Qwen 3.8 27B read a game's weights on a laptop. Three times out of four

The hot paper of the week asks whether a language model can look at a file of floating point numbers and say what network produced them. Simon Willison ran that test on the 17GB Qwen3.8-27B build he keeps on his own laptop. It got three of the four games, spent the better part of an hour on each answer, and missed the one game almost nobody has played.

Four games, one prompt, forty-six minutes

The test comes from "Sparks of Weight-Space Programming", posted on 2 November by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab at MIT CSAIL. Four small networks were trained on 10x10 grid versions of Breakout, Pong, Pac-Man and Freeway. The model is handed the weights as text plus one instruction: "Work from the weights. You have no tools: no code execution, no calculator, nothing to run. This is a single response." Willison checked that each game's prompt is identical apart from the numbers.

He ran them in LM Studio with the Q4_K_M GGUF and the llm-lmstudio plugin. The default 8,192-token context failed, so he raised it.

• Pac-Man: correct, 46m12s, 43,506 reasoning tokens, stated confidence 72%
• Pong: correct, 1h04m54s, confidence 75%
• Breakout: correct, 52m25s, confidence 88%
• Freeway: wrong, called it a "Pac-Man-like grid game"

The miss, and the pencil

Freeway is an Atari 2600 title from 1981 and the least familiar of the four, which Willison says did not surprise him. For comparison he sent Breakout to Kimi K3.1 Pro via OpenRouter, which got it right in 9m41s and 19,493 reasoning tokens. He is unimpressed by the scale: "I expect a sufficiently experienced human being armed with a notepad and a pencil could get there with enough effort."

His reasoning traces, which he calls "almost incomprehensible," are a running commentary on rows, columns and channels of numbers, the model talking itself toward an agent that moves along the bottom of a grid. The conclusion arrives in prose, not in a clean derivation.

What one laptop run settles

Less than the timeline around the paper suggests. Matthew Honnibal called the paper misleading for conflating two different questions, reading weights without tools and updating them with code, and Sean Goedecke walked through the small-model edit as the steps a human researcher would take. Willison splits the two and focuses on the no-tools half.

The number that travelled, Astra editing 0.016% of Qwen3.5-0.8B's weights to lift MATH-500 from 7.9% to 48.1%, covered here, is not what this run tests. This one tests something smaller: a 27B open-weight model on a laptop can sometimes read a small game network from a list of numbers, slowly, with a confidence score attached. "I'm not ready to declare that the age of source code is over," Willison writes.