The Pattern We Cannot See
Verified cases of machines finding predictive structure beyond human sight — and what they change about knowing
The pattern does not sit above the waterline. It sits below the depth an unaided eye can reach.
The claim
Every earlier information technology carried a human bottleneck. A person had to see the pattern, state it, and convert it into instructions before any machine could use it.
Machine learning removed the bottleneck. A model can now find and act on predictive structure that no human has seen, stated, or understood — and this book assembles the verified evidence.
We used to have to know the pattern before the machine could act. The machine now acts on patterns we cannot see.
Part I / The dependency
Sight before software
For the entire history of computing before machine learning, the pattern had to pass through a human mind on its way to the machine. Begin with that dependency.
CH / 01
The rule we had to write
Consider what a payroll program is. Somewhere, a person understood a regularity: hours worked relate to wages owed by a rule involving rates, overtime thresholds, and tax brackets. The person wrote the rule down. A programmer translated the rule into code. The computer then executed the rule millions of times without error and without fatigue.
Notice what the computer contributed and what it did not. It contributed speed, scale, memory, and reliability. It did not contribute the rule. The pattern connecting hours to wages was discovered, understood, and articulated by humans before the machine ever ran. The machine amplified a discovery. It made none.
This was not a limitation of payroll software. It was the operating constitution of information technology for its first seven decades. A database returned what a schema designer anticipated. A search index ranked by criteria engineers chose. Early chess programs played chess by evaluating positions with formulas that chess-playing humans wrote, encoding human judgments about material, mobility, and king safety. Expert systems of the 1980s were the constitution made explicit: engineers interviewed specialists, extracted their rules, and stored the rules as if-then statements.
Each of these systems could exceed human performance in execution — faster, broader, tireless. None of them could exceed human performance in seeing. The machine’s ceiling of insight was the insight of the people who programmed it.
Every arrow in that sequence except the last passed through a person. The world offered regularities. A human noticed one, compressed it into language or mathematics, and handed the compression to the machine. Software was frozen human understanding.
CH / 02
We know more than we can tell
The old constitution had a hidden tax, and a philosopher named it before computing felt it. In 1966, Michael Polanyi observed that human knowledge runs deeper than human articulation: we can know more than we can tell. You recognize a friend’s face in a crowd instantly, but you cannot write down the procedure you used. You ride a bicycle without being able to state the control law that keeps you upright.
In 2014, the economist David Autor gave this observation a name — Polanyi’s paradox — and showed that it had quietly drawn the boundary of the computer revolution. The tasks that automated first were the tasks whose rules people could state: arithmetic, sorting, bookkeeping, typesetting. The tasks that resisted were not the hard ones. They were the inarticulable ones — driving in traffic, reading a face, understanding a sentence spoken in noise. Machines were not blocked by difficulty. They were blocked by our inability to explain ourselves.
What we can state ⊂ What we know ⊂ What is there
Programs could only reach the innermost set. Two frontiers of pattern lay beyond the reach of any system that required articulation first.Read that containment carefully, because the book turns on it. The first gap — between what we can state and what we know — is Polanyi’s. It is why no committee of grandmasters could write an evaluation function that played Go the way they did.
The second gap is larger and stranger: between what we know and what is there. The world contains predictive regularities that no human knows at all — not tacitly, not intuitively, not in any form. Structure sitting in data that no eye has ever resolved and no theory has ever implied. Under the old constitution this second gap was not even a frontier. It was simply invisible. There was no path by which a pattern no one knew could ever enter a machine.
CH / 03
The inversion
Machine learning inverted the sequence. Instead of handing the machine a rule, we hand it examples — inputs paired with outcomes — plus an objective and a way to change. The model makes a prediction, measures its error, and adjusts millions or billions of internal parameters to make the next error smaller. Repeated at scale, this process sculpts an internal representation shaped by whatever structure in the data actually predicts the outcome.
