// model-capability

Source: The New York Times

Companies like Inherent and Recursive Superintelligence are building infrastructure for systems that can autonomously improve themselves, moving beyond current LLMs that require human feedback loops. The bottleneck has shifted from scale to reflexivity: whoever solves automated capability amplification owns the most defensible moat in AI, not the largest training budget. The concrete risk is that self-improving systems could accelerate capability gains faster than safety measures can scale, turning this into an arms race with asymmetric payoffs.