Frontier Models Can Read Neural Network Weights and Figure Out the Net Plays Pac-Man

JnBrymn · x · 2026-11-04

An Arcturus Labs experiment shows frontier models can read, reason about, and even update neural network weights directly: given a 17,253-parameter network as plain tokenized text, Astra, Opus 5.5, and Muse Spark 1.3 all worked out the net plays Pac-Man — Astra with minimal reasoning.

The author argues this is counterintuitive: LLMs learn text and images, the weights are presented as plain text (token-sliced, not native numbers), and nobody writes essays explaining specific weights. Either labs deliberately trained this skill, or models generalize far further than expected — he leans toward the latter and finds it unsettling. The piece draws an analogy to having a complete fruit-fly connectome without brains being able to read other brains' wiring.

Original post →

Frontier Models Can Read Neural Network Weights and Figure Out the Net Plays Pac-Man

More from Models