Karl Schwarzschild worked out the physics of a black hole while dying at the Russian front. He sent the calculation to Einstein by letter, from a military hospital, and was dead within months. He had not been looking for it. The war had given him time between shellings, and the equations had simply arrived, more finished than he was.
Benjamín Labatut’s When We Cease to Understand the World is built from men like Schwarzschild: minds that reached a truth too large for the body carrying it. Fritz Haber. Werner Heisenberg. Alexander Grothendieck, who built an entire architecture of mathematics and then walked away from it, distrustful of the discipline he had made. The book’s thesis sits in its title. Understanding, past a certain point, does not accumulate. It ruptures.
The question worth asking of this book now is whether it is right to leave its scientists so passive. Labatut’s scientists do not choose their catastrophes. Knowledge finds them and breaks them. But history is not only a record of minds standing alone against ideas too large to hold. It is also, here and there, a record of structures built to hold the idea together with the people who carry it. The word hold will change its meaning three times before this essay is done. First a mind holding an idea. Then a society holding the consequences of one. Finally, whether we can hold something built well enough to hold us back.
Bali’s subak system coordinates irrigation across an entire volcanic watershed: hundreds of rice terraces, timed against pest cycles and water scarcity, managed through water temples centred on Pura Ulun Danu Batur. The anthropologist Stephen Lansing, in Priests and Programmers, modelled the system and found something close to an optimisation engine — upstream and downstream farmers synchronising their planting to control pests better than any farmer could manage alone. The authority running it was religious, not administrative. The priests who set the calendar were the same body that bore the consequence of getting it wrong, since the water they blessed was the water their own fields depended on. Lansing’s account has its critics — some anthropologists argue he overstated how deliberate the coordination was — and the causes of what followed are contested. But the broad shape is not seriously disputed: when the Indonesian government’s Green Revolution replaced coordinated planting with uniform high-yield varieties in the 1970s, the synchrony collapsed, water shortages and pest outbreaks followed, and the system was substantially reinstated by the 1990s.
Tamil Nadu’s eri tanks worked on a related principle, closer to home. The village sabha held both the technical decision — when to release water — and the religious calendar that made desilting a festival obligation rather than a chore. That fusion thinned under colonial administration, which drew tank maintenance into engineering departments answerable to the state rather than the village. Restoration efforts now are an attempt to rebuild what thinned, rather than what was cleanly severed.
The Bishnoi go further still. In 1730, at Khejarli, Amrita Devi and 363 others died embracing khejri trees to stop them being felled for a maharaja’s palace. This was not policy. It was identity: a community for whom harm to the trees was harm to themselves, with no distance left between the two to argue across.
What each of these shares is not an institution but an exposure. The subak priest is exposed to the consequences of the water he releases. The sabha is exposed to the consequences of the tank it manages. The Bishnoi are so identified with the trees that there is no self left standing apart from the harm done to them. None of them received a tool from outside and then had to decide, under pressure, whether to use it responsibly. The capability and the restraint were never two different people to begin with.
Fritz Haber shows what happens without that exposure. He was not blindsided by what his chemistry could do. He understood, with total clarity, that the same process feeding the world could also gas men in trenches. He made the choice anyway, inside an institution built precisely to reward it. No scaffold surfaces a danger that has already been surfaced and set aside. Haber’s failure was not cognitive. It was exposure: the chemist was never made to stand where the gas landed. The war ministry decided, and the trench absorbed it, and the two were never the same body.
South Africa’s Truth and Reconciliation Commission and Rwanda’s Gacaca courts are not water temples. They manage nothing ongoing, and they answer a different kind of question — not how to hold restraint and capability together over time, but how a society metabolises harm once it has already happened. Still, the underlying instinct echoes. Amnesty in the TRC was conditional on full disclosure: the discovery of what happened and the judgment of what to do about it, produced in the same public act, by the same people, at the same time. Gacaca went further, handing judgment to elected neighbours rather than an external legal class. Both refuse to let the people closest to the truth be separated from the people with authority to act on it.
Modern India has comparatively little of this. Partition, the anti-Sikh violence of 1984, and the ruptures since have mostly been handled through inquiry — commissions that investigate but do not reconcile. This is a real gap, and it says something about which kind of scaffold a society reaches for first.
Every example here rests on a premise this essay has been quietly relying on throughout: a human hand still on the instrument, and still standing close enough to feel what the instrument does.
