I’m a freshman studying Computer Science at the University of Washington. I like learning about ML systems, continual learning, multi-agent reinforcement learning, and mechanistic interpretability. Previously, my research focused largely on applications of machine learning in biology, including assessing the biological risks of genomic language models, with work presented at NeurIPS and CVPR workshops; forecasting malaria incidence using machine learning, published in IEEE; generative drug discovery with GANs; biologically inspired spiking neural nets; and developing language models to better understand VDJ recombination.

Some things I am interested in:

Intelligence. How can we get LLMs to learn continuously like humans and be sample-efficient? And what new infrastructure do we need to build to support this? Is this solvable through Test-Time Training, or do we settle for one-time compact insertions of large corpora into a model’s KV cache? Can we do this with SGD and backprop, or is some new learning algorithm needed? Speaking of which, why are humans so sample-efficient in the first place -- what is it about our neurons that allows us to learn from small amounts of data?

Augmentation. For all of human history, automation has allowed humans to get more things done. Why is this, and how can we ensure it continues as AI eats at the cognitive layer of decision-making?

Personalization and Interfaces. In the 1960s, users would interact with a mainframe from their little terminals. Access to this giant computer was limited to a chosen few, bottlenecked by time-sharing. Then came Licklider and his dream of personal computing, sparking the creation of the Macintosh and the GUI, making computers accessible to billions. It feels like we’re in the 1960s for AI. We prompt large, centralized models given a slice of compute, and it’s impossible to use them forever (rate limits). How can we give everyone their own personal models that continuously adapt to them? And what is the equivalent of AI’s GUI, just as prompting is the equivalent of terminal commands?

Telepathy. The amount of progress we have made with AI is astonishing and superlinear. Superintelligence will dawn on us within the next 15 years, whether through LLMs or not. Will we be to ASI as ants are to us? It is already getting extremely hard to control these systems and get them to do what we want: Pandora’s box is open, and there’s no going back. Today's clunky interfaces don't do justice to the raw, tangled intelligence LLMs have. The only way we can keep up with the pace of LLMs (quite literally, tokens/sec) is through our brains and our thoughts. New models that can continuously learn, decode, and encode neural data noninvasively, combined with LLMs’ capabilities, will usher in a new era of personalization and self-extension, keeping humans in the loop and making human-AI interaction frictionless.

Note: I shamelessly stole this website template from Patrick Collison.