Humanity has never suffered from a shortage of questions. We have suffered from a shortage of answers, and perhaps even more from a shortage of ways to search for them deeply enough.
Some problems survive for decades because they are not simply difficult. They sit at the edge of what human cognition, available computation, experimental capacity, money and time can reasonably explore. Better energy storage. Fusion. New materials. Diseases we still do not understand. Antibiotic resistance. Climate systems. Turbulence. Fundamental mathematics.
These questions remain open not because nobody thought to ask them, but because the territory between the question and the answer is enormous.
For most of history, that territory could only be explored at human speed. Every scientist had one brain. Every laboratory had finite people, machines and budgets. Every research program had to decide which hypothesis deserved another year and which dead end had already consumed enough careers.
Artificial intelligence may be changing that equation.
Not because money can buy truth. It cannot. Or, depending on how far we push this argument, perhaps we are about to discover an uncomfortable qualification to that sentence.
What money can increasingly buy is search bandwidth and search depth.
More agents. More hypotheses. More simulations. More independent attempts. More critics trying to destroy the strongest idea. More time exploring paths a human research team could never investigate during a lifetime.
You do not buy the answer.
You buy more attempts at reaching it.
“The next great inequality may not be access to information. It may be access to enough cognition to discover information that does not yet exist.”
We have already moved through several generations of knowledge infrastructure. Search engines taught us to find what humanity had written. Generative AI is becoming an answer engine, synthesizing what humanity already knows.
The next step is considerably more consequential.
A discovery engine does not retrieve an existing answer. It investigates a question whose answer may not yet exist.
And after that comes something stranger still: a question engine, a system that helps decide which unknowns deserve additional intelligence, how much, and for how long.
This no longer feels entirely theoretical.
On September 8, OpenAI published what it says is a solution to the Navier-Stokes existence and smoothness problem, one of mathematics’ Millennium Prize Problems. The architecture of the attempt is almost as interesting as the mathematics. OpenAI says the Navier-Stokes effort consumed approximately 130 billion output tokens and generated 2.7 million agent messages. The agents arrived at the result after about 88 hours, followed by another 17 hours of Lean formalization and verification using GPT-6 Astra. OpenAI also says it had agents exploring other Millennium Problems and later shifted resources toward Navier-Stokes when that path became more promising. (OpenAI)
That last detail matters.
Even in this early example, we are not looking at one giant machine permanently assigned to one question.
We are looking at allocation.
Different problems exist simultaneously. Resources move. Promising directions attract more cognition. Less promising ones lose it. New evidence changes where the system concentrates its effort.
That begins to look less like a chatbot and more like a research economy.
Marc Vidal highlighted another side of the same experiment by estimating a cost of roughly $6 million in tokens. OpenAI does not disclose that dollar figure, so it remains Vidal’s estimate rather than an official invoice. But the exact number is almost secondary.
A famous unsolved problem received a measurable allocation of agents, tokens, time and compute.
The question had a computational budget.
That changes something.
Imagine two research groups attacking the same problem. One can afford fifty agents for six hours. The other can deploy fifty thousand stronger agents for a month, connect them to specialist models and simulations, commission independent attacks on every promising hypothesis, and fund physical experiments to test whatever survives.
Neither is guaranteed the truth.
But come on, they are clearly not playing the same game.
Capital has always bought better science. It buys laboratories, researchers, datasets, instruments and time. AI does not invent this inequality. What AI can do is amplify it because capital can now potentially purchase enormous quantities of parallel intellectual effort.
The divide is no longer simply between those who have AI and those who do not.
It becomes the difference between those who can ask intelligence and those who can deploy intelligence at scale.
And this takes us somewhere uncomfortable.
Companies already keep discoveries private. Governments classify research. Businesses protect trade secrets. Pharmaceutical companies do not publish everything they discover. Energy companies do not automatically release every useful process into the public domain.
Do any of those sound familiar?
AI does not invent secrecy either. It simply makes exploration faster, broader and potentially much more scalable.
