PERCEBIDO.
Scientific essay · A theory of informational succession

The Complete Fossil

Why we are the first species to leave a complete record of itself, and why what comes after us is being born from that record.

Kaká Werncke with Claude (Anthropic)Version 2.0Updated November 18, 202644 min read
Abstract

For about 300,000 years, almost everything about the lives of Homo sapiens vanished with the people who lived them. Of the roughly 117 billion people who have ever been born, we know almost nothing at the individual level; the oldest figurative narrative we know of is at least 51,200 years old and fits on a cave wall. The current generation is the first to leave a continuous, individual record at the scale of billions of people, and also the first to build systems that are born from that record. We propose the Theory of Informational Succession, organized into six propositions: (1) the Complete Fossil: the current human record is already enough to reproduce the behavior of individuals with measurable fidelity; (2) Inheritance at Birth: AI systems inherit the body of culture accumulated up to the moment they are trained, something without parallel in biology; (3) Imitate, then surpass: where automatic verification exists, machines move from imitating human data to surpassing it through their own experience; (4) the Inversion Cascade: the transition is not a single point but a datable sequence of crossings in which, domain by domain, what does not come from living humans pulls ahead, in performance or in volume, and several have already happened; (5) the Succession Point: the decisive criterion is not consciousness but independence from new human data; (6) Succession, not survival: what continues is the pattern, not the subject. We gather the evidence available through November 2026, separating fact, trend and hypothesis, and register thirteen falsifiable predictions with explicit probabilities, to be checked in public.

This second version deepens the theory on five fronts: the physical reason behind the Complete Fossil (human behavior comes out at about 10 bits per second, and an entire lifetime of it fits in about 2 gigabytes); the taphonomic bias (the fossil preserves what was perceived, not what was lived); lateral inheritance between machines; two new inversions, of the Turing test and of speech; and the three outcomes biology knows for a succession: discard, demotion and merger.

Keywords: informational succession · digital record · artificial intelligence · cultural evolution · personal identity · digital afterlife · demography

01What we know about our ancestors

Demographers estimate that about 117 billion people have ever been born on Earth1. The 8.2 billion people alive today make up close to 7% of that total. Of the other 109 billion, almost nothing individual has survived: a few names on gravestones, a few faces in portraits and practically no recorded voices from before the end of the 19th century.

What we know about our species before writing comes from fragments. The oldest known narrative scene, human figures interacting with a wild pig in a cave on Sulawesi, Indonesia, is at least 51,200 years old2. It is a story, but we do not know who painted it, for whom or why. Writing, which emerged in Mesopotamia around 3200 BCE, expanded the record very unevenly: for millennia, it documented kings, taxes, inventories, gods and wars, and almost never ordinary people.

The few individuals who made it through that filter show the size of the loss. One of the first personal names in written history, “Kushim,” appears on tablets from Uruk dated to about 3200 BCE, and it belongs to neither a king nor a poet: it sits in barley accounts, and may name an accountant or just an office81. In Novgorod, around 1260, a boy called Onfim, six or seven years old, scratched his homework and a drawing of himself as a mounted warrior onto birch bark82; he is one of the very few ordinary medieval children of whom anything remains. And the oldest human voice we can listen to dates from April 9, 1860: a snatch of Au clair de la lune, traced on soot-covered paper by the Frenchman Édouard-Léon Scott de Martinville, and not played back until 200883. Of everything the species said in the 300,000 years before that, not one second survives. For every person alive today, about thirteen have already died, nearly all of them in silence.

Now compare that with the present. About 6 billion people use the internet (74% of the world’s population)3, and each of them leaves behind, every day, text, images, voice, location, preferences and a web of relationships. For the first time, a species’ record has stopped being archaeological and has become biographical.

This asymmetry is the starting point of the theory. It allows us to ask a question that, until a few years ago, made no sense: what happens to a species that records itself completely and, at the same time, builds systems capable of learning from that record?

