After 11 years straight pushing the frontier from inside the big AI labs, I am spending some time working as an independent researcher in New York City.
Mar. 2025 – Jul. 2026
I was part of the founding team at Thinking Machines Lab. I worked on high-compute RL science, as well as an uncategorizable long-term research bet.
Sep. 2022 – Feb. 2025
At OpenAI, I landed contributions to most parts of the frontier deep learning stack:
- [2024-2025] in fundamental architecture research that bore fruit, creating a novel inference-efficient architecture component that later landed in a frontier base model.
- [2024] in pre-training, co-leading the GPT-4o bigrun.
- [2024] in mid-training, as overall tech lead of GPT-4o mini. I pitched this run, specced it for inference, executed its bigrun(s), and helped de-risk bleeding edge synthetic data recipes.
- [2023] in product x research co-design, conceiving of, training, and beta-releasing an early high-compute agentic search system which was perhaps the first “Deep Research”.
- [2022-2023] in post-training as part of the small research team that created ChatGPT, where I worked on RL science, co-designed key aspects of the tool-use ML, and helped tame the training effort into a repeatable “mainline” model training machine. My name is on some of the first GPT-4 post-trained checkpoints that served prod ChatGPT traffic.
I spun up, recruited, and briefly managed the London unit of OpenAI’s Training team, serving as TLM of an inference-efficient architecture research effort.
Sep. 2015 – May 2022
DeepMind is where I came of age as a researcher. I was the padawan of Alex Graves, who is one of the field’s true greats. I collaborated on the earliest language model scaleups, making fundamental research and engineering contributions, for example inventing an efficient technique for large-scale model sharding on a whole TPU pod. Before the industrialization of LLMs, I drove fundamental research on deep generative models, neural network architecture, and information-theoretic learning algorithms. See my Google Scholar page for published work. Some of this was undertaken as part of DeepMind’s joint PhD program with University College London (UCL).