Import AI 465: Open vs closed gaps; Kimi K3; Demis' big policy plan

The UK government's AI Security Institute has found that the gap between open and closed weight models in cybersecurity is closing. This gap was previously measured at 6 to 10 months, but has now narrowed to less than 7 months. Specifically, GLM-5.2 is nearly as strong as Claude Opus 4.6, which was released 4.3 months earlier, while DeepSeek V4-Pro is between Claude Opus 4.5 and GPT-5, released in different months.

The institute plans to test Kimi K3, a 2.8 trillion parameter model, once its weights are publicly available. This development implies that cyber defenders must prepare quickly for the possibility of powerful AI becoming accessible without the same safeguards as proprietary companies. Meanwhile, China's Kimi K3 model, with 2.8 trillion parameters, has been developed to match or surpass leading proprietary models like Claude Fable 5 and GPT 5.6 Sol.

Although it may have some brittleness, Kimi K3's performance is at a frontier level. Additionally, Kimi has demonstrated the ability to create a compiler and design a chip using only open-source EDA tools, suggesting that future AI systems may be able to self-improve and create new hardware. Finally, DeepMind founder Demis Hassabis has proposed a regulatory framework for artificial general intelligence (AGI), modeled after a standards body like FINRA, to oversee and test frontier AI systems.