Artificial intelligence tools created in recent years have the potential to revolutionize many aspects of human life.

The capabilities of Large language models (LLMs) and other deep learning approaches have surprised even their creators. The AI revolution is only taking its first steps; One can only imagine where it will stand in five or ten years from now.

Still, the use of AI tools in many fundamental and basic science fields remains limited, creating a substantial gap between expectations and reality. The challenges scientists face when attempting to harness AI's potential are unique to each field, with some being more prevalent than others.

AlphaFold, the protein-folding prediction system, is considered the most significant achievement to date in leveraging AI to advance science. Nevertheless, the success of AlphaFold and LLMs only emphasizes the common challenges in employing AI across many scientific disciplines. These models rely on large (and in the case of LLMs, incredibly large) annotated datasets and are designed to perform specific tasks. 

In many fundamental science fields, numerous challenges exist in data collection. Furthermore, when the data is plentiful, the main hurdle often becomes deciding what to examine, identifying what it signifies, and annotating it.

An additional common challenge is the immense intricacy of real-life phenomena. Even when computers learn to cope with this complexity and generate accurate predictions about future occurrences, the crucial question in science often shifts from knowing what will happen, to understanding why. This necessitates the development of tools to scrutinize the AI 'black box'.

At the Weizmann AI Research Institute and AI Hub, we are identifying these challenges of integrating AI into different fields of fundamental science. We are developing new tools and models to manage and overcome these obstacles, and training the next generation of scientists who will carry forward the AI revolution in fundamental science in the future.