Frontier Radar #4: China has caught up, so what's left of the Western AI lead?
decoder ·
The article examines how Chinese AI models like Kimi K3 and GLM-5.3 have narrowed the gap with top Western models, making raw model performance hard to defend as a business advantage. Western labs accuse Chinese rivals of using distillation and benchmark tuning, but the author argues that a model lead can no longer be sustained because capabilities sold through APIs inevitably migrate to competitors. The remaining Western edge is limited to three areas: abstract specialty tests, reliability (measured by repeatability on agentic benchmarks), and cybersecurity capabilities, which are largely restricted from public access. The reliability gap is commercially most significant but unstable, as even newer Western models can fall behind their predecessors. The cybersecurity gap is shrinking quickly, with GLM-5.3 nearly halving the distance to US leaders in a month. Distillation is plausible because Chinese labs allegedly ran millions of interactions through Western models to gather training data, though no verifiable evidence has been published. Ultimately, the article concludes that the industry's moat has shifted from the model itself to the broader system and ongoing work that produces future models.
decoder