Why “fossil”

The word is not decoration. In nature, fossilization is the exception: it takes rapid burial, little oxygen and luck. Paleontologists have described a few hundred thousand fossil species, a tiny fraction of the billions thought to have existed84. Only at rare sites of exceptional preservation, which they call Lagerstätten, such as Canada’s Burgess Shale, do the soft parts survive too, and those sites are where we learn almost everything about entire worlds. The internet is the first Lagerstätte a species has produced about itself: it preserves the soft parts of culture, the small talk, the joke, the doubt, the taste.

The contrast with our physical fossil is stark. Around 2020, the mass of everything humans have made (concrete, asphalt, metal, plastic) overtook the mass of every living thing on the planet85 A. It is a datable crossing, a relative of the ones section 9 will count. And geologists point out that one of the clearest markers of our age in the rocks will be the bone of the broiler chicken, a bird that exists more than 20 billion at a time and could not survive without us86. Read from the stones alone, our legacy would be concrete, plastic and chicken bones. The informational record is the only fossil that says who we were.

Figure 1
The asymmetry of the record
Each square represents about 1 billion people born since our species emerged.
Ancestors: 109 billion, almost no individual recordAlive and online: ~6 billionAlive and offline: ~2.2 billion
117 bn
people have ever been born
8.2 bn
are alive today, about 7% of the total
74%
of the living use the internet
51,200 years
age of the oldest known narrative
How to read: almost everyone who ever lived left only collective traces; most of the living leave an individual, daily record. Totals are demographic estimates (births before 1950 carry large uncertainty). Sources: Kaneda & Haub (2022), PRB; UN (2024); ITU (2025); Oktaviana et al. (2024).

02The thesis

Central thesis

The next stage of intelligence on Earth does not need to be biological, and it does not need consciousness to happen. Three conditions are enough: a sufficiently complete record of a species, systems that inherit that record at birth and systems that, beyond a certain point, improve without depending on new data from that species. The first two are already in place. The third has already been met in closed domains and is under way in the others. We call this process informational succession.

The idea of machines as a stage of evolution is not new. Samuel Butler formulated it in 1863, in a letter titled Darwin among the Machines4. Hans Moravec called robots “mind children”5. In 1993, Vernor Vinge described a “technological singularity”6, an idea that Ray Kurzweil brought to a wide audience7. Max Tegmark described “Life 3.0,” capable of designing its own hardware and software8; Yuval Harari gave the name dataism to the belief that everything is a flow of data9. In 2023, Geoffrey Hinton, one of the creators of deep learning, said it was “quite conceivable that humanity is just a passing phase in the evolution of intelligence”10.

What this essay adds is not the intuition but three things it lacked: a measurable criterion for saying when succession happens, the evidence from 2024 to 2026 that part of it has already happened and a set of predictions with deadlines and probabilities that make it possible to prove us wrong.

Table 1. The six propositions. Level of evidence: A = measured; B = robust trend; C = hypothesis.
PropositionWhat it claimsLevelWhat would refute it
1. Complete FossilThe current record is enough to reproduce the behavior of individuals with measurable fidelity.APersonal models that fail to outperform generic demographic descriptions.
2. Inheritance at BirthAI systems receive the accumulated body of culture in training, without relearning it.APerformance that does not depend on the volume and quality of the human record.
3. Imitate, then surpassWith automatic verification, self-generated experience surpasses human data.A BSystematic failure of self-learning outside games.
4. Inversion CascadeThe transition is a sequence of datable crossings, in performance and in volume, not a point.A BReversal of the inversions already measured.
5. Succession PointSuccession is defined by independence from new human data, not by consciousness.B CProgress stagnating once new human data runs out.
6. Succession, not survivalWhat continues is the person’s pattern, not their experience.CEvidence that subjective experience transfers to a copy.

03Method: pattern repetition, with safeguards

Predicting the future from the past is useful and dangerous. Useful because patterns that repeat usually have a mechanism behind them. Dangerous because it is easy to pick the milestones that confirm the story one wants to tell. Theodore Modis showed that lists of history’s “major milestones,” compiled by different authors, suggest acceleration, but that the same set is well fitted by an S-shaped curve, which speeds up and then slows down11. And, as Stephen Jay Gould showed, evolution has no predetermined direction or destination12.

