OpenAI has put hundreds of AI-produced maths manuscripts online for people to inspect.

On October 6, it announced results from an unreleased internal model. The current collection lists 719 manuscripts grouped into 372 related families. Those numbers describe a collection of papers and supporting arguments; they do not mean 719 separate, confirmed breakthroughs. OpenAI's repository

Still, this is worth paying attention to. AI producing research at this scale raises a question that matters well beyond maths: how do we keep up with checking its work?

I am Alex. Welcome to ShortCu8 by Innov8.

Let's dive in.

When someone says AI solved a problem, ask what was checked.

The useful detail is the evidence behind the answer. Look for a proof, an independent review, or a test you can inspect.

The model was given roughly 4,000 problems during its evaluation. Some results build on earlier model outputs, and related manuscripts are grouped together. The collection is at different stages of verification. Repository

OpenAI says an average result used compute equivalent to about three hours of ChatGPT Pro thinking. It also released ten summaries of the model's reasoning. The model itself remains unreleased. Release announcement

So I would be careful with “AI solved 500 questions.” There is a substantial research collection here, but a headline number cannot tell us how much of it will survive scrutiny.

On October 7, OpenAI's correction log recorded a sign error in one manuscript. It invalidated an argument and a construction used by two other papers. All three were withdrawn. Correction log

You do not need to understand the maths to understand the consequence. Later work relied on an earlier result. When that foundation failed, the dependent work needed to come down too.

That is a useful reason to publish the workings. People can inspect an argument, identify the gap, and see the correction. The withdrawals matter. So does making them visible.

The October 7 update says 300 of the 719 top-line results, about 42%, have been formalized for computer checking. Correction log

The computer checks the precise claim written in its language. A person still needs to check that this matches what the paper claims to prove, including its assumptions. Lean's explanation

That does not make the remaining results automatically wrong. OpenAI explicitly warns that some results without formalizations could have issues. Repository

It does mean we should keep the verification status attached to the result. “Produced by AI” tells us where an answer came from. It leaves the question of correctness open.

We can inspect the manuscripts, proof artifacts and revision history ourselves. That makes this release more useful than a claim with nothing behind it. Public collection

My view: the exciting possibility is that AI gives researchers more worthwhile arguments to investigate. The unresolved part is how much checking those arguments require.

For your own work, borrow one habit: before relying on an important AI answer, open its original source and check the assumption your decision depends on. Another confident answer from the same chatbot is not an independent check.

I would rather read an AI research release that shows its mistakes than one that only shows its successes.

OpenAI has given us a large collection to investigate. The number worth watching next is how much of it holds up.

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Found the right person, but they’re in a country where you don’t have an entity? Setting one up can take months and significant cost.

Remote removes that barrier by becoming the legal employer through our own entities — handling compliant contracts, local benefits, tax setup, and onboarding for you. In fact, an employee is onboarded to Remote every 7 minutes.

Once they’re hired, the same in-house teams that support employment locally also run payroll — so you’re not bouncing between disconnected providers. Less setup, less complexity, and less time between finding the right person and getting them started.

See how it works

#AD2

Parallel, Turbopuffer, and Wordsmith run their entire GTM motion on Attio, with agents that chase every buying signal, build pipeline, and move deals forward, 24/7. Try Attio now.