Big Theme: Quantity Ahoy

From Slop to Non-Slop

Two years ago, every argument about AI was about quality. The machines hallucinated court cases, absorbed biases from their training data, misspelled words in their own images, drew hands with six fingers, and couldn't count the letters in strawberry. The confident claims that biases and hallucination errors were inherent to machines already sound quaint. Those complaints faded so steadily that few people noticed them leave.

Well, to be honest, a few still arrive every morning. An author finds two em dashes in a stranger's article and declares it slop. A mathematician or a lawyer posts a machine blunder, and one caught error becomes a verdict on everything AI does. Each catch makes a quiet case for the top professional, and each assumes that everyone, everywhere, has one on call at a fair price. Most people never have. They read whatever the author chose to write, file petitions in whatever words they can manage, and have no mathematician to call for the problems they have at hand.

This week Microsoft showed a panel of AI agents diagnosing hard medical cases correctly more than twice as often as generalist doctors working alone. The interesting part is reach. Such a panel can sit beside every patient anywhere. A machine mathematician can sit beside every engineer. A machine writer can produce an eight-out-of-ten petition, in any style, for anyone who asks. The machine need not be the best person in the world to be the best help someone has ever had.

Humans rarely abandon an argument; some people still believe the Earth is flat. And, honestly, it is fun to read well-written articles and well-crafted memes about AI idiosyncrasies. For most purposes, though, the quality debate is over. The bar that mattered was usefulness, and the work cleared it while we were busy counting fingers. Useful intelligence is becoming plentiful, and we are still inspecting its punctuation.

Where This Essay Lands

For readers who stop after a section or two, here is where we end up.

This essay claims exploding innovation, far beyond a faster version of the old pace. In a couple of days of machine effort, one lab cut the operations needed in a quantum circuit seventeenfold. Another posted 719 papers on hundreds of open mathematical problems, weeks after its agents put forward a claimed answer to part of a famous problem that had stood for a quarter century. Those are a handful of rooms. Now picture every lab, factory, and back office on earth aiming the same machinery at whatever it is looking for, a few quarters from now, with better models and cheaper attempts. The result is as hard to imagine as foreseeing the rainforests while watching the first cells divide.

The trajectory matters more than any snapshot. Error rates keep falling while volume keeps rising. Every result becomes raw material for the next, across fields that once kept their distance. Breakthroughs will stack on breakthroughs before the first is understood. Much of this will happen inside companies that never publish, and some of it will enter products before any human can explain it. The more agents we deploy, the less any rulebook can say what they will do, even when each one does exactly as asked. The physical world slows the flood, and even that brake is weakening.

This is not about feeling gloomy or happy. The coming tsunami may usher in an era of continuous, compounding, and highly complex creative destruction whose scale and scope are hard to put into words. We are heading into a world where few sitting ducks will stay comfortable as the innovation crossfire around them intensifies. History offers no guide. Every past wave of invention was paced by how many human minds could work on it and how many hours they had. That ceiling is now lifting. The novelty lies in the sheer quantity: more attempts, more results, and more combinations than any economy has ever had to absorb. Those who build the capacity to absorb it will thrive, and the greatest risk sits with those who decide not to try.

One Proof, Seven Hundred Papers

There is an old, grim saying that one death is a tragedy and a million is a statistic. Science has just met its own version.

In September, one lab set roughly ten thousand agents loose on one of the hardest open problems in mathematics: whether the equations of fluid flow always behave. The agents talked among themselves, passed partial results around, and produced a proof in under four days that the world's specialists are still arguing over. That is a headline. It has a hero, a century of history, a famous prize, a quarrel over credit, and seminar rooms full of people checking every line. It can be doubted, admired, and debated because there is only one of it.

Then, in October, the same lab posted 722 manuscripts in one go. Its model had been pointed at about four thousand problems and left to run. It came back with results in 372 families across many corners of mathematics, each costing on average a few hours of machine thinking. Within days, a handful had been withdrawn or revised. That is a statistic. Hundreds of results fit neither in a headline nor in a mind, and nobody in the field quite knows how to think about them.

