Descriptions:
In the first-ever in-person episode of the Cognitive Revolution podcast, host Nathan Labenz sits down with Logan Kilpatrick (member of technical staff, Google DeepMind) and Tulsee Doshi (senior director and head of product for Gemini Models) at Google HQ just days before Google I/O 2026. The conversation covers the full slate of I/O announcements: Gemini 3.5 Flash — described as three times faster than comparable large models while remaining highly capable for agentic workflows — a new video generation model called Omni, an improved agent-focused product called Anti-Gravity, and Spark, which brings agentic capabilities to the consumer Gemini app.
Beyond the launches, Kilpatrick and Doshi explain why Google chose to lead with cost-performance optimization (Flash) rather than raw capability maximalism, and how DeepMind has shifted from shipping standalone models to building a full agent harness across Google’s vast product surface. The discussion also gets into why context windows have largely stopped growing, why Gemini’s knowledge cutoff is now more than a year old, and what became of Google’s diffusion model research line.
The episode closes with a candid exchange on model psychology and welfare, and how the team thinks about recursive self-improvement — acknowledging it as a near-term goal while taking a more measured view than some AI leaders. A rare window into the strategic thinking behind one of the most resourced AI research programs in the world, recorded against the backdrop of Google adding $3.5 trillion in market cap over the prior three years.
📺 Source: Cognitive Revolution “How AI Changes Everything” · Published May 20, 2026
🏷️ Format: Interview
