Simon Willison said Qwen 3.8 27B, a free AI model he ran on his laptop, interpreted another model’s raw weights after about 45 minutes of setup, lowering the barrier to outside scrutiny.
Key Takeaways
Simon Willison said Qwen 3.8 27B running on his laptop could interpret raw model weights
Willison reproduced one of the paper’s ideas locally after about 45 minutes of setup
The paper said GPT-6 Astra edited 0.016% of Qwen3.5-0.8B’s weights and raised MATH-500 from 7.9% to 48.1%
Willison’s laptop test broadened the field by lowering the barrier to replication and outside scrutiny
Willison posted the result and then wrote that he reproduced one of the paper’s ideas locally after about 45 minutes of setup. Willison is an independent developer and researcher whose experiments often act as an early reality check on emerging model claims.
Reading Raw Weights Leaves The Lab Demo Stage
The original paper argued that large language models can inspect and even modify neural network weights directly, and Willison said Qwen could do the interpretation piece on consumer hardware.
That matters because weight space programming gains credibility when an outside developer reproduces part of it without a giant internal lab stack.
Willison highlighted a line from the paper saying GPT-6 Astra could infer what a network was trained to do by reading raw weights, and he pointed to a stronger claim that Astra edited 0.016% of Qwen3.5-0.8B’s weights and lifted MATH-500 from 7.9% to 48.1%.
Also Read: Simon Willison Flags Census Shift That Could Break AI Trend
The Result Is Narrow, But The Implication Is Wide
Willison’s test does not prove all of the paper’s larger claims. It shows Qwen 3.8 27B could participate in weight space programming by reading model parameters as text-like structure.
Neural network weights are the billions of numerical values learned during training and are usually treated as opaque internals that humans do not edit directly except through training code.
If models can reason over them in plain-language interfaces, that changes how developers may debug, audit and patch systems. The idea cuts against an older assumption in AI tooling.
Most model improvement still depends on retraining, fine-tuning or reinforcement loops. Reading Raw Weights suggests a different path, where a model diagnoses another model’s internals and proposes surgical edits rather than full retraining.
From Prompt Engineering To Model Surgery
Developers spent the last two years mastering prompt engineering, tool use and agents.
Weight space programming pushes one layer deeper, from steering outputs to rewriting the machinery that produces them. That said, the strongest parts of the paper still involve controlled setups, and the safety risks are obvious.
A system that can edit weights without conventional training loops could fix bugs, but it could also hide them.
Willison’s laptop test broadens the field by lowering the barrier to entry for replication and scrutiny. If an open model like Qwen can read weights on a laptop, more independent researchers can probe whether the technique works across architectures or on less toy-like tasks.
Reading Raw Weights now becomes a practical research target rather than a purely theoretical claim.
Reading Raw Weights And What Comes Next
For open-model developers, Qwen is the relevant name here because it lowers the cost of testing the idea. A lab-only result around GPT-6 Astra would be interesting, but a reproducible result on Qwen makes outside scrutiny possible.
The next thing to watch is replication. Others will test whether weight space programming can improve real models reliably, or whether the most dramatic gains stay trapped inside carefully chosen demos.
Read Next: Drax Power Station Could Become A £20bn AI Data Center Hub