GPT-6 Astra can supposedly read a neural network's weights and rewrite them. The headline math result, though, used Python, practice problems and 50 test runs.
The day after the paper went viral, Simon Willison pasted the raw weights of tiny game-playing networks (17,253 floats each) into Qwen 3.8 27B on his laptop. It identified Breakout, Pong and Pac-Man. It missed Freeway.
If you build on these models, the useful question isn't whether the new tier can do it. It's how much of the demo survives one attempt without the scaffold.
Full breakdown of the fine print: GPT-6 Astra Can Edit a Model's Weights. Read the Fine Print First.
