Ross Taylor is co-founder and chief executive of General Reasoning, a London-based AI research company building models capable of sustained work and decision-making. He previously co-founded Papers with Code, led development of Galactica, and led reasoning research at Meta AI, where his team contributed to post-training Llama 2 and Llama 3.
Taylor began his career advising the UK government before moving into quantitative finance and sports betting, where he developed an interest in statistics, forecasting, and time-series analysis. His open-source pyflux library provides Python tools for probabilistic time-series modeling.
In 2018, he met Robert Stojnic at a startup incubator, and together they founded Papers with Code to connect machine-learning papers with their implementations, datasets, benchmarks, and results. Meta acquired the company in 2019. There, Taylor expanded from research infrastructure into language-model development; he also co-authored AxCell, which automated the extraction of experimental results from academic papers.
Taylor was first author and research lead of Galactica, Meta’s 2022 scientific language model. The project explored curated scientific training data, repeated training on high-quality material, and intermediate reasoning tokens. Its troubled public launch sharpened his distinction between strong benchmark performance and reliable product behavior: a capable base model can still generate persuasive falsehoods without effective post-training.
He subsequently worked on Llama 2 and Llama 3 post-training and reinforcement-learning approaches to mathematical reasoning. Early experiments combined scientific continued pretraining, verifiable rewards, and value models; Taylor concluded that stronger reasoning also depends on sufficient base-model capability, compute, and context length. His account of that work treats reinforcement learning from human feedback as essential to converting raw language-model capability into dependable behavior.
Building agents that can handle uncertainty
Taylor founded General Reasoning with Chengxi Taylor to tackle long-horizon sequential decision-making: tasks that demand sustained judgment as conditions change and feedback arrives late. Its work includes:
- KellyBench: A benchmark Taylor co-authored requiring agents to predict Premier League matches, manage bankrolls, and adapt throughout a simulated season. Every frontier model tested lost money on average, exposing the gap between fluent analysis and effective decisions.
- OpenReward: General Reasoning’s shared environment platform supports training and evaluating language-model agents across reusable reinforcement-learning environments.
Taylor also advocates sovereign AI at the model layer, arguing that countries cannot build durable AI strategies solely around applications when advanced technological capabilities depend on underlying model intelligence. His public argument reflects his experience developing language models in the United Kingdom and building a British AI research company.