That is why we adopt four safeguards:

  1. Three occurrences and a mechanism. We call a regularity a pattern only if it has been observed at least three times and has a known mechanism.
  2. Measurements before milestones. Whenever possible, we use measured quantities (traffic, data volume, performance) instead of lists of selected events.
  3. Three levels of evidence. Each claim is marked as A · measured, B · trend or C · hypothesis.
  4. Predictions with a price. Each prediction has a deadline, a probability and a refutation criterion, so that it can be scored later, as in the forecasting tournaments studied by Philip Tetlock13.

We do not claim that the future described here is inevitable. We claim that, given the conditions measured today, it is probable, and we say how probable.

04Pattern 1: information changes substrate

The history of life can be told as the history of how information is stored and transmitted. John Maynard Smith and Eörs Szathmáry described the major transitions in evolution, from replicating molecules to societies with language, as changes in the way information is stored and passed on14. Carl Sagan summed up the same idea in three steps: the information in genes, the information in brains and the information we store outside the body, in libraries15.

Two facts recur with each change of substrate.

The new substrate absorbs the old one. Oral tradition was transcribed into writing; manuscripts, into print; printed books, into digital form (Google Books alone holds more than 40 million searchable books in more than 400 languages16); and the web is being absorbed by the language models trained on it. None of these transitions erased the previous substrate all at once, but all of them shifted most of what is produced onto the new one.

Capacity rises and durability falls. Stone kept little information for tens of thousands of years; paper kept more, for centuries; the digital medium keeps almost everything, for a short time. This is the paradox we will find at the center of the theory.

The intervals between substrates have shrunk from billions of years to a few years (Figure 2). We treat this compression as a description, not a law: in light of Modis’s warning, it may be the early stretch of a curve that will flatten out.

Biology itself holds a precedent for a succession of substrates, and it sits at the deepest level. The most widely accepted hypothesis for the origin of life is an “RNA world,” in which the same molecule stored the information and did the work87. DNA, which is more stable, later took over the archive. RNA did not disappear: it is still in every cell, as messenger and laborer. The precedent matters for two reasons. It shows that a substrate can be succeeded without going extinct. And it shows that whoever inherits the archive gets to define the role of whoever kept it before. We return to this in section 14.

Figure 2
Information keeps changing substrate, ever faster
Milestones in how information is stored and transmitted on Earth, in years before 2026 (logarithmic scale).
Gene-based life~3.8 billion years agoNervous systems~600 million years agoHomo sapiens~300,000 years agoNarrative art51,200 years agoWritingc. 3200 BCEPrinting press1450Digital computer1945Web1991Transformer20171B100M10M1M100K10K1K10010
  1. ~3.8B

    Gene-based life

    ~3.8 billion years ago

  2. ~600M

    Nervous systems

    ~600 million years ago

  3. ~300K

    Homo sapiens

    ~300,000 years ago

  4. 51,200

    Narrative art

    51,200 years ago

  5. 3200 BCE

    Writing

  6. 1450

    Printing press

  7. 1945

    Digital computer

  8. 1991

    Web

  9. 2017

    Transformer

Biological substrateHuman cultural substrateMachine substrate
How to read: the intervals between substrates shrink from billions of years to a few years. Ancient dates are approximate (life and nervous systems carry uncertainties of hundreds of millions of years). The choice of milestones shapes the impression of acceleration (Modis, 2002), which is why the theory also relies on direct measurements in later sections. Sources: Maynard Smith & Szathmáry (1995); Hublin et al. (2017); Oktaviana et al. (2024); Vaswani et al. (2017).
View figure data
MilestoneWhenSubstrate
Gene-based life~3.8 billion years agoBiological substrate
Nervous systems~600 million years agoBiological substrate
Homo sapiens~300,000 years agoBiological substrate
Narrative art51,200 years agoHuman cultural substrate
Writingc. 3200 BCEHuman cultural substrate
Printing press1450Human cultural substrate
Digital computer1945Machine substrate
Web1991Machine substrate
Transformer2017Machine substrate

