{"id":1330,"date":"2026-08-17T00:50:00","date_gmt":"2026-08-16T15:50:00","guid":{"rendered":"https:\/\/onepress.co.kr\/?post_type=briefing&#038;p=1330"},"modified":"2026-08-17T00:55:30","modified_gmt":"2026-08-16T15:55:30","slug":"2026-08-17-ai-erdos-unit-distance-disproof-en","status":"publish","type":"briefing","link":"https:\/\/onepress.co.kr\/index.php\/briefing\/2026-08-17-ai-erdos-unit-distance-disproof-en\/","title":{"rendered":"AI breaks 80-year-old Erd\u00f6s guess, verified by humans after just one question"},"content":{"rendered":"<p><strong>2026-08-17 00:50 KST<\/strong><\/p>\n<p>After being asked a single question, the AI \u200b\u200bmodel destroyed nearly 80 years of conjectures related to Erd\u00f6s&#8217;s unit circle distance problem with a counterexample.<\/p>\n<p>This does not mean that all difficult problems in mathematics have been automatically solved. The results were reviewed by company researchers and independent mathematicians, and details of the models and learning materials used were not disclosed.<\/p>\n<h2>What was the problem?<\/h2>\n<p>This is a combinatorial geometry problem that asks how coloring and structure are limited in a unit circle distance graph that connects points on a plane at a constant distance. Mathematician Paul Erd\u00f6s had speculated since the 1940s that a certain form would always hold, and for decades there were no general proofs or counterexamples.<\/p>\n<h2>What did AI do?<\/h2>\n<p>Once the OpenAI researchers presented the problem in natural language, the internal reasoning model developed counterexample construction in a previously uncommon direction. The key was to find finite points and connecting structures that satisfied the conditions and broke the conclusion of the guess. The results were organized into human-readable arguments.<\/p>\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" src=\"https:\/\/onepress.co.kr\/wp-content\/uploads\/2026\/08\/ai-erdos-unit-distance-disproof-en.png\" alt=\"AI breaks 80-year-old Erd\u00f6s guess, verified by humans after just one question\" loading=\"lazy\" \/><figcaption class=\"op-briefing-image-caption\">This is a generated image created to illustrate a topic and is not a photo of an actual scene or observation.<\/figcaption><\/figure>\n<h2>How did you check the correct answer?<\/h2>\n<p>Mathematicians outside the company independently reviewed the construction and logic to confirm that the counterexample met the conditions. Mathematics has the advantage that parts can be double-checked by calculations or formal verification. Still, it was not resolved solely through model output, and people participated in problem definition, proof theorem, and public verification.<\/p>\n<h2>What wasn&#8217;t revealed<\/h2>\n<p>The company disclosed the proof and explanation, but did not disclose which internal model it used or all learning materials. Therefore, it is difficult to fully assess from the outside the impact of existing literature and the reproducibility of different models. The mathematics community points out that attribution, agreement and verification rules must keep up with the pace of technology.<\/p>\n<h2>Why it&#8217;s important<\/h2>\n<p>This is an example that shows that AI can go beyond summarizing known solutions and contribute to the search for new counterexamples. At the same time, being able to come up with a good answer once is different from having consistent research skills. Publicly available tools, reproducible experiments, and rigorous peer review are the next criteria.<\/p>\n<h2>official source material<\/h2>\n<p><a href=\"https:\/\/openai.com\/index\/model-disproves-discrete-geometry-conjecture\/\" target=\"_blank\" rel=\"noopener noreferrer\">OpenAI result and proof<\/a><\/p>\n<p><a href=\"https:\/\/www.nature.com\/articles\/d41586-026-01651-0\" target=\"_blank\" rel=\"noopener noreferrer\">Nature independent report<\/a><\/p>\n<p><a href=\"https:\/\/www.nature.com\/articles\/s42256-026-01269-x\" target=\"_blank\" rel=\"noopener noreferrer\">Nature Machine Intelligence editorial<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>OpenAI&#8217;s internal inference model proposed constructing a counterexample to Erd\u00f6s&#8217;s conjecture regarding the unit circle distance graph, and independent mathematicians confirmed the argument, but details of the model and training materials were not made public.<\/p>\n","protected":false},"featured_media":0,"template":"","meta":[],"class_list":["post-1330","briefing","type-briefing","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/onepress.co.kr\/index.php\/wp-json\/wp\/v2\/briefing\/1330","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/onepress.co.kr\/index.php\/wp-json\/wp\/v2\/briefing"}],"about":[{"href":"https:\/\/onepress.co.kr\/index.php\/wp-json\/wp\/v2\/types\/briefing"}],"wp:attachment":[{"href":"https:\/\/onepress.co.kr\/index.php\/wp-json\/wp\/v2\/media?parent=1330"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}