April 2026: an amateur solves a 60-year-old math problem with AI, and OpenAI shuts down Sora

The amateur mathematician story

The problem in question was a combinatorics question involving the structure of specific graph configurations that had been open in the mathematical literature since the 1960s. Professional mathematicians had made partial progress on it over the decades but the full proof had not been found.

The person who produced the proof was not a professional mathematician. He had a background in computer science and described himself as an interested amateur. His approach was to use a combination of AI reasoning tools, primarily Claude and Wolfram Alpha, to explore different proof strategies, check intermediate steps, and identify connections between the problem and results in adjacent areas of combinatorics.

The proof was submitted to a mathematics journal and, after peer review, was accepted in late April. This was not AI producing a proof autonomously. It was a human using AI as an amplifier for mathematical reasoning in a way that gave him access to capabilities that would previously have required graduate training and years of immersion in the literature.

The professional mathematics community's response was thoughtful rather than defensive. Several senior mathematicians noted that the interesting question is not whether the proof was done with AI assistance, but whether the proof is correct and illuminating. It was both.

What this says about AI and expertise

The combinatorics proof is one of the clearest examples yet of AI working as a genuine intellectual partner rather than as a text generator or information retrieval system. The amateur mathematician used reasoning models to explore a large space of possible proof strategies that would have taken a professional mathematician years to work through manually.

This is precisely the use case that reasoning model proponents had been pointing to since DeepSeek-R1 in January 2025. Extended thinking, systematic exploration of logical consequences, and the ability to hold large amounts of mathematical context in working memory are exactly the capabilities where AI was supposed to augment human experts. The April 2026 proof was evidence that the augmentation was working at a level serious enough to contribute to open research problems.

The implication for how we think about expertise is significant. Expertise has traditionally meant accumulation of domain knowledge over years of study and practice. AI tools give determined amateurs access to a breadth of domain knowledge they could not otherwise have. The remaining variable is the quality of questions being asked and the judgment to evaluate whether an AI-generated reasoning step is actually valid, which still requires significant intellectual development.

OpenAI shuts down Sora

Sora launched publicly in December 2024 and was shut down eighteen months later, in April 2026. The announcement was brief: OpenAI stated that Sora's technology was being integrated into other products and that the standalone product was being discontinued. No specific successor product was announced at the time.

From a business perspective, the shutdown was unsurprising. Sora had strong technical reviews but a limited user base relative to ChatGPT. The cost of running video generation at scale is significantly higher per query than text generation. Competing products from Google (Veo 3) and several specialised video generation companies had caught up in quality while pricing more aggressively.

The pattern of rapid obsolescence

Sora's eighteen-month lifespan mirrors GPT-4o's eighteen months as a flagship product. This is now the expected lifecycle for a frontier AI product in a competitive category. Capabilities that were genuinely impressive at launch are matched or exceeded within a year, and the economics of maintaining a dedicated product become unfavourable relative to folding the technology into a broader platform.

For users this means that any product built on a specific AI capability has to plan for that capability either disappearing or being commoditised. For companies this means that competitive advantages from AI technology are measured in months. The durable advantages are in distribution, proprietary data, trust, and integration depth, not in model quality alone.