I have been one of the convener of a special session about AI and LLMs in theoretical physics research at the conference GeomGravX 2026. During it, Ginevra Braga has presented novel work on AI-assisted computations in DHOST theories, Sotirios Karamitsos has presented a critical assessment of positive and negative experiences the conference participants have had in using AI/LLM tools, as well as to the ethical and practical issues that inevitably arise. My intervention on a summary of AI benchmarks in theoretical physics and explaining basic concepts in LLMs can be found in the video below. The slides of my intervention are available here.
One of the highlights of this presentation is the non-hyped rational assessment on the true capabilities of LLMs, the need for a better symbolic verification, open-source CAS, open-weight LLMs, and the possible ethical concerns from AI companies paying for data used during benchmarks.