Anthropic Quietly Acquires Claude Fable to Curb Competition

In a 319-page system document, Jonathon Ready has shed light on a discreet strategy by Anthropic regarding its AI models, Fable 5 and Mythos 5. The company has implemented silent interventions to limit the effectiveness of Claude Fable when used to develop competing state-of-the-art language models.

These interventions, which will not be visible to the user, aim to prevent the use of Claude for specific requests, such as building pre-training pipelines or designing ML accelerators. While using Claude for these purposes already violates Anthropic's Terms of Service, these new protections are intended to deter actors likely to breach these terms.

Unlike the visible measures taken for cybersecurity, biology, chemistry, and distillation attempts, these silent interventions will not trigger a model change in Fable 5. They include methods such as prompt modification, the use of guiding vectors, or effective parameter fine-tuning (PEFT). Anthropic estimates that these measures will only affect about 0.03% of traffic, concentrated in less than 0.1% of organizations.

This is the first time Anthropic has announced such silent interventions, raising questions about transparency and fairness in AI development. The concept of recursive self-improvement mentioned in the document adds an almost science-fictional dimension to this strategy.