OpenAI’s GPT-5 Pro demonstrated groundbreaking creativity by solving an open problem in convex optimization, producing a novel and mathematically verified proof that extended known bounds beyond what human researchers had achieved. This milestone highlights AI’s emerging ability to contribute original scientific insights rapidly, fostering a collaborative future where AI and humans accelerate research breakthroughs together.
OpenAI recently released GPT-5 Pro, which has demonstrated a groundbreaking capability in mathematical research. A researcher named Sebastian Bubbeck tested GPT-5 Pro by feeding it a new mathematical paper on convex optimization containing an open problem—an unsolved challenge in the field. Contrary to expectations that the AI would merely regurgitate existing knowledge, GPT-5 Pro not only solved the problem but improved the known mathematical bounds in an original way. It took just 17 and a half minutes to produce a novel proof that extended the boundary from 1/L to 1.5L, a result that human researchers had not achieved. This proof was manually verified and found to be mathematically sound.
The significance of this discovery lies in the nature of the problem itself. Convex optimization is a fundamental concept in machine learning and AI, involving finding the best solution by minimizing a convex function, often visualized as rolling a ball to the bottom of a bowl. The research focused on the step sizes used in gradient descent, a core algorithm in AI training, and whether the optimization curve remains convex for larger step sizes. While the original paper guaranteed convexity only up to 1/L, GPT-5 Pro proved it could be extended to 1.5L using advanced mathematical tools like Bregman divergence inequalities and co-coercivity, demonstrating a deep understanding of optimization theory.
This breakthrough has profound implications for AI’s reasoning capabilities. Unlike previous models that primarily performed pattern matching and recombination of existing knowledge, GPT-5 Pro exhibited genuine creativity by devising a new mathematical proof not present in its training data. OpenAI’s president Greg Brockman described this as a “sign of life,” suggesting emergent behaviors that hint at true understanding rather than mimicry. The speed of the AI’s solution—17 and a half minutes compared to the 25 minutes human researchers took to verify it—highlights the potential for AI to accelerate scientific discovery dramatically.
The interaction between AI and human researchers following this discovery also points to a collaborative future. Inspired by GPT-5 Pro’s proof, the original authors revisited their work and extended the bound further to 1.75L, showing that AI and humans can push each other to new heights rather than compete. This event has sparked debate in the AI community about whether this marks a step toward artificial general intelligence (AGI) or is simply an impressive but narrow achievement. Regardless, it represents a milestone in AI’s ability to contribute original insights to complex scientific problems.
Looking ahead, this development suggests a future where AI systems could routinely conduct original research, advancing human knowledge at unprecedented speeds. While some mathematicians express concern about the role of human researchers in such a future, the current example shows a complementary relationship where AI inspires and enhances human work. As AI models become more specialized for scientific discovery, we may soon see AI-generated proofs and research becoming commonplace in academic journals, potentially revolutionizing the pace and nature of scientific progress worldwide.