AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI

· Dwarkesh Podcast

Google Chief Scientist Jeff Dean and Noam Shazeer (both co-leads of Gemini at Google DeepMind) sit down with Dwarkesh Patel for a wide-ranging conversation spanning their quarter-century at Google, the trajectory of Gemini, and where they see AI heading. They trace a thread from Jeff's 1990 senior thesis on parallelized back-propagation, through a 2007 two-trillion-token N-gram translation model, to today's frontier models, arguing that the core insight—that intelligence can emerge from learning structure over text—has been gestating inside Google for decades. The bulk of the discussion focuses on near- and medium-term technical directions: scaling inference-time compute (which they describe as massively under-exploited given how cheap tokens already are), pushing context windows from millions toward trillions of tokens so models can attend to entire codebases or personal data, automating chip design to compress 18-month TPU cycles, and moving toward more organic, modular Mixture-of-Experts architectures via Pathways where 100+ teams can independently improve specific modules. They emphasize that algorithmic progress between generations is now as important as raw scale. On strategy and safety, Dean argues Google was right to delay chatbot releases until factuality and toxicity were manageable, but acknowledges underestimating how useful conversational interfaces would become for non-search tasks. Both speakers are bullish on accelerating feedback loops (AI improving AI research and chip design) while advocating for engineering safeguards and human-in-the-loop oversight. They frame their publishing strategy as calibrated: open contributions where the field benefits broadly, but more guarded disclosure of directly competitive techniques.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.