Experts say Kimi K3 didn't steal from Anthropic's Fable
technical experts dispute white house claims that moonshot ai's kimi k3 relied on distilling anthropic's fable model, noting the timeline makes complex data theft impossible and that advanced reinforcement learning is likely the real driver.
us officials including michael kratsios and scott bessent accused chinese firm moonshot ai of building the kimi k3 model by extracting watermarks and propriety data from anthropic's fable, but experts like braden hancock and nathan lambert reject this narrative. they emphasize the physical impossibility of distilling enough data, training, and releasing a model of k3's caliber in the two weeks since fable's launch, noting that simple supervised fine-tuning cannot replicate frontier-level reasoning. while anthropic previously identified systematic distillation attempts, researchers suggest the model's performance likely results from heavy investment in reinforcement learning infrastructure, not cloned data. the dispute highlights growing us-china tensions over ai IP, even as experts warn that api-based distillation is becoming an outdated explanation for rapid model improvements.