The Day Problems Became Property

Mathematicians do not need to read this article. But if you run agents at a company, the most important news this morning actually came from mathematics.

The status of unsolved problems has changed. The first step for a question that was public to become a private asset was taken this week.

There are three pieces of evidence. OpenAI released 700 papers at once. Researchers, working with an AI agent, obtained a 69.13% improvement in OpenAI’s integer multiplication result in a single day. And the mathematician Hugo Duminil-Copin warned that mathematicians may stop sharing research problems publicly before AI solves them first.

Tied together in a single line, this morning’s news is the day the price of problems changed. The value of what is unsolved was placed on the same footing as the value of solving it.

The article’s core concept, visualized.

A Day in Integer Multiplication

Integer multiplication is the most basic of computations. It is the operation we perform at every moment, and lowering the number that measures this operation’s efficiency has been a task left open in mathematics for a long time.

Today’s report is that this number moved by AI’s hand. From OpenAI’s integer multiplication result, the conditional exponent saving value rose to about 0.001289283116. The improvement was about 69.13%, and the period was one day.

The way of an agent working through the night and a person choosing in the morning now returns as day-unit results. What changed is not only the speed of discovery. The hand that handles the problem changed.

In mathematics, the standard is not the size of an improvement but the time it takes. One paper, one conference, one academic year. The seasons of mathematics have usually flowed like this. For the same number to be written again in a single day means the unit of research has changed.

The hand that filled that day was the agent. What was reported is that researchers worked with an AI agent. The structure is one where a person sets a hypothesis, the agent runs variations, and the person checks the result in the morning. The improved exponent saving value is a product of that loop. Only someone who knows the speed of mathematics can say exactly how large an achievement a 69.13% margin is. What matters is not the margin but the cycle.

When the cycle becomes a day, the review cycle becomes a day too. The result that came out overnight is checked in the morning, and that check starts the next loop. When discovery and review turn at the same speed, the unit of mathematics itself shrinks to a week.

There is another reason this number draws attention. Exponent saving is a result that gets updated in the public record. Today’s 69.13% improvement, and the next improvement stacked on top of that record, both come out on shared numbers. This is why the door opens in the next section. Mathematics that works on public records is, at the same time, mathematics that was trying to protect those records.

The Door Mathematics Closes

It began with 700 papers. Right after OpenAI released 700 papers, Duminil-Copin said that mathematicians may stop publishing research problems so as not to have them solved by AI.

700 is a scale set down all at once. A single paper appearing and 700 appearing at once are different things. It is not that one individual problem was newly created, but that the field’s agenda was rewritten in one stroke. Duminil-Copin’s remark came as a response to that very scale.

Choosing which door to close is a decision each researcher makes when posting a problem. The cost of such decisions is borne by the field. When one researcher stops posting, another researcher cannot tell whether the problem was solved or simply never posted. It is a structure where silence itself becomes the problem.

In mathematics, problems have played the role of currency. Circulating between the public forum, conferences, and papers, the number of people working on them together grew. Trust built through openness led to the next collaboration. A published problem also carries free surveillance. Another person points out a wrong part, and another person finds a shorter path. Mathematical verification has operated on top of this surveillance.

Now that currency has become a leak. The moment it is published, the side that can solve it sees it first. Duminil-Copin’s warning is not a declaration that the entire mathematics community has already closed the door. But it is a clear signal of direction: a door that was open is tilting toward the side that can close it.

The paradox is here. The actor trying to close the door is looking at the ability of a model trained on the literature that kept that door open. It is a composition in which intelligence grown from public results ends publishing itself.

If results no longer flow to the public, the fresh material the next generation of models can learn from shrinks. The paradox goes one step further. The moment mathematics retreats into private work, the field bids farewell to the way it has kept itself public.

Let us hold onto the remark again. It is the phrase that they may stop. It is not the word that they already stopped. It is a warning that things can tilt toward closing. What the warning points to is not the door but the slope. If the slope continues, somewhere a quiet room will take the place where problems used to be shared.

The Community That Opens the Door

The community’s answer came from the opposite side. Hugging Face turned OpenAI’s math problem dataset into open-source RL environments. By attaching the conditions of reinforcement learning, it becomes possible for anyone to train a model on those problems.

The form of an RL environment matters. A problem is originally text. Once it becomes an open-source RL environment, the problem becomes an executable task. The answer can be verified, and that verification serves as the signal for learning. Storing a problem and making a problem run are different things. What Hugging Face built is the latter. It is the side that makes problems run.

And the cost of running that playground moved to the community’s share. As long as it is open-source, the cost of running the training environment sits with the person who wants to run it, not the person who made the problem. Problems have become terrain that asks again whose problem they are.

