Ross Taylor
I am the co-founder and CEO of General Reasoning. We are a new research company building long-run intelligence: AI systems that learn over extremely long time horizons.
Previously I led the reasoning team at Meta AI, working on Llama 3 and Llama 2. We achieved the first models internally that beat GPT-4 on reasoning using an early RL recipe. I also spun up the initial Llama post-training team with my former co-founder Robert Stojnic, and made the key hires for the team who delivered Llama 2.
Before that, I was the research lead and first author on the Galactica language model in 2022. The model was state-of-the-art on several benchmarks, and introduced innovations such as thinking tokens and data-efficient multi-epoch pre-training. It had a notorious base model demo - ChatGPT came out two weeks later with the crucial RLHF ingredient and the rest was history.
Before language models, I co-founded the startup behind Papers with Code, which scaled to a million monthly active users and was acquired by Meta in 2019. In my more distant past, I worked at a legendary Hong Kong betting syndicate building quantitative models for sports, built open source time series for Python, and much more. I believe strongly in the power of logarithmic objectives and collective scenius.
I grew up in Ipswich, UK and have lived in London since 2013. I studied at Warwick and Cambridge. I enjoy surfing, football, and Kipling poems.
Research
Interviews and Talks
- Scaling to Long Horizons - July 2026
- Interview on etn. - June 2026
- The RL Renaissance - September 2025
- Interconnects with Nathan Lambert - July 2025
- The Hitchhiker's Guide to Reasoning - November 2024
- Interconnects with Nathan Lambert - August 2024
- Teaching LLMs to Reason - July 2023
Press
Posts
- Scaling RL Compute - 23 March 2025
- Notes on the Membrane - 9 December 2024
- QK Norm and the Curious Case of Logit Drift - 3 October 2024
- Synthetic Data: Escaping the Ouroboros - 21 April 2023