Andrey Gromov

I used to think about AI as a scientific problem: how intelligence emerges from compute, data, and sampling, and how far simplified models can take us in explaining the behavior of larger systems. Today I mostly think about how to design and build a system that will answer these questions for me.

• Autonomous AI science.
• Science of LLMs: initialization, scaling laws, post-training, evaluation, interpretability, sampling.
• Science of data: toy models, artificial data, model collapse, curricula.
• Science of reasoning: inference and pre-training trade-offs.
• Science of efficiency: pruning, quantization, distillation, efficient architectures.

The unifying theme of my physics research was emergence in classical and quantum systems. I was particularly interested in phenomena where topology, geometry, and interactions between constituents conspire to create unique quantum material properties.

• Topological phases of matter and fractional quantum Hall effect
• Fractons and topological glass
• Active matter
• Correlated systems in curved space