AI may not allow that premise. Schwarzschild’s calculation was too large for the man who produced it, but it was still his. It did not calculate him back. Haber’s danger, too, was visible and permitted: he knew what the chemistry would do, and did it anyway. A predictive machine can fail differently, because it can complete the separation Haber only began.
Consider what happened to biometric authentication in India’s public distribution system. Aadhaar, the country’s biometric ID, was linked to ration shops nationwide on the promise of cutting fraud: a fingerprint or iris scan, matched against a central database, before a family could collect its subsidized grain. For most people this worked as designed. For manual labourers whose fingerprints wear down with the years, for the elderly, for anyone in a village where the network drops, it did not — the machine returned no match, and match is the only language the system speaks. The government has disputed direct causation in some of them, the way Lansing’s critics dispute how deliberate the subak’s coordination really was, and the individual causal chains are genuinely contested. But the structural pattern under the disputes is not: the engineers who built the authentication system are not exposed to a failed match. The ration dealer running the handheld device is not exposed to it — he is following a rule, not making a judgment. The official who mandated biometric authentication nationwide is several offices removed from any single machine on any single day. And the person standing at the counter, whose fingerprint the machine does not recognise, is exposed to all of it, with no one else in the chain positioned to notice, let alone answer for, what just happened to her. This is not a problem unique to India but Aadhaar’s scale, wired directly into the machinery of who eats, is where the stakes come closest to Haber’s, and to the Bishnoi’s own history of consequence with nowhere to hide from it.
It does not take intent to hold us this way. It only takes enough precision, applied at enough distance from consequence, when prediction stops assisting a decision and starts resembling one. If that is what this is, then no water temple can be built around it exactly as built before, because a water temple is built around a river, and a river does not predict the priest.
And yet the subak priests still watch the water. The sabha still keeps its calendar. None of them believed, at the outset, that the scaffold would hold forever. Only that it was worth building anyway, for as long as a hand remained on the instrument.
Perhaps the honest response is not a finished framework but the unglamorous work of building enforced exposure where it no longer exists by default — and some of that work, however imperfectly, has already started. The Reserve Bank of India’s 2022 digital lending guidelines require that a regulated lender, not the app or the algorithm behind it, be the named party responsible for a borrower’s key disclosures and for any recovery practice carried out on its behalf — an attempt to drag exposure back onto an institution that can be summoned and answered to. India’s Digital Personal Data Protection Act, passed the following year, gives people a right to grievance redress against automated processing, at least on paper. Neither is Khejarli. The RBI’s rule only reaches lenders it regulates, and the harassment-by-algorithm cases that prompted it mostly involved apps operating outside that net, pulled from app stores only after the harm was already done. The DPDPA carves out broad exemptions for government use of data — precisely the register, the state’s own automated systems, where Aadhaar’s failures occurred. These are early, gameable versions of the idea the subak priest embodied without needing a statute at all: exposure has to be structural, or it is theatre. But they are at least an admission that the old premise — a hand on the instrument — now has to be engineered rather than assumed.
That watchfulness comes in more than one register. There is the kind with no audience, that needs none — the water either rises or it doesn’t, and the watching is done whether or not anyone sees it done. And there is a louder kind, performed at scale, precisely so that power cannot pretend not to be seen. The Internet Freedom Foundation’s Project Panoptic, a public registry tracking facial recognition deployments by police and government bodies across India, works on this principle: no agency gets to quietly retire a surveillance system from public memory, because the record of it lives with a civil liberties group rather than with the agency itself. Set this against China’s own answer to a similar problem — a filing requirement under which companies must register their recommendation algorithms, and since 2023 their generative AI systems, with the Cyberspace Administration. That, too, is exposure. But it is exposure to a single regulator rather than to the public, which solves the problem of no one watching by creating one more watcher who is not, in turn, watched by anyone else. The loud register only works if there is no single throat to silence. A registry only the state can see returns the whole arrangement closer to Haber’s war ministry than to Khejarli.
The slow kind catches what accumulates unnoticed. The loud kind catches what happens too fast for the first to notice at all. Neither one, as it currently stands, would have caught what happened at that ration counter in time. It may take both — not because the danger has been solved, but because standing watch, this time, has to be designed rather than inherited: the only way left to hold something built well enough to hold us back.