Cloud AI

ChatGPT-5.4 Pro Breaks 60-Year Mathematical Barrier: How AI Discovered What Human Mathematicians Couldn’t

On an ordinary Monday in April 2026, 23-year-old Liam Price entered a single prompt into ChatGPT-5.4 Pro and solved a mathematical problem that had eluded the world’s brightest minds for 60 years. What happened next would force mathematics to take AI seriously in a way no previous breakthrough had.

Price, an amateur with no advanced mathematical training, guided ChatGPT to prove Erdős’s primitive set conjecture – a problem involving special sets of integers where no number divides another. Fields Medalist Terence Tao confirmed the result: “All previous researchers went astray from the very beginning and fell into a fixed thinking trap.”

This wasn’t just another AI success story. It represented something fundamentally different: a new methodology, a cognitive leap that humans hadn’t considered, and what may be the first truly original mathematical reasoning from artificial intelligence.

The Erdős Legacy: Mathematical Problems That Define Our Era

Paul Erdős, the legendary “mathematician of the century,” left behind a collection of over 1,500 unsolved problems that have shaped mathematical research for decades. These aren’t mere puzzles; they represent fundamental questions about the nature of numbers, sets, and mathematical truth itself.

Erdős Problem #1196 – the one Price and ChatGPT solved – concerns “primitive sets”: collections of integers where no number divides another. This generalizes the concept of prime numbers (which are primitive sets of size 1) to arbitrary collections. Erdős proposed that for any primitive set, the sum of reciprocals of elements grows no faster than 1 + o(1) as numbers become large.

The significance extends beyond number theory. Primitive sets appear in cryptography, complexity theory, and even computational biology. Understanding their fundamental properties could impact multiple fields simultaneously.

What Erdős Problem #1196 Actually States

The conjecture deceptively simple: for any primitive set A ⊂ [x,∞) (where no distinct elements divide each other), does the sum Σ(1/(a log a)) for a∈A converge to 1 + o(1) as x→∞?

Previous mathematicians had made progress. Stanford’s Jared Duker Lichtman proved an upper bound of approximately 1.399 + o(1) in 2023, but the exact 1 + o(1) limit remained out of reach. The problem had resisted attacks from multiple directions for six decades.

The breakthrough came not from deeper specialization, but from a different perspective entirely – one that no human researcher had considered.

How ChatGPT-5.4 Pro Solved What Humans Couldn’t

The AI’s approach was fundamentally different from human mathematical reasoning. Where humans tend to follow established patterns and search for incremental progress, ChatGPT-5.4 Pro made a conceptual leap that bypassed decades of conventional thinking.

Key technical elements of the breakthrough:

• Novel Formula Application: The AI applied a known formula from related mathematical fields in a way no human had considered for this specific problem type
• Unconstrained Reasoning: Unlike human researchers who work within established frameworks, the AI explored mathematical territories that seemed “illogical” to trained mathematicians
• Rapid Pattern Recognition: The model identified connections between primitive sets and other areas of mathematics that human researchers had treated as separate disciplines

“There was kind of a standard sequence of moves that everyone who worked on the problem previously started by doing,” Tao explained. “The LLM took an entirely different route, using a formula that was well known in related parts of math, but which no one had thought to apply to this type of question.”

The “Vibe Mathing” Revolution: Democratizing Mathematical Discovery

Price’s approach represents a paradigm shift in how mathematical research might work in the AI era. He describes his method as “vibe mathing” – an exploratory approach that relies more on intuitive questioning and repeated trials than deep theoretical preparation.

This raises profound questions about the future of mathematical research:

1. Accessibility: Can AI systems lower the barrier to entry for mathematical discovery, allowing non-specialists to contribute to advanced research?
2. Collaborative Intelligence: What does human-AI mathematical collaboration look like when the AI contributes genuinely original insights?
3. Knowledge Representation: How do we conceptualize mathematical knowledge when AI systems can make connections that humans can’t explain or understand?

The implications extend beyond pure mathematics. If AI can help solve problems that require thinking outside established frameworks, similar approaches might work in theoretical physics, computer science, engineering, and other fields where human thinking tends to become entrenched.

Technical Validation: From Raw AI Output to Rigorous Proof

The AI’s raw output wasn’t immediately publishable. As Lichtman noted, “The raw output of ChatGPT’s proof was actually quite poor. So it required an expert to kind of sift through and actually understand what it was trying to say.”

The validation process involved multiple steps:

• Initial Verification: Price and collaborator Kevin Barreto (University of Cambridge undergraduate) confirmed the logical structure
• Expert Analysis: Tao and Lichtman refined the proof, extracting the key insights
• Formalization: The result was formally proven in Lean theorem prover, ensuring mathematical rigor
• Peer Review: Publication on Erdosproblems.com with community validation

This process reveals an important truth: AI can generate novel insights, but human expertise remains crucial for validation, refinement, and ensuring mathematical rigor.

