a16z Frontier: When AI can solve mathematical problems, what truly counts as your rare ability?
Author: Robert Hackett, a16z Frontier Editor, Co-founder of Fortune Crypto
Compiled by: Gu Yu, ChainCatcher
Mathematics is facing an existential crisis due to advancements in AI, but the a16z Frontier research team is more optimistic about the future of this field than ever before.
If AI solved a problem you spent years researching, would you want to celebrate?
Since reading mathematician Kirwin Hampshire's “The Dark Night of Mathematics” in July, I have been pondering this question repeatedly. Hampshire has a vision of the life he is training for. As AI models begin to find counterexamples to long-standing conjectures, he wonders if he will ever have the opportunity for his own discoveries. "I am going through a profound existential crisis," he wrote. "I have been screaming internally for days. It feels like I am living in a nightmare."
And this was before OpenAI released hundreds of mathematical manuscripts in October. Since then, members of the research community have been expressing shock, excitement, and confusion, contemplating what it means to conduct mathematical research in an AI-accelerated era and what that feels like.
One of the results from OpenAI provided a proof that solved a problem that Cornell University mathematician Justin Moore has repeatedly returned to and studied for 19 years. In 2010, Moore bought a bottle of XO cognac in the suburbs of Marseille, planning to share it with others who had worked on the problem when the answer arrived. He told colleague Steven Strogatz, that once he and his peers understood the proof and unanimously agreed it was correct, he still planned to open the bottle. But the answer arrived along with hundreds of other results, presented, in Strogatz's words, "like a severed rhinoceros head." Moore anticipated that the toast would be "more like a wake than a celebration."
Researchers at a16z Frontier know what it feels like to study a model that might solve a problem. They widely use AI in cryptography, theoretical computer science, economics, and mathematics, tackling stubborn old problems, testing ideas, and constructing proofs. We gathered them together to exchange notes------a recurring question: what happens when the speed of proofs arriving exceeds our ability to understand them?
Our senior research advisor and head of the Stanford Applied Cryptography Group, Dan Boneh, proposed a thought experiment: imagine a model proves that P does not equal NP, one of the Millennium Prize Problems, considered one of the most important questions in all of computer science. (If true, this means there exists a class of problems for which solutions can be efficiently verified, but no algorithm can efficiently solve every instance.) Now imagine this proof is a billion pages long. Could any one person understand it? Assuming that is even possible, how long would it take to absorb, interpret, and teach it?
Two days after our conversation aired, mathematician Levent Alpöge announced that he believed he had proven the Riemann Hypothesis with the help of AI. This conjecture about the distribution of prime numbers, proposed in 1859, is another Millennium Prize Problem, with hundreds of results in number theory assuming it to be true. In his announcement, Alpöge said, "It is natural to feel this is an endpoint. For me, it is a loss of a major guiding light, but what I have seen in the process is so astonishing that I believe this is the dawn of something far grander in all of science."
The verification problem—what it takes to check someone else's work—is a question we have spent a lot of time thinking about. Alpöge published his proposed proof using Lean. Lean is a programming language that allows computers to check every logical step. But even if you could check every step of Dan's billion-page proof, you still face the task of making it comprehensible—he said this task requires "an army of mathematicians and computer scientists" and would take a very long time. (At the very least, that's decent job security.)
a16z Frontier researcher and computer scientist Ittai Abraham (also the mind behind the Decentralized Thoughts blog) believes there are two frontiers in mathematics. The first marks the range of discoveries made. The second marks the level of human understanding of what has been discovered.
Abraham pointed to mathematician William Thurston's 1994 paper, which argued that mathematics is a social pursuit, and its progress cannot be measured solely by the absolute number of theorems. "What we are producing is human understanding," Thurston wrote. In his view, explaining ideas, questioning arguments, and helping others understand is at the core of the mathematical pursuit.
This is precisely the condition under which Moore would open that bottle of cognac: understanding the proof with his peers. A week after OpenAI's release, Sang-hyun Kim and Yash Lodha published a simplified exposition of the same proof. Their paper did not claim any new discoveries but clarified for other mathematicians why the result is true and how its reasoning can be applied elsewhere.
Our a16z Frontier research colleague, Scott Duke Kominers, a professor at Harvard Business School and an avid enthusiast of mathematics (and puzzles!), knows well the joy Moore would experience in saving that bottle of cognac for this moment. Kominers said that AI helped him solve a problem he has been returning to since around 2007. When he wrote to those mathematicians whose work he had followed for years but had never met, they responded with delight. According to Scott, no one arrived there with any tools and felt less excitement. Scott now envisions researchers using AI to build an entire mathematical system. (He also wrote about this trend earlier last year.) He believes that as tools become more powerful, our ambitions should grow accordingly.
Mathematician Frank Vallentin, in a recent paper, described the joy brought by this collaboration in an AI disclosure at the end. For years, he carried a worn copy related to a problem that has fascinated him since his student days. Imagine a repeating point pattern called a lattice, where a sphere is placed at each point with the same radius centered. Can these spheres cover the entire space, or are there gaps somewhere? Given the lattice and radius, Vallentin wondered how difficult it would be for a computer to answer this question as the number of dimensions increases.
On a Sunday in September, an AI model found a set of arguments for another problem about lattices. Vallentin wondered if the same set of arguments could also be applied to the problem he had long cherished. After some back-and-forth contemplation, he found that it could indeed be applied. The resulting proof precisely characterized how difficult the covering problem is. Vallentin simplified the proof, found it beautiful, and took responsibility for its correctness, even though he was uncertain how to allocate credit. However, he wrote that he was very happy to be the "messenger of the proof."
So, what does this mean for someone aspiring to be a mathematician? Will there still be things left for them to do?
Our researchers believe there is still a lot of work ahead: absorbing, refining, and expanding new results, explaining these results to others, and advancing human understanding. Boneh has a brilliant metaphor for this: he imagines AI as a giant tree, dropping juicy fruits for us to examine and enjoy.
For those hoping to make discoveries themselves, this may sound like cold comfort. But these tools also provide an opportunity to attempt work that was once out of reach—just as Kominers envisioned, building entirely new mathematical systems. Or as Alpöge wrote: "I hope this proof allows a glimpse of the wonders ahead and shows us the joy and amazement we are about to discover in the simplest of things."
People who are currently working and studying will also have the opportunity to decide which issues are worth investing all their efforts from the very beginning. Justin Thaler is a research partner at a16z Frontier and an associate professor in the Department of Computer Science at Georgetown University, where he teaches computer science. He points out that no matter how powerful machines become, people will still want to make their own judgments. To make good decisions, you must understand mathematics and its consequences, especially when building systems that others will trust. In cryptography, judgments about which problems will continue to be difficult to solve have already influenced the plans to protect the internet from quantum computer threats.
Mathematicians have spent a long time trying to hit difficult targets. In the future, there will be more targets that move faster and farther. Let us welcome more precise aiming.