Notes on atoms, permissions, and trust.

This past September, OpenAI assigned roughly 10.000 agents to one of the oldest open problems in mathematics, the behaviour of the Navier-Stokes equations that describe how fluids move, and produced a result that may settle it. What matters to me here is what the result does not actually do: even if the proof holds, it does not, by itself, make a single wing, hull, or turbine easier to build. A question not able to be solved for 90 years, now answered in weeks, at almost no cost. Intelligence seems like it will keep cutting the cost of knowing, of designing, perhaps even of building. But what it has not cut is everything a product must still pass through to reach the world: the test, the certification, the factory, the customer.

In my second piece (“When intelligence becomes abundant”) I argued that when something becomes abundant, its value does not disappear, it moves to whatever the abundance cannot produce - the new scarcity. Here we recognize scarcity showing up in the physical world. The intelligence is now the cheap part, in money and above all in time, the actual wing is not. My first piece (“When intelligence is done by a machine”) asked what happens to work, the second, where value goes. This third one asks what happens when that intelligence reaches the world it is meant to change.

Convergence, the idea that frontier technologies now push each other forward, has been the backbone of our thesis since I produced the firm's teaser in 2024, and has been a recurrent perspective I’ve been writing about in an upcoming report about it we’ll release at a later point. What I want to add here are thoughts on some specifics about how discovery may now move at machine speed, but the world it has to enter does not. And the next decade belongs to whoever understands the barriers.

1. Two speeds

A protein structure that once took a laboratory years to resolve now takes seconds. The clinical trial that turns it into a medicine still takes the years it always took. That is the situation: discovery runs at one speed, deployment at another, and unless we innovate on deployment itself, on the way intelligence enters the world, there will be increasing friction (or perhaps serious mismanagement) between what we are capable to do, and what we are implementing.

In my last piece I wrote that the new scarcity comes in three kinds: physical, institutional and human. And it is those three scarcities which are actually the framework to understand the three barriers where fast discovery meets a slower world: The atoms themselves, the permission to act, and the trust of the people affected. Each behaves differently and deserves its own look.

2. Atoms: the physical barrier.

NVIDIA's own engineers, writing this year about AI models that simulate physics, titled a post "Don't yet trust the model, test the physics," and cautioned that a model can score well on its error metrics "while still violating important physical constraints." That sentence is worth pausing on. The model is intelligent and cheap, and it can still describe a flow that cannot physically exist. Between a good score and a working turbine sits the entire discipline of engineering, and engineering does not compress the way analysis did.

Energy makes the same point at scale. The delay to connect new power generation to the American grid held roughly 2,061 gigawatts at the end of 2025 (1.5x the country's installed capacity), and the median project connected last year had waited 61 months (Lawrence Berkeley National Laboratory). We can now design a thousand candidate batteries in a week, but the factory that produces the winning one waits 5 years for its grid connection. Most of the work remains on this side of the economy: JPMorgan Asset Management estimates the world needs roughly $90 trillion of infrastructure investment by 2040. So it seems the physical world is back at the center of the problem, and it has its own speed.

3. Permission: the institutional barrier.

The second barrier is the one the public debate keeps looking for in the wrong place. The argument about regulating AI is almost entirely an argument about models: what they may say, how large they may become, what the labs must report. But when intelligence enters the physical world, the rules that pace it are mostly not AI rules at all: A drug discovered by a model is approved on the same trial timelines as any other, an autonomous system flies under aviation rules, connects under grid rules, and crosses borders under trade law. So the question worth asking is not the one the debate keeps asking, but whether the institutions that already govern everything intelligence will touch are ready for what is arriving, and what ready would even mean.

Behind those rules sits something more than procedure. Institutions are how a society, acting as a body larger than any of its members discerns what it holds to be right and wrong, and enables, paces or restricts what enters on that judgment - for reasons social, economic and environmental as much as technical. Seen this way, permission is less bureaucracy than collective judgment, and collective judgment has a speed of its own (how that judgment is formed and swayed through political dialogue is another story, one my first piece alluded to whether our sociopolitical systems can carry the consequences of intelligence done by a machine and made abundant).

Autonomous driving shows one side. Waymo began as the Google self-driving car project in 2009, and 17 years later it is a working service of roughly half a million paid rides a week, in one country, with every new jurisdiction a separate negotiation. Social media is the opposite side of the experiment: it was deployed on entire populations before any rules arrived, and now being regulated after the fact, when it can no longer be withdrawn. Anyone who has watched bureaucracy delay something good, or been quietly grateful that it delayed something bad, knows both failures from daily life. The lesson is not that institutions should simply move faster. Speed and reversibility have to be judged together, because an experiment that can be stopped is a fundamentally different thing from one that cannot.