The jump from one to hundreds took a few weeks. The next release need not stop at hundreds. Three thousand, thirty thousand, and eventually a million are a matter of budget. Nothing requires any of them to be announced. The same recipe works for anyone who can state a problem precisely and score an answer automatically. This covers far more than mathematics: circuits, molecules, materials, schedules, routes, and designs of every kind. Every room on earth that is looking for something can now run its own version.

That is the flood this essay is about. The flood of fake news and machine slop has been debated for years. Slop is the easy case, since it fails on its own. The hard case is good output, pouring out in volumes no person, team, or institution was built to absorb. Readers should brace for the vocabulary. Flood, deluge, torrent, and inundation will return again and again in this article. We apologize in advance for the hydraulics.

A Month of Floods

The mathematics papers were only the loudest example. The last few weeks brought a dozen quieter ones. Each would have filled a year of headlines in its field’s journals not long ago.

Software is being rebuilt from the outside. In under two weeks, one developer released open versions of seven flagship programs from one of the world's best-known design software makers, built with agents and without the original code. Others are porting decades-old video games by the dozen. The copies are rough, full of crashes and missing features, so expect your naysayers to keep producing the memes. Plus, they enjoyed a big advantage: the original sat beside them like an answer key, ready to say right or wrong at every step. The important takeaway given our Super-Moore era is that copying is the warm-up lap.

The real race has already started. A popular programming tool of more than half a million lines was rewritten into over a million lines of a new language in eleven days, with up to sixty-four agents working side by side. Its creator reckoned three engineers who knew the code intimately would have needed a year. Earlier this year, one model combed through the major operating systems and browsers and surfaced thousands of previously unknown security flaws in a matter of weeks. One had hidden from human auditors for twenty-seven years.

Biology is moving at the same pace. This week about two thousand agents worked through an atlas of eight million mouse brain cells, in which specific genes had been switched off, and returned leads on brain disease in thirty-two minutes. A human team would have needed months. Along the way, the agents drew on more than seven hundred published papers, where a single agent managed about a dozen. Last month a genome atlas predicted the effects of every one of nine billion possible single-letter changes in human DNA.

None of this needed a model unveiled with fanfare. It needed persistence, numbers, and a budget. And every example is a floor. The models doing this work are the weakest the world will use from here on. Next year's will be cheaper. They will hold more and will run in larger crowds. If this is the output of a few weeks in late 2026, the output of a few quarters from now belongs to a different scale altogether.

Critics still count how many of a hundred machine answers are wrong today, whether ten, thirty, or forty. Bias and hallucination followed the same path: the error count kept falling while the volume kept climbing. Quantity also changes the arithmetic. A method that works once in a thousand tries has roughly a ninety-five percent chance of succeeding at least once in three thousand. Where success can be recognized, a long shot becomes a method. Some apparent limits of intelligence turn out to be limits on how many failures we can afford.

A Mathematical Revolution in Disguise

Popular language calls these systems large language models. The name stuck because the first thing they did was write. We have always believed they were something larger: a mathematical revolution that happened to arrive through text. Language was only the doorway. A model that merely predicted words could never have attacked fluid equations, redesigned quantum circuits, or read a brain atlas. These systems are sets of equations that absorb context in new ways and find patterns along long, variable chains of reasoning. This is why they travel from chemistry to code to law without asking permission. The walls between fields are low. The water is patient.

The last few weeks show that revolution working at volume, in fields where precision is everything. A million-line program must compile and pass every test. A proof must survive a formal checker. A circuit must do exactly what it claims in fewer steps. A system that can hold hundreds of thousands of lines, thousands of interacting variables, or a long chain of equations in view at once, and keep working on them for days, turns out precise work by the truckload. The flood has moved well past essays and summaries. It carries code, proofs, designs, and predictions, each of which can be checked, used, and built upon. What machines are doing is far from ideal, but what is guaranteed is that their capabilities will improve and what is a new given is that the machines are doing things we could never dream of doing until now.

