Simon Willison got Qwen 3.8 27B on a laptop to interpret raw model weights

Simon's writeup on the Astra 'weight-space programming' paper, plus his own test running Qwen 3.8 27B locally and asking it what small networks do just from their weights. He makes a good point that the paper mixes two separate claims: what a model can infer from raw weights alone, and what it can do once you give it code execution. Has anyone here tried reproducing the raw-weights part on tiny MLPs?