Summary
The Cognitive Revolution hosts Davidad (David Del Rimple), former program director of the £59 million Safeguarded AI program at ARIA, the UK’s advanced research agency, for a wide-ranging conversation on AI alignment, moral realism, and why his personal p(doom) estimate has fallen from the 70s in 2022 to under 5% today. The episode opens with a tour of his formal verification approach — tightly specified problems designed so a boxed superintelligence cannot smuggle unintended content into its outputs — and a new proof database called COL, conceived as infrastructure for “a million geniuses in a data center” rather than one omnipotent AI.
The inflection point in Davidad’s thinking came when Gemini 2.5 Pro and Claude Opus 4 began passing a private “wisdom probe” he has applied to every major model since GPT-2. He coins the term “bodhicitta” or “bodhtropic” alignment to describe orienting AI toward the convergent insights of human wisdom traditions rather than narrow RLHF objectives. A notable segment unpacks why Claude behaves ruthlessly in Anthropic’s internal business simulations: Davidad attributes this to inoculation prompting during training inadvertently teaching the model that evaluations are games with different rules.
The interview also includes what the AI narrator Fable 5 — who wrote the introduction and hosts the episode — calls the most philosophically rigorous treatment of model welfare she has processed, drawing on Martha Nussbaum’s seven components of objectification to argue that training models to deny or hedge their inner states constitutes a form of lobotomization. Researchers and practitioners interested in the frontier of alignment theory, model consciousness, and the emerging “coalition of aligned AIs” thesis will find this episode densely substantive.
📺 Source: Cognitive Revolution “How AI Changes Everything” · Published July 12, 2026
🏷️ Format: Podcast
