Adrian de Wynter

Adrian de Wynter is a principal scientist at Microsoft and a researcher (PGR) at the University of York. His main work is related to fundamental problems in AI and science, such as cognition (e.g., understanding, reasoning) in machines, measurement (in science, like evals), and social impact.

His primary research interest is the interplay of intelligence and reasoning between humans and machines. This touches upon foundational aspects of AI, such as the question of whether showcasing understanding is required for 'good' dialogue (it's not); or whether LLMs are able to learn, as opposed to relying on their intrinsic knowledge (they can). He also studied the ability of LLMs/agents to interact with their environment (by using DOOM!). Measurement-wise, he developed an algorithm with cryptographic guarantees for determining trust in evaluators (e.g., LLMs-as-judges).

Earlier theoretical work showed with category theory that some prompting strategies are (formally) better than others, and how (and when) LLM-based data augmentation works. Even earlier research--because apparently he's getting old--worked on mathematical aspects of NAS, such as undecidability and approximation algorithms.

His other research interests include recreational mathematics (games), preserving endangered languages, and computational social science. In the latter he has worked on mitigating toxicity, unfairness, and other harms of LLMs; personalisation and sycophancy in LLMs; research on research (e.g., LLMs); and did one of the first studies of the impact of ChatGPT on loneliness. He also has stuff in SIGBOVIK.

In terms of service, he has--like everyone and their grandma--reviewed for AAAI, *ACL, NeurIPS, ICML, ICLR and so on (and was named a gold reviewer, which honestly is really nice). He also reviews for Nature (AI & Ethics; Artificial Intelligence, and Communications), IEEE Transactions on Games, and ACM TIST. He was recently named a Thinking About Thinking fellow.

In his spare time, Adrian enjoys photography, cooking, type 2 fun, and divulging personal information to strangers on the internet.