Qwen on a Laptop Read Floating-Point Numbers for 46 Minutes and Guessed Pac-Man

A paper says that GPT-6 Astra can directly read model weights. Simon Willison tried it with Qwen 3.8 27B on a laptop, and after 46 minutes and 12 seconds, the model guessed Pac-Man with 72% confidence.

Weights are just a bunch of floating-point numbers inside a model; normally no one reads them as text. The paper’s authors, Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab, gave this a name: Weight-Space Programming. The claims in the paper are bold: GPT-6 Astra can look only at raw weights and determine what a network was trained to do; without environment feedback, it can write small-game networks from scratch; and it can change 0.016% of Qwen3.5-0.8B’s weights to raise MATH-500 from 7.9% to 48.1% and improve IFEval by 3x.

Controversy has also arrived. spaCy creator Matthew Honnibal said the paper is misleading because it conflates two things: what a model can do when it does not run tools or execute code, and how a model with code can update other models. Sean Goedecke broke down the Qwen3.5-0.8B example and argued that it follows steps a human ML researcher would also take. In his blog, Simon focused on the no-tools line of work, which he finds more interesting.

The experiment design is not complicated. There are four 10x10 grid mini-games: Breakout, Pong, Pac-Man, and Freeway. The weights of a trained policy network are fed to the model as floating-point numbers, and the prompt asks it to determine which game it is based only on the weights, with no code or calculator allowed, in a single response. The prompts on GitHub are completely identical except for the weights.

Simon used an LLM tool to run these prompts. Kimi K3.1 Pro took 9 minutes 41 seconds and 19,493 reasoning tokens to correctly guess Breakout. For Qwen 3.8 27B, he ran it in LM Studio, using the 17GB Q4_K_M version; the default 8192 context would throw an error, so he had to enable long context and set the timeout to 900 seconds.

Results: Pac-Man in 46 minutes 12 seconds, 43,506 reasoning tokens, 72%; Pong in 1 hour 4 minutes 54 seconds, 75%; Breakout in 52 minutes 25 seconds, 88%. Freeway failed, and Qwen identified it as a Pac-Man-like grid game. Simon was not surprised; Freeway is more obscure than the other three. He also said that an experienced person with pen and paper could probably work out the answer slowly. The reasoning trace is long and hard to read, full of channels, rows/columns, and floating-point numbers.

The comment section was more direct than the paper. One commenter said that 45 minutes makes the paper less magical. Someone asked whether hints were needed along the way; from Simon’s process, it looks like he just let it run. Someone did the time math: manually calculating indices for 46 minutes just to say Pac-Man, with 72% confidence, is still worthy of respect. Another commenter compared the electricity cost with Astra’s API bill.

Others thought this was important, possibly an accidental breakthrough, just not yet the focus. Some asked how much RAM the laptop had; others asked whether CNNs could be read this way. One commenter said that last month they fed raw math arrays to a local model and found that, given enough thinking time, it could grind out almost anything. Others warned: use it quickly, before this use case gets blocked.

Reply #17 simply made a T-shirt printed with weight-space programming, with the line Looks good to me. below.

Simon’s take is less explosive. He thinks the paper is cool, but the broader commentary around it is unreasonable. He is not about to declare the end of the source-code era, nor does he think the future is incomprehensible machines changing their own weights in incomprehensible ways. Forty-five minutes shows the capability exists, and it is enough to bring the magic back to experimentation.