Replicating weight-space programming: models' written nets decompile to identical programs

kalyan_kpl · x · 2026-11-18

A replication and extension of Astra's weight-space programming work, with three key findings: (1) writing weights reproduces — cloned Breakout nets scored 514/550/307 vs MIT's 517 and held up on modified variants, while Claude Opus 5.5 can also write working nets but often hits output-token limits; (2) the genuinely new capability is translating programs into weights — models write nets that behave like compiled programs, with two models independently decompiling Astra's Breakout net into Python matching its actions in all tested states; (3) reading weights is mostly reading the input layer — game identification is near-perfect, but recovery of spatial layouts on three new self-trained games is limited.

Original post →

More from Models

Models channel →