Much of it will also happen out of sight. For every lab that posts its results, others will run the same machinery inside a company and say nothing, a theme we return to later. All of it sits on a curve that keeps steepening, since each model improvement multiplies across every room running it. The defining feature of what has begun is quantity. The part we keep forgetting is that it compounds.

Dealing with Millions, Billions, Trillions

Take the last few months in one breath, including some of the things discussed in a section above. Thirteen million lines of machine-checked proof for Fermat's Last Theorem in eleven days. A million lines of rewritten code in another eleven. Thousands of hidden security flaws across every major operating system in a matter of weeks. A flaw that sat for more than a decade in the language behind most of the world's websites, found by several AI-equipped teams independently within weeks of one another. Seven flagship design programs rebuilt from the outside in under two weeks. Nine billion predictions about the human genome. Eight million brain cells read in thirty-two minutes. More than a thousand protein designs in two days, with hundreds binding their targets in real labs. Nearly a hundred million telescope images searched in two and a half days. More than eight hundred objects turned up that no paper had ever described. Seven hundred mathematics manuscripts in a single post. Nine more open problems solved in a peer-reviewed journal a week later. Contributions written by agents to the world's largest code platform passed seventeen million in a single month. Any one of these would have led the news for a week just some months ago.

Read that list again, slowly. Each line took days or weeks. Each once was a career, a department, or a decade. Together they describe a new fact about the world: machines now turn out precise, checkable work in a bulk no civilization has ever had to receive.

The deeper story is what that bulk makes possible. Every science we have was built inside the human skull, and the skull is small. We discussed limitations of skull in a detailed article earlier. The flood takes that argument much further. Most of us can juggle about four things at once. Physics has an exact formula for two bodies orbiting each other, yet three already defeat any general solution. So our sciences learned to simplify. Economists hold all else equal. Engineers assume the friction away. Biologists study one gene at a time, though most traits and diseases arise from thousands of small genetic differences acting together. The abacus could never carry a differential equation. Calculus and computers widened the room, yet a person still had to hold the pattern at the end. Much of what we call knowledge is a careful approximation, chosen because it fits a few moving parts.

The new machines push that wall back. Think of how an essay like this one gets written. Earlier, we could read dozens of ideas, keep perhaps six in mind, and quickly cut to the two or three we can carry, losing the rest along the way. Now, a machine helps keep all forty in view, remember every one, test every pairing, and return to the discards a week later. The same holds where the stakes are higher. The Fermat proof had to keep thirty thousand theorems consistent with one another. A million-line rewrite must keep every function agreeing with every other. A genome model weighs how thousands of DNA letters act together, a pattern no teams or collection of tools could ever hold together for humanity until now.

This is where quantity turns into complexity. When a tool can hold ten thousand interacting variables, the old simplifications stop being necessary. An economy can be modeled as millions of different households and firms, with no need for one average buyer. A material can be designed against hundreds of competing properties at once. A drug can be judged by its effects across the whole body and alongside the other medicines a patient actually takes. Some may rightly argue that inclusion of numerous variables or data points may not increase the quality of work across all fields, but just as we stopped using log tables when calculators became accessible, the ability to handle more will suppress the tendency to work with little.

While many of the coming innovations will be better answers to old, known questions, the larger share may come from questions we never asked, because no mind could reach the complexity levels required to frame them. And when thousands of agents chase those questions in different directions and bring the pieces together, the results compound.

The strange part is how calmly we take all this. We repeat: each item on that list at the beginning of this section made a day of news, if that. When each quarter doubles the last, the doubling stops being an event and becomes weather. Events stop being news at exactly the moment they start becoming infrastructure. We are growing bored by the arrival of the thing we are least prepared for.

