AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Google DeepMind's 50X Growth & AI Lab Strategy with Logan Kilpatrick

· The Cognitive Revolution

Logan Kilpatrick (Google DeepMind) returns to The Cognitive Revolution for a wide-ranging discussion of the May 2025 AI release avalanche and Google's transformation into a top-tier AI powerhouse. He reveals that Google AI usage has grown 50x year-over-year, from 10 trillion to 500 trillion tokens per month—roughly 50,000 tokens per person on Earth monthly—and credits this to the mid-2023 merger of Google Brain, parts of Google Research, and DeepMind under Demis Hassabis's leadership. The organizational unification, paired with Google's structural incentive to embed AI across billion-user products (Search, Workspace, YouTube, Cloud) and its TP/compute advantage, has closed the gap with competitors he says was visible from the outside. On competitive dynamics, Kilpatrick expects more divergence among labs going forward because "low-hanging fruit has been captured" and structural advantages (infrastructure, distribution) will increasingly differentiate winners; he sees room for startups that pick a focused wedge but warns against competing head-on in foundation models without serious capitalization. He is skeptical of scenarios where labs hoard their best models (e.g., AI 2027), citing multiple game-theoretic incentives to ship externally, and Google's Cloud mandate to serve external developers. He also discusses Anthropic's Windsurf cutoff as defensible business logic, Google's diffusion language model demo as a potential new paradigm enabling personal generative UI, Gemini 2.5 Pro's exceptional long-context performance on OpenAI's MRCR benchmark, and the ongoing simplification of agent scaffolding as models absorb reasoning. On AGI and the future of work, Kilpatrick predicts the AGI "moment" will be a product experience that weaves memory, long context, and reasoning together rather than a single model release. He expresses a strongly human-centric philosophy: he writes his own emails and tweets without AI assistance, believes people will continue to value authentic human perspective (including podcasts), and views NotebookLM-style on-demand content as net-additive rather than substitutive for human-created work. He closes with an open invitation to email him directly (lkilpatrick​@google.com) for early-access program consideration.

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