Qwen 3.8 27B on a laptop reading raw model weights - big deal or party trick?

Simon Willison tried an example from the new 'weight-space programming' paper and found a 27B open model running locally could interpret what a small network was trained to do just from its raw weights. There's already a rebuttal going around arguing the original claims are overstated. For those of us who left big labs partly over closed models, it's interesting that this is reproducible on consumer hardware at all. Does local reproducibility change how much you trust results like this?