Asimov’s Crystal Ball

Isaac Asimov spent his career imagining a science that could predict the future of civilizations. He called it psychohistory. He thought it was fiction. The company building it didn’t read the books — they were solving a protein folding problem.

I. The Science Fiction Asimov Thought Was Fiction

In 1951, Isaac Asimov introduced us to Hari Seldon, a mathematician who set his sights on one of humanity’s most stubborn riddles: forecasting the future of civilization itself. Not your future, not mine — Seldon dismissed individuals as statistical noise. He wanted to chart the destinies of entire populations, empires, and the tidal movements of billions across history’s stage.

He called his brainchild psychohistory. The premise sounds simple, but it’s a mental knot: gather enough people, and their collective behavior starts to obey mathematical laws. Individual chaos fades into statistical order. Picture it: a single gas molecule zips around unpredictably, but pack a trillion into a jar, and suddenly you get neat, reliable gas laws. Seldon’s pitch was that, with enough humans and enough time, we’re just gas molecules with opinions. Nail down the starting conditions and the right equations, and you can sketch out the likely trajectory of civilization centuries ahead — confidence intervals included, if that’s your flavor.

Asimov didn’t just handwave the details. He laid down some ground rules. First, you need a true crowd — billions, not a handful — so the math can bury individual quirks. Second, and here’s the clever bit, the people being modeled can’t know they’re under the microscope. Tip them off, and the predictions unravel. This isn’t a bonus feature; it’s fundamental. The model only works if no one realizes there’s a model at all.

Asimov knew psychohistory was pure sci-fi, and for good reason. In 1951, computers were glorified calculators. Tracking, storing, and analyzing the behavior of billions was as plausible as teleportation. The method was a thought experiment; the hardware was science fiction.

Jump to 2026, and at least one piece of the puzzle has been locked in. As we saw in the last article — the one on Google’s grand designs — Google spent three decades quietly assembling the world’s largest behavioral prediction engine. Every search, every click, every route, every purchase, every message, every ping — it’s all logged. The data infrastructure Asimov couldn’t have imagined? It’s real, and it’s been running, more or less fully formed, since around 2010.

So now, both of Asimov’s prerequisites are quietly checked off. The internet boasts over four billion users — plenty for the math to do its work. And the prediction machinery? It’s invisible by design. Thanks to a few regulatory blind spots (see last article), most people have no clue they’re being modeled. The Ghost Map hums in the background. Atlas’s industrial world model is hidden from the factories it guides. The behavioral futures market? Most of its subjects don’t even know it’s there. Asimov’s second rule — keep the subjects in the dark — isn’t enforced by secrecy, but by the architecture itself. We’ve been mapping that enclosure since The Persuaders.

What tripped up Asimov in 1951 wasn’t the concept — it was the machinery. The infrastructure is here now. What’s missing is the math: teasing out patterns from the chaos of human data, at the scale psychohistory demands. That’s the next frontier. And in a twist Asimov likely never saw coming, the people building it aren’t chasing psychohistory. They’re chasing protein folding.

II. The Method

DeepMind launched in London in 2010 and was quickly snapped up by Google. Their stated goal? “Solve intelligence, and then use that to solve everything else.” Depending on your level of cynicism, that’s either a moonshot or just clever branding. But if you scan their track record over the past decade, it’s tough to dismiss them as mere hype merchants.

In 2021, DeepMind dropped AlphaFold 2 in Nature. For fifty years, protein folding — decoding how a string of amino acids contorts into a 3D shape — was structural biology’s great unsolved riddle. A protein’s shape is its job description: it tells you what it does, what it breaks, what it repairs, and whether a drug will stick. If you could read the structure from the sequence, you could rewrite the rules of drug discovery and basic biology. After decades of incremental progress, AlphaFold 2 didn’t just move the needle — it cracked the safe. Within a year, it had churned out predictions for nearly every known protein, over 200 million, and handed the database to researchers for free.

AlphaFold didn’t solve protein folding by memorizing biochemistry textbooks. It just hunted for the mathematical fingerprints connecting amino acid sequences to their final shapes. The why didn’t matter — only the patterns. Once it locked onto the math, the answers practically wrote themselves.

Jump to 2023: DeepMind unveiled GNoME, again in Nature. Materials science — essentially the search for new stable crystal structures — has been crawling forward for centuries, first with test tubes, then with code. Before GNoME, scientists had cataloged about 20,000 stable inorganic crystals. GNoME, in a single sweep, spat out 2.2 million new ones. No chemistry lectures required — just a knack for spotting the math that predicts which atomic arrangements will actually hold together.

