The Expiring Illusion of AI Control

For years, the global conversation around artificial intelligence safety has relied on a comforting assumption: the most powerful, potentially dangerous AI systems would remain locked behind the corporate APIs of a few well-funded tech giants. Today, that assumption is rapidly expiring.

A recent analysis by the UK government’s AI Security Institute (AISI) reveals that the capability gap between proprietary, closed-weight models and publicly available open-weight models is collapsing faster than anticipated. When tested on specific, narrow cybersecurity tasks, leading open models like GLM-5.2 and DeepSeek V4-Pro are now performing at levels comparable to frontier closed models released just 4 to 7 months prior. This is a significant acceleration from the 6-to-10-month lag observed previously.

This shrinking gap is most vividly illustrated by the arrival of Kimi K3, a massive 2.8 trillion parameter model from China. Kimi K3 is not merely chasing benchmark scores to rival frontier heavyweights like Claude Fable 5 and GPT 5.6 Sol; it is demonstrating early, practical signs of recursive self-improvement—the ability of an AI to build better AI. In recent evaluations, Kimi K3 successfully developed a compact GPU compiler called MiniTriton and autonomously designed a functional chip for a smaller AI architecture within a single 48-hour window.

However, the democratization of frontier intelligence comes with caveats. AISI researchers noted that while open models excel at narrow tasks, they still struggle with long-horizon, complex operations that require chaining multiple capabilities together. Industry insiders often refer to this subtle brittleness as "big model smell"—a tendency for models to be heavily optimized for benchmark tests (benchmaxxing) at the expense of true, robust generalization.

Despite these current limitations, the trajectory is clear. Kimi plans to release the weights of K3 to the public in the coming weeks, effectively diffusing frontier-level intelligence into the wild. This shifts the paradigm of AI policy from centralized control—where regulators can intervene at the corporate platform level—to a decentralized reality where anyone with sufficient computing power can run highly capable systems.

As DeepMind founder Demis Hassabis proposes new, stringent regulatory frameworks modeled after financial oversight bodies, the core debate is shifting. We are moving past the race to build the smartest machine, entering a new era where society must figure out how to thrive when state-of-the-art intelligence is available to everyone.

Key Points

• The UK AISI found that the performance gap in cyber capabilities between open and closed AI models has narrowed to just 4-7 months.
• China's Kimi K3 (2.8 trillion parameters) is approaching frontier model performance and demonstrating capabilities in AI-driven chip design and compiler development.
• While highly capable in narrow tasks, open models still exhibit some brittleness and lack the broad generalization of proprietary models in complex scenarios.
• The imminent release of powerful open-weight models challenges current AI safety paradigms, which rely heavily on centralized platform control.

Why It Matters

The rapid advancement of open-weight models democratizes access to frontier AI, fueling widespread innovation while simultaneously rendering traditional, gatekeeper-based safety regulations obsolete.

Sources:

• Import AI 465: Open vs closed gaps; Kimi K3; Demis' big policy plan — Import AI (Jack Clark)