Comparison of Human vs AI Mathematical Approaches

Broader Implications for AI and Scientific Discovery

This breakthrough forces us to reconsider several fundamental assumptions about AI and scientific research:

Implementation Playbook for AI-Assisted Mathematical Research

1. Problem Selection: Choose problems where established approaches have failed but cross-domain connections might work
2. Prompt Engineering: Use minimal, open-ended prompts that encourage exploration rather than specific directions
3. Human-AI Collaboration: Have both mathematical expertise and open-minded exploration
4. Validation Framework: Implement rigorous validation processes that maintain mathematical standards
5. Pattern Recognition: Look for novel approaches that can be generalized to other problems

“We have discovered a new way to think about large numbers and their anatomy,” Tao stated. “It’s a nice achievement. I think the jury is still out on the long-term significance.”

The Future of Mathematical Research in the AI Era

What does this breakthrough mean for the future of mathematics and computational science? Several key questions emerge:

First, it suggests that AI might be particularly valuable for problems where human thinking has become entrenched. When mathematicians work on a problem for decades, they develop deep expertise but also deep biases about what approaches might work. AI systems, lacking this historical baggage, can explore radically different approaches.

Second, the “vibe mathing” approach suggests a new model for mathematical education and research. Rather than focusing solely on deep theoretical knowledge, future mathematicians might need skills in human-AI collaboration, prompt engineering, and recognizing valuable AI insights.

Finally, there’s the question of whether this approach can be generalized. If AI can solve Erdős problems in novel ways, can similar methods work in other areas of mathematics? What about theoretical physics, computer science, or engineering problems where human approaches have reached diminishing returns?

Practical Applications Beyond Pure Mathematics

The implications of this breakthrough extend far beyond number theory. Primitive sets and similar mathematical structures appear in:

• Cryptography: Understanding the properties of sets with divisibility constraints
• Computer Science: Algorithms for working with mathematical structures
• Theoretical Physics: Mathematical foundations of quantum field theory
• Data Science: Understanding mathematical properties of large datasets

The novel approach discovered by ChatGPT-5.4 Pro – using known mathematical formulas in unconventional ways – might have applications across these fields. When faced with problems where conventional approaches have failed, AI-guided exploration might provide new solutions.

FAQ: Understanding the Breakthrough

What exactly is a primitive set?

A primitive set is a collection of integers where no number divides another. For example, {2, 3, 5, 7} is primitive because no number in the set divides another. The set {2, 4, 8} is not primitive because 2 divides 4 and 8 divides 4. Prime numbers are naturally primitive since they have no divisors other than 1 and themselves.

Why was this problem important for 60 years?

Erdős problems are fundamental questions about the nature of mathematical structures. Understanding primitive sets helps us understand the fundamental properties of numbers and their relationships. This specific problem was particularly challenging because it required proving a precise limit on the growth rate of sums over primitive sets – something that resisted multiple approaches from prominent mathematicians.

How did the AI come up with a new approach?

Unlike human mathematicians who tend to build on existing approaches and work within established frameworks, ChatGPT-5.4 Pro was able to make connections across different areas of mathematics. The AI applied a known formula from related mathematical fields in a way that no human researcher had considered for this specific problem. This represents what may be the first genuinely original mathematical reasoning from AI.

Does this mean AI will replace mathematicians?

No, but it does suggest a fundamentally different way of doing mathematics. The AI’s raw output required human expertise to understand, validate, and refine. Future mathematical research will likely involve human-AI collaboration, where AI systems generate novel approaches and human mathematicians provide the expertise needed for validation and application. The key shift is from “human vs AI” to “human with AI.”

Can this approach work on other mathematical problems?

Potentially. The breakthrough suggests that AI-guided exploration might be particularly valuable for problems where human thinking has become entrenched. When mathematicians work on a problem for decades, they develop deep expertise but also deep biases about what approaches might work. AI systems, lacking this historical baggage, can explore radically different approaches. However, this approach is likely most valuable for problems that involve connections across different mathematical domains.

What are the limitations of this approach?

The approach has several important limitations. First, the AI’s output requires significant human expertise to understand and validate. Second, the method works best when there are existing mathematical frameworks that can be applied in novel ways, rather than entirely new mathematical territory. Third, while the approach generated a novel solution, the long-term significance and impact remain to be seen. Finally, the approach requires human guidance to ensure the AI explores promising directions rather than mathematical dead ends.

Sources and References

1. Erdős Problem #1196 – Official Statement. Erdős Problems. https://www.erdosproblems.com/1196, accessed 2026-05-01
2. Howlett, Joseph. “Amateur armed with ChatGPT ‘vibe maths’ a 60-year-old problem.” Scientific American, April 24, 2026. https://www.scientificamerican.com/article/amateur-armed-with-chatgpt-vibe-maths-a-60-year-old-problem/
3. Anonymous. “23-Year-Old Amateur Solves 60-Year-Old Math Problem with ChatGPT, Terence Tao Says Previous Researchers Went Astray from the Start.” 36Kr, April 27, 2026. https://eu.36kr.com/en/p/3784815604817154
4. Lichtman, J. D. “On the distribution of primitive sets.” Doctoral thesis, Stanford University, 2022.
5. Tao, Terence. Comments on Erdős Problem #1196. MathOverflow and Erdős Problems Forum, April 2026.