LLMs edit other models' weights

Researchers used a large LLM (GPT‑6 Astra) to read, reason about, and change the numeric parameters of a smaller open model (Qwen), dramatically improving benchmark math performance by touching a tiny fraction of weights. The method, labelled 'weight‑space programming', bypasses conventional retraining and lets an AI directly repair or augment another AI. — If scalable, this creates a new class of AI capability‑multipliers and a governance problem: models that can build, fix, or weaponize other models without human‑in‑the‑loop training.