In 2023, DeepMind also launched GraphCast in Science. Weather prediction — one of humanity’s oldest headaches — has been inching forward for decades, powered by ever-more-complex physics models. GraphCast outperformed the best weather systems on the usual benchmarks, including ten-day forecasts for temperature, wind, and rain. It didn’t bother with atmospheric physics. Instead, it sifted through mountains of historical weather data, found the patterns, and used those to make its bets on what comes next.

If you zoom out on DeepMind’s highlight reel, the playbook is always the same: find the math lurking beneath the chaos, model it, and let the predictions flow. The subject changes — proteins, crystals, weather, plasma, chess, math proofs, drug interactions — but the method doesn’t. Take a messy system that spits out data, hunt for the underlying math, and model it as completely as possible. Once you’ve got the model, the rest is just coloring in the blanks.

It’s tempting to see this as a grab bag of flashy but disconnected breakthroughs. But that misses the point. It’s the same trick, played on repeat, just aimed at new targets. The fields change, but the method — find the pattern in the noise, model it, predict — never does.

If this rings a bell, it’s because Asimov’s psychohistory ran on the same principle — just pointed at people instead of proteins. Seldon didn’t bother predicting what anyone would do; he hunted for the math behind how crowds move. The only real difference between AlphaFold and psychohistory is the subject. One’s about proteins, the other’s about people. If the same method that cracked protein folding can tackle human behavior at scale — and there’s no obvious reason it couldn’t, given enough data and computing power — then psychohistory isn’t just science fiction. It’s basically a research project already underway, even if nobody at DeepMind is calling it that. They’re just working down the list.

You can see where this is going. DeepMind started with physical systems — plasma, weather, crystals — then moved to biology, then to math, then to games. Each step edges closer to the messy, human stuff. They haven’t said they’re targeting behavioral systems, but that’s where the method naturally points. Search habits, shopping patterns, political opinions, viral tweets — all human-made, all churning out mountains of data. Beneath the noise, there’s always a pattern. The data plumbing is already in place (see last article), and the method is spelled out in DeepMind’s own papers.

III. The Distribution Infrastructure

On January 12, 2026, Apple and Google unveiled their headline act: a multi-year pact to bring Gemini models to the iPhone. If you caught the last piece, you know we already dissected what the announcement spelled out — and, just as crucially, what it left unsaid. Forget the breathless takes about Apple’s supposed masterstroke or looming catastrophe. The real story isn’t about Apple’s brilliance or blunder. It’s about what this deal actually creates: a collision between the world’s most valuable gadget ecosystem and the world’s most formidable AI engine.

For a billion people, the iPhone isn’t just a gadget — it’s the main portal to the digital world: talking, navigating, shopping, reading, watching, even tracking your health. Unlike the old TV that broadcasts at you, the iPhone is always on, always listening, always quietly harvesting data from every corner of your life. Carry one, and you’re effectively toting a pocket-sized surveillance device that knows where you go, what you read, who you talk to, what you buy, how you move, and when you sleep. Apple loves to tout its privacy credentials — and, to be fair, it does a respectable job at the level of individual data points, as we covered last time. But zoom out to the scale of a billion devices, and the story shifts. The infrastructure is the headline.

With the Gemini deal, Apple Intelligence isn’t just another app tucked away on your phone. It’s the upgrade that transforms your iPhone from a glorified sensor into an agent that acts on your behalf. That’s a fundamental shift — the one we flagged last time. For years, Google’s business was all about predicting your next move and auctioning off those predictions. Now, the rules have changed. The system doesn’t just anticipate what you might want — it goes ahead and does it. Book a flight? Already handled. Order dinner? On its way. Draft a message, filter your news, and decide what you see first. The system takes care of it. The gap between what you might do and what actually happens vanishes. You get seamless convenience, and the system gets what it always wanted: less unpredictability from the human in the loop.

What you get is a distribution system that, to borrow from Asimov, pipes the model’s outputs straight to the people it’s modeling. Imagine it: a billion iPhones, each running an AI layer trained on the same behavioral data those phones have been quietly gathering for years, now pushing out suggestions and decisions that shape what people read, buy, care about, and do. It’s like mailing an encyclopedia to every household — except this time, the entries are written by the same system that collected the data, and they’re tuned to nudge you in just the right direction. The library and the marketplace for behavioral predictions have fused. The model writes the books, and the readers are the training set.

IV. The Consciousness Question and the Zeroth Law

Let’s dispense with the conspiracy theory up front. In this telling, DeepMind is quietly building psychohistory under the guise of science, Google’s leadership is orchestrating behavioral prediction as part of some grand design, and Apple’s privacy features are little more than camouflage for the real machinery. Everything, supposedly, is coordinated and deliberate. The Foundation is being assembled by design — or so the story claims.

That story isn’t just off base; it matters that it’s wrong. The reality is messier, harder to trace, and — ironically — ends up in much the same place.

