Two Minute Papers · 2026-11-07 · notable
Two Minute Papers — 'OpenAI's Astra Just Opened The Black Box Of AI'
In short
Two Minute Papers covers Weight-Space Programming, the MIT CSAIL study where GPT-6 Astra reads raw neural network weights and edits them by hand, with no training loop.
Two Minute Papers · 2026-11-07 · 3 sources checked
Two Minute Papers explains how GPT-6 Astra reads a network's weights and rewrites them directly.
What is it?
'OpenAI's Astra Just Opened The Black Box Of AI' went up on the Two Minute Papers channel on 7 November 2026. The video's description points to the Weight-Space Programming blog post by MIT CSAIL researchers, which tests whether GPT-6 Astra can work out what a network computes just by looking at its weights.
How does it work?
In the study, Astra gets the weights as text or as tensors in a sandbox where it cannot run the network, plus a small budget of forward passes. In one experiment it found a direction behind Qwen3.5-0.8B's looping on math problems and subtracted it, lifting MATH-500 from 7.9% to 48.1% by changing 0.016% of the weights.
Why does it matter?
Reading and repairing a model without data or gradient training would be a new tool for interpretability. The video gives a short visual walk-through of a result that has drawn both excitement and pushback this week.
Who is it for?
anyone curious about interpretability research