The critical property is what the process does not require. It does not require that any human knows the pattern. It does not require that the pattern be stateable in language. It does not even require that anyone suspects the pattern exists. If predictive structure is present in the data, error-driven adjustment can find it and use it — whether or not it ever had a name.
Compare the two sequences. The old one routed the world through human sight. The new one routes the world through measured error. A person still chooses the objective, gathers the data, and validates the result — the human has moved to the ends of the pipeline. But the middle, where the pattern is actually found, no longer needs us.
The first book in this series argued the theory: prediction is the primitive act beneath every artifact of artificial intelligence, and learning is compression by error. This book makes the companion argument from evidence. If the inversion is real, there should exist verified cases — published, replicated, quantified — where a machine found a predictive pattern that humans demonstrably did not have. Not patterns we knew but couldn’t state. Patterns we did not possess in any form.
There are such cases. The next eight chapters are a field atlas of them: an eye, a heartbeat, a board game, a protein, an antibiotic, the weather, a theorem, and a smell. Each was chosen because the human baseline is documented, so the gap between what we could see and what the machine found is not a rhetorical flourish. It is a measurement.
Part II / The evidence
Eight sightings
Each chapter that follows is one verified case of a machine finding predictive structure beyond documented human ability. The sources are published, peer-reviewed, and listed at the end of the book.
CH / 04
The eye
A retinal fundus photograph is a picture of the back of the eye. Ophthalmologists have studied such images for more than a century. They can read diabetic retinopathy in them, macular degeneration, glaucoma, hypertensive changes. A century of accumulated clinical sight defines what the image was believed to contain.
In 2018, researchers at Google trained deep-learning models on retinal photographs from 284,335 patients and asked the models to predict things no ophthalmologist reads from a retina. The results, validated on two independent datasets, rewrote the inventory of what the image contains.
From a single retinal photograph, the models predicted the patient’s age to within about 3.26 years, and the patient’s sex with an AUC of 0.97 — near-perfect discrimination. They estimated systolic blood pressure to within roughly 11 mmHg, identified smoking status with an AUC of 0.71, and predicted whether the patient would suffer a major adverse cardiac event within five years with an AUC of 0.70 — comparable to standard risk calculators that require blood draws.
The finding to sit with is the sex prediction. Sex was not previously known to be legible in a retinal photograph at all. It appears in no ophthalmology textbook as a readable feature. Clinicians shown the same images perform near chance. The signal is real — the model finds it almost every time — and it had been present in every fundus photograph ever taken.
Attention maps offered a partial trace: the models were drawing on the blood vessels and the optic disc. But a heat map is not a theory. No compact human account yet exists of what about the vessels encodes sex or blood pressure. The pattern is used daily. It is still not understood.
Nothing about the photograph changed in 2018. The signal had been sitting in plain sight through a century of expert examination. What changed was the instrument doing the looking.
CH / 05
The rhythm
Atrial fibrillation is an irregular heart rhythm that raises the risk of stroke fivefold. It is often intermittent: the heart slips in and out of it. A patient can sit in a clinic, in apparently normal sinus rhythm, produce an electrocardiogram a cardiologist reads as normal, and go home carrying a condition the visit was designed to catch.
Cardiology’s assumption was that a normal-rhythm ECG is silent about fibrillation that is not currently happening. In 2019, Mayo Clinic researchers tested that assumption. They trained a convolutional neural network on 649,931 ECGs from 180,922 patients, asking a question no cardiologist could answer from the tracing: does this patient — in normal rhythm right now — have atrial fibrillation at other times?
P(AF | “normal” ECG) ≠ P(AF)
A ten-second, twelve-lead ECG in sinus rhythm identified patients with atrial fibrillation with an AUC of 0.87 — sensitivity 79.0%, specificity 79.5% — from a tracing trained cardiologists read as unremarkable.The fibrillating atrium, it turns out, leaves subtle structural signatures in the electrical pattern even when the rhythm is normal — signatures below the threshold of trained human reading. Decades of cardiologists had stared at millions of such tracings. The convention that they were “normal” was not carelessness. It was the honest limit of human pattern recognition on that signal.