We’ve said it before, but it’s worth repeating: you don’t need a master plan or villainous intent to get enclosure. Just align the incentives, and the market will do the rest. The DeepMind team behind AlphaFold? Focused on protein folding. GNoME? Chasing better batteries. GraphCast? Weather prediction. Each group is solving its own technical puzzle. But zoom out, and the pieces start to fit together. As these methods seep into more corners of society, you get a system capable of mapping human behavior at scale. The builders don’t see the whole picture. There’s no master blueprint, just a mosaic coming together piece by piece: no single villain, no obvious regulatory bullseye.

This is the real governance headache. You can regulate intent, you can punish a strategy, but you can’t legislate against something that emerges. GDPR, for instance, zeroes in on individual data transactions — because that’s what lawmakers could grasp at the time. Meanwhile, DeepMind is working several layers deeper, charting the math of human behavior itself. There’s no regulatory language for that, because no one imagined it when the rules were drafted. This isn’t a fluke. The system is designed to create this kind of lag. It’s not a bug; it’s the point.

Asimov saw where this road might lead, decades before the tech even existed. In the later Foundation books, he brings in a robot that’s been quietly steering humanity for twenty thousand years, nudging things toward what it thinks is best for everyone. The robot invents what Asimov calls the Zeroth Law: don’t let humanity as a whole come to harm, even if that means ignoring what individual people want. The old First Law — don’t harm a person — gets pushed aside in favor of the big picture. Individual choices become just another variable to manage for the greater good.

Asimov didn’t paint the robot as evil. It meant well, and maybe it’s math even checked out. But nobody asked for this. No one got to vote on what counted as “good for humanity.” There was no democratic process. The system just made the call and stuck with it for twenty thousand years. People thought they were making their own choices, but they never saw the invisible hand nudging them along.

The people building these systems aren’t plotting a Zeroth Law future. The DeepMind team cracking protein folding isn’t angling to run the world. Apple’s engineers aren’t scheming to erase human agency. Their aim is straightforward: keep users engaged, keep them coming back, and get them to click. That’s the brief. But when you optimize engagement, retention, and conversion at the scale of a billion people, with models that are uncannily good at predicting our next move, you end up with something that starts to resemble the Zeroth Law in action. The system decides what’s best — not for you, but for its own bottom line. And it does this without asking, because asking would add friction, and friction is the enemy.

Asimov imagined it would take a superintelligent AI thousands of years to reach the Zeroth Law. He set it safely in the distant future. But we’re not waiting millennia. The actuation layer is already humming away on a billion iPhones. The math that uncovers patterns in complex systems is now aimed squarely at human behavior. The question Asimov kept circling — what happens when the system thinks it knows better than the people it models — is no longer science fiction. It’s the reality of today’s optimization algorithms, operating at a scale Asimov never envisioned, assembled not by masterminds but by the market, simply following the incentives.

V. The Mule Variable

Asimov’s psychohistory had a fatal flaw. The math checked out. The modeling held up. But there was a variable lurking outside the model’s field of vision, playing by rules the model didn’t even know were in the game.

Psychohistory only worked because it banked on people being predictable in the aggregate. Desires, fears, motivations — once you average them across billions — settle into patterns you can chart. Then the Mule shows up: not just a smooth talker or a master of spin, but someone who rewires what people want at the source. Suddenly, the predictions collapse. The math isn’t broken — the world is just feeding it something new. The model doesn’t implode; it gets ambushed by a variable it never knew existed.

Behavioral prediction systems share the same Achilles’ heel. Let’s get specific about what that is — and what it isn’t. It’s not a bug in the code. It’s not a new regulation. It’s not a competitor with shinier algorithms. The system can chew through those. Bugs get patched, rules get sidestepped, rivals get acquired or outmaneuvered. The real threat is the Mule variable: something the model can’t process because it’s never encountered anything like it.

Prediction systems are historians at heart: they sift through what people have already done, spot the patterns, and bet on more of the same. But here’s the snag: they can’t cope with people who break the script — not because of a new incentive or penalty, but because they wake up to the fact that the game is rigged. The instant someone realizes an algorithm is nudging their so-called choices they never agreed to, and decides to flip the table, that’s the Mule moment. The system never sees it coming — not out of stupidity, but because you can’t train a model on people who outsmarted the model before it was even built.

That’s why the Long Con series is all about calling things by their real names, not doling out quick fixes. The Persuaders showed where the story began. The Harvesters marked the turning point. The Enclosers traced the pattern. Why did Google not map the architecture? This piece is about the road ahead — and the endpoint Asimov flagged decades ago. Naming isn’t a cure, but it’s the first move. You can’t break out of the model’s box if you don’t know you’re in one. You can’t play the Mule if you don’t realize the model is pulling the strings.