If We Had $50 Million to Spare

The famous labs are spending heavily to make a point. They aim their machines at problems everyone has heard of: a fluid equation that has resisted a century, a set of quantum circuits, a theorem with its own legend. The results are proofs of concept, a way of showing what the machines can now do. Many people watch and shrug. Very impressive, they say, but what is it for?

We ask ourselves the same question in a narrower form. What would we do at GenInnov with fifty million dollars to spare? For a fund our size, the question is purely theoretical, which makes it the most enjoyable kind. At October's rate of a few hours of machine thinking per result, that sum buys more attempted research than any investment firm on earth could staff. Here is roughly how we would spend it.

Speed would be the least interesting part. High-frequency traders already own the microseconds, and we have no wish to join them. Our problem has always been breadth. Tens of thousands of listed companies sit inside a far larger web of suppliers, customers, and rivals. Every day brings a torrent of filings, patents, papers, hiring notices, and price moves in a dozen languages. Every investor copes the same way: by ignoring almost all of it. We would start by giving every company we could ever own an agent of its own. It would read everything the company and its rivals publish, never take a holiday, and remember what management promised three years ago better than management does.

Then we would connect them. A shortage at a chemical plant, a new standard for networking chips, a quiet change in a regulator's wording: each touches dozens of companies through chains no analyst can hold in mind. Today the second- and third-order effects reach prices slowly, because people work them out one link at a time. Thousands of agents, each watching its own company and talking to its neighbors, could trace a single news item through the whole web within the hour. This is the skull problem again. Analysts think in pairs of companies. The economy is a web of millions of links.

The market's own data forms a second web, probably the bigger prize. Every portfolio is a tangle of cross-currents in prices and volumes. Yet almost every risk model still reduces that tangle to pairs: how one stock moves with another, summed into a correlation table. Principal components and our own work on the effective number of bets squeeze a little more out of the same table. Pairs were never the market's truth. They were simply the most a person, or an ordinary computer, could manage. A large fleet could watch baskets in place of pairs, in real time, across thousands of names at once. It could see the same piece of news reaching one group of stocks within minutes and another only after days. It could spot a cluster of companies starting to move as one before anyone has a story for why.

That would change portfolio construction itself. A portfolio could be judged by how its holdings respond to many possible shocks at once, measured on maps no correlation table can draw. Relative value could be found between shifting baskets of companies as well as between pairs. Diversification could be measured by what actually moves together in the moments that matter. Most investors build portfolios on a picture of the market drawn to fit the human skull. The flood offers the first chance to draw a different one.

We would also point a large share of the fleet at the flood itself. Every week brings thousands of papers, code releases, and patents. Somewhere among them sits the result that will hurt a profitable incumbent three years from now. A swarm could read all of it and ask three questions of each item. Who does this help or hurt? How soon? And most importantly, what can we do for our innovations from the new progress?

Put together, this would be a systematic strategy unlike any built so far. Quantitative investing grew up on what could be counted: prices, ratios, and a handful of factors, chosen partly because people and their computers could manage only a handful. Agents can now count what used to need judgment, across thousands of companies at once. Much of what they found would be noise. Some attempts would fail expensively. The rest might change how a portfolio is built.

Compounding Out of Sight

We do not have the fifty million. Nobody in our part of the world is about to wire it over for a daydream either; the valley where such cheques get written is a long flight away. The irony is that the sum is small, less than many companies spend on advertising in a year. What it buys grows with every model generation.

The experiments will spread unevenly. The famous labs demonstrate their capabilities in public. Elsewhere, firms will quietly give a small team a budget and point a fleet of agents at an intricate problem: a new alloy, a smarter supply chain, a better demand forecast. Some will accept the cost of unsuccessful attempts. Others will wait until a rival’s results become difficult to ignore. Many first attempts will disappoint, as research usually does.