I believe it is worth saying where I stand in this matter, even if it may evolve over time. I do not believe the choice is between laissez-faire and asphyxiation. Regulation, done well, is an alignment of incentives that everyone involved can live with and sustain. The real test of a rule in this era is not whether it permits or forbids, but how fast it can understand its implications, recognize a mistake, in either direction, and correct it. A rule that cannot be revised when the evidence changes has stopped being caution and become a habit through unsupervised inertia.

A question hides under all of this, and I raise it because it challenges my thinking: if the test of government is how fast it corrects mistakes, and machines may one day be better at applying intelligence than parliaments, should AI run government? The thought gets easier to take seriously once you notice how much of politics, as practiced, is the management of emotion (disguised as interpreting context): voters, coalitions, consent. But that observation has some caveats: in a system that rules by consent and managing emotion is in large part the job, a machine that administers flawlessly without winning the consent of the governed would not be governing. On a darker note, if a machine learned to win it, it could manage a society's emotions at machine scale. In my first piece I quoted the CEO of Google DeepMind saying democracy "might have to give way to something better." Who should decide, and by what right, could deserve a piece of its own (and give me the chance to talk about philosophy which I deeply enjoy).

For my work at 1200vc, the consequence of all this is quite apparent: reading policy is now part of pricing a physical technology, and the causality runs both ways: sometimes the rule paces the technology, sometimes the technology forces the rule to change. The line that matters is between anticipating that interaction and capturing it.

4. Trust: the human barrier.

The third barrier is the least discussed and the hardest to see, because it rarely looks like a barrier. It actually  looks like adoption.

Societies rarely refuse a technology outright. They accept it quickly, by default even. Social media, again: it arrived as a way to stay connected, and a decade later we noticed it had quietly renegotiated what may count as true, what attention is worth, and what a childhood looks like. Dating apps arrived as a convenience and rewrote courtship. None of this was decided anywhere. It was accepted, and called evolution, and the price became visible only even before we could understand it. The human barrier is not resistance to technology, but rather the slow discovery of what was agreed to without deliberation. Intelligence is being accepted on the same terms, and what it does to truth, to judgment, to meaning, we are set to learn the same way: in hindsight. 

This transition will also naturally divide, although differently than evolution has done before. The great divides of the industrial and internet eras ran mostly between societies: the industrialized and the not quite, the connected and the not quite. This one runs through them, between the people, firms and countries that use these super intelligent systems intensively and those that do not, with competitiveness and relevance accruing to the first. And alongside the divide by exclusion, I’d expect a divide by choice: people who deliberately remain closer to the human, the way craft survived the factory not out of nostalgia but as a statement about what they are unwilling to hand over and value. How much cohesion a society keeps across the speed of the technological advances we are going to be experiencing may matter as much as any productivity gain.

The sharpest description I have heard of the difference between a human and these systems is this: a human will, at some point, say stop. The machine, left to itself, will keep going. In my second piece I counted judgment among the new scarcities. Restraint may be its final form, the oldest of the virtues: knowing where the point is, and holding it. And that may be the key to the third barrier: trust is not won by what a technology can do, but won by what the people behind it are visibly willing not to do.

5. The open questions

As in my previous pieces, I would rather end with the questions that I keep asking myself rather than with the answers I do not have.

• Which barrier hits first, and where? Energy in North America, permission in Europe, trust in different places for different technologies. Deployment, and its consequences, good and bad, will move toward whoever releases theirs first.

• Can the form of government learn the new pace? Our institutions were built to keep the peace and uphold a social contract, and for long stretches they did both. The question is not deregulation but design: institutions able to understand, test, measure, and adopt or discard quickly and visibly, distinguishing the experiments that can be stopped from those that cannot, without losing the oldest job of all, the wellbeing of the governed.

• Who underwrites the years in between? Between machine-verified knowledge and working physical capability sits a gap measured in permits, connections, trials and public patience. That gap is where strong technologies die quietly. Whose capital, and whose institutions, are built for it?

• What does diligence become? If what decides a physical technology's future is also a grid connection, a regulation, and a society's acceptance, then the analysis that matters is no longer only technical and commercial. How can, or should even, investors underwrite these barriers?

• And what are we accepting without deciding? Which of today's defaults, about truth, meaning, and who falls behind, will look in 20 years like decisions nobody made?

I’m excited to talk about these topics, and more, to some of the humans involved in the space and in a position to do something about across national security, critical industries, energy and physical infrastructure - at our upcoming Annual Summit next week, whose theme this year is, quite fittingly chosen, The Convergence.

____________________________________________________________________________

I write from a seat that looks at frontier technologies daily: new materials, energy storage, biotechnology, industrial automation, water, advanced manufacturing. The science passing through our desks has never been stronger, and the distance between that science and the world it is meant to change has never been more visible. Intelligence has become abundant, the world it must enter has not, and that, more than any model release, is what the next decade will be about.

Paul Clastre Autumn 2026

Disclaimer: Text written by human, research supported by AI, cover image generated by AI