Skip to content

Neural networks

1 unrolled thread about neural networks, each one readable on a single page.

Rikiya Takehi@rikiyatakehi

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x...

SummaryResearchers demonstrate that frontier LLMs like GPT-6 Astra can directly read, understand, and edit neural network weights as if they were code, enabling weight-space programming. By modifying just 0.016% of Qwen3.5-0.8B's weights, they fixed reasoning loops and improved math performance 6x and instruction following 3x without gradient-based training.Nov 2, 20267 tweets♥ 10.8K