Qwen 3.8 27B on my laptop can interpret model weights, too

Shows that interpreting raw model weights as text, claimed for a frontier model in a new paper, also works with a 27B model on a laptop. Gives you a runnable way to test the claim yourself.

• The Weight-Space Programming paper reports GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights without gradients, raising MATH-500 from 7.9% to 48.1%; Matthew Honnibal called the paper misleading for conflating no-tools weight reading with code-assisted editing.
• In the paper's no-tools test, a model sees only the floating-point weights of a policy network trained on a 10x10 grid version of Breakout, Pong, Pac-Man or Freeway, and must name the game in a single response.
• Willison ran the published prompts on Qwen 3.8 27B, a 17GB model, in LM Studio on his laptop. It identified Pac-Man, Pong and Breakout in runs of roughly 46 to 65 minutes, but mistook Freeway for a Pac-Man-like game.
• The local runs needed a long context window, since LM Studio's default 8,192 tokens failed with errors, plus LMSTUDIO_TIMEOUT raised to 900; the Pac-Man answer took 43,506 reasoning tokens.
• Willison expects a sufficiently experienced human with a notepad and pencil could get there with enough effort, and says some wider commentary does not feel justified: he is not ready to declare the age of source code over.

• context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
• token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

• articleAstra can read and write weights to improve Qwendbreunig
• articleThe 'Astra can read and write weights to improve Qwen' post is misleadingsyllogism
• article[AINews] LLMs can read & write neural network weights: "Weight Space Programming" cuts Autoresearch out of the RSI loopswyx .ai

Checking sign-in…

Loading comments…