Ching Lam CHOI.

I’m interested in cryptographic approaches to AI safety and robustness: property testing to certify model behaviour, one-way functions for the safe deployment of open-source models, and zero-knowledge proofs to probe convergent and emergent multi-agent systems.

I’m also working towards Platonic world models that reason across modalities, including sim-to-real systems in which agents perform experiments following the scientific method. My recent work explores the structure of synthetic data, reliable representations, and what makes data useful for learning.