Fellowship Handbook

Before joining Iliad, I studied Mathematics and Cognitive Science at Williams College, graduating in June 2026. By the time I applied, I knew I wanted to pursue abstractions research: understanding how agents organise what they know about the world, and whether there are underlying structures that different kinds of agents have in common.

I joined the Iliad Fellowship because it could support that direction. Very few organisations help junior researchers work on agent foundations, and many of the usual routes into AI safety focus on other approaches. For me, finding a programme where these questions were welcome was a major part of the appeal.

My research asks how agents break their knowledge of the world into modular, reusable pieces, and how much of that structure is shared across different agents and cognitive architectures. During the Fellowship, I focused on two frameworks in particular: Wentworth and Lorell’s natural latents, and Eisenstat’s condensation. These gave me concrete approaches to investigate the broader questions I was interested in.

My mentor and I developed a great working relationship, and continuing into an extension felt like a natural next step. That relationship was central to how my Fellowship developed. Looking back, I would encourage prospective Fellows to pay attention to the fit between their interests and their mentor’s, alongside the research topic itself.

I’m now continuing this work during the extension, applying condensation to a range of toy settings. The aim is to understand its shortcomings more precisely and work out how to address them. Iliad has given me a way to keep pursuing a research direction that can be difficult to find support for early in a research career.

Before moving into AI safety, I was a postdoctoral researcher in physics. There weren’t many jobs in my particular subfield, so I started looking at what a transition out of academia could look like. I wanted to keep doing intellectually interesting work, but I also cared about what that work would contribute to. Quant finance and projects aimed mainly at making AI more capable didn’t particularly appeal to me. My interest in safety came partly from wanting a different direction, and from concerns about whether the organisations developing AI would use it responsibly.

I had already done MATS before joining Iliad in summer 2026. In my experience, there was a strong focus at MATS on producing experimental results: “make a plot,” then keep moving. I wanted more room to step back and ask whether the assumptions behind an experiment made sense, or whether we were pursuing the right question. Iliad gave me that freedom, alongside a research direction closer to my theoretical interests.

I worked with Charles, another Fellow, with Dmitri Vaintrob as our mentor. Charles brought extensive AI knowledge, and we spent much of the first three months reading papers and exploring ideas together. It was valuable, but also frustrating. We weren’t finding original results, and we sometimes got on each other’s nerves. Looking back, I wish I’d better understood how much time building expertise in a new area could take. By the end of those three months, I had developed a much deeper understanding of the literature and the questions we wanted to pursue.

Things started coming together almost immediately after we began the extension, with promising early results in a direction connected to our earlier explorations. A talk at the Iliad conference helped us connect those interests to robustness and error correction in neural networks. We’re now investigating why networks can develop discrete structures in their internal representations, where substantial changes to an internal activation can leave the output largely unaffected. And whether understanding those structures could make networks easier to interpret.

The work is still exploratory, but we now have a direction we’re excited about. What Iliad gave me was the time to think deeply, a collaborator to work through ideas with, and a community of peers and senior researchers. Together, those helped turn a frustrating period of learning into a research project I want to keep pursuing.

There is a great diversity of experiences and stories in AI Safety. I came to Iliad after a long career in academia, working at the intersection of mathematics and theoretical physics (most recently as a professor at the University of Edinburgh). Much of my work in the past two decades has focused on translating between these two fields, harnessing intuition from one to produce surprising results and better understanding of the other. In recent years, I wanted to explore new research areas and was inspired by the amazing developments in AI for science. I also found the problems of interpretability and alignment to be both scientifically fascinating and incredibly important. I decided I needed to quickly learn more and try to help.

I signed up for the Iliad Intensive in April 2026, and with Iliad’s help organized an AI Safety event for our math and physics students (and faculty) in Edinburgh in May. At the same time, I got in touch with former colleagues working in AI safety, most notably at Timaeus. I was very strongly drawn to the perspective Timaeus took in its research, combining a creative application of techniques from geometry and physics with careful experimental validation, all focused on characterizing and controlling generalization in models. When offered an Iliad Fellowship, I reached out directly to Daniel Murfet at Timaeus to ask if he could serve as my mentor, and he graciously agreed.

My Fellowship research in the summer of 2026 was a combination of technical experimentation, theory development, and exploration of new ideas. I was given substantial leeway in charting my own course, which (in my case) was hugely beneficial: I could find projects that were best suited to my tastes and background. My main technical work was in determining how best to apply “patterning” methodology from SLT to steer for desired behaviors in reinforcement learning, and to prevent reward hacking. I worked on this with a small team of Timaeus researchers, as I got to know the company better. I also learned a great deal from conversations with other fellowship participants and researchers at LISA; a group of us set out to learn about physics-inspired approaches in alignment, which turned into a new Physics of Deep Learning day at the Iliad Intensive.

It became increasingly clear during this summer that I wanted to make alignment research a full-time pursuit. The science was fascinating, the community was enthusiastic and supportive, and the goals were becoming critical. It had been a long time since my research moved this quickly and could have such a large potential impact. As Timaeus transitioned to part of Resolution, I applied for a full-time position with them. I’ve gone on leave from my university job and am now working at Resolution. I’m doing research across several areas, while promoting connections with the senior academic community.

I feel incredibly fortunate to have been able to make this change. My route to the Iliad Fellowship was somewhat nonstandard, having joined with substantial senior research experience but working in a completely new area. At no point were transitions automatic, from Intensive to Fellowship, Fellowship to a full-time job. However, the experience, knowledge, and contacts I gained during the Fellowship did open up a vast array of opportunities to work in AI safety, and I feel strongly that something would have worked out no matter what. Be it leading a research group in alignment back at my university, starting my own nonprofit, or joining a growing organization like Resolution.