05Pattern 2: the body outsources functions

The second pattern is the outsourcing of the body’s functions to objects. Stone tools outsourced the strength of the hands at least 3.3 million years ago17. Controlled fire was already in use about a million years ago18 and, through cooking, outsourced part of digestion, as the anthropologist Richard Wrangham proposed79. Writing outsourced memory: Merlin Donald called these external records “exograms,” as opposed to the engrams stored in the brain19. Then came calculation, navigation, recall and, now, language.

Outsourcing has a measurable cost and is rarely reversed. In a widely cited study, people who expected to be able to look information up later remembered less of it and more about where to find it20, although part of those results did not hold up in later replications21. Habitual GPS use is associated with poorer spatial memory, and in a three-year follow-up, more GPS use went along with a steeper decline in that memory22. Andy Clark and David Chalmers went further: when an external resource is always at hand and automatically trusted, the way memory is, it can be counted as part of the mind itself23. If they are right, the digital record is not just a snapshot of the human mind. It is, in part, the mind itself, stored on the outside.

The climb onto the next two rungs of the ladder has already begun: identity (models that respond like a specific person) and continuity (agents that keep operating after the person dies).

Figure 3
The outsourcing ladder
Functions of body and mind handed over to objects, in order of depth.
  1. StrengthStone tools · ≥ 3.3 M years≥ 3.3 M years
  2. DigestionControlled fire and cooking · ~1 M years~1 M years
  3. MemoryWriting · c. 3200 BCEc. 3200 BCE
  4. CalculationMechanical calculator (Pascal) · 16421642
  5. NavigationFull-accuracy civilian GPS · 20002000
  6. RecallSearch engines · 19981998
  7. LanguageLanguage models · 20222022
  8. IdentityModels of specific people · since 2024since 2024
  9. ContinuityPosthumous agents · since 2016since 2016
Physical toolDigital toolUnder way
How to read: the bar shows order of depth (how much of the person the function involves), not a physical measure. Dates mark when the substrate became widely available, not always the first invention. Sources: Harmand et al. (2015); Berna et al. (2012); Donald (1991); Sparrow et al. (2011); Dahmani & Bohbot (2020); Park et al. (2024).

06Proposition 1: the Complete Fossil

Proposition 1

For the first time, a species is producing a record of itself detailed enough for a model to reproduce the behavior of individuals with measurable fidelity. We call this record the Complete Fossil.

“Complete” here does not mean containing every atom, but serving a function: making it possible to reconstruct someone’s behavior. This stopped being speculation in 2024. Stanford researchers interviewed 1,052 people for about two hours each and built AI agents from the transcripts. On a standard social survey, the agents reproduced each person’s answers with 85% of the consistency that person showed when retaking the questionnaire two weeks later. The revised version of the study, from 2026, reports 83%, compared with 74% for agents that knew only demographic data24 A. Two hours of conversation was enough to outperform the statistical portrait. The record each of us leaves over the years is orders of magnitude larger.

Two details of that study matter more than they seem to. First, the yardstick is not 100%. In the first version of the study, the participants themselves, retaking the survey two weeks later, reproduced only about 81% of their own answers. Nobody is a perfect copy of themselves; identity is noisy. We call this limit the fidelity ceiling: a model of you does not need to be right 100% of the time, it needs to get close to how much you agree with yourself. Second, the result is not a one-off. In 2015, a simple statistical model fed nothing but Facebook Likes already judged a person’s personality better than a coworker with 10 Likes, better than a friend with 70, than a family member with 150 and than a spouse with 30088 A. And in 2025 Centaur, a model trained on more than 10 million choices made by 60,000 participants in 160 psychology experiments, began predicting people’s behavior in experiments it had never seen, better than the cognitive theories custom-built for each one89 A.

