Sholto Douglas

Sholto Douglas is an AI researcher whose work spans language-model training, multimodal models and reinforcement learning. He studied space engineering at the University of Sydney and machine learning at Tsinghua, and was a top-50 world fencer. He taught himself machine-learning research while working as a McKinsey consultant, then coauthored research on chain-of-thought, Gemini and Gemma, and a systems guide to scaling language models. At Anthropic, he works on scaling reinforcement learning and discusses Claude models and agentic AI on podcasts.

Now at: Anthropic
Works: 10
Citations: 1,740
h-index: 5

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2019

Space engineering and machine learning

In 2019, he is a mechatronic engineering student at the University of Sydney, pursuing a Bachelor of Engineering Honours with a Space Engineering Major and a Bachelor of Commerce. During a nearly five-month stay at Tsinghua University, he studies machine learning, computational biology, mathematics, and developmental economics while training with China's national fencing team.
November 2023

Latent-variable chain-of-thought training

In November 2023, he coauthors Training Chain-of-Thought via Latent-Variable Inference, a fine-tuning method that typically improves held-out accuracy on GSM8K and BIG-Bench Hard tasks.
December 2023–March 2024

Gemini model reports

In December 2023, he coauthors Google's Gemini report introducing Ultra, Pro and Nano sizes; in March 2024, he is an author of Gemini 1.5, whose report describes models that recall and reason over millions of tokens, and appears with Trenton Bricken on the Dwarkesh Podcast. By March 2024, he has moved from McKinsey into Google's Gemini research. At McKinsey, he teaches himself machine-learning research, reads Gwern's scaling hypothesis and uses a grant from the TPU access program to scale large multimodal models; James Bradbury notices his online questions and recruits him to Google as an experiment in pairing high enthusiasm with strong engineers.
March 2024

Open Gemma models

In March 2024, he is among the Gemma Team's named authors; the report introduces open 2-billion- and 7-billion-parameter models built from Gemini research and technology.
February–May 2025

Scaling language models and reinforcement learning

In February 2025, he coauthors How to Scale Your Model, a systems guide to how TPUs and GPUs run language models and how to parallelize training and inference. By May 2025, he is at Anthropic and focuses on scaling reinforcement learning. In May 2025, he returns to the Dwarkesh Podcast with Trenton Bricken and joins Jacob Effron's Unsupervised Learning to discuss Claude 4 and coding as an early signal of model progress.
October 2025

Sonnet 4.5 on The MAD Podcast

In October 2025, he joins Matt Turck on The MAD Podcast to discuss Claude Sonnet 4.5, why reinforcement learning on language models starts working in 2024, and how agents stay coherent through 30-hour coding sessions. He describes his path from a top-50 world fencer through Google's Gemini team to Anthropic.
February 2026

Agentic AI on TBPN

In February 2026, he appears on TBPN as a Member of Technical Staff at Anthropic who specializes in scaling reinforcement learning and agentic AI. He discusses a new model that tackles complex problems through extended reasoning and test-time computation.
As of October 2026

Reinforcement learning at Anthropic

As of October 2026, he is a tech lead in reinforcement learning at Anthropic. He grows up in Sydney and studies robotics before working at DeepMind. In October 2026, he joins Nicholas Marwell on Joe Lonsdale's American Optimist podcast to discuss why they believe AGI is likely within the next couple of years.