My central interest is emergence: how molecules assemble into cells, how neural circuits give rise to thought, how order arises from local interactions.
I study deep learning systems empirically and theoretically, with a focus on explaining puzzling phenomena in how they learn.
Much of my work so far has been on how diffusion models learn to factorize and compose concepts — the mechanisms behind compositionality in generative models.
Currently I'm studying evolution strategies for finetuning large language models, and how they compare to gradient-based methods.
Biological connectomes are shaped by simple, local growth rules: each neuron forms synapses guided by spatial gradients, cell-type-specific motifs, and plasticity constraints. Collectively, these processes produce networks capable of memory, stability, and computation.
Artificial networks, by contrast, are engineered by global optimization, which hides how structure and function co-develop. I build models of how computational properties emerge from cortical wiring rules.
Beyond intelligence, I'm interested in self-organization and self-replication: what does a system need in order to make stable copies of itself?
I use artificial life systems — Lenia in particular — as a testbed, and evolution as a tool, to look for the conditions under which self-replication emerges.
How language models store, retrieve, and reason about their own knowledge — and where the geometry of memory breaks down into conflict and confident hallucination.
Synthetic neural networks that grow rather than are designed — governed by probabilistic rules parameterized by biologically inspired gradients and inhibitory motifs.
Why evolution strategies with surprisingly small populations can finetune LLMs of billions of parameters.