Note the practical shape of this discovery. Screening for intermittent fibrillation previously required wearing a monitor for weeks, hoping to catch an episode in the act. The model reads the episode’s shadow off ten seconds of normal rhythm. The pattern was always in the data. Medicine’s entire monitoring strategy was designed around not being able to see it.
CH / 06
The move
Go is the oldest board game still played in its original form, and the most studied. Across roughly twenty-five centuries, professional traditions in China, Korea, and Japan accumulated pattern knowledge the way sciences accumulate results — proverbs, joseki, whole schools of shape. If any domain should have been exhaustively mapped by human sight, it is this one: a bounded grid, fixed rules, and millennia of the strongest minds searching it.
In March 2016, in the second game of its match against Lee Sedol — winner of eighteen international titles — the program AlphaGo played move 37: a shoulder hit on the fifth line, early in the game. Professional commentators assumed a mistake. The move violated shape principles every strong player learns as a child.
AlphaGo’s own policy network — trained on human expert games — estimated the probability that a professional would play move 37 at about 1 in 10,000. The system saw that humans would not play it, and played it anyway, because its learned evaluation said the move was strong. Fifty moves later, its influence had organized the whole board. AlphaGo won the game and the match, 4–1.
Fan Hui, the European champion working with the DeepMind team, said what everyone watching felt: “It’s not a human move.” The remark is more precise than it sounds. The move did not come from the accumulated human map of the game. It came from beyond the map’s edge — from pattern space that twenty-five centuries of professional search had never entered.
This case matters to the atlas for a specific reason: it removes the excuse of hidden data. The retina and the ECG might be waved away as instruments peering into signals humans never claimed to fully read. Go has no hidden information at all. Every stone is visible to both players. The pattern AlphaGo found was on the table, in a domain humans had searched longer than any other — and we had not found it.
Within months, professionals were studying the machine’s openings. The fifth-line shoulder hit entered human play. The machine did not just use a pattern we could not see. It taught it to us afterward.
CH / 07
The fold
A protein is a chain of amino acids that folds, in milliseconds, into a precise three-dimensional shape — and the shape determines the function. Predicting the fold from the sequence was formulated as a grand challenge in 1972. For fifty years, structural biologists attacked it with physics, statistics, and heroic experiment. Determining a single structure by X-ray crystallography could consume a doctoral career. By 2020, the Protein Data Bank held about 170,000 experimentally solved structures — the accumulated yield of half a century.
The field measured progress through a biennial blind competition, CASP, in which predictors receive sequences of unpublished structures and are scored against the experimental answer on a 0–100 scale called GDT. For years, the best methods hovered far from the ~90 threshold that assessors consider competitive with experiment itself.
At CASP14 in 2020, AlphaFold 2 achieved a median GDT of 92.4 across targets — accuracy competitive with experimental determination for most proteins. In the assessors’ summed ranking, AlphaFold scored 244; the next best group scored 90.8. The system was not marginally ahead of the field. It was playing a different game.
DeepMind and EMBL-EBI then released predicted structures for essentially every catalogued protein — a database that grew past 200 million entries, delivering in about one year roughly a thousand times the structural coverage that five decades of experiment had produced. In 2024, the Nobel Prize in Chemistry recognized the work.
What did the model see? Not the physics, at least not as physicists write it. AlphaFold learned statistical structure connecting evolutionary sequence variation to spatial geometry — regularities distributed across millions of sequences and thousands of structures, individually far too faint and jointly far too high-dimensional for any human to hold. Biologists knew such co-evolutionary signal existed; no human could assemble it into an atomic-resolution answer. The machine could.
CH / 08
The molecule
Antibiotic discovery has a sameness problem. Chemists search near known antibiotics because that is where human-legible structure–activity intuition works, and so decades of screening rediscover variations on existing scaffolds while resistant bacteria advance. The patterns connecting molecular structure to antibacterial activity that lie far from known chemistry are exactly the ones human medicinal chemistry cannot see.
