Two Futures for LLMs in Mathematics (wiredream.com)

This week, two major developments involving large language models (LLMs) in mathematics emerged from Anthropic and OpenAI, showcasing contrasting approaches in the realm of theoretical computer science. Anthropic collaborated with researchers Josh Alman and Virginia Williams to refine matrix multiplication techniques, leading to a groundbreaking reduction in the time complexity of the long-standing 3SUM problem, known for its computational hardness. Their work demonstrated that this classical problem can be solved in a time complexity of O(n1.9992), challenging previously held assumptions about its difficulty and hinting at a broader potential to refine our understanding of similar computational problems. Conversely, OpenAI released a repository of over 700 papers, which included ambitious claims regarding matrix multiplication, notably proposing a sharp reduction of the exponent of matrix multiplication (ω) to 2.25. If validated, this would mark a significant advancement, presenting the largest reduction since the 1980s. However, the quality of submissions in the repository has raised concerns, as many lack rigorous human review and contain unclear notations that can confuse readers. The stark contrast between the two approaches highlights the critical role of clarity and accuracy in research dissemination, particularly when leveraging advanced AI technologies to contribute to academic progress.