OpenAI published 719 mathematical manuscripts on October 6, 2026, from an internal model aimed at about 4,000 problems. A typical result took around three hours of compute. I recognized the argument I have been making at work for a year. This is progress. It showed up in the field people name when they want progress to sound impossible.

The example everyone was saving

When someone tells me knowledge work will be fine, they do not start with status reports. They start with a research claim, or a legal theory that turns on a distinction most readers will miss, or a design that has to be right because the system around it is expensive to undo. If they want the example to end the conversation, they pick mathematics. Few people can read the page, a wrong step ruins whatever gets built on it, and the best practitioners can spend years on one question and come back with nothing to show.

That example was the shelter. Software and drafting could speed up, the obvious tasks could leave the building, and real thinking would stay on a human pace, because real thinking looked like a proof: scarce people, long attention, a standard that does not grade on a curve. October 6 is that shelter producing a catalogue. One lab, one model the rest of us cannot yet run, and a day’s output that a strong department would not recognize as a day’s work.

Some of those manuscripts will be wrong. New work at this volume shows up before anyone has finished checking it.

It is the same task

Look at the work itself. Someone takes an open problem, holds a huge amount of prior work in mind, and comes back with a structured claim that other experts can attack. That is a proof. It is also the research memo, the regulatory analysis, the clinical review, the architecture decision, and the strategy document a company will spend a quarter carrying out. The documents have different names. The task is the same: read the record, form a claim, show the reasoning, and survive a check.

We sorted those tasks by how smart the person had to be, and we treated the top of the sort as a different economy. Faster tools for junior work, supervision for senior people, and a calendar kept normal by the scarcity of whoever was smartest. Mathematics was the cleanest version of that story, because there the smartest people are the profession. There is no layer of business context under the thinking. Being difficult is the whole job. If a model can spend an afternoon on a problem of that kind, and can do it thousands of times, then a hard problem is no longer a promise that the work will take a year.

I watch the smaller copy inside the organizations I run. The work we still hand to the most senior people is mostly reading, judging, and writing something down so everyone else can move: a design review, the narrative after an incident, the memo that says whether a regulated workflow is allowed to change. We describe that layer as safe because the people in it are good. They are good. Good no longer describes the pace. The pace was how long a serious attempt took, and a serious attempt just got cheap in a field stricter than any review I sit in.

Every industry after this one

I welcome the progress because it did not show up as a demo of an easy task. It showed up as mathematics. The industries still describing their best people as the bottleneck are describing a bottleneck that already failed somewhere harder.

There is always another group that claims to be the exception. Scientists, counsel, principal engineers, the specialists who “actually understand the domain.” Each group watches the previous one speed up and decides that this one is made of judgment. I have sat in those conversations this year, on both sides of the table, and the structure never changes. The work is particular, the context is heavy, and a mistake is costly. All of that was true of the problems in the catalogue, and none of it kept the attempts on a human schedule.

What I want from the teams I am responsible for is to retire the search for the group that stays slow. Knowledge work is going to move at the speed of an attempt, including the work we reserve for the smartest people we can hire. That covers the labs, the law firms, the banks, and the product organizations that still talk about their senior people as a natural limit on how fast the work can go. The attempts are coming either way. Our job is to decide which claims we are willing to stand behind, and to be glad the attempts exist. I have wanted that acceleration in software for a year. Seeing it arrive in mathematics makes me want it everywhere else.