Simon Willison · 2026-11-03 · notable
Simon Willison — Qwen 3.8 27B reads model weights on a laptop
In short
Simon Willison ran Qwen 3.8 27B locally and it named the game behind three of four policy networks from raw weights. It failed on Freeway. Kimi K3.1 Pro got Breakout right in about 10 minutes.
Simon Willison · 2026-11-03 · 3 sources checked
A local 27B model can do a slow version of the weight-reading task from the GPT-6 Astra paper.
What is it?
Simon Willison tested whether a local model can reproduce the no-tools weight-reading task from the paper "Sparks of Weight-Space Programming". The task is to name which of four games (Breakout, Pong, Pac-Man, Freeway) a policy network was trained on, using only its weights as text.
How does it work?
He ran a 17GB Q4_K_M build of Qwen 3.8 27B in LM Studio through the llm-lmstudio plugin. Qwen named Pac-Man, Pong and Breakout with 72% to 88% confidence, taking 46 to 65 minutes each, and got Freeway wrong. Kimi K3.1 Pro, run via OpenRouter, identified Breakout in about 10 minutes.
Why does it matter?
Willison argues the no-tools weight-reading question is more interesting than the weight-editing one. He also says a patient human could probably do the same, so the result does not show machines beyond comprehension.
Who is it for?
ML researchers and local-model users
Try it
llm install llm-lmstudio