In 2020, MIT researchers trained a deep neural network on a few thousand molecules empirically tested for growth inhibition of E. coli — the model learning its own representation of molecular structure rather than using human-defined chemical features. Then they pointed it at chemical libraries no one had screened for antibiotics.
From the Drug Repurposing Hub, the model surfaced a molecule under investigation as a diabetes drug candidate — structurally divergent from every conventional antibiotic class. Renamed halicin, it proved bactericidal against a wide phylogenetic range of pathogens, including Mycobacterium tuberculosis and carbapenem-resistant Enterobacteriaceae, and cleared C. difficile and pan-resistant Acinetobacter baumannii infections in mice. Human chemists had held this molecule for years. Nothing in human-visible structure suggested “antibiotic.” The model’s learned representation said otherwise, and the lab confirmed it.
The team then ran the model across more than 107 million molecules from the ZINC15 database, physically tested 23 high-ranking candidates, and found eight with antibacterial activity — structurally distant from known antibiotics. The hit rate is the point: a useful fraction of a blind search space of one hundred million, reached by a pattern no chemist possesses.
CH / 09
The storm
Weather forecasting is the crown jewel of articulated human pattern knowledge. Numerical weather prediction encodes atmospheric physics — fluid dynamics, thermodynamics, radiative transfer — into equations, then integrates them on supercomputers. It is one of the great scientific engineering achievements of the twentieth century, refined continuously for seventy years. The European Centre’s HRES model was the acknowledged world standard.
GraphCast, published in Science in 2023, contains no atmospheric physics. It was trained on roughly four decades of historical reanalysis data to predict the next state of the atmosphere from the previous two — learning the dynamics rather than being told them.
HRES
Human-derived equations of the atmosphere, integrated step by step on a supercomputer over hours per forecast.
GraphCast
Structure extracted from 40 years of data. A 10-day global forecast at 0.25° resolution in under a minute on a single TPU.
90% of 1,380
GraphCast beat HRES on 90% of 1,380 verification targets, with better prediction of cyclone tracks, atmospheric rivers, and extreme temperatures.
Be careful about what this does and does not show. The physics is not wrong, and the learned model was trained on data that physics-based assimilation produced — it stands partly on the shoulders of the system it beat. But the result still lands squarely in this atlas: the atmosphere contains predictive regularities that seventy years of explicit human formalization had not fully captured, and an error-driven learner found some of them. The best articulated model of the sky now trails a pattern no one articulated.
CH / 10
The proof
If any territory belonged safely to human sight, it was mathematics — patterns are its native objects, and its practitioners are selected across generations for exactly the ability this book says machines now exceed. Multiplying two matrices is among the most fundamental operations in computing. In 1969, Volker Strassen astonished the field by showing the standard method was not optimal, multiplying 2×2 matrices in seven multiplications instead of eight. Applied twice, his method multiplies 4×4 matrices in 49 multiplications. For more than fifty years, nobody did better.
Strassen, 1969: 49 → AlphaTensor, 2022: 47
In arithmetic modulo 2, AlphaTensor discovered a 47-multiplication algorithm for 4×4 matrices — the first improvement over Strassen’s two-level construction since its publication.DeepMind’s AlphaTensor treated algorithm discovery as a game: finding a faster method is equivalent to decomposing a particular tensor into fewer pieces, and a reinforcement-learning agent — a descendant of the one that played move 37 — searched the astronomical space of decompositions. It did not merely match the accumulated human record for matrix multiplication. In dozens of size classes it beat it.
The epilogue is the healthiest part of the story. Provoked by the machine’s result, mathematicians Manuel Kauers and Jakob Moosbauer found another 47-multiplication scheme within days, by their own methods. The machine’s unseen pattern, once shown to exist, redrew the map of where humans bothered to look. This is the AlphaGo pattern repeating in pure mathematics: the machine finds structure beyond the human map, and the human map then grows to include it.