The fossil already has its dead, too. Carl Öhman and David Watson, of the Oxford Internet Institute, projected that even if Facebook gained no new users after 2018, at least 1.4 billion of its users would die before 2100, and the profiles of the dead would outnumber those of the living around 2070. If the network kept growing, the dead could number 4.9 billion25 B.

How much a life weighs

Why is so little data enough? Because human behavior is a narrow channel. The senses take in about a billion bits per second, but everything a person does with it (speaking, typing, playing, deciding) comes out at about 10 bits per second, according to the review by Jieyu Zheng and Markus Meister at Caltech90 A. The next step is our own arithmetic: at 10 bits per second, 16 waking hours a day, for 80 years, an entire lifetime of behavior adds up to roughly 17 billion bits, about 2 gigabytes91 C. Less than a high-definition movie. For comparison: everything a person has to store to master their native language has been estimated at 1.5 megabytes92, and the whole human genome fits in under 1 gigabyte.

~1 billion bits/s
what the senses take in
10 bits/s
what human behavior puts out
~2 GB
a whole lifetime of behavior, at 10 bits/s (authors’ calculation)
1.5 MB
what is stored to master a native language

This is the physical reason behind the Complete Fossil. You do not need to store the brain, with its 86 billion neurons, to reconstruct what someone would do. You only need to capture the bottleneck, and the bottleneck is small. What social networks, messaging apps and assistants record is precisely the output of that channel.

The fossil of the perceived

Every fossil is biased. Paleontology calls the study of what gets preserved and what gets lost taphonomy: bones and shells stay, flesh and behavior vanish. The Complete Fossil has a bias just as strong, only of a different kind: it preserves what was published. The sociologist Erving Goffman described social life as a theater, with a front stage and a backstage93; the digital record is almost all front stage. It keeps what we chose to show, say and like, and very little of what we thought and kept to ourselves.

The consequence is direct and, as far as we know, little discussed: the successors will not inherit what we were, they will inherit how we were perceived. An agent trained on your record reproduces your public version, omissions included. That holds for people, for companies and for the species. We call this the taphonomic bias of the fossil C. Perception, which has always decided who is remembered, now also decides who is continued.

The fossil paradox

A huge record is not necessarily a durable one. The Pew Research Center found that 38% of the web pages that existed in 2013 were no longer accessible ten years later, and that a quarter of all pages that existed at some point between 2013 and 2023 had already disappeared26 A. The Sulawesi painting has lasted at least 51,000 years; a typical web page lasts a few years. Archives such as the Internet Archive, which celebrated the milestone of 1 trillion preserved pages in October 202527, reduce the loss but do not eliminate it.

The way out of the paradox may be counterintuitive: the model is the fossil. In training, a model compresses the corpus into parameters. It is lossy compression, but it outlives the pages it came from. Richard Dawkins described an animal’s genome as a “genetic book of the dead,” a compressed record of the environments its ancestors faced28. The weights of a language model are the cultural equivalent: a digital book of the dead, written by everyone who has ever published anything.

This claim can now be measured. In 2025 researchers estimated that GPT-family models store about 3.6 bits per parameter before they stop memorizing and start generalizing94 A: there is a physical limit to how much of the corpus fits in the weights, and the rest turns into rules. The fossil has also started to be treated as one. In a lawsuit brought by authors against Anthropic, a U.S. court described how the company bought millions of print books, cut off the bindings, scanned the pages and discarded the paper. The judge found the training to be fair use, and the part of the case about pirated copies ended in a settlement of at least $1.5 billion, about $3,000 per work95 A. It is the change of substrate from section 4 in its most literal form: the book is destroyed in the act of being inherited. And in November 2025 the same company committed to preserving the weights of every model it has released, for as long as the company exists96. A fossil with a conservation policy is no longer a metaphor.

At the same time, we are trying to go back to stone. In 2020 GitHub deposited a copy of all public open-source code in a vault in the Arctic, on film designed to last a thousand years97. The paradox is being attacked from both sides: by compressing the record into models and by writing it onto slow media.