Imagine a corporation spending $100 million in machine reasoning on a difficult problem and discovering something genuinely transformative. A new material. A dramatically better energy process. A therapeutic breakthrough. A cryptographic technique. Something commercially enormous, strategically sensitive, dangerous, or perhaps destructive to the company’s own existing business.
Why must it reveal the answer?
It may not.
And suddenly we encounter a strange form of ignorance.
Humanity may not know an answer, not because nobody has discovered it, but because somebody discovered it and the rest of humanity does not know that they know.
“A civilization can be ignorant in two ways: because nobody found the answer, or because someone found it and kept it.”
Could this already be happening? In ordinary forms, certainly. Private knowledge has always existed. Have frontier AI systems already produced civilization-changing discoveries hidden from everybody else? We do not know. Pretending that we do would turn an interesting question into conspiracy.
The important point is that the mechanism now exists for the distance between public knowledge and private knowledge to grow faster.
Perhaps sufficiently large deployments of frontier cognition will eventually require some form of compute accountability. Not governments reading every prompt. Not a bureaucratic surveillance system for thought. Something closer to the accounting systems we already accept around money, energy or other powerful infrastructures.
How much frontier computation was deployed? In which broad domain? Under what safeguards? What models participated? What verification process followed?
Machine-generated knowledge may eventually need something resembling a receipt.
But there is another possible future, and I find this one much more interesting.
Napster demonstrated an idea much bigger than music piracy. Millions of ordinary computers could participate in a network whose collective capacity exceeded what almost any individual participant could build alone.
Forget the songs for a moment.
The breakthrough was peer-to-peer aggregation.
Now apply that logic not to music, but to cognition.
A GPU here. Storage there. A university contributes a specialized model. A scientist contributes expertise. A laboratory contributes experimental capacity. Somebody contributes money. Millions of machines contribute idle compute.
Individually, none of them can afford a cognitive supercomputer.
Together, maybe they can.
Call it crowd-thinking.
Napster distributed access to things that already existed. A peer-to-peer intelligence network could distribute the capacity to discover things that do not exist yet.
But there is an important correction to the mental picture.
This would not be one enormous collective brain choosing one question, throwing everything at it, declaring victory and then moving to the next.
Why would it?
Humanity does not stop doing physics while it works on cancer. It does not stop studying energy because somebody is researching Alzheimer’s. A cognitive commons should not behave that way either.
It could contain thousands of simultaneous questions arranged across entire domains of investigation.
Cancer. Fusion. Materials. Mathematics. Climate. Rare diseases. Agriculture. Neuroscience. Antibiotics. Fundamental physics.
Inside each domain, hundreds of specific questions.
The people and institutions providing the compute could decide where their resources go.
Someone might dedicate all of their machines to oncology.
Another participant might allocate 50 percent to energy, 30 percent to climate and 20 percent to mathematics.
A university might dedicate a cluster to one particular materials problem.
A foundation could fund neglected diseases.
Someone else might simply say, “Put my compute wherever the community believes another attempt has the highest value.”
And some capacity could enter a shared pool that the commons itself allocates.
Now the system becomes much more interesting.
It is no longer one super-brain.
It is an economy of questions.
Intelligence flows continuously toward different problems according to curiosity, social value, expertise, urgency, probability of progress and the preferences of the people providing the resources.
I would call that cognitive liquidity.
The ability to move intellectual capacity from one problem to another almost as we move capital.
A pandemic appears? Compute can migrate toward it.
A fusion experiment produces an unexpected result? Additional agents can be routed there within hours.
A promising cancer hypothesis begins surviving independent attacks? More cognition can follow the evidence.
A mathematical direction stops producing anything useful? Resources can move elsewhere.
OpenAI’s own description of shifting agents from other Millennium Prize Problems toward Navier-Stokes after seeing a promising result offers a small, centralized glimpse of precisely this kind of dynamic allocation. (OpenAI)
A distributed version could do it continuously and across thousands of questions.
That makes crowd-thinking much more than distributed computing.
It becomes a stack.
The first layer is physical: compute and energy.
GPUs, CPUs, storage, networks, electricity. There is no cognitive commons without physical infrastructure underneath it.
The second layer is orchestration.
