ANT'.

Related Resources

• https://simonwillison.net/2026/Nov/3/qwen-can-interpret-model-weights/
• https://github.com/RikiyaT/lm_understand-logs