OpenAI has announced new results on ten long-standing open problems in mathematics and theoretical computer science. The advances, detailed in a company blog post, span fields from geometry and cryptography to algorithm design. This development matters less for the specific solutions found and more for the method used to find them. It demonstrates a powerful new workflow for scientific research where human experts guide large language models (LLMs) to perform complex computational searches, dramatically accelerating a process that could otherwise take years. This is not a story of an AI independently "solving" math, but one of AI acting as a formidable tool for human-led discovery.
The New Workflow: From Prompt to Progress
The core of OpenAI's achievement is a refined methodology for human-AI collaboration. The process sidesteps the current limitations of LLMs in abstract reasoning by leveraging their primary strength: code generation. Researchers did not ask a model like GPT-4 to produce a formal proof from scratch. Instead, they described a mathematical problem and a high-level strategy for finding a solution or counterexample in natural language.
The model, a fine-tuned version of a GPT-4-class system, would then translate this prompt into a working Python program. This program was designed to search for the desired mathematical object—be it a denser lattice packing, a more efficient algorithm, or a function that disproves a conjecture. The researcher would execute the generated code, analyze its output, and then refine the prompt to steer the AI in a more promising direction.
This iterative loop combines the best of both worlds. The human provides the domain expertise, intuition, and strategic direction, identifying which avenues are worth exploring. The AI handles the complex, often tedious task of translating that strategy into error-free, efficient code capable of exploring vast search spaces. This approach effectively lowers the barrier for researchers who may be experts in their theoretical field but not in high-performance scientific computing. It transforms the process from manual coding and debugging to a dialogue with a capable programming assistant.
Key Advances: From Geometry to Cryptography
The ten results presented by OpenAI are not minor tweaks; they represent genuine progress on problems that have occupied researchers for years. The problems can be broadly grouped into finding new constructions, optimizing existing algorithms, and finding definitive counterexamples.
Finding Better Constructions
A significant portion of the advances involves finding new, improved mathematical "objects" or constructions. In the field of Ramsey Theory, which studies the emergence of order in large random structures, the model found a new construction related to Schur's Theorem. It produced a way to color integers up to 160 with five colors such that no three integers x, y, and z satisfying x+y=z are the same color. This improves the lower bound for the Schur number S(6), a long-standing challenge.
Similarly, in geometry, the team tackled the Cap Set Problem. This problem asks for the largest possible set of points in a high-dimensional space where no three points lie on a single line. The AI-generated program found a new, larger cap set in dimension 9, consisting of 512 points. While a small improvement, progress on this problem is famously difficult. The model also found a denser packing for the Leech lattice in 11 dimensions, a result relevant to information theory and physics.
Optimizing Complex Algorithms
Perhaps the most widely applicable results came in algorithm design. OpenAI's model discovered a new algorithm for multiplying 4x4 matrices using only 47 multiplications. This is a new record, improving upon the 49 multiplications required by the foundational Strassen algorithm adapted for this size. Since matrix multiplication is a fundamental operation in nearly every field of scientific computing, from graphics and machine learning to physics simulations, even small efficiency gains can have a massive impact at scale. This result is particularly notable as it surpasses work done by specialized systems like DeepMind's AlphaTensor, suggesting the generality of the LLM-based approach.
The model also improved the efficiency of a sorting network for 13 elements. A sorting network is a fixed sequence of comparisons used to sort a list of items, and finding the most efficient network for a given number of inputs is a classic computer science problem.
Disproving a Decades-Old Conjecture
One of the most definitive results was the disproof of a 30-year-old conjecture in complexity theory. The Beigel-Reingold-Spielman (BRS) conjecture from 1994 dealt with the properties of threshold functions, which are simple models of computation. For three decades, the conjecture was assumed to be true, but a formal proof remained elusive.
Using the collaborative workflow, a researcher prompted the model to search for a function that would violate the conjecture's predictions. The AI generated a program that successfully found such a function, providing a concrete counterexample that disproved the BRS conjecture. This is a powerful demonstration of the method's utility, as finding a single counterexample is often more difficult than formulating a general proof, requiring an exhaustive and cleverly guided search.
Community Reaction: A Powerful Tool, Not a Replacement
The announcement has been met with both excitement and careful analysis. In discussions on platforms like Hacker News, developers and researchers largely agreed that the results are significant. However, the consensus is that this marks the maturation of AI as a research assistant, not as an autonomous scientist.
Commenters repeatedly emphasized the "human-in-the-loop" nature of the work. The successes were not achieved by simply feeding a problem statement into a black box. They were the product of a dialogue between an expert who knew what questions to ask and a tool that could rapidly explore possible answers. The AI's role was to automate the translation of human intuition into executable search programs.
A key distinction was drawn between problems of "search" and problems requiring abstract "proof." The ten advances are all fundamentally search problems—finding a specific object, a better configuration, or a counterexample. The LLM proved exceptionally capable at this. However, the work does not demonstrate an ability to generate novel, abstract mathematical proofs, which require a different, more profound form of reasoning. This aligns with the understanding that current LLMs are powerful pattern-matching and generation engines, not yet abstract thinkers. Nonetheless, the ability to accelerate the search component of research is a major contribution, freeing up human researchers to focus on higher-level strategy and insight.
What to Watch Next
This work from OpenAI provides a clear validation of a specific methodology for AI-assisted research. The immediate future will likely see this workflow adopted more widely by mathematicians and computer scientists tackling similar combinatorial search problems. The success here serves as a blueprint for how to productively collaborate with current-generation AI models.
The key question is how far this approach can be pushed. Its applicability to problems requiring less search and more abstract, multi-step reasoning remains an open question. Future progress will depend on improving the models' core reasoning capabilities and their ability to understand more nuanced and complex scientific prompts.
Beyond pure mathematics, the next frontier will be the application of this code-generation workflow to other scientific domains. Fields like materials science, drug discovery, and genomics are rich with problems that involve searching for optimal structures or configurations. The techniques demonstrated here could prove to be a powerful accelerator across the entire scientific landscape, changing not what questions we ask, but the speed and scale at which we can find their answers.