Willison reproduced the weights trick on a laptop with Qwen 3.8 27B

Simon Willison took the buzzy paper about GPT-6 Astra interpreting model weights directly, ran the same idea against Qwen 3.8 27B on his laptop, and says the small model got there too. It took more than 45 minutes.

Simon Willison

@simonw

X

I tried it with Qwen 3.8 27B on my laptop and found that, after 45+ minutes, that model could achieve the same thing

The paper's claim, as Willison describes it, is that a model can directly interpret model weights handed to it as prompts full of floating point numbers. That is the part that made people sit up, and it is the part he says is not exclusive to a frontier model. We covered the original result earlier.

What the reproduction actually shows

Willison frames his post as deflating the buzz rather than confirming it. He is not saying the paper is wrong. He is saying that if a 27B model you can run locally reaches the same outcome, given enough time, then the interesting thing is the capability, not the lab that demonstrated it.

That distinction matters for how you read the next round of papers like this one. A result presented as a property of a specific frontier model invites a specific conclusion, that somebody has pulled ahead. A result that replicates on a laptop invites a different one, that the behaviour was probably sitting in models of all sizes and nobody had pointed a prompt at it yet.

The 45 plus minutes is worth holding onto. Willison does not report what the frontier model took, so the honest summary is that the small model matched the outcome, not the speed. Whether the gap is seconds or hours is not in his post.

This is one engineer running one informal test on one machine, which is exactly the weight he gives it. No pass rates, no repeated trials, no comparison harness. It is a data point of the most useful kind, cheap to produce and easy for anyone else with a laptop to check, and it arrived fast enough to land while the original claim was still circulating.

For working engineers the practical read is to treat "only the big model can do this" claims as unverified until somebody points a local model at the same prompt. In this case that took one person, one laptop, and under an hour of compute.