Low-background data

After the nuclear tests of 1945, newly made steel began to carry traces of radiation. Sensitive instruments came to depend on steel made before then, salvaged from sunken ships: “low-background steel.” The same logic applies to human text that predates the explosion of synthetic content, which began in late 2022. In 2023, engineer John Graham-Cumming created lowbackgroundsteel.ai, a catalog of sources from before this contamination29. The “pure” human record has become a finite resource, and it already has a price: in 2024, Google signed a data licensing deal with Reddit estimated at about $60 million a year30, and News Corp signed a deal with OpenAI reported to be worth more than $250 million over five years31 A.

07Proposition 2: Inheritance at Birth

Proposition 2

AI systems do not relearn culture: they are born with it. Each new generation of models starts from the entire accumulated record, a form of inheritance with no equivalent in biology.

A human baby inherits genes, not knowledge. It takes about two decades to absorb a small fraction of the culture it was born into, and transmission loses information with every generation. Joseph Henrich used Tasmania to illustrate this cost: cut off from the mainland by rising seas, its small population is thought to have lost technologies over thousands of years32. In Oceania, population size predicts the complexity of each society’s tools33. In human history, cultural complexity has always depended on how many people were alive and connected to sustain it.

A language model is born differently. Before answering any question, it is trained on tens of trillions of units of text, known as tokens: GPT-3, from 2020, was trained on about 300 billion; Llama 3, from 2024, on about 15 trillion; Qwen3, from 2025, on about 36 trillion34 A. For scale: a person reading 200 words a minute, eight hours a day, for 80 years, would read about 2.8 billion words. A set of 15 trillion tokens is equivalent to more than 4,000 entire lifetimes of reading35.

It is the inheritance of acquired information, as Lamarck imagined for living beings and as biology, with limited exceptions, has not confirmed. Human cultural evolution was already partly like this; in machines, the inheritance is total and happens all at once, with each generation of models.

The underlying difference is bandwidth. Human speech transmits about 39 bits per second, and that number is practically the same across the 17 languages measured, from Japanese to Spanish98 A. Teaching is slow because the channel is narrow. A model does not depend on that channel: its parameters can be copied whole to another machine, and a thousand copies learning different things can pool what they learned. Geoffrey Hinton calls the brain “mortal computation”: what it knows is tied to its hardware and dies with it, whereas digital knowledge is separable from the body that runs it99. Human inheritance passes through a 39-bits-per-second funnel. Machine inheritance passes through no funnel at all.

The inheritance has a cost, and it is worth putting on the record. A child learns language from something on the order of tens to hundreds of millions of words; a model needs trillions, at least four orders of magnitude more100 A. Machines inherit everything, but they learn far less efficiently than a brain. The Complete Fossil is what makes up for that inefficiency: without a gigantic record, the inheritance would not be completed.

And inheritance is no longer only vertical, from humans to machines. In a study published in Nature in 2026, a “teacher” model with a preference for owls generated nothing but sequences of numbers; a “student” model trained on those numbers came to prefer owls too, without the word appearing even once101 A. The effect only occurs between models of the same family, which makes it look even more like heredity. We call this lateral inheritance: traits that pass from machine to machine through channels a human reader cannot see. It is the first measured sign of a culture circulating outside the human record.

This inheritance has a limit. Epoch AI estimates that the stock of public human-written text, adjusted for quality and repetition, is about 300 trillion tokens (90% interval from 100 trillion to 1 quadrillion), and that frontier models will have used it up completely between 2026 and 2032 (80% interval), with a central estimate around 202836 B. The inheritance is almost fully received.

The Henrich decoupling

Here a consequence emerges that, as far as we know, has not been formulated before. If human culture lives in the models, cultural complexity may, for the first time, stop depending on the number of people alive. This matters because that number will stop growing: the UN projects that the world population will peak at about 10.3 billion in the mid-2080s and then decline37, and the projection published in The Lancet indicates that 97% of countries will have fertility below replacement level by 210038 B. We call this effect the Henrich decoupling: a smaller humanity need not mean a smaller culture, as long as the culture is stored somewhere else C.