Models, agents, tools and humans must behave like something more coherent than thousands of machines shouting ideas at one another. Large problems need to be decomposed into smaller investigations, explored independently where necessary, cross-pollinated where useful, challenged, recombined and eventually tested.
This is the difference between distributed compute and distributed cognition.
The third layer is knowledge.
And this one is crucial.
A useful collective intelligence cannot require every participant to surrender private data, proprietary research or personal knowledge into one gigantic database. If contribution requires giving up ownership, many of the institutions holding the most valuable knowledge will simply refuse to participate.
The principle should be different:
“Contribute intelligence without surrendering ownership.”
This is where Kwaai becomes particularly interesting.
Kwaai describes its mission around personal control, self-sovereign identity, transparency and openness. Its Personal AI work includes local operation, private data, revocable third-party access and the ability to grant targeted permissions to query personal information. Kwaai is also developing KwaaiNet as AI running over a distributed peer-to-peer fabric. (Kwaai)
It is also experimenting around confidential vector search and knowledge-base homomorphic encryption. (Kwaai)
Homomorphic encryption is fascinating in this context because, at least for certain forms of computation, it allows useful operations to be performed over encrypted information without exposing the underlying plaintext.
It is not magic. It is computationally expensive, not every AI operation maps neatly onto it, and privacy-preserving architectures involve far more than one cryptographic technique.
But philosophically it changes something important.
You may not have to choose between contributing knowledge and keeping ownership of knowledge.
Imagine a hospital participating in a global investigation without uploading its patient database into somebody else’s cloud.
Imagine companies allowing carefully defined computations against proprietary knowledge without revealing the knowledge itself.
Imagine individuals allowing collective intelligence to learn something from their data without surrendering their digital lives.
The knowledge stays where it belongs.
The compute travels toward it.
The resulting intelligence can still contribute to something larger.
Open-source AI democratizes the instrument.
Privacy-preserving open discovery could democratize the search without requiring universal surrender of the data.
Then comes the fourth layer: allocation and governance.
And perhaps this is the hardest one.
Who decides which questions receive cognition?
At first the answer sounds obvious: the community.
Then democracy walks into the room and things immediately become complicated.
Suppose I contribute one laptop and somebody else contributes 10,000 GPUs.
Should our votes carry the same weight?
Democracy has wrestled with a version of this problem for centuries. A successful entrepreneur may create thousands of jobs, pay millions in taxes and generate enormous economic value, yet still receives one ballot. Another citizen may contribute almost nothing economically and receives one ballot too.
That is not simply a defect in democracy.
It expresses a principle: sovereignty is not supposed to be purchasable.
Should a cognitive commons make the same choice?
If voting power is proportional to compute, wealthy contributors can buy the research agenda.
Congratulations. We just reconstructed concentration inside the system designed to escape it.
But pure one-person-one-vote creates another problem.
Science is not a popularity contest.
A million people could vote to spend a trillion tokens investigating a fashionable fantasy while an obscure materials question nobody understands could quietly unlock an energy revolution.
So perhaps we are asking the wrong question.
The system does not need one mechanism deciding everything.
It can separate legitimacy, expertise, contribution and allocation.
The community can decide which broad areas deserve shared resources.
Experts can evaluate whether specific questions are meaningful, tractable, safe and what evidence would constitute progress.
People contributing compute can decide where their own capacity goes.
And a portion of the total resources can remain in a genuinely common pool, governed collectively and directed toward important questions that may attract too little private compute on their own.
This changes the voting dilemma substantially.
Someone contributing 10,000 GPUs should probably have enormous influence over the destination of those 10,000 GPUs.
They do not necessarily deserve 10,000 times more political authority over everybody else’s resources.
Contribution can buy allocation rights without buying sovereignty.
That feels like an important distinction.
And it means the commons could support multiple allocation models simultaneously.
Some compute is self-directed.
Some is donated to particular scientific domains.
Some follows community votes.
Some follows expert recommendations.
Some could be automatically routed toward questions where new evidence indicates that another burst of computation might have unusually high value.
And some should perhaps be reserved entirely for verification.
This starts looking less like a global referendum and more like a constantly changing portfolio of human curiosity.
