Could you tell us a little about yourself and about your studies before coming to Imperial?

I completed my undergraduate in mathematics at Cardiff University, with a year-long placement at the University of Warwick and CERN doing some particle physics.  I also undertook an internship, at Cardiff, working on adding functionality to a game-theoretic Python package.

After graduating (in 2018), I spent nine months traveling Europe in a van and thinking about what to do with my life.  I read some articles on “AI alignment”, and decided that this was probably both the most important and most interesting problem on which I could be working. (Note that these articles no longer represent my understanding of the risks from advanced AI.)

Immediately before the MSc in AI, I worked at the University of York on interpretable machine learning from a safety-engineering perspective.

What attracted you about the MSc in AI?

I applied to a handful of MSc degrees, but chose the MSc in AI at Imperial because of the following, relevant advantages.

• The MSc is one year, allowing quick upskilling.
• It’s located in London which is one of the best places to be for proximity to important organizations and events.
• The degree is open to students with strong maths backgrounds but doesn’t have extensive CS prerequisites.
• The module choice is flexible, and the range of modules on offer is broad (one student might specialize in, e.g., computer vision and deep learning, whilst another might specialize in verification and symbolic techniques, and yet another in NLP).
• Both the group project and individual project allow students great freedom over the project topic and direction, with world-leading mentorship from experts in the area.

What did you enjoy the most?

One of the best things about the MSc is the quality of the students admitted.  You will be surrounded by smart and motivated people from a variety of backgrounds.  In addition, the group coursework and group project give you the opportunity to work with your classmates, get to know each other, and learn from one-another.

What did you find more challenging?

The workload is intense, and there was a lot more coursework than I was expecting.  This MSc is not for the faint of heart!  However, be assured that if you are admitted then you will be able to get through it, and that on the other side you will be a high-functioning and capable expert in AI.

Could you tell us about some of your achievements on the MSc that make you proud?

I learned a lot on the MSc, both technical material and, more generally, how to get tasks done.  I’m particularly proud that I proposed my own individual project, on a topic that I was interested in (incorporating symbolic methods into reinforcement learning from human feedback), and feel that I grew as a researcher by doing so.

What did you do in your spare time?

I don’t recall having all that much spare time on the MSc AI!  I did manage to fit in a week-long climbing holiday in the Christmas break, and otherwise we spent numerous evenings in the post-grad bar chatting about AI.

Could you tell us about your individual project?

I proposed my own project, on “neuro-symbolic reward learning”.  In short, the aim was to attack the AI alignment problem, i.e., the problem of building AI systems that are trying to do what we want them to do (or that are “aligned” to our values).  Technically, the idea was to supplement reinforcement learning from human feedback with techniques from computational argumentation.  At a high-level, this should enable the system to reason logically about human preferences whilst utilizing reinforcement learning to train agents who try to satisfy those preferences.  I recently published a paper summarizing my MSc thesis, which can be read here.

On the one hand, I feel that proposing my own project helped me to grow as an independent researcher.  On the other, as an inexperienced student, my research ideas were not particularly good, and there are surely benefits from working on more concrete and well-defined projects.

What have you been doing since you graduated?

I am still at Imperial, pursuing a PhD with the Safe and Trusted AI CDT!  My focus has shifted and my technical work is now on manipulation and deception in AI systems, from a game theory and causal incentives perspective.  As well as my research, I do a bunch of stuff at Imperial and within the broader “AI Safety” community.

• In the first year of my PhD I undertook an internship at The Center on Long-term Risk, working on AI manipulation.
• I do a range of teaching, including giving a guest lecture on the Ethics, Privacy, AI in Society module on “Long-Term AI Risk”.
• I run a reading group and seminar series on the topic of AGI Safety (i.e. the safety of artificial general intelligence, or smarter-than human systems).

See my website for more, and more up-to-date, info. about me and what I’m up to.

Do you have any advice for prospective students?

I believe that it’s likely that we will build transformative, smarter-than-human AI some time this century.  In addition, I think there is a high chance that this will constitute an existential risk for humanity.  Furthermore, I think that there are technical and policy interventions that can be advanced now that will reduce this risk.  I therefore advise prospective students to work on this problem, if they feel motivated and able to do so.

Otherwise, the MSc in AI undoubtedly gives its graduates a solid foundation on which to build a career in industry or academia, doing either technical or policy work, with an eye to either contemporary or longer-term concerns.

It won’t hurt to get in a bit of Python practice before you start, too…