The AI Maths Revolution Has Begun

Quick Overview

The purported AI math revolution, evidenced by GPT-5 solving long-unsolved Erdős problems, is immediately complicated by the website maintainer clarifying that the AI only found existing references, not new proofs, causing key figures like Demis Hassabis to call the initial claims embarrassing.

Key Points: GPT-5 was initially reported to have solved 10 previously unsolved Erdős problems, sparking claims of an AI math revolution. Mark Sellke, a researcher involved, later clarified that GPT-5 actually found existing references that solved the problems, which he was personally unaware of, leading to initial overstatement. Thomas Bloom, maintainer of erdosproblems.com, called the initial report a "dramatic misrepresentation," stating that 'open' status only means he is personally unaware of a solution. Demis Hassabis called the situation "embarrassing," implying OpenAI should have verified the claims before public announcement. The video highlights that while AI is useful for searching literature (like finding the counterexample {1, 2, 4, 8, 13} to a conjecture), the initial hype about autonomous proof discovery was premature. The speaker emphasizes that AI math tools currently function better as advanced search tools rather than autonomous theorem provers, citing that their reasoning capabilities are still weak. The video concludes by promoting an 'Outskill Generative AI Mastermind' course, offering access to AI knowledge and tools for $0 for the next 48 hours.

Context: This video from 'Science News with Sabine Hossenfelder' critiques the recent sensationalized reports regarding AI, specifically GPT-5, solving complex, long-standing mathematical problems like those posed by Paul Erdős. The context revolves around the gap between AI's actual capability (finding existing literature) and the public perception (autonomous breakthrough discovery), which the speaker argues is driven by hype and risks commoditizing mathematics.

Detailed Analysis

Sabine Hossenfelder discusses recent reports suggesting Artificial Intelligence, specifically GPT-5, solved ten previously unsolved Erdős problems, a feat framed by some as the beginning of a 'science acceleration' or even the 'singularity.' She immediately challenges this narrative by presenting tweets showing the clarification: researcher Mark Sellke confirmed GPT-5 found references to solutions that he personally was unaware of, not that the AI independently proved them. This led to criticism, including a tweet from Thomas Bloom, the maintainer of erdosproblems.com, calling it a "dramatic misrepresentation," and a comment from Demis Hassabis labeling the situation "embarrassing." Hossenfelder argues that the AI acted as a powerful literature search tool, finding a counterexample to one conjecture that had been missed by mathematicians for decades, but this is not equivalent to original mathematical discovery. She contrasts this with the Millennium Problems, which remain unsolved because they require deep, novel insight, not just reference retrieval. She concludes that current LLMs, while capable of complex computation and avoiding certain errors, exhibit weak logical reasoning and are still far from replacing human mathematicians for fundamental breakthroughs, comparing the hype to a 'tip of the iceberg' moment before a potential crash. The video ends with an advertisement for an 'Outskill Generative AI Mastermind' course, offering a limited-time free access to learning about 30+ AI tools.