The value of a playground shows in repetition. When many people run the same problem, many trajectories are left behind. Those trajectories become material for the next generation of models. Publishing is an ethical stance, and it is also a question of supply.

The side that wanted to protect problems and the side that expanded problems into a playground moved in the same week. At one end a door closes, and at the other a door opens. Problems now exist in two forms.

The interesting point is that both are now handling the same problems. The question of which side wins comes after that. One problem has taken two forms: a private asset and a public training ground.

The two paths look like this.

flowchart TB
    P["One public problem<br/>integer multiplication task"] --> A["closed door path<br/>private asset: one company's model material"]
    P --> B["open door path<br/>public training ground: open-source RL environments and community trails"]

The meaning of this split is that the problem itself has become an asset. An asset comes with a price. A problem that stays private becomes material that grows one company’s model. A problem that is published becomes the training ground of every model. A structure has emerged in which the same equation carries two values.

This split has already seeped into how the next generation of models is built. Sakana AI’s new framework helps go beyond the limits of human evaluation by deploying virtual subagents when running open-ended research tasks. The 27B model’s performance went up 1.6x. In the seat that had been evaluating on behalf of humans, a model made by a model was placed. Open-ended tasks are a domain where scoring an answer is hard, so human evaluation was the bottleneck. In a research loop where agents help each other, the seat of the bottleneck moves. The seat that was a human hand shifts toward the execution environment.

A Company’s Problems Are Now Assets

The same compression is happening inside companies. Moving that agent loop, which produced a day-level improvement, to internal tasks such as optimization and data analysis is enough. At that moment, the problems a company has not solved occupy the same grade as the results being solved.

Company problems already have that shape. An optimization problem defined by internal data, an operational bottleneck not yet resolved, an algorithmic task you want to solve before a competitor solves it first. Such problems do not leave as papers. For that reason, up to now they were protected only under the name of data. Now they must be protected under the name of problems too.

When an agent solves that problem, the problem leaks out through several paths. Conditions inside the prompt, intermediate results of experiments, records of improvements. In a day-cycle loop, the amount that leaks grows on a day cycle, and the targets to audit accumulate on a day cycle.

What to note is that what leaks is not always the problem itself. The shape of the problem leaks along with it, and an experiment record written with how far things have come as a condition is in a form an agent can read before a person does. Who improved what, with which model, under which conditions, through which path. Without this record, you do not know the change history of the asset.

So the question moves one step beyond speed. Where do you run that problem, with whom, and while leaving what record.

ThakiCloud’s agent-native cloud Paxis is a product that has officially shipped with v1.1 GA. In Paxis, the whole of agent operation is platform resources. Skills, Tools, Policies, and Audit Logs are managed as first-class resources, and autonomy is bounded by a governance of L0 through L3 that sets the limits of each task.

If problems have become assets, the seat where you run agents is a decision of the same grade as assets. Paxis stands the entire platform inside a company as a sovereign, on-prem Kubernetes deployment. An execution environment in which unsolved problems and experiment records do not cross to an external API is the first safety device of the problem-asset era.

Execution isolation is in the same context. The moment you hand a problem to an agent, the range of tools that agent can reach becomes the size of the leak. Paxis’s isolated sandbox execution runs an agent inside the fence of a single task. MCP connectors and the skill market lend only the tools a task needs. It is a structure of working inside the fence, carrying only the tools required.

The longer the loop, the more cost is recalculated. Paxis’s CostRouter mixes grades of models within the same loop by per-task model selection. Policy gates and audit logs leave a record of which path a day-level improvement came out on.

On the morning it became property, a company chooses one of two seats. The side that guards its own problems, the side that opens its own problems as a training ground. That choosing itself becomes a capability of the agent era.

Listing the morning’s three numbers again, they all point to the same thing. The 700 papers are results published all at once. The 69.13% is a record updated in a single day, and the 1.6x is a result that went past even the bottleneck of evaluation. The remaining question is one. Where a company runs its problems is, in the end, the answer. The signal of a door closing and the signal of a door opening arrived on the same morning. Because the morning when the price of problems is set is right now.

References

This article was written by synthesizing the news below.

• HuggingNews, Researchers Raise OpenAI Integer Multiplication Exponent Saving 69% in a Day
• HuggingNews, Math Researchers Abandon Openness After OpenAI’s 700 Paper Release
• HuggingNews, Hugging Face Turns OpenAI Math into Open Source RL Environments
• HuggingNews, Sakana AI’s New MASS Method Boosts 27B Model Performance 1.6x