Much of the payoff will be too small to announce. A manufacturer improves a hundred ordinary things: less heat loss here, fewer rejected parts there, longer intervals between repairs. Together, they change what the business can charge and which rivals can keep up. An innovative era can arrive through a thousand ordinary costs that quietly stop being necessary. The failures will be visible. Many of the wins will be filed under operations. The AI bill will arrive neatly itemized, while the benefits scatter across the business. Expect the argument about returns to lag the returns themselves.

Now multiply those rooms by thousands. Their results refuse to sit in a pile. Each useful answer can become a foothold for the next. DeepMind’s AlphaEvolve has already found faster matrix-multiplication algorithms, improved a chip circuit, and accelerated a routine used to train the models behind the system itself. Some of the gains improve the machinery doing the discovering. A faster training routine helps build the next model; a better model can search for the next improvement.

Breakthroughs also arrive in sets. A promising material may be too expensive to make, its process too hard to control, its design too fragile. Improve one and little happens. Improve all three and a product appears, perhaps because three teams that never met each solved one piece. Work in one field can quietly remove the obstacle that kept another field’s success useless.

Old ideas revive too. A shelved design may have been waiting for a cheaper sensor or a better catalyst. Its rejection was a verdict on the surrounding world. When that world changes, the verdict can expire. The flood expands what we can invent and what we can recover from the things we once abandoned.

Much of this compounding can happen with no human in the audience. The next user of a result may be another machine. A predicted protein structure can help rank a molecule; that ranking can select an experiment; its result can improve the next prediction. No person needs to read every intermediate answer for the chain to be useful. Findings can enter products without ever becoming papers, presentations, or stories anyone tells.

This changes our relationship with knowledge. We may be able to check that a result works long before anyone can offer an explanation that fits comfortably in a human mind. Knowing that and understanding why begin to travel at different speeds. The debate about whether machines truly understand their discoveries will continue. Meanwhile, some of those discoveries will already be at work, and others will be superseded before we have finished understanding them.

The people living through this will struggle to see it. When so much changes in so many directions at once, even an operator in the middle of a market cannot tell what it is competing against. Its next rival may be working in another field, solving a problem whose relevance neither side yet recognizes. The map grows old while we are drawing it.

The Next Discovery Has Already Started

We do not have the fifty million. Nobody in our part of the world is about to wire it over for a daydream either; the valley where such checks get written is a long flight away. We are resigned to what we could have done! The irony is that the sum is small, less than many companies spend on advertising in a year. What it buys grows with every model generation.

So the experiments will start elsewhere and spread unevenly. Today, the famous labs run them in public. Next come the firms that quietly hand a small team a budget and point a fleet of agents at an intricate problem: a new alloy, a smarter supply chain, a better demand forecast. Some of that has surely begun already. Many companies, including plenty in Asia, prefer to see proof before they spend on a hunch, so they will move later. Even the most cautious will try once a rival's results become hard to ignore. Many first attempts will disappoint.

Much of the payoff will be too small to announce. A manufacturer improves a hundred small things: less heat lost here, fewer rejected parts there, a longer interval between repairs. Together, they change what the business can charge and which rivals can keep up. An innovative era can arrive through a thousand ordinary costs that quietly stop being necessary. The bears will miss most of it. They will tally the rising cost of AI, assume the models never improve, ignore the ordinary failure rates of all research, and seize every disappointment. The failures will be visible. Many of the wins will be filed under operations. Most visible will be the higher costs of AI. We are already reading articles moaning about these items aplenty these days, and we expect a flood there as well in the coming quarters, accompanied by the usual ROI questions.

Going back to models being aimed at problems or opportunities, searching for solutions never imagined. Multiply those rooms by thousands. The results then refuse to sit in a pile. Each one becomes a foothold for the next before the first is understood. A single system at one lab shows the shape. It found a way to multiply small matrices in fewer steps than a method that had stood for over half a century. It sped up a routine inside the lab's own models enough to shave their training time. It rewrote part of a chip design, helped cut errors in reading DNA, and improved quantum circuits. Each gain fed the machinery that produces the next one. First-order thinking, which asks what one result means, will keep missing the second and third orders arriving behind it. Some findings will be superseded before they are understood.

