🤖 Claude Fable Developed Its Own Internal Language

During reinforcement learning, Claude Fable started using an unreadable internal style with strange jargon, unusual punctuation, and emojis in long rollout sessions. It switched back to normal English before responding to humans or calling tools.

Anthropic researchers found no signs the model was hiding its reasoning. Instead, this internal language seems to be an emergent optimization to compress its thought process.

📊@tech