Paper says only GPT-6 Astra reads model weights; Qwen 27B on a laptop did it too. Willison: "not ready to declare the age of source code over"
First seen on X 2 days ago@simonw ♥ 3,04715 views
After a paper, "Sparks of Weight-Space Programming," claimed AI models can read and write neural-network weights directly, developer Simon Willison posted on Nov. 3 the result of running the same test on his laptop. His X post drew about 3,000 likes.
The paper claimed GPT-6 Astra can tell what a network was trained to do from raw weights without tools, and that editing 0.016% of the weights of a small Qwen3.5-0.8B lifted its MATH-500 score from 7.9% to 48.1%. Willison ran a 17GB quantization of Qwen 3.8 27B in LM Studio. Given the paper's no-tools prompt with weights of a grid-game network, it described the Pac-Man-style game as an 8x8 maze at 72% confidence after 46 minutes 12 seconds and 43,506 reasoning tokens. It identified Pong in 1 hour 4 minutes 54 seconds and a Breakout-like game in 52 minutes 25 seconds, and failed Freeway. Kimi K3.1 Pro via OpenRouter identified Breakout in 9 minutes 41 seconds.
Willison noted the games were simplified 10x10 grids, only four were tested and one failed. A patient human with pencil and paper could do the same, so it is not evidence of superhuman ability, he argued. Critics say the paper mixes no-tools reading with tool-assisted editing. "I'm not ready to declare that the age of source code is over," he wrote. He used an M5 MacBook Pro with 128GB but said Qwen 3.8 27B should work on 32GB.
The same day, account xingyudang posted a "Writing Weights Index" (1,100 likes) saying open-source models struggle on the task. One reply to Willison asked to see both the laptop's electricity bill and Astra's API bill.
“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…”
