Distilling the Moat

AI companies train models by spending massive amounts on GPU costs and labor, and this investment is believed to form a 'moat' that competitors cannot easily catch up to. However, an LLM is essentially like a web service that responds to queries and exposes some internal data in the process, so in principle it is possible to replicate a model through repeated queries.

This replication technique is exactly distillation, a standard technique in the industry. In court, Elon Musk admitted that xAI used the distillation technique on OpenAI's models to train Grok, and Anthropic has accused three Chinese AI companies—DeepSeek, Moonshot AI, and MiniMax—of using more than 24,000 fake accounts to generate over 16 million conversations with Claude in order to improve their own models.

The reason distillation works so well is that the victimized company subsidizes the cost of its service on a massive scale. AI models are not subject to copyright protection and cannot be patented, so they are only protected as trade secrets, and a mere violation of terms of service provides weak legal protection. It is also hard for AI companies to claim that 'crawling is illegal,' since they themselves built their models through crawling.