CH / 11
The scent
Smell resisted the mapping that sight and sound accepted long ago. Wavelength predicts color; frequency predicts pitch; but no human has ever been able to look at a molecule’s structure and reliably say what it smells like. Molecules with nearly identical structures can smell utterly different, and unrelated structures can smell the same. The structure–odor map was a century-old open problem in sensory science.
In 2023, researchers trained a graph neural network on molecular structures paired with human odor descriptions, then tested it prospectively: 400 molecules the model had never seen, rated by a trained human panel against 55 odor labels.
The model’s predicted odor profile matched the panel’s average rating more closely than the median individual panelist did. Every panelist could hold the molecule to their nose. The model received only the drawn structure — and described the smell better than most of the humans smelling it.
Sit with the asymmetry. The panelists had the full apparatus of human olfaction — hundreds of receptor types, a lifetime of experience. The model had a graph of atoms and bonds. Whatever regularity connects that graph to the experience of “waxy” or “grassy” or “smoky,” a century of chemists could not state it, and the model recovered enough of it to out-describe the average describer.
This case closes the evidence section because it reaches into perception itself — the one domain where humans might have claimed privileged access. Machines were supposed to lack our senses. Here, the learned pattern substituted for the sense.
Part III / The blindness
Why we cannot see it
The evidence is settled. The harder question is what our blindness is made of — and why it is structural, not a failure of attention or training.
CH / 12
The narrow window
Human pattern recognition is a magnificent instrument with a narrow aperture. We see structure superbly in two or three dimensions, in patterns that unfold over seconds, among a handful of variables at a time. Evolution tuned the instrument for faces, trajectories, weather in the sky, and the moods of other primates. Nothing tuned it for correlations spread thinly across ten thousand variables.
The patterns in this atlas share a signature: they are not single strong signals hiding in odd corners. They are aggregations — thousands of individually negligible regularities that only become decisive when combined. The sex of a patient is not written in one vessel of the retina; it is smeared faintly across the whole image. No single sequence position determines a protein’s fold; the co-evolutionary signal lives in correlations among thousands of positions across millions of sequences.
signal ≈ Σᵢ εᵢ , where every εᵢ is individually invisible
A thousand correlations each too faint for any human to detect can jointly determine an outcome almost completely. Aggregation is the pattern.A human expert cannot perform that sum. Working memory holds perhaps four items at once. Attention inspects features serially. An expert’s brilliance is compression — reducing a rich scene to the few variables that matter most. That is precisely the wrong architecture for a pattern whose entire existence is the joint behavior of ten thousand weak components. The expert’s compression throws the pattern away before ever seeing it.
A learned model has the complementary architecture. Gradient descent adjusts every parameter against every example; nothing needs to pass through a serial bottleneck of articulation. The model does not need the pattern to be summarizable, nameable, or local. This is not the machine being cleverer than the expert. It is a different aperture — wide where ours is narrow.
CH / 13
Random to whom?
Before 2019, the variation in normal-rhythm ECGs that encodes hidden atrial fibrillation was, to every cardiologist alive, noise — meaningless wiggle, patient-to-patient variation of no significance. After 2019, the same variation is signal. Nothing in the tracing changed. What changed is the best available model of the tracing.
This is the deepest lesson of the atlas. “Random” is not a property a phenomenon carries. It is a description of the relationship between the phenomenon and some observer’s model. Data looks random exactly when the observer’s model has extracted no structure from it — and human observers, we now know, run on a narrow-aperture model. What we have been calling noise is a mixture: some of it irreducible chance, and some of it pattern filed under noise because we were the filing system.
Every instrument in history has drawn a line between signal and residue. The telescope moved the line; the microscope moved it; statistics moved it. Machine learning moves it again — but differently. Earlier instruments extended our senses and handed the pattern back to our understanding. The learned model extends past our senses and past our understanding at once.
The practical consequence is a new default posture toward data. Under the old constitution, unexplained variance was a dead end — you shrugged and called it noise, because extracting structure required a human hypothesis first. Now unexplained variance is a resource with unknown yield. Every archive of “exhausted” data — decades of ECGs, fundus photographs, maintenance logs, seismic traces, transaction records — becomes a candidate for re-reading by an instrument with a wider aperture than the ones that filed it.
