Weight-Space Programming shows GPT-6 Astra reading, repairing and writing neural network weights directly with no gradient descent, including a 0.016% weight edit that lifted Qwen3.5-0.8B on MATH-500 from 7.9% to 48.1%.
Thinking Machines launched Inkling 2 Small, a 594B-parameter MoE available now with closed weights, with Apache 2.0 weights promised next week after external safety testing and red-teaming.
OpenAI put offensive security in the spotlight: GPT-6 Cyber went to approved Daybreak Red customers, and a prototype pipeline produced 948 Linux kernel local privilege escalation exploits.