Research
Current focus
I’m joining the AI Hub at BNY to work on agentic software engineering and trustworthy AI systems. The questions I care most about right now:
Long-horizon agents & continual learning
How do we build agents that carry work through over long horizons, and keep learning as they go? I’m interested in memory, credit assignment over long trajectories, and continual learning that lets an agent adapt to new tasks without forgetting old ones.
Scalable RL environments
Agents are only as good as the worlds they train in. I’m working on ways to create RL environments at scale that are diverse, verifiable, and open-ended enough to keep teaching. I’m also interested in semi-verifiable domains, where outcomes can only be partly or noisily checked, and in algorithms that can still learn reliably from them.
Benchmarks & evals
Measuring what agents can actually do: benchmarks and evaluations that hold up over long horizons, resist saturation and gaming, and say something about real-world reliability.
Trustworthy AI
Making agent behavior reliable, legible, and safe to deploy where the stakes are high: evaluation, oversight, and the engineering that lets people depend on what an agent does.
Alignment & AI governance
Keeping increasingly capable systems aligned with what people intend, and the institutions, policy, and legal frameworks needed to govern them well.
Frontier directions
Research paths toward AGI and ASI, and what it takes to get there well.
Also exploring
Meta-RL & open-ended adaptation
Agents that make sense of loosely specified environments, work out what they need to learn, and ask questions when they should, rather than waiting for a perfectly specified task.
Regenerative neurobiology
Neural regeneration, and what it would take to repair the brain and spinal cord.
Brain-computer interfaces & neuromechatronics
Neural decoding for robotic actuation and prosthetic control. Closed-loop Sim2Real neural decoding and spinal cord stimulation for restoration of limb control in paralysis.
Synthetic & engineering biology
Building an equivalent of AlphaGenome for bacterial genomes. Can we predict functional properties of microorganisms from their DNA genomes? Can we infer their metabolic reaction networks? And building on this, can we propose genome edits to have desired functional effects?
Philosophy
Philosophical questions raised by AI, engineering, and neuroscience; computational methods in philosophy; philosophy of physics, computational metaphysics, and more speculative approaches. Aristotle and Olaf Stapledon are perennial favorites.
Law & generative AI
Copyright law and computational frameworks for dealing with new challenges posed by generative AI. What constitutes fair use? What new paradigms of patent law are enabled by large AI systems?
Previously: Mount Sinai
At Mount Sinai I worked on applications of machine learning to neural signal processing, behavioral neuroscience, and functional genomics,1 and built foundation models for genomic DNA sequence-based tasks. Examples of that work include:
- Domain-generalized deep learning for detecting neuronal spikes in whole-cell patch-clamp postsynaptic current traces
- Forecasting lab animal behavioral events with computer vision and sequence modeling
- Predicting neuroanatomical sources of detailed electrophysiology spike traces2
- Foundation models for genomic DNA sequences
- Benchmarking of self-supervised learning methods for mammograms