GPT-6 Astra Directly Rewrites Model Weights, Achieving 'Weight-Space Programming'
Understand the story
On November 4, 2026, The Decoder reported that researchers enabled GPT-6 Astra to directly read and generate the weights of other AI models without conventional step-by-step mathematical optimization, a method called 'weight-space programming.' According to the report, Astra can identify the games, graphics, and trained language that a network has learned from the weights alone. In practical application, a network of roughly 1.5 million parameters written by Astra for 9x9 Go defeated a similarly sized model trained on about 20 million games. This is the latest report on the event; no other related reports were previously recorded, and the event is currently in its early disclosure stage.
AI overview of reporting · Updated 3 h前
Reporting timeline
Explore different aspects of the story through its reporting.
- The DecoderGPT-6 Astra sees the matrix and reprograms other AI models directly through their weights
Researchers let Astra directly read and generate model weights without conventional step-by-step mathematical optimization, an approach they call "weight-space programming." From weights alone, Astra can identify the games, graphics, and training language a network has learned; a network of about 1.5 million parameters it wrote for 9x9 Go beat a same-sized model trained on roughly 20 million games.
Story heat trend
Current heat9·Comparable peak10（11/5 01:00）·Comparable change over 24 hours –
The trend compares the same continuously observed sources, which may cover fewer sources than the current heat score. Hover or select the chart to inspect each hour; use the left and right arrow keys to navigate.