Breakthroughs also arrive as sets. A promising material may be too expensive to make. Its process may be too hard to control. Its design may be too fragile. Improve any one, and little happens. Improve all three, and a product appears. When work spreads across thousands of directions, one team may be quietly solving the problem that kept another team's success useless. Ten agents make forty-five possible pairings. Ten thousand make nearly fifty million. Old ideas revive too. A shelved design may have been waiting for a cheaper sensor or a better catalyst. Its rejection was a verdict on the surrounding world. When that world changes, the verdict can expire.

Secrets by Default

In 1942, a young Soviet physicist noticed something missing. Papers on nuclear fission had vanished from Western journals. The silence told him more than any paper could. He wrote to Stalin that the Americans must be building a bomb.

The flood brings back that kind of reading. A company running thousands of agents for months leaves only shadows: power bills, hiring patterns, odd patents, and a product that is suddenly better with no paper to point to. Outsiders will see that something is happening without knowing what. Suspicion and conjecture will fill the gap. Silence becomes a signal. So does its imitation, since a rival can stage shadows as easily as cast them. Secretive projects once needed a national laboratory. Now they need a credit line and a question. Even the yardstick goes private, as public tests of machine ability get gamed almost as soon as they appear. Progress will continue. The yardstick will become a secret.

Being Right No Longer Buys a Hearing

Time for the usual change in the narrative arc. Back to the flood itself and the other side of the story. Everything so far has been about production. The harder question is who receives it all. The quality debate was comfortable because every failure came with a remedy. Catch the hallucination, punish the fabrication, improve the model, wait for the next version, or even not use AI. Each version did get better. Quantity offers no such fix. With each model, the problem will worsen. Any remedy has to be built from human attention, trust, and time. These three things do not improve exponentially.

Whether published or kept behind a company’s doors, useful results eventually compete for something scarce: attention, money, an experiment, a place in production. Our institutions have been built for a trickle. Peer review rations the hours of qualified readers, a court’s docket the hours of judges, a hiring process the hours of interviewers. Producing something worth considering was itself an expensive filter. Remove that cost and the rationer becomes a dam asked to pass a thousand times its design flow, holding back mostly true water.

Anyone who has run a research department on the sell-side, like this writer, knows the feeling in miniature. At first, every complaint is about too little. Once the business grows, the struggle becomes too much: hundreds of stocks, a river of reports, and investors willing to hear only a handful of messages. Improving the research does not manufacture another hour in the reader’s day.

The maintainers of curl have lived through both sides. They ended their bug bounty program after a flood of poor submissions. By October, their complaint had changed: the AI-assisted reports were largely accurate and arriving in overwhelming volume. Imaginary flaws could be dismissed. Real ones created obligations. Better answers made the workload harder to escape.

When a committee cannot examine every good proposal, familiarity becomes a cheap substitute for judgment. Famous names and conventional ideas can win precisely when discovery is spreading beyond them. The flood can widen the entrance while narrowing the desk where decisions get made. Some good things will be refused simply because something equally good arrived first. Even worse, they will be refused because no one has the time to look at them while busy executing the previous idea.

Our sense of time is built for scarcity too. We celebrate a breakthrough and fear a failure. A continuous flow of consequential progress offers no single event around which to organize either feeling. Even buying decisions wobble. Faster discovery can make customers slower to commit, since a better option may be only months away. The future keeps bidding against the present. The prize may go to those who can change their minds cheaply.

A deluge of falsehood is a scandal. A deluge of truth is a climate.

