Publications

GradDFT is a JAX-based library enabling the differentiable design and experimentation of exchange-correlation functionals using machine learning techniques.

This article highlights limitations in the general-purpose application of language models like GPT-3 when directly utilized by humans, emphasizing the need for a novel evaluation framework to discern between failures due to capability constraints versus misunderstandings of user intent, ultimately indicating that these models currently lack comprehensive understanding of human commands and are far from achieving widespread general-purpose utility in real-world scenarios.