research: empirics, theory of learning algorithms in interactive, stochastic environments; building efficient, scalable real-world systems. some recent themes include:

measuring representation quality for model selection in self-supervised learning
offline reward modeling, reinforcement learning with limited, suboptimal data
algorithms for low-latency inference in ml pipelines
structured and compositional generative models (for code, images, audio)
previously, i've worked on fast, simple algorithms for off-policy robot learning, generative models for music, program synthesis.