The delicious irony of Anthropic bemoaning distillation
Anthropic has complained about competitors using distillation (training smaller AI models on outputs from larger ones), but critics note that Anthropic itself used outputs from other AI systems during training, highlighting the perceived hypocrisy in its stance.
Background
- **Anthropic** is an AI safety company founded by ex-OpenAI employees, best known for its Claude family of large language models. It has been one of the most vocal advocates for "responsible AI" and against the unregulated use of frontier models.
- **Distillation** (or "model distillation") is a technique where a smaller, cheaper model is trained to mimic the outputs of a larger, more capable model. It's how many companies (including OpenAI and others) create efficient, low-cost versions of their flagship systems — and also how third parties sometimes copy a model's behavior without permission.
- Recently, Anthropic has publicly complained about rivals using distillation to replicate Claude's capabilities, calling it unfair and a threat to safety norms. The irony: Anthropic itself reportedly fine-tuned earlier models on outputs from OpenAI's GPT-4 — a form of distillation — to bootstrap Claude's performance when it had little proprietary training data of its own.
- The post highlights this inconsistency: Anthropic built its product partly by leaning on a competitor's work, then turned around and condemned others for doing the same thing once it had a market lead.