When AI Stopped Assisting Science and Started Doing It

Something happened that deserves more careful attention than it’s getting.

On October 6, OpenAI published 722 mathematical manuscripts to a public GitHub repository. The papers were produced by an internal frontier model that has not been released and has no name the public knows. The results span 17 fields of mathematics, organized into 372 distinct families. Some of the claims, if verified, would represent genuine progress on problems that have been open for decades.

The immediate reaction split into two camps, as it always does. One said this is extraordinary and transformative. The other said, slow down; none of this is peer-reviewed, and we don’t know if any of it is correct.

Both reactions are right, but they’re also both missing the more important question: are they correct and who’s verifying that answer?

Two Questions. Only One Has Been Answered.

Let me separate two things that the coverage has been packaging together, but deserve further analysis because they have completely different implications.

The first question: are these mathematical results correct? That is a question for mathematicians, and it will take months to answer. About a fifth of the 722 manuscripts have Lean proofs, meaning machine-checked formal verification. The remaining roughly 580 do not, and OpenAI itself acknowledges in the repository README that some unformalized results could have issues. The mathematical community will work through this carefully. Some results will hold. Some probably won’t. That process is essential and it hasn’t happened yet.

The second question is different: what does it mean that an AI model produced 722 mathematical manuscripts spanning 17 fields, using the computational equivalent of roughly three hours of thinking per result?

An AI system operated as a genuine mathematical discovery engine, which posed itself approximately 4,000 problems and produced novel mathematical work across number theory, geometry, theoretical computer science, mathematical physics and logic. The verification question matters for the study of mathematics, but the feat and capability matters for everything else.

The Model Nobody Can Use

The detail I find most striking isn’t the 722 papers, but the model that produced them, which remains nameless and without an assigned API. It’s not available to any customer, any researcher outside OpenAI, or any company trying to build on it. OpenAI published the outputs of a capability that the market has not yet seen and cannot yet access.

In investment terms, that is a very specific signal. The company has demonstrated a capability at the frontier that is, by definition, ahead of what the market has priced. The distance between what OpenAI’s internal models can do and what the market understands those models can do widened with this accomplishment.

However, that gap doesn’t stay wide forever. It narrows when the model releases, when competitors catch up, or when the results inform new benchmarks that everyone starts optimizing against. But right now, in October 2026, there is a documented public demonstration of a mathematical capability in an AI system that nobody outside the company can use or fully evaluate.

What This Changes About the Investment Thesis

In Investing in Revolutions, I describe the Convergence Effect as the principle that when transformational technologies combine, the result is not additive but multiplicative. Two plus two doesn’t equal four. It creates something that neither technology could produce alone.

I’ve been writing about AI plus scientific discovery as a forward-looking investment theme for two years. Drug discovery timelines collapsing from years to weeks with hybrid AI and quantum computing. Materials science advancing through AI-generated molecular simulations. Cryptography being reshaped by AI-assisted analysis of mathematical foundations.

Those were predictions grounded in early signals. This week, one of them came to fruition.

An AI model doing genuine mathematical discovery at the level of 722 manuscripts across 17 fields is a categorically different relationship between AI and scientific knowledge than the one we’ve been operating under. The fields where this matters most are exactly the ones I’ve been watching. Number theory is the mathematical foundation of cryptography, which is the foundation of every blockchain system and secure communication protocol. Mathematical physics underlies materials science and quantum computing hardware. Theoretical computer science is the backbone of algorithm development. Progress in any of these fields propagates downstream into products, companies and entire industries.

When AI can move those fields faster than human researchers working alone, the downstream propagation accelerates. The question for investors is what is going to accelerate, what are the accelerants, how fast, and who is positioned in the path of that acceleration.

The Verification Economy That Doesn’t Exist Yet

There’s a specific investment implication in this story that I haven’t seen anyone name clearly.

162 of the 722 manuscripts have machine-checked Lean proofs. The rest don’t. OpenAI relied on an advisory group at the Institute for Advanced Study to provide initial oversight. The group described the release as a first step that begins rather than completes the work of incorporating results into mathematical knowledge.

What that means practically: we now have a world where AI systems can produce mathematical and scientific claims faster than existing human infrastructure can verify them. That gap between production and verification is not a temporary problem. It scales with the capability. Every improvement in AI’s ability to produce scientific results widens the verification gap unless something is built to close it.

That something is a new category of infrastructure: Formal verification tools, AI-assisted proof checking, automated pipelines for testing mathematical claims against known results and internally consistent frameworks. The Lean proof system is a beginning, but it covers 22% of this release and wasn’t designed for the volume that’s coming.

The verification economy for AI-generated science is going to be significant and barely exists today. The companies that build it will be doing something that is genuinely necessary rather than merely useful, which is the most durable kind of infrastructure to build.

The Line That Has Been Moving for 80 Years

When computers first did numerical calculations faster than humans, most mathematicians dismissed it as computation rather than mathematics. Machines could calculate. They couldn’t discover. That was the line.

The line has been moving ever since. Computer-assisted proofs of major results like the four color theorem generated controversy in the 1980s and 1990s because the proof couldn’t be fully checked by a human. Automated theorem provers advanced through the 2000s and 2010s. Each time the line moved, the same argument appeared: this isn’t real mathematics; it’s just computation.

The line moved again, more visibly than it has in years and in a way that is harder to dismiss because the output is sitting in a public GitHub repository that anyone can examine.

I’m not claiming all 722 results will survive scrutiny. I’m claiming that the capability demonstrated in producing them represents a genuine and durable shift in what AI can do in scientific domains.

The Inflection Point for Investors

In Investing in Revolutions, I argue that the greatest investment opportunities emerge at inflection points, when transformational technologies cross a threshold that fundamentally changes what becomes possible. This is one of those moments.

AI’s emerging ability to generate original scientific discoveries signals a shift in how innovation happens and where value will be created. The implications extend beyond AI developers to drug discovery, advanced materials, energy, cryptography and industrial R&D, where accelerated research could compress development timelines, lower costs and unlock entirely new markets.

For investors, the opportunity is to recognize this inflection point before its commercial implications become obvious. That means identifying the infrastructure required to validate AI-generated discoveries and the companies best positioned to turn those discoveries into proprietary technologies, scalable businesses and lasting competitive advantage.

The individual findings still require verification, but the broader signal is clear: AI is crossing from assisting scientific discovery to potentially driving it. The investors who understand the significance of that shift, and where its effects will converge, will be best positioned to capture the value it creates.