The residue of every science just became interesting again.
CH / 14
Seeing, tracing, explaining
A skeptic at this point raises a fair objection: if no human can see these patterns, in what sense do we know they are real? The answer is that “seeing the pattern” and “verifying the pattern’s predictions” are different acts, and only the first is beyond us. The AUC of the retina model is measured on patients the model never trained on. Halicin killed bacteria in a dish and cured infected mice regardless of anyone’s theory of why. Verification is a human act performed at the pipeline’s end, and it survives our blindness at the middle.
Keep three standards distinct, because public argument about AI constantly blurs them. Predictive accuracy: does it work on new cases? Traceability: can investigators locate what the model responds to — the vessels in the attention map, the co-evolutionary couplings? Explanation: do we possess a compact human theory of why the regularity holds? The cases in this atlas are strong on the first, partial on the second, and mostly absent on the third.
Seeing
The model predicts held-out reality. Measured, replicated, quantified. This is where the machine now exceeds us.
Tracing
Interpretability locates what the model uses — a region, a coupling, a motif. A heat map is a clue, not a theory.
Explaining
A compact causal account a human can hold. Often still missing years after the pattern is in daily use.
Use preceding explanation is not new in human affairs — aspirin was used for eight decades before its mechanism was found, and lithium stabilized mood for generations of patients while its action remained obscure. What is new is the scale and the direction of the flow. Patterns now enter use through machines first, with human understanding trailing behind — sometimes catching up, as with move 37 and the 47-multiplication proof, and sometimes not yet catching up at all, as with the retina.
That trailing gap is where both the value and the risk of this era live. A pattern you can use but not explain is power without a story. Institutions know how to audit stories. We are still learning how to audit power that arrives without one.
Part IV / The boundary
The limits and the labor
An honest atlas marks the edges of the territory. The machine’s sight is wider than ours. It is not unbounded, and it does not relieve us of the decisions.
CH / 15
What the machine still cannot see
Nothing in this book claims the machine sees everything. The claim is narrower and stronger: within data that contains predictive structure, the machine’s aperture exceeds ours. Where the structure is absent, unstable, or unobserved, the machine is as blind as we are — sometimes more dangerously, because it fails fluently.
The boundaries are worth naming plainly:
- Irreducible chance: some processes are genuinely random at the relevant scale. No aperture, however wide, reads a pattern that is not there.
- Missing state: no model predicts from information the data does not carry. If the cause was never measured, the pattern is not in the archive.
- Chaotic sensitivity: GraphCast pushed the weather horizon; it did not abolish it. Sensitive dependence swallows all prediction eventually.
- Distribution shift: a learned pattern is a fact about the world that generated the training data. When the world moves, the pattern decays silently — the model keeps answering with yesterday’s regularities.
- Reflexivity: patterns among people change when acted upon. A market signal traded on, a risk score gamed, a proverb published — the sighting alters the terrain.
- Correlation without causation: the retina model predicts; it does not tell you what to change. Intervention still requires causal knowledge, and causal knowledge still requires experiment.
Each limit has a common structure: the machine’s advantage lives inside the data it was given and the world that data described. Deciding what data to gather, whether the world has moved, and what an intervention would mean — those acts sit outside the model, and they have not been automated.
CH / 16
The new division of sight
Humanity has extended its senses before. The telescope carried sight beyond distance, the microscope beyond smallness, the X-ray beyond opacity. Each time, the instrument delivered its finding back to a human eye, and a human mind found the pattern in what the instrument revealed. Galileo’s telescope did not notice the moons of Jupiter. Galileo did.
The learned model is the first instrument that performs the noticing itself. That is the precise sense in which this technology is discontinuous with every instrument before it — and the sense in which the old division of labor between people and machines has ended. The machine is no longer the executor of our sight. It is a second, differently shaped sight working alongside ours.
