How this editorial can be challenged
Who pays for the independent work that turns an AI-produced result into knowledge other people can use?
Producing a result and making it intelligible are different activities. If rewards and resources favor the first while treating the second as spare-time service, explanation, scrutiny and broad participation can become bottlenecks even as machine output expands.
Companies can reasonably organize initial expert collaborations to make difficult work manageable, and direct corporate funding can help researchers move faster. Demanding an elaborate new funding system could slow useful discoveries.
Coordination and funding are valuable. The proposal is to separate financial support from control over conclusions and access, using independent allocation and public outputs rather than assuming either unpaid work or corporate selection is the only option.
We have not independently verified the scientific claims, estimated the total cost of understanding them or measured whether existing funding is sufficient. The workforce experiment is a bounded analogy, not direct evidence about mathematical research institutions.
Broad independent access, sustained support for researchers outside initial collaborations and widely reusable explanations would show that current arrangements are solving the problem. Evidence that a proposed fund distorts priorities would require redesign.
There is another kind of infrastructure
When I read a spectacular scientific claim, I want to know what it lets another person understand. Can a graduate student learn the idea? Can a skeptical researcher test the assumptions? Can somebody use it to ask a better question? These are not demands to make science less ambitious. They are how an impressive result becomes useful beyond the moment of its announcement.
AI infrastructure is usually pictured as buildings, chips and power. I think we need to include something less photogenic: the time of people who can turn unfamiliar output into explanations others can inspect and reuse. More machine-produced material does not automatically create more of that time. It can make the need for it more urgent.
Give the second achievement a name
A new result and a new understanding are related, but they are not interchangeable. A researcher might reconstruct an argument, locate the assumption that does the real work, find a simpler route or reveal where an advertised conclusion is too broad. That contribution can matter even when the researcher was not first to announce the result.
The workplace study in today's briefing offers a limited analogy: improved assisted performance did not establish a measured gain in unaided technical knowledge. It does not prove anything about the economics of mathematics. It does remind us to avoid awarding one achievement on the evidence of another. A finished output can be valuable while the work of understanding remains unfinished.
The funding is not zero—and that matters
OpenAI describes its initial academic collaborations as independent and unpaid. That should not be paraphrased as an absence of support. The Simons Foundation has announced $15 million in combined funding and resources with Anthropic and OpenAI, with awards managed by the foundation. It includes an open proposal route for relevant activities. This is a constructive institutional response worth recognizing.
What the announcement cannot yet tell us is whether support will be sustained, broadly accessible or adequate for the work researchers actually encounter. We have not calculated that need, so we should not invent a funding gap or a magic percentage. The useful next evidence will come from what is funded, who can participate and what reusable understanding emerges.
Access changes who gets to contribute
The independent academic advisory group AGMAI has objected to company-selected early access and asked for the physics artifacts to be released. Its concern makes the funding question more interesting. A researcher can have time and expertise but still be unable to contribute if the relevant material is unavailable. Money and access are complementary, not substitutes.
There is also a career dimension. If the first group invited receives most of the opportunity to explain a result, people outside that circle may struggle to see a worthwhile role. This is a potential institutional effect, not a finding that any named researcher acted improperly. A fair process should make room for independent approaches and give visible credit to useful explanations, corrections and negative results.
Coordination is not the enemy
The strongest defense of a small initial collaboration is practical: difficult material needs organization, and an experienced group can make the first public account more coherent. Corporate money can also support work that would otherwise wait. Insisting on perfect institutions before any research proceeds would create its own cost.
My preference is neither a ban on collaboration nor a demand that every expert work for free. Support independent allocation, release inspectable materials, disclose relevant interests and protect researchers' freedom to publish unwelcome conclusions. Fund explanation as an output in its own right. Then judge the arrangement by participation and useful work produced, rather than by how warmly the research community greets a launch.
Choose what kind of abundance we want
There is a tradeoff here. Every hour spent explaining an existing output is an hour that researcher cannot spend pursuing a different question. Paying for explanation does not remove that opportunity cost; it makes it visible. Independent institutions should help decide when understanding a promising result is worth the effort and when a field should preserve space to follow its own curiosity.
I would welcome a future in which AI gives us many more fruitful ideas. I would be less impressed by a future filled with claims that only a small invited group can interpret. The choice is not between machines and people. It is between treating human understanding as a scarce input someone else will somehow supply, and deliberately building the conditions in which more people can share it.
Read the reporting
Opinion is ours. The factual record is linked below.
OpenAI — scientific collaboration and release arrangements AGMAI — access, independence and support recommendations Simons Foundation — $15 million combined funding and resources Nature Human Behaviour — assisted performance and unaided knowledge