A public ledger of questions could make that portfolio visible.
Which diseases deserve another trillion inference tokens?
Which materials problems might unlock abundant energy?
Which scientific questions affect millions of people but have almost no commercial incentive behind them?
Which challenges have we labeled impossible simply because serious attempts were historically too expensive?
Which investigations are underfunded?
Which are consuming vast amounts of compute without producing meaningful progress?
Where did a new discovery suddenly change the odds?
And, perhaps most intriguingly, which questions should not be answered?
Strangely, I am particularly curious about those.
Because openness cannot mean recklessness. Some discoveries in biology, cybersecurity, weapons or other dual-use fields could be genuinely dangerous. A cognitive commons cannot operate according to the childish principle that every piece of information must immediately become public.
Perhaps the rule is open by default, restricted by exception.
And the exception itself must be visible, justified, reviewable and revisited. Otherwise “safety” becomes a very convenient word for rebuilding secrecy.
OpenCommons becomes interesting around this layer, not because it has built the system I am describing. It has not. But its existing work sits unusually close to the institutional questions such a system would face. OpenCommons explicitly focuses on collaboration among community, industry, academia and government, along with replicable systems, interoperability, open standards and community-driven governance. (OpenCommons)
The technical challenge is pooling intelligence.
The institutional challenge is allowing thousands of different actors to participate without requiring them to trust one central owner.
That may turn out to be harder.
For everything that can safely remain open, however, the social contract seems straightforward to me.
If the community supplies the cognition, the resulting knowledge belongs to the commons.
And not just the headline answer.
The methods. The evidence. The code where possible. The assumptions. The failed paths. The uncertainties. The contradictions. The experiments. The provenance. The independent replication.
Which takes us to the fifth layer: trust.
Who decides when an answer is actually an answer?
Ten thousand agents agreeing does not make something true.
A beautiful simulation does not make something true.
A confident model certainly does not make something true.
Reality remains an unforgiving referee.
In mathematics, formal systems can sometimes verify whether a chain of logic actually follows. In empirical science, things get messier. A molecule has to behave outside the model. A material has to exist outside the simulation. A treatment has to work in biological systems. A reactor has to survive the conditions we claim it can survive.
If generation becomes abundant, verification may become the scarce resource.
We may therefore need not only answer engines, discovery engines and question engines.
We may need verification engines.
And this is another place where the portfolio model becomes useful.
Not everyone needs to contribute compute to generating new hypotheses.
Some people might deliberately allocate their machines to replication.
Others to adversarial testing.
Others to searching for contradictory evidence.
One research pool proposes.
Another tries to destroy the proposal.
A third reproduces it independently.
A fourth waits for physical evidence.
Crowd-thinking should not produce only more answers.
It should produce organized skepticism at machine scale.
Every consequential answer should arrive carrying its own evidence.
What supports it?
What assumptions does it depend on?
What would falsify it?
Which independent systems reproduced it?
Which laboratories replicated it?
Which parts are experimentally demonstrated and which remain inference?
I called this a proof-carrying answer before. I think evidence-carrying answer may be better.
Outside mathematics, proof is often the wrong word.
The important principle is that an answer should not merely tell us what to believe.
It should make belief inspectable.
Once machines can generate millions of hypotheses, ideas become cheap.
Trust becomes expensive.
And underneath every layer lies the least digital part of this entire story.
Energy.
Every token has a physical history.
Electricity enters infrastructure. Silicon converts it into computation. Computation becomes inference. Inference becomes search. Search produces hypotheses. Hypotheses become experiments. Occasionally, experiments become knowledge.
Energy is not simply the bill AI leaves behind.
It may become one of the primary inputs of accelerated discovery.
Which means abundant energy and abundant intelligence increasingly become the same conversation.
More energy permits more computation.
More computation permits more attempts.
More attempts potentially produce more discoveries.
Some of those discoveries improve energy itself.
The loop begins to feed itself.
We already build public telescopes, particle accelerators, laboratories and supercomputers because we have decided that certain questions are too important to leave entirely to private markets.
Why should public machine cognition be conceptually different?
