SPARKS OF WEIGHT-SPACE PROGRAMMING

Summary from the TLDR Dev newsletter

Experiments show GPT-6 Astra reading and directly editing neural network weights without gradient descent, including writing small game-playing networks from scratch. Editing 0.016% of Qwen3.5-0.8B's weights reduced reasoning loops and raised MATH-500 accuracy from 7.9% to 48.1% with 50 forward passes for feedback. The gains removed an obstacle to existing ability, while similar edits failed to meaningfully improve a Gemma model that rarely looped.