Sam Deverett

prof_pic.png

I focus on helping clarify our understanding of the risks and opportunities from AI.

As a Research Engineer at the UK AI Safety Institute, I evaluate the capabilities of advanced AI systems. The goal is to equip policymakers, industry leaders, and the wider research community with an empirical understanding of the risks from AI so that we can develop it responsibly and regulate it appropriately.

I am also interested in the use of AI for promoting democracy. I believe language models have potential to scale deliberation and enable more representative, transparent, and accountable processes for collective decision-making. Projects in this vein include democratizing the governance of AI as well as automatically mapping the opinion landscape on a given topic.

Links to my relevant pages are in the icons above and below. I’m always happy to connect.

timeline

2024 - Technical Staff, Research Engineer @ UK AI Safety Institute
Evaluating the capabilities of frontier AI systems.
2024 - 2024 Research Fellow @ AI Futures Fellowship
Prototyped LLMs’ potential to scale deliberative democracy, with a focus on finding the best questions to ask people in order to surface points of consensus or cruxes of disagreements. I was supervision by Jonathan Stray and collaborated with the AI Objectives Institute.
2023 - 2023 Researcher @ MIT, Fraenkel Lab
Built an LLM for metabolomics research using retrieval-augmented generation.
2023 - 2023 Researcher @ AI Safety Hub
Produced workshop paper on defining and preventing collusion in multi-agent reinforcement learning settings. I served as the team lead, and we were supervised by Lewis Hammond.
2022 - 2023 Data Science Engineer @ DraftKings
A “Data Science Engineer” is a Data Scientist + Machine Learning Engineer.
I worked on the Risk & Fraud team. The job was full-stack: formulating business cases, creating data pipelines, developing machine learning models, deploying them to production, setting up monitors, and responding to errors. The two projects I worked most on were quantifying the skill of a bettor and catching identity fraud during a registration attempt.
2021 - 2022 Product Analyst, Data Science @ DraftKings
I worked as an analyst for the data science team. This invovled tasks like designing KPIs, running experiments to quantify the impact of new models, conducting deep dives to understand customers’ behavior, building dashboards, and presenting insights to stakeholders from around the business.
2017 - 2020 Student @ UC Berkeley
Bachelor’s in Data Science
Regents’ and Chancellor’s Scholar
Certificate in Entrepreneurship & Technology

publications

2023

  1. Defining and Mitigating Collusion in Multi-Agent Systems
    Jack Foxabbott*, Sam Deverett*, Kaspar Senft*, and 2 more authors
    NeurIPS Multi-Agent Security Workshop, Dec 2023

selected repositories