Frontier LLMs can natively read and write neural network weights, outperforming PPO-trained nets

RobbinDeBank · reddit · 2026-11-03

A Reddit share of the blog "Sparks of Weight-Space Programming" reports that frontier LLMs are surprisingly good at natively reading and writing neural network weights, beyond humans or conventional algorithms — potentially a big deal for optimization and interpretability (unreviewed blog claim). Key results from the figures: GPT-6 Astra identified what a network was trained to do from raw weights alone (99% confidence on Breakout from 17,253 weights, no tools); it wrote small game-playing networks from scratch, with no environment feedback, that outplay PPO policies trained for 100M steps (fixing a failing car task with a one-shot 3.5% weight edit, 0%→100% success); and when asked to make Qwen3.5-0.8B better at math, it located a reasoning-loop weight issue and fixed it by natively editing 0.016% of weights, raising MATH-500 from 7.9% to 48.1%.

Original post →

Frontier LLMs can natively read and write neural network weights, outperforming PPO-trained nets

More from Models