OpenAI announced yesterday that its o3 reasoning model has disproved a central conjecture in discrete geometry that has stumped mathematicians since 1946. The claim would be easy to dismiss as more AI hype, except this time the academics who previously called out OpenAI’s mathematical overreach are vouching for it.

The conjecture in question comes from discrete geometry, a field that studies geometric problems with finite or discrete elements rather than continuous spaces. OpenAI hasn’t released the full technical details publicly yet, but they’ve shared their findings with mathematicians who specialize in the field.

This matters because it’s not OpenAI’s first rodeo with grand mathematical claims. Earlier this year, they announced a breakthrough that mathematicians quickly debunked. The company took heat for overselling capabilities, and the incident became a case study in AI vendors overpromising on reasoning tasks.

This time, they appear to have learned their lesson. The mathematicians who exposed the previous claim have reviewed o3’s work and confirm it holds up. That’s significant validation.

What o3 Actually Did

Reasoning models like o3 work differently from standard language models. They spend more compute time on each problem, effectively “thinking” through multiple approaches before settling on an answer. OpenAI’s o3 model, which isn’t yet in general availability, represents their latest push in this direction.

The discrete geometry conjecture that o3 disproved had resisted human mathematicians for eight decades. The model generated a counterexample that proves the conjecture false. In mathematics, that’s how you disprove something: you find a single case where the rule doesn’t hold.

Counterexamples can be simple or complex, but finding them often requires checking vast solution spaces that humans can’t feasibly explore. That’s where compute-heavy reasoning models have an edge. They can systematically test possibilities that would take human mathematicians years to work through by hand.

Why This Isn’t Just Model Hype

When OpenAI makes claims like this, skepticism is warranted. AI companies have incentives to oversell capabilities, especially in areas like mathematical reasoning where most people can’t verify the claims themselves.

But the independent confirmation changes things. The mathematicians involved aren’t OpenAI employees or partners with a stake in hyping the model. They’re the same people who previously pushed back on inflated claims.

If o3 can tackle problems like this, it suggests reasoning models are approaching a threshold where they’re useful for novel mathematical research, not just rehashing known solutions. That’s different from models that can solve calculus homework or pass standardized tests. Those are impressive feats of training data coverage and pattern matching, but they don’t push the boundary of human knowledge.

What’s Still Missing

OpenAI hasn’t released o3 to the public yet. Most of their I/O-style announcements this cycle have been “coming soon” rather than “available now.” That limits what independent researchers can verify or build on.

The technical paper with the full proof and counterexample also isn’t out yet. Until that’s published and survives peer review, there’s still room for surprises. Mathematical proofs need to be checked thoroughly. History is full of claimed proofs that fell apart under scrutiny.

And discrete geometry, while an important field, isn’t the same as tackling unsolved problems in number theory or topology that have million-dollar prizes attached. The Riemann Hypothesis isn’t falling to o3 tomorrow. But disproving an 80-year-old conjecture is still a legitimate achievement, assuming it holds up.

The Bigger Picture

If reasoning models can contribute to mathematical research, that opens doors beyond math itself. Physics, cryptography, computer science, and material science all rely on solving problems that involve exploring massive search spaces. A model that can systematically test possibilities faster than human researchers might accelerate work in those fields.

It also raises questions about how mathematicians will work in the future. If models can generate counterexamples or proofs, does that change what humans focus on? Do mathematicians shift toward posing better questions and interpreting results, while models do the grunt work of searching solution spaces?

Those are longer-term concerns. For now, the immediate takeaway is that OpenAI made a claim, backed it up with independent verification, and appears to have learned from past mistakes. That’s progress.

Other Tools News

Stability AI released Stability Audio 3.0, a small model that can generate two-minute audio tracks and run on-device. The company says it can create six-minute songs with the larger version. This follows the pattern of audio models getting longer context windows, though most generated music still sounds like background tracks rather than something you’d actually listen to on purpose.

Figma added an AI assistant to its design canvas. You can use natural language prompts to generate designs, edit existing ones, or automate iteration tasks. It’s the predictable next step for design tools, and it’ll probably save time on grunt work like creating variations. Whether it produces designs worth using is another question.

And GitHub confirmed a security breach affecting 3,800 repositories via a malicious VSCode extension. The extension was designed to exfiltrate code and credentials. This isn’t an AI story, but it’s a reminder that supply chain attacks on developer tools are real and getting more sophisticated. If you install extensions, check what permissions they’re requesting.