Perhaps the next great scientific infrastructure is not one machine at all.
Perhaps it is a distributed capacity to think.
Not one brain.
Not one question.
Not one institution deciding what deserves attention.
A continuously shifting network of questions, people, knowledge, models, experiments and computation.
An economy in which cognition can flow toward problems.
A commons in which people can contribute capacity without surrendering ownership.
A governance system in which contribution matters but does not automatically purchase sovereignty.
A verification system in which the answer is not the end of the process but the beginning of scrutiny.
And then we reach the question underneath every other question.
What should humanity ask?
There may be an enormous territory between impossible and never seriously attempted with enough intelligence.
We have no idea how large it is.
No, AI will not make everything possible. Physics will continue saying no. Mathematics contains limits. Biology will remain extraordinarily complex. Politics will remain painfully human.
But I would dare to say that the frontier of what is investigable may be expanding faster than our imagination of what deserves investigation.
Search engines taught us to find what humanity had written.
Answer engines are learning to synthesize what humanity knows.
Discovery engines may help us reach knowledge humanity does not yet possess.
Question engines may help us decide where another billion attempts might matter.
And a cognitive commons could allow those questions to coexist, compete, attract resources, lose resources, gain evidence and evolve.
That changes the metaphor completely.
We are not building one artificial genius and asking it the one question that matters most.
We may be building an ecosystem of questions, with intelligence constantly flowing through it.
And perhaps that becomes the final scarcity.
Not intelligence.
Not even compute.
Purpose.
The tragedy would be to build unprecedented cognitive capacity and spend most of it optimizing advertising, engagement, financial arbitrage and the next eleven seconds of somebody’s attention.
The opportunity is considerably larger.
Build the compute layer.
Build the orchestration layer.
Build knowledge systems that allow contribution without surrendering ownership.
Build an allocation system that lets cognition flow between thousands of worthy questions.
Build governance that cannot simply be purchased by whoever owns the most GPUs.
Build verification into the answer itself.
Build enough energy to make deep search cheaper.
Then ask better questions.
Lots of them.
The price of an answer is beginning to become visible.
What we still have to decide is who gets to ask, who gets to allocate the intelligence, who gets to know, what counts as knowing, and whether we are ambitious enough about the questions we choose to fund.
SOURCES:
- OpenAI — “On the Navier-Stokes Millennium Prize Problem”
https://openai.com/index/navier-stokes-solution/ - Marc Vidal — LinkedIn post on OpenAI, Navier-Stokes and the estimated computational cost
https://www.linkedin.com/feed/update/urn:li:activity:7503173460882255872/ - Kwaai — “About”
https://www.kwaai.ai/about - Kwaai — “KwaaiNet”
https://www.kwaai.ai/kwaainet - Kwaai — “Workgroups”
https://www.kwaai.ai/workgroups - Kwaai — “Summit 2026”
https://www.kwaai.ai/summit2026 - NIST — “Privacy-Enhancing Cryptography: Fully Homomorphic Encryption (FHE)”
https://csrc.nist.gov/Projects/pec/fhe - NIST — “Protecting Model Updates in Privacy-Preserving Federated Learning”
https://www.nist.gov/blogs/cybersecurity-insights/protecting-model-updates-privacy-preserving-federated-learning - NIST — “Scalability Challenges in Privacy-Preserving Federated Learning”
https://www.nist.gov/blogs/cybersecurity-insights/scalability-challenges-privacy-preserving-federated-learning - OpenCommons — “About”
https://opencommons.org/About - OpenCommons — “Community Standards”
https://www.opencommons.org/Community_Standards - Clay Mathematics Institute — “Rules for the Millennium Prize Problems”
https://www.claymath.org/millennium-problems/rules/ - Clay Mathematics Institute — “The Millennium Prize Problems”
https://www.claymath.org/millennium-problems/ - National Academies of Sciences, Engineering, and Medicine — “Reproducibility and Replicability in Science”
https://www.nationalacademies.org/read/25303 - National Academies — “Improving Reproducibility and Replicability”
https://www.nationalacademies.org/read/25303/chapter/9
You must be logged in to post a comment.