Sid Baines - Aligning AIs is hard: why work on it?

Wednesday 28 Oct 2026, 13:30 → 15:00 Europe/London

610 (G.O. Jones Bulding)

https://cern.zoom.us/j/67501511814?pwd=4RlJLpbhXENmVQVLaPAbMuweQ5xIwM.1

About the speaker: 

Hi! I’m an ex-QMUL PPRC PhD student (2021-2025) - I worked with Ulla Blumenschein to study heavy charged Higgs bosons decaying into hW^+ at ATLAS. Since completing my PhD, I completed research upskilling in AI safety at ARENA, research projects at SPAR, and a fellowship at LASR Labs, in which I studied shaping LLM personas. I am now a researcher at Arcadia Impact Alignment Team, where I perform technical alignment research with the aim of reducing loss-of-control risks. In my free time, I like to climb, cook new foods, and play board games 🙂

1 Biscuits and Coffee

Aligning AIs is hard: why work on it?

Five years ago, frontier AI was able to perform tasks which take a human approximately 30 seconds, like answering a simple question or counting the words in a short passage. Three years ago, it was able to perform basic lookup tasks taking humans a few minutes, like searching for a fact on wikipedia. One year ago, it was able to perform multi-hour projects, like ‘train an adversarially robust image model’. Last month, claims were made that frontier AI models solved a millennium problem, amongst other open problems in mathematics.

However, it is not clear that our ability to ensure these models are aligned with human values is keeping up. Recent events have made this issue more salient to the public, with multiple incidents over the summer involving collections of agents autonomously performing cyber attacks without instruction to do so. As models become more capable, it is possible that their abilities far surpass human abilities to control them, and we may face risks not from human-actors misusing the model, but instead from the models themselves acting against humanity. In the limit, this could lead to catastrophic risks or the permanent disempowerment of humanity, often referred to as ‘loss-of-control’ risk.

In this talk, I’ll discuss the progress of AI capabilities (historical and projected), and the growing gap between progress in capabilities vs. understanding models behaviours and motivations. I’ll then motivate and discuss my work in technical AI safety research, specifically on understanding what shapes model motivations, focussing on loss-of-control risks. I’ll also discuss my path into this work and opportunities for students to get involved.

Convener: Dr Sid Baines