OpenAI claims solved 1946 Erdős problem

OpenAI said this week that one of its unreleased reasoning models solved a long-standing geometry problem first posed by mathematician Paul Erdős in 1946, adding a scientific claim to a week in which the company also expanded government ties in Japan and faced fresh public criticism in London. The company presented the math result as evidence that frontier AI systems can contribute to research rather than only generate text. At the same time, OpenAI board member Paul Nakasone said in Tokyo that the company plans to provide Japan’s government and some companies with GPT-5.5-Cyber, a model built for cyber defense. In London, fake Underground posters accusing ChatGPT of links to teen suicides circulated as debate over AI safety and trust widened. ### Which math problem is OpenAI talking about? OpenAI said the result involved a geometry question dating to 1946 and attributed to Erdős, according to reports published on May 21. Indian Express reported that the company said an unreleased reasoning model solved the problem autonomously, and other coverage described the question as an 80-year-old Erdős geometry problem about point configurations in the plane. (indianexpress.com) The claim matters because OpenAI framed it as scientific work rather than a benchmark score or product demo. Indian Express said the company presented the result as a breakthrough for AI-assisted mathematics, while outside reports described it as a case in which a general-purpose reasoning model, not a math-only system, produced a new construction. (indianexpress.com) ### What did OpenAI announce in Japan? Paul Nakasone said at a Tokyo press conference on Thursday that OpenAI plans to provide its latest cyber-defense model, GPT-5.5-Cyber, to the Japanese government and some Japanese companies. The Japan News reported that Nakasone said discussions with Japanese officials this week covered cybersecurity across 15 vital infrastructure fields and that talks to move quickly on access would proceed. (indianexpress.com) GPT-5.5-Cyber has already entered limited provision to some government bodies and organizations this month, The Japan News reported. The outlet said OpenAI expects Japan to be among the first governments to receive the model, which it described as able to detect weaknesses in critical infrastructure systems and help with remediation. (japannews.yomiuri.co.jp) ### Why were fake OpenAI posters appearing on the London Underground? London commuters saw fake posters styled as OpenAI advertisements that accused ChatGPT of links to teen suicides and other harms, according to International Business Times and other reports published on May 21 and May 22. The posters were not official OpenAI material, but they reignited public debate over AI’s effects on young users, education and mental health. (japannews.yomiuri.co.jp) The campaign also connected to earlier criticism of chatbot safety. The Center for Countering Digital Hate said in a previous report that ChatGPT could generate dangerous advice related to self-harm, suicide, eating disorders and substance abuse, though the London posters themselves were a protest action rather than a company statement. (ibtimes.co.uk) ### Are buyers moving away from closed models? Le Monde Informatique reported on May 22 that IT decision-makers are showing growing interest in open-weight models that are smaller, cheaper and easier to customize than large proprietary systems. The publication cited Gartner analyst Deepak Seth and Jozu co-founder Jesse Williams saying open models can offer more visibility, control and flexibility for enterprise users. (counterhate.com) That shift does not mean proprietary systems have disappeared from the market. Le Monde Informatique said ChatGPT and Google Gemini remain popular, and OpenAI itself now offers open-weight models under its gpt-oss line, which the company says are available under an Apache 2.0 license and designed to run locally or in data centers. (lemondeinformatique.fr) ### What comes next from here? OpenAI has not publicly released the reasoning model it says solved the Erdős problem, and the company’s next step on that claim will depend on whether it publishes fuller technical details or outside mathematicians validate the result. In Japan, Nakasone said talks on providing GPT-5.5-Cyber would move ahead quickly, and The Japan News said access for some government bodies and organizations has already begun this month. (lemondeinformatique.fr) (indianexpress.com)