Martin Burke PhD
Quantitative Researcher | AI, Model Validation & Computational Science | PhD, Imperial College London
Martin Burke is a quantitative researcher working at the intersection of AI and quantitative finance. Currently researching AI-safety debate frameworks for the UK AI Safety Institute — with broader interests across mechanistic interpretability and applied AI — he previously spent a decade validating derivatives and counterparty-credit-risk models at banks including Citi. He holds a PhD in Theoretical Chemistry from Imperial College London, where he pioneered the application of sum-of-squares optimisation — a technique from real algebraic geometry — to molecular energy minimisation.
Now focused on AI — building practical AI applications, with a particular interest in AI safety and mechanistic interpretability.
- AI research (safety, mechanistic interpretability, applied AI / software development)
- Quantitative finance (model validation, derivatives)
- Computational / theoretical chemistry
About
Two themes have run through my whole career: I'm fascinated by using computers to predict future behaviour and to model what can't easily be measured, and I see connections others miss. Early on I realised most technical fields are really the same mathematics applied to different problems. That same interest also draws me to applications such as algorithmic trading, where prediction has to contend with noisy data, changing incentives, and systems that react to being modelled.
My path runs from building a venture-backed start-up's R&D team from scratch to AI research today. Throughout my career I noticed that the same handful of equations resurface in different disguises.
- The finite-element solvers I used to predict how coronary stents would behave during a ten-year service life are, mathematically, the very PDE methods I'd later apply to price derivatives.
- The coarse-grained molecular dynamics I ran to model fibril aggregation in Huntington's disease (published in PNAS) propagates the Langevin equation — which, as the Ornstein–Uhlenbeck process, is how interest rates are modelled in finance.
- The convex relaxations of non-convex problems I built to predict molecular structure in my PhD thesis rest on the same optimisation and linear algebra that now trains deep networks for next-token prediction.
- Those very relaxations are now an AI-research tool in their own right: they can certify that no adversarial input will fool a trained neural network (Raghunathan et al., NeurIPS 2018).
Currently my focus is Large Language Models. Beyond next-token prediction, chain-of-thought reasoning offers, for the first time, a computational model of human reasoning — something never previously measurable. I'm currently researching recursive debate frameworks for the UK AI Safety Institute — constructing obfuscated arguments to probe scalable oversight — and I'm increasingly drawn to mechanistic interpretability and to building practical AI applications.
I believe large language models are reshaping how every discipline works with data of all kinds — knowledge and expertise are being commodified, and the future belongs to those who wield these tools effectively, and to those making sure we can trust them.
Read more about my PhD research →
“It is by logic that we prove, but by intuition that we discover.” — Henri Poincaré. In the original French: “C’est par la logique qu’on démontre, c’est par l’intuition qu’on invente.”