Inside the black box of AI: How AI models actually think | Emmanuel Ameisen (Anthropic)

Interpretability research shows LLMs form general concepts and plan ahead internally, reframing both capability jumps and AI safety work.

• Models are 'grown' not built: Anthropic trains Claude via pre-training (predict next word on massive text corpora) then fine-tuning on curated dialogue examples, producing opaque internal math no one designed.
• Interpretability finds real planning: Looking at intermediate activations, researchers found Claude plans rhymes in poetry many tokens ahead (e.g., deciding to end a line with 'rabbit') without ever stating the plan.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.