Import AI 465 Discusses Open vs Closed Gaps, Kimi K3, and Demis Hassabis' Policy Plan

Summary

The article emphasizes that the performance gap between open-weight and closed AI models is narrowing, shows China's Kimi K3 approaching frontier technology, and proposes a FINRA-like regulatory body for AGI oversight.

Why it matters

A narrowing gap could shorten the preparation time that advocates have before powerful capabilities become publicly available, potentially destabilizing cyber attack-defense dynamics. The spread of high-performance open models like Kimi K3 could democratize AI power but also introduce unknown risks. As Demis Hassabis suggested, a coordinated regulatory framework could provide a unified approach to assessing and mitigating frontier AI threats.

Highlights

• AISI reports that GLM-5.2 and DeepSeek V4-Pro now match closed models from 4–7 months ago in narrow cyber tasks.
• China's Kimi K3 (2.8T parameters) demonstrates frontier-level performance and AI-focused self-improvement, e.g., creating a Triton-like compiler.
• Demis Hassabis proposes a FINRA-style standards body, led by the US, for voluntary testing and final regulation of frontier AI systems.

Critical take

The article's performance claims may be inflated due to benchmark-focused tuning ('benchmaxxing'), and the proposed regulatory model could face challenges in implementation across diverse international AI actors.