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

	Qwen 3.8 27B on my laptop can interpret model weights, too(simonwillison.net)
	721 points by theanonymousone 8 days ago \| 211 comments
	tl;dr: A recent paper claimed frontier models like GPT-6 Astra can interpret and edit raw neural network weights without tools ("Weight-Space Programming"), and the author replicated the game-identification experiment locally using Qwen 3.8 27B, which correctly identified Pac-Man, Pong, and Breakout from weights alone (but failed on Freeway). While the results are impressive, the author pushes back on broader hype, noting the paper conflates weight-reading with tool-assisted weight editing, and that a determined human with pencil and paper could likely do the same analysis.
	HN Discussion: ↑Qwen 3.8 27B rivals frontier models when guided properly, making expensive APIs unnecessary ↑Skeptical about whether the model truly interprets weights or relies on training/prompt leakage ↑Questions whether the interpretations are actually correct, echoing author's skepticism •Curious about practical applications like weight editing vs retraining or LoRAs •Requests for further experiments or reproductions to validate claims