Simon Willison Says Qwen 3.8 27B on His Laptop Matched the Viral GPT 6 Astra Weight Interpretation Result
There's a buzzy new paper about GPT-6 Astra being able to directly interpret model weights (prompts full of floating point numbers) - I tried it with Qwen 3.8 27B on my laptop and found that, after 45+ minutes, that model could achieve the same thing
Context
If a small open model can read model weights like GPT 6 Astra, the viral result may say more about the method than about one frontier model.
A widely shared paper claimed GPT 6 Astra can directly read and edit the weights of another model, the raw numbers that make it up. Simon Willison (@simonw) is a developer and blogger known for testing AI tools hands on. The Nov 2 deck covered the original claim.
Deep dive
Simon Willison (@simonw) posted on Nov 3 about a buzzy new paper on GPT 6 Astra being able to directly interpret model weights, which he described as prompts full of floating point numbers. He wrote that he tried the same idea with Qwen 3.8 27B on his laptop, and found that after more than 45 minutes that model could achieve the same thing. His post links to a write up on his blog.
The original claim, covered in the Nov 2 deck, was that GPT 6 Astra edited a small fraction of the weights of a small Qwen model to fix reasoning loops and improve math, without normal training. That post drew more than 8 million views.
Willison's test is an informal reproduction by one person, not a peer reviewed study, and the full method and results are in his blog post. Still, it changes how the original result reads. If a mid sized open model running on a laptop can do the same kind of weight interpretation, the ability may not be unique to a frontier model, and the setup of the task may matter as much as the model doing it.
The post passed 150,000 views and spread among researchers who had debated the original paper.
What to expect next
Watch for the paper's authors to respond, and for other researchers to repeat the test with different open models and tasks. Careful comparisons of accuracy, time and cost between GPT 6 Astra and smaller models would settle how special the result is. If the method works widely, expect new tools for inspecting and editing model weights directly. Willison may also publish follow up tests on his blog.