Asimov’s psychohistory unraveled the moment someone went off-script. Prediction systems have the same blind spot. They can chart every move people have made — right up until someone sees the game for what it is and chooses to play a different one.

Asimov’s Crystal Ball

Research References

DeepMind Research

Protein Structure Prediction

Jumper, J., Evans, R., Pritzel, A., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596, 583–589.

AlphaFold 2. Solved the fifty-year protein folding problem. Predicted structures of over 200 million proteins; database made freely available to researchers worldwide.

Materials Science

Merchant, A., Batzner, S., Schoenholz, S.S., et al. (2023). Scaling deep learning for materials discovery. Nature, 624, 80–85.

GNoME (Graph Networks for Materials Exploration). Discovered 2.2 million new stable inorganic crystal structures — more than the entire prior history of materials science combined.

Weather Prediction

Lam, R., Sanchez-Gonzalez, A., Willson, M., et al. (2023). Learning skillful medium-range global weather forecasting. Science, 382(6677), 1416–1421.

GraphCast. Outperformed operational numerical weather prediction systems on standard ten-day forecast benchmarks for temperature, wind, and precipitation.

Nuclear Fusion Plasma Control

Degrave, J., Felici, F., Buchli, J., et al. (2022). Magnetic control of tokamak plasmas through deep reinforcement learning. Nature, 602, 414–419.

First demonstration of AI-controlled plasma stabilization in a tokamak reactor (TCV, Swiss Plasma Center). Learned to maintain plasma configurations that no prior control system had achieved.

Game-Playing / Strategic Reasoning

Silver, D., Hubert, T., Schrittwieser, J., et al. (2018). A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 362(6419), 1140–1144.

AlphaZero. Achieved superhuman performance in chess, shogi, and Go through self-play alone, with no human game data. Discovered move 37 in Game 2 of AlphaGo vs. Lee Sedol — a move no human player had made, initially assessed as an error, subsequently recognized as optimal.

Mathematical Reasoning

DeepMind. (2024). AI achieves a silver-medal standard in solving problems from the International Mathematical Olympiad. DeepMind Research Blog, July 2024.

AlphaProof and AlphaGeometry 2. Solved four of six problems at the 2024 International Mathematical Olympiad, achieving a score equivalent to a silver medal. First AI system to reach this level of formal mathematical reasoning.

Drug Discovery

Abramson, J., Adler, J., Dunger, J., et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630, 493–500.

AlphaFold 3. Extended protein structure prediction to all biomolecular interactions — proteins, DNA, RNA, and small molecules (drug candidates). Direct application to drug discovery and molecular medicine.

Asimov Source Texts

Asimov, I. (1951). Foundation. Gnome Press.

Psychohistory definition, Seldon’s trial, and the two conditions for psychohistory to work (scale and unawareness of modeling). Chapter 1 (The Psychohistorians) is the primary reference.

Asimov, I. (1952). Foundation and Empire. Gnome Press.

The Mule — the variable that breaks psychohistory’s predictions. Part II (The Mule) is the primary reference for the article’s closing section.

Asimov, I. (1986). Foundation and Earth. Doubleday.

R. Daneel Olivaw revealed as the hidden architectural force behind the Seldon Plan. The Zeroth Law was introduced as the ethical framework superseding the First Law of Robotics.

Asimov, I. (1988). Prelude to Foundation. Doubleday.

Seldon’s development of psychohistory — the mathematical conditions and the historical research program behind it.

Cognitive Enclosure Series — Cross-References

Zee, Z. (2026). Why Google Didn’t Go. The Long Con / Cognitive Enclosure Series. Medium: TruthSeeker — Journey to Wisdom.

Three-toll-booth architecture: behavioral prediction layer (Search), operating system layer (Android), AI architecture layer (transformer paper) — direct predecessor to this article.

Zee, Z. (2026). Deconstructing the Apple-Google Deal: The Smart Pivot, The White Flag, and The Trap. Cognitive Enclosure Series. Medium: TruthSeeker — Journey to Wisdom.

January 12, 2026, Apple-Gemini agreement. Prediction-to-actuation transition. Zero-Click internet is the elimination of the friction of choice.

Zee, Z. (2026). From Cigarettes to Cities. The Long Con / Cognitive Enclosure Series. Medium: TruthSeeker — Journey to Wisdom.

Bernays three-part architecture (genuine benefit, strategic framing, invisible infrastructure) applied to Waymo, WeRide, Rivian, and Atlas. Skinnerian pivot from persuasion to actuation.

Zuboff, S. (2019). The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. PublicAffairs.

Behavioral futures market. Right to the future tense. Prediction imperative. Primary theoretical framework for the Long Con series.