Descriptions:
Sam Witteveen analyzes Anthropic’s published report accusing three Chinese AI labs — DeepSeek, Moonshot AI (makers of the Kimi K2 series), and MiniMax — of running industrial-scale distillation operations against Claude. Anthropic’s figures: 24,000 fake accounts generating 16 million API exchanges, each tailored to extract specific model capabilities including reasoning, tool use, coding, and refusal handling.
Witteveen breaks down what each lab allegedly targeted: DeepSeek (150,000 exchanges, under 1% of the total) focused surgically on rubric-based grading to use Claude as a reward model for reinforcement learning; Moonshot AI (3.4 million exchanges) targeted general reasoning, computer use, and agent development; MiniMax (over 13 million exchanges) concentrated on agentic orchestration and coding. He notes the timing is conspicuous — the accusations coincided with all three labs releasing significant new models, and potentially just ahead of a major new DeepSeek release.
The video also covers the broader controversy: a parallel Google report, a leaked OpenAI memo making similar claims, Elon Musk’s counterargument that Anthropic itself paid a $1.5 billion settlement for training on copyrighted books, and a pushback from tinygrad’s George Hotz arguing Anthropic effectively surveilled paying customers. Witteveen explains the technical mechanics of distillation and why these extraction patterns are detectable, while remaining agnostic on the ultimate truth of the accusations.
📺 Source: Sam Witteveen · Published February 24, 2026
🏷️ Format: News Analysis
