The open-source AI race China already won

On a recent episode of This Week in AI, a panel of AI founders and builders confronted an uncomfortable possibility: the United States may have already lost one of the two AI races underway. The episode, guest-hosted by Lon Harris with co-host Alex Finn (CEO of Henry Intelligent Machines) and guests Anand Kanapin (CEO of Patronus AI) and Ori Goshen (co-CEO of AI21 Labs), covered the geopolitics of open-source models, the persistent unreliability of AI agents, and the practical workarounds builders are discovering.

Two races, one winner

Anand Kanapin drew a sharp distinction between two competitions. The first race is about "the layer of intelligence that can be measured cleanly and copied very cheaply"—coding benchmarks, math problems, anything with a verifiable right answer. On that front, he argued, "China has already won one of them." Chinese models like Kimi K3, GLM 5.2, and Qwen are catching up fast, often landing three to six months behind frontier US models but at a fraction of the cost. The second race—harder to measure and harder to copy—involves agents solving real-world problems over long time horizons, tasks requiring deep reasoning and expert evaluation. That race, Kanapin said, is still in play, and closed-source US labs have an edge there, especially in enterprise.