Lenny's Podcast Interview with Anthropic Interpretability Researcher Emmanuel Ameisen: How AI Models Actually Think and Plan
Lenny's Podcast released a new episode on October 18, 2026, featuring Emmanuel Ameisen, an interpretability researcher at Anthropic. The discussion centers on the internal mechanisms of large language models, exploring why models are described as 'grown' rather than 'built,' and how interpretability research reveals capabilities like advance planning, cross-language concept sharing, and potential structural similarities to consciousness. This resource offers specific insights from a frontline researcher for professionals focused on AI underlying logic and safety alignment.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
Lenny's Podcast published a deep-dive interview on October 18, 2026, focusing on AI's underlying mechanisms with guest Emmanuel Ameisen, an interpretability researcher at Anthropic. Before joining Anthropic, Ameisen spent years in applied ML at Stripe and Insight and is the author of 'Building Machine Learning Powered Applications.' His current work involves looking inside LLMs like Claude to understand how they think and plan.
According to the official episode description, the discussion challenges traditional views of LLMs. Key topics include why the 'next-token predictor' framing undersells how models actually plan and think, why anthropomorphizing AI can be the right approach, and why models are 'grown,' not built. The episode also covers what the 'J space' reveals about model self-awareness and the smooth scaling curve behind apparent sudden capability jumps.
For AI practitioners and researchers, the value of this selection lies in connecting frontier interpretability research with AI alignment and safety practices. Ameisen's research has produced striking findings, such as models forming general representations of concepts, planning many tokens ahead without explicit statements, and potentially exhibiting structures similar to conscious/unconscious processing distinctions. While this recommendation is based on verified publisher metadata and topic lists rather than a full transcript review, the guest's authoritative role and the technical depth make it suitable for readers interested in internal model logic beyond external performance.