# The AI Maths Revolution Has Begun Source: https://www.youtube.com/watch?v=IKjfrFMjz08 Recap page: https://rapidrecap.app/video/IKjfrFMjz08 Generated: 2026-02-08T16:33:35.137+00:00 --- ## 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. ![Screenshot at 00:38: The video shows a tweet from Kevin Weil claiming GPT-5 found solutions to 10 previously unsolved Erdős problems, which is immediately contrasted by an update from Mark Sellke clarifying that GPT-5 only found existing references, suggesting the initial claim was an overstatement.](https://ss.rapidrecap.app/screens/IKjfrFMjz08/00-00-38.jpg) **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. ### The AI Math Hype Cycle - GPT-5 allegedly solved 10 Erdős problems - This claim was immediately walked back by researchers noting the AI found existing literature, not new proofs - Demis Hassabis called the situation 'embarrassing'. ### The Role of AI in Mathematics - AI excels at literature search, finding a counterexample to an Erdős conjecture missed for decades - AI reasoning is weak compared to finding novel solutions to hard problems like the Millennium Problems. ### Economic & Societal Impact - Employees using AI tools earn 40% more than peers - The speaker suggests the commoditization of math knowledge is not yet happening, but the shift toward AI reliance is evident. ### Course Promotion - Outskill offers a 2-Day Generative AI Mastermind - Covers foundations, prompt engineering, building AI bots, and automations - Free for the next 48 hours, usually priced at $395. ![Screenshot at 00:05: The title card showing the topic "AI in Maths" against a background of mathematical formulas.](https://ss.rapidrecap.app/screens/IKjfrFMjz08/00-00-05.jpg) ![Screenshot at 00:20: Visual metaphor comparing AI math assistance to putting math problems into a shopping cart, questioning "Maths as Commodity?"](https://ss.rapidrecap.app/screens/IKjfrFMjz08/00-00-20.jpg) ![Screenshot at 00:38: Tweet from Kevin Weil claiming GPT-5 solved 10 Erdős problems, setting up the central conflict of the video.](https://ss.rapidrecap.app/screens/IKjfrFMjz08/00-00-38.jpg) ![Screenshot at 00:55: Tweet from Sebastien Bubeck confirming the event and noting Mark Sellke joined OpenAI, adding context to the researchers involved.](https://ss.rapidrecap.app/screens/IKjfrFMjz08/00-00-55.jpg) ![Screenshot at 01:12: Screenshot of the erdosproblems.com website listing various unsolved problems categorized by mathematical field.](https://ss.rapidrecap.app/screens/IKjfrFMjz08/00-01-12.jpg)