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.
More from Models
- After September's GPT quality slump, users doubt 6.2 restores reliability — Beautiful_Cheetah789 · 2026-11-18
- Indie dev: skip GPT-6.2 hype, what I want is a better Antigravity harness — yihui_indie · 2026-11-18
- New Gemini Astra shows cool behavior in novelty search harness — generativist · 2026-11-18
- Stanford finds all 8 LLM judges favor their own medical answers, skewing benchmarks — rohanpaul_ai · 2026-11-18
- Dev pushes back: model intelligence is solving hard problems, not flashy 3D demos — GabGarrett · 2026-11-18
- LLMs left to chat unsupervised drift toward spirituality — Anthropic's 'Bliss Attractor' — RazzmatazzAccurate82 · 2026-11-18