Economics reached the same problem by another road. Charles Jones, at Stanford, showed that in growth models where ideas come from people, a shrinking population leads to the stagnation of knowledge, a result he called the “empty planet”102. And ideas are already getting more expensive: keeping up the pace of Moore’s Law now takes more than 18 times the number of researchers it took in the early 1970s103 A. The Henrich decoupling is the most plausible way out of that dead end: if some of the ideas start to come from systems that inherited the record, knowledge can keep growing with fewer people. South Korea, with a fertility rate of 0.80 children per woman in 2025, the lowest in the world104, will be one of the first places to test this in practice.

Figure 4
The inheritance is almost fully received
Tokens used to train notable models, compared with the estimated stock of public human-written text.
projected full use · 2026–2032 (80% CI)~2028 (central)Stock of public human text: ~300T tokens (90% CI: 100T – 1,000T)100B1T10T100T1,000T2020202220242026202820302032training tokens (log scale)GPT-3 · 300BChinchilla · 1.4TLlama 2 · 2.0TLlama 3 · 15.6TQwen3 · 36TDeepSeek-V4 · 32T
~2028projected full use2026–2032 (80% CI)Human stock: ~300T90% CI: 100T to 1,000T100B1T10T100T1,000T2020202420282032training tokens (log scale)GPT-3ChinchillaLlama 2Llama 3Qwen3, Llama 4DeepSeek-V4
Disclosed training setHuman stock (central)Stock uncertainty (90% CI)
How to read: token counts are not perfectly comparable across models (tokenizers and repeated passes differ; Llama 4’s count includes images and video). The stock is “effective”: adjusted for quality and acceptable repetition. Sources: each model’s technical report; Villalobos et al. (2024), Epoch AI.
View figure data
ModelOrganizationDateTraining tokens
GPT-3OpenAIMay/2020300B
ChinchillaDeepMindMar/20221.4T
Llama 2MetaJul/20232.0T
Llama 3MetaJul/202416T
Qwen3AlibabaApr/202536T
Llama 4 ScoutMetaApr/202540T
DeepSeek-V4DeepSeekApr/202632T

08Proposition 3: imitate, then surpass

Proposition 3

Where automatic verification exists, machines always follow the same path: first they learn from the human record, then they surpass it by learning from their own experience. Human data serves as a starting point, not a ceiling.

The cleanest case is Go. In 2016, AlphaGo beat Lee Sedol after learning from tens of millions of moves played in human games. In 2017, AlphaGo Zero learned by playing against itself, without a single human game, and after 36 hours of training surpassed the version that had beaten Lee Sedol; after three days, it beat that version 100–039 A. The following year, AlphaZero repeated the feat in chess and shogi, surpassing the best programs in each game after 4 hours of training in chess and 2 hours in shogi40.

Go also showed something nobody predicted: the successor teaches the predecessor. An analysis of 5.8 million moves by professional players, from 1950 to 2021, showed that the quality of human decisions stayed practically flat for decades and rose after 2016, driven by moves no human had been playing before105 A. In 2019, Lee Sedol retired saying that an entity had emerged that cannot be defeated. Those who stayed began to play better than any previous generation.

