Large language models (LLMs) have shown remarkable ability in generating mathematical proofs, such as finite-time blow-up solutions to the Navier–Stokes equations and proposed proofs of the quasi-Riemann Hypothesis, signaling potential for accelerating discovery across science and technology. However, their integration into STEM ecosystems presents a critical choice: a 'golden age' where LLMs empower a new generation of scientists, develop new expertise, and preserve open, collaborative inquiry, or a 'dark age' where short-term gains in output come at the cost of weakening independent human scientific activity, concentrating knowledge production in few hands, and ultimately creating knowledge alien to human understanding. The author argues that reality will likely blend both scenarios.
A key challenge lies in LLMs' struggle to translate complex formal proofs (e.g., in Lean) into comprehensible natural language — despite generating correct code-like proofs hundreds of thousands of lines long, their natural language renderings are often messy and require intensive study to grasp. This raises a fundamental question: if humans cannot understand the proofs, what is the value of mathematics beyond benchmarking LLMs? Mathematics, deeply rooted in human cognition — from innate concepts of number, space, and time to culturally evolved systems — risks becoming disconnected from human sense-making if not carefully stewarded. The piece concludes with eight points examining the choices shaping this outcome and calls for actionable proposals to steer toward a future where technology augments, rather than replaces, human mathematical insight.
Source: Hacker News · Summarized by HeadlinesBriefing