Hi! My name is Guy (pronounced /ɡiː/ as in ghee). I am currently in the Finetuning team at Anthropic, making Claude more helpful and aligned with human values. Previously, I was scaling ML risk systems at TikTok and trained agents to collaborate and behave more like humans at Microsoft Research. I graduated from Imperial College London with an MEng in Mathematics & Computer Science.
Contributing to the development of safe and capable AI systems by making Claude more helpful and aligned with human values as part of the Finetuning team.
Mitigated spam, bots, fake engagement, inauthentic and harmful behaviours as part of the Risk Systems ML foundation team.
Conducted research on human-like and multi-agent collaboration in the Deep Reinforcement Learning for Games team.
Avanade, joint venture between Microsoft and Accenture, provides IT consulting and services focused on the Microsoft ecosystem.
Two friends and I wanted to attend talks and workshops about blockchain technologies. We realised Imperial College was missing such society, so we decided to create it.
Research in computer science at Imperial College London to automate proofs in number theory using the Lean Theorem Prover supervised by Prof. Kevin Buzzard. Open source project available on GitHub.
Mercuris Conseil is a startup specialised in developing web solutions for small and medium size businesses.
Obtained First-Class Honours.
Master's thesis: Multi-agent Deep Reinforcement Learning for Anatomical Landmark Detection, supervised by Dr. A. Alansary and Prof. D. Rueckert.
In this project in collaboration with World's Edge studio, we train Multi-Agent Reinforcement Learning (RL) naval units in Age of Empires IV . The aim, unlike related work, is not to reach superhuman performances and beat the top players. The aim is to build a more immersive and enjoyable experience for the players by creating more human-like opponents. This work was presented at GDC AI summit '22, “Age of Empires IV: Machine Learning Trials and Tribulations”. My co-speaker Peter Chan speaks about how battle outcomes are predicted using supervised learning. Disclaimer: the RL naval units are not in the released version of the game.
View talk (requires GDC subscription)
How well do people assess human-likeness in human- and AI-generated behavior? We present a qualitative study of hundreds of crowd-sourced assessments of human-likeness of behavior in a 3D video game navigation task. We focus on an AI agent that has passed a Navigation Turing Test. We give insights on the characteristics that people consider as human-like. Understanding these characteristics is a key first step for improving AI agents in the future.
CHI '22 publication
A key challenge on the path to developing agents that learn complex human-like behavior is the need to quickly and accurately quantify human-likeness. We present an automated Navigation Turing Test. We demonstrate the effectiveness of our automated NTT on a navigation task in a complex 3D environment. Our best models achieve high accuracy when distinguishing true human and agent behavior.
ICML '21 publication
My Master's thesis, supervised by Dr. Amir Alansary and Prof. Daniel Rueckert, achieves state-of-the-art accuracy in anatomical landmark detection. we explored approaches involving multiple cooperating agents with a focus on their communication in order to improve performances. The work was subsequently published in the MICCAI 2020 MLCN workshop. You can also read the original report here.
Integrated court detection, ball tracking, and multi-person pose estimation to gather metrics on tennis players such as step count, shot hit map, ball speed and shot types, in Java in a team of 5.