“The long-term vision is: foundation models that acquire reasoning from fully synthetic data, then learn semantics from a small, curated corpus of natural language. This would help us build models that reason without inheriting human biases from inception.”