In recent years, Artificial Intelligence, notably through advancements in Large Language Models (LLMs) such as GPT, has garnered significant attention both within academia and the broader public sphere. However, these general-purpose LLMs have been criticized for their tendency to produce spurious or "hallucinated" information when grappling with specialized or technical domains.

In this seminar, Yuan-Sen Ting of the Australian National University and the Ohio State University presents how he and colleagues address this limitation with the first astronomy-centric LLM—AstroLLaMa—that can produce text completion and embedding that outperform GPT models. They show that LLMs can generate scientific hypotheses of a complexity comparable to those produced by human experts through techniques such as in-context prompting and fine-tuning on domain-specific literature. Their mission is to democratize the field of astronomy by developing public-facing, AI-driven large language model tools specialized for this discipline. They posit that these specialized foundational models can revolutionize the methods we employ for literature searches and the tracing of intellectual developments within the field. They argue that the physical sciences, particularly astronomy, serve as an ideal test bed for investigating the potential of modern LLMs. This inquiry stands to fundamentally reshape our understanding of both artificial and human intelligence and the boundaries of accumulated knowledge.