Anthropic is a hypocrite.

Anthropic is a hypocrite.

Anthropic is claiming that Alibaba’s Qwen AI lab used almost 25,000 Claude accounts to help improve the performance of the Qwen models.

This type of technique is called distillation. The premise is that you have a “teacher” model (think Claude Opus) and a student model (in this case, most likely one of the Qwen models). The teacher model is prompted, and its responses are used to train and refine the student model.

The distillation will improve the performance of the smaller model, at a much lower cost than traditional model training.

Here is the ironic part for me.

There are a lot of differing opinions on the types of data used to train Claude in the first place. Anthropic web scraped public internet data and spent millions of dollars acquiring books to scan and train models with.

If I am an author, I may not want my work being used to train AI models. The individuals who contributed to open source may not want their hard work being used to train coding agents either. Anthropic used this data because it was a cheap and quick way to get the datasets required to train such a large frontier model.

Kind of similar to what Qwen is also doing. Paying for access to data, to train a model on, without Anthropic consenting to it. An interesting parallel.

Here is my take: Anthropic’s concern is less about a competitor model getting closer to the performance of Claude, it’s the fact that Qwen is a free and open weight, which is a direct threat to Anthropic’s entire business model.