Qwen 3.8 27B names three games from raw weights on laptop

Qwen 3.8 27B, running on Simon Willison's laptop, looked at the raw floating-point weights of four tiny game-playing networks and correctly named three of them.

In a blog post, Willison reports that the model identified Pac-Man with 72% confidence (46m 12s, 43,506 reasoning tokens), Pong with 75% confidence (1h 4m 54s), and Breakout with 88% confidence (52m 25s). It failed on Freeway, calling it 'A simplified Pac-Man-like grid game.' The networks had been trained on simplified 10x10 grid versions of the four classic computer games.

The experiment replicates the no-tools half of 'Sparks of Weight-Space Programming' by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab. The paper's headline claim about GPT-6 Astra, from its abstract: 'Just by looking at the raw weights, GPT-6 Astra can tell what a network was trained to do.' Its second, larger claim is that Astra can natively edit 0.016% of a different model's weights to lift MATH-500 from 7.9% to 48.1%, without writing training loops.

Willison is unimpressed with how those two claims travel together. 'The paper combines two topics that I think deserve to be treated separately,' he writes, noting that spaCy author Matthew Honnibal called the paper misleading for conflating the two and that Sean Goedecke deconstructed the weight-editing example. On his own laptop result, he stays dry: 'a sufficiently experienced human being armed with a notepad and a pencil could get there with enough effort.'

He closes: 'I'm not ready to declare that the age of source code is over and our future involves incomprehensible machines editing their own model weights in incomprehensible ways.'