Import AI 465: Open vs closed gaps; Kimi K3; Demis’ big policy plan
| Source: Import AI (Jack Clark)
Tags: UK AISI, DeepSeek, Kimi K3, GLM-5.2, open-weight models, cybersecurity, Moonshot AI, Import AI
The UK AISI reports the cyber-capability lag between open and closed AI models compressed from 6-10 months to 4-7 months, with GLM-5.2 and DeepSeek V4-Pro now rivaling frontier closed models — while Kimi K3 (2.8T parameters) signals China's push to the closed-model frontier tier.
Details
Import AI Issue 465 covers three developments in one newsletter. The lead story is the UK AI Security Institute's first public analysis of how open-weight models trail closed-weight models on cybersecurity tasks. The gap has narrowed: open models now lag 4-7 months, down from 6-10 months measured through most of 2025. On 70 narrow cyber evals, GLM-5.2 (the leader among open models) is closest to Claude Opus 4.6, released just 4.3 months earlier. DeepSeek V4-Pro falls between Claude Opus 4.5 and GPT-5. The gap widens on long-horizon cyberrange tasks — GLM-5.2 reaches Opus 4.5 performance on end-to-end hacking scenarios, but proprietary models still lead. AISI warns: 'cyber defenders have a short window to prepare before today's frontier cyber capabilities may become accessible without the same safeguards.'\n\nThe second story covers Kimi K3, a 2.8 trillion parameter model from Moonshot AI (China), described as the clearest evidence yet that Chinese firms are closing the gap on Western frontier models — not just open-weight leadership. AISI intends to benchmark Kimi K3 on its cyber eval suite once weights are released.\n\nThe third story covers Demis Hassabis' policy plan, though this section was unavailable in the extracted content. The title suggests it addresses AI governance views from Google DeepMind's CEO.