Google DeepMind researcher, Gemini team
Sholto Douglas
Profile
Sholto Douglas is a research scientist who spent three years as one of the most important engineers on Google DeepMind’s Gemini effort, and in 2025 moved to Anthropic to lead scaling of reinforcement learning. He’s one of those rare researchers who became an insider voice on frontier model training by force of self-study — the often-told story is that he was doing AI research from 10pm to 2am every night before getting noticed by Google engineers who couldn’t figure out who this person asking sharp questions online was.
At DeepMind he worked on PaLM and then became a lead architect on the Gemini family, contributing to training infrastructure, inference systems, and the research direction that closed Google’s gap with OpenAI. Noam Brown has publicly called him one of the most important people behind Gemini’s success. At Anthropic he’s now focused on pushing RL for agentic capabilities — making models that can chain long sequences of actions reliably, which he argues (rightly) is the actual bottleneck for agents, not context length.
What makes him matter for developers learning AI is the clarity with which he talks about what’s actually happening at the frontier. His X posts and long-form podcast appearances with Dwarkesh Patel are the closest thing the public has to ground-truth commentary on how frontier labs think about scaling, compute, data, and training. When he says something about reliability scaling with model size, or why RL will keep working, those aren’t hot takes — they’re hypotheses from someone who has actually been in the loop at two of the three frontier labs.
If you’re trying to build a mental model of where capabilities are going over the next few years, Douglas is a better signal than almost any analyst or journalist. He’s technical, calibrated, and willing to reason out loud.
Key Articles & Papers
Sholto's Blog Gemini: A Family of Highly Capable Multimodal Models PaLM: Scaling Language Modeling with PathwaysVideos
Spotify Podcasts