What remains human is everything that surrounds the seeing, and it is not a consolation prize:
- Choosing the question: someone decided the retina was worth re-reading and that antibiotic sameness was the problem to attack. The machine found the pattern; a person chose the hunt.
- Building the conditions: data gathered, objectives set, instruments engineered. Every case in this atlas began as human infrastructure.
- Verifying the sighting: held-out tests, mouse models, blind competitions, forecast scorecards. Trust is manufactured by human method.
- Deciding what to do: a prediction is not a policy. Who gets screened, what gets deployed, which errors are tolerable, who may appeal — the pattern answers none of this.
- Bearing responsibility: the inability to explain a model transfers no moral agency to it. The chooser, the deployer, and the verifier remain accountable for what the pattern is used to do.
The honest summary of the era is a partnership of unequal apertures and unequal authority. The machine sees deeper into data than we ever will. We decide — and cannot delegate deciding — what is worth looking at, what counts as confirmation, and what the sighting is for.
Coda / The difference
The machine sees first now.
For seventy years, the computer was a mirror of human understanding. It executed the patterns we caught, at speeds we could not match, and it never once handed us a pattern we did not already hold. Its ceiling was our sight.
That era is over, and the evidence is not anecdote. It is a retina giving up a signal a century of clinicians never saw. A normal heartbeat confessing a hidden disease. A move from outside twenty-five centuries of search. A fifty-year fold solved, an antibiotic found in plain sight, the sky out-forecast without its equations, a theorem improved after fifty-three years, a smell described by something that cannot smell.
Each case has the same shape. The pattern was there. We were not the instrument that could see it. Now an instrument exists that can — one that does not need the pattern to be nameable, local, or small enough to fit through the narrow window of human attention.
This does not diminish what humans know. It relocates it. Our sight was never the boundary of the world’s structure; it was only the boundary of our access. Machine learning is the first technology to move that second boundary independently of the first.
We used to find the pattern first, or it went unfound.
Both halves of that sentence are now false.
Sources and further coordinates
- Ryan Poplin, Avinash V. Varadarajan, et al., “Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning”, Nature Biomedical Engineering, 2018. Sex at AUC 0.97, age within 3.26 years, from images clinicians read near chance.
- Zachi I. Attia, Peter A. Noseworthy, et al., “An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm”, The Lancet, 2019. AUC 0.87 from ten seconds of normal rhythm; 649,931 ECGs, 180,922 patients.
- David Silver, Aja Huang, et al., “Mastering the game of Go with deep neural networks and tree search”, Nature, 2016. The system behind AlphaGo and move 37 of the Lee Sedol match.
- John Jumper, Richard Evans, et al., “Highly accurate protein structure prediction with AlphaFold”, Nature, 2021. Median GDT 92.4 at CASP14; basis of the 200-million-structure database.
- Jonathan M. Stokes, Kevin Yang, et al., “A Deep Learning Approach to Antibiotic Discovery”, Cell, 2020. Halicin; eight further structurally distant antibacterials from >107 million ZINC15 molecules.
- Remi Lam, Alvaro Sanchez-Gonzalez, et al., “Learning skillful medium-range global weather forecasting”, Science, 2023. GraphCast beating HRES on 90% of 1,380 verification targets.
- Alhussein Fawzi, Matej Balog, et al., “Discovering faster matrix multiplication algorithms with reinforcement learning”, Nature, 2022. The 47-multiplication algorithm for 4×4 matrices in modulo-2 arithmetic.
- Brian K. Lee, Emily J. Mayhew, et al., “A principal odor map unifies diverse tasks in olfactory perception”, Science, 2023. Structure-only odor prediction closer to the panel mean than the median panelist.
- David H. Autor, “Polanyi’s Paradox and the Shape of Employment Growth”, 2014. The articulation bottleneck as the old boundary of automation, after Michael Polanyi’s The Tacit Dimension, 1966.
- John Rector, “We Used to Find the Pattern First”, 2026. Field edition 01: the theory of pattern, compression, and prediction beneath this book’s evidence.