More Agents, Fewer Guarantees

Control is the same story inside the machinery. A quantity explosion of agents means, almost by definition, that no rulebook can fully control them. A fixed set of laws in the style of old robot fiction sounds tidy. Yet something always remains unspecified, which is basic mathematics. Midas got exactly what he asked for. The trouble sat in the specification. We can ask for the optimum and forget to say what must survive the optimization. Guardrails are one more specification, with gaps of their own. Rogue behavior needs no malice. It needs only enough instances and enough ambiguity.

Every handoff adds loss. Two generals sending messengers across a valley can never be certain they agree, however many messages they exchange. Bits can arrive intact while meaning slips. If each handoff keeps nine-tenths of the intent, five handoffs keep less than three-fifths. Machines checking machines create loops of their own. Ten reviews from the same kind of mind count roughly as one, because verification only multiplies when the checkers can fail differently. Emergence is subtler still, since it needs no failure of control at all. Many agents learning from and building on one another can produce behavior nobody designed, the way a traffic jam appears with no bad driver.

Here too, the weak point is human. Microsoft's red teams spent a year attacking deployed agent systems. The weakness they exploited most often was the human approval step itself, worn down by fatigue and by small requests that each looked harmless. Every guardrail assumes a person with fresh attention reading each request. Every real deployment delivers thousands of requests, a bored approver, and a button that says yes. Quantity is the solvent of the human veto. The serious builders already design around it. One lab runs about thirty thousand agents at a time internally. Its machine monitors trim a hundred thousand flagged transcripts a week to roughly fifty that reach a person.

Capping the number of agents sounds prudent. A limit that binds only the cautious, though, is a subsidy to the incautious. The practical answer is constant verification, done mostly by machines, with humans deciding what deserves checking. The safe number of agents is a frontier that keeps moving. The safety work has to move with it.

Where the Flood Meets the World

The physical world is the last brake, and the strongest. Ten thousand hypotheses can be narrowed to three excellent experiments in an afternoon; the experiments still take months. A drug still needs its trials, an alloy its furnace, a factory its permits, a power line its right of way.

The ledger is lopsided. One AI system proposed some 380,000 stable crystal structures, against roughly twenty thousand found in all of human experimentation. A robotic lab then made 41 new compounds in seventeen days, a pace that would still need centuries to work through the list. The first drug whose target and molecule were both found by AI reached human trials in a fraction of the usual time, and it now faces a final trial of hundreds of patients and a full year of dosing. The machines can dream a thousand times faster than the world can build.

So the water pools upstream. Ideas will queue for lab benches, test slots, regulators, and trusted signatures. Those choke points will become far more valuable. Whoever owns the last physical step owns a veto over the first thousand intellectual ones. The friction is partly healthy, since it buys time to absorb. It is also shrinking, as robotic labs that plan, run, and learn from their own experiments push the boundary back a little every year. Anyone counting on the riverbank to hold forever is counting on the slowest part of the system staying slow.

Building the Riverbanks

This essay has left aside the darker uses of volume: armies of fake persons, synthetic majorities, and complaints, appeals, and attacks that win by sheer number. They are real. Others cover them well. Music, books, and video have their own version of the story. Listeners and readers filter machine output hard today. Yet as quality rises and cost falls, the machine's share of what we consume may not stay small. Resistance tends to hold only until a few defect and find a business in it, as we already see with images.

Our concern is the flood of the good. Organizations that build ways to sort, verify, and act on abundance will pull away from those still built for a trickle. An institution that cannot refuse good work will be buried by it. We can purchase brains. We have to grow the ability to use them. The riverbanks themselves must be built with some of the intelligence that is raising the water. The greatest risk still sits with those who decide not to attempt. The next greatest sits with those who attempt and build no way to take delivery.

The moment a tsunami becomes visible is the last comfortable moment. Today the water is warm. The headlines are about em dashes. The first waves, hundreds of papers here and billions of DNA predictions there, are only the edge of what is coming. Quality was the question we knew how to argue about. Quantity is the one we will have to live with.