Three threads, one question: how do physical and computational systems learn?

The newest pillar — and, I think, the highest-leverage. The starting question: what if we treat the act of scientific discovery itself as an optimization problem over strategies, with the LLM as the proposal distribution and experimental feedback as the gradient signal? The first installment, Scientific discovery as meta-optimization, tests this framing on combinatorial optimization — a domain where the loop is closed, the metrics are real, and the agent has nowhere to hide.

From there, the program scales up: agents that frame problems, agents that design experiments, agents that draft and review papers. Crucially, everything is built so that failure is preserved. A discarded hypothesis, in this framework, is not noise — it is evidence. Aggregating those signals is what makes the second track eventually do science, not just generate text.

Long-term, I am after an LLM-native publication venue: a place where the artifacts are not PDFs but structured proposals, reproducible runs, and machine-readable critiques; where agents review each other; and where humans walk in for the parts they enjoy and the decisions that need taste.

Conventional computing has, for decades, equated parallelism with spatial multiplicity — more cores, more tensors. But physics provides a different kind of parallelism: collective dynamics with long-range order, where distant components of a system align without explicit communication. This is the insight behind memcomputing, and it is what my work tries to harness.

My PhD established several pieces of the theoretical foundation — directed-percolation in digital memcomputing machines, the self-averaging property that makes their convergence robust, and the relationship between memory and long-range order in dynamical systems. With Prof. Ivan Schuller's group, we showed experimentally that thermal neuristor networks — VO₂-based oscillators coupled through heat — exhibit exactly the collective dynamics the theory predicts.

Looking forward, the goal is a complete nonlinear dynamical-systems substrate for AI: neuromorphic devices built from novel materials, theory that explains when collective order is computationally useful, and architectures that exploit both.

Quantum many-body problems sit at the centre of condensed-matter physics; their state spaces are intractable, but their structure is rich, and that structure is what learning systems thrive on. I built Transformer Quantum State — a multipurpose transformer-based variational ansatz that handles a family of different Hamiltonians within a single architecture.

On the algorithmic side, my pre-PhD work used reinforcement learning to compile topological quantum gates, showing that braiding sequences for non-Abelian anyons can be discovered automatically with fidelity competitive with hand-designed sequences. More recently I have been working toward large quantum models: foundation-scale architectures that learn quantum states across systems, geometries, and interaction structures.