I am a PhD Student studying AI Safety at the University of Toronto with my supervisor David Duvenaud. I am broadly interested in the alignment problem — the problem of aligning the values and objectives of artificial agents to those of our own. I am currently working on goal specification using the assistance formalism. Instead of attempting to directly train an agent on a hard-to-formalize objective, is it easier to train a reinforcement learning agent to solve a general assistance problem and then have the agent interact with a human to learn the objective? ¯\_(ツ)_/¯. I also enjoy software design and have been developing relearn, a Rust library for reinforcement learning. Check out my GitHub profile for other projects.
My other research includes work on agent incentives as part of the Causal Incentives Working Group. Specifically, on formalizing incentives and analysis of incentives when the MDP assumption is violated. Before starting my PhD I worked on machine learning at Thalmic Labs (a.k.a. North) for the Myo gesture-recognition armband and the Focals smart glasses then moved into AI research with the Google Brain Residency program.