The field of Artificial Intelligence (AI) has taken great strides in recent years, particularly with advancements in Large Language Models (LLMs) like ChatGPT. This progress has stirred much public curiosity. However, general-purpose language models like ChatGPT sometimes 'hallucinate' or invent non-existent information when handling specialized subjects. This tendency has led to a common belief that while these models can handle everyday tasks efficiently, they falter when dealing with more specific or niche subjects. Our latest research challenges this perception. We demonstrate that through specialized methods such as in-context prompting and fine-tuning using astronomy literature, LLMs can generate scientific hypotheses with a complexity comparable to those conceived by human experts. We also introduce a technique involving 'adversarial prompting,' where two AI systems engage in a conversation, with one creating scientific hypotheses and the other critiquing them. This dialogue enhances the abilities of both models, leading to the generation of increasingly sophisticated hypotheses. We argue that the field of physical sciences, specifically astronomy, provides a safe playground to explore the capabilities of modern LLMs. This exploration could fundamentally redefine our perceptions of intelligence and knowledge.
Host: Shude Mao
