Anthropic interpretability researcher explains how language models plan, think, and form concepts

In a podcast interview, Emmanuel Ameisen, an interpretability researcher at Anthropic, breaks down how LLMs like Claude operate internally—arguing that the "next-token predictor" framing misses how models actually plan multiple tokens ahead, form general concept representations, and exhibit structures that may parallel conscious versus unconscious processing in humans. The conversation covers how models are "grown" rather than built, what interpretability research reveals about model self-awareness and scaling, and how this work connects to AI safety and alignment.