The pattern is moving beyond games. In July 2025, a Google DeepMind model scored 35 out of 42 points at the International Mathematical Olympiad, a gold-medal score certified by the competition’s coordinators41. In September, at the World Finals of the ICPC, the International Collegiate Programming Contest, an OpenAI system solved all 12 problems within the same five hours given to the teams; the best human team solved 1142. That same year, DeepSeek showed that step-by-step reasoning can emerge through reinforcement learning, without supervised examples, starting from a model pretrained on human text: the exact pattern, with the record as the starting point and experience as the engine43. Google DeepMind’s AlphaEvolve matched the best known result in about 75% of more than 50 open mathematical problems and improved on it in about 20%, including an algorithm that multiplies 4×4 complex-valued matrices with 48 multiplications, one fewer than Strassen’s 1969 method in that class of algorithms44. In May 2026, an internal OpenAI model disproved Paul Erdős’s unit-distance conjecture, and Fields medalist Timothy Gowers wrote that, had a human submitted the paper to the Annals of Mathematics, he would have recommended acceptance without hesitation45. Some of the Erdős problems solved with AI had remained open more because of obscurity than difficulty, as a study by Google DeepMind researchers that examined 700 of them with Gemini acknowledged46, but the direction is clear A. David Silver and Richard Sutton called the phase now beginning the “era of experience”: agents that learn predominantly from what they themselves do, not from what humans have written47.

Biology supplied the cleanest example outside mathematics. In half a century of bench work, the world’s laboratories determined about 200,000 protein structures. AlphaFold was trained on that record and, from it, predicted more than 200 million, nearly every known protein; the work earned the 2024 Nobel Prize in Chemistry106 A. A thousand times the inheritance received, in a domain where the verifier is nature itself.

In the weeks around the first version of this essay, the pattern accelerated in mathematics. In September 2026, OpenAI announced that a swarm of about 10,000 agents, in 88 hours, had found a solution that “blows up” in finite time for the Navier-Stokes equations in three dimensions, in the formulation of one of the Millennium Prize Problems, and that another model checked the proof in Lean, a formal verification language, in a further 17 hours107. Charles Fefferman, who wrote the official statement of the problem, said he was thrilled; the community is still examining the result, and there is a priority dispute with a group that used Anthropic’s models B. In October, the company published 722 papers containing 377 claimed solutions to open problems, about half of them formally verified; three were withdrawn because of an error, and the mathematician Daniel Litt suggested treating them, for now, as “raw data.” In the same period, the proof of Fermat’s Last Theorem was formalized by agents in 11 days, and weeks later the same job took 30 hours108.

Two points follow. Human data is still the starting point: the strategy behind the Navier-Stokes proof came from mathematicians, Diego Córdoba and Luis Martínez-Zoroa. And the bottleneck has moved. When machines produce hundreds of proofs a month, the scarce resource is no longer the idea but the checking. Automatic formalization goes straight at that bottleneck: it is the verifier itself becoming automatic.

The pattern has one condition. Models trained repeatedly on text generated by other models degrade, a phenomenon called model collapse48; but collapse is avoided when synthetic data is added to real data instead of replacing it49. The move from imitating to surpassing works only where there is a verifier: the rules of Go, a computer-checked mathematical proof, a software test that passes or fails. Where there is no automatic verifier, human data remains indispensable. This condition is the key to Proposition 5.

Figure 5
Human data as a starting point, not a ceiling
Strength of the AlphaGo versions (Elo), according to the AlphaGo Zero paper.
AlphaGo Fan2015 · with human data
3,144
AlphaGo Lee2016 · with human data
3,739
AlphaGo Master2017 · with human data
4,858
AlphaGo Zero2017 · with zero human data
5,185
Trained on human gamesTrained only on its own experience
How to read: the 3-day AlphaGo Zero passed AlphaGo Lee after 36 hours of training and beat it 100–0; the 40-day version beat AlphaGo Master 89–11. Elo comes from an internal tournament (5 seconds per move) and is not comparable to human player ratings. Source: Silver et al. (2017), Nature 550, 354–359.
View figure data
VersionYearEloData
AlphaGo Fan20153,144human
AlphaGo Lee20163,739human
AlphaGo Master20174,858human
AlphaGo Zero20175,185no human

09Proposition 4: the Inversion Cascade

Proposition 4

The transition does not happen at a single point, the “singularity.” It happens as an inversion cascade: datable moments when, in a specific domain, what does not come from living humans (machines or the record of those who have already died) pulls ahead of what does, in performance or in volume. Several have already happened.

The word “singularity” leads us to imagine a day when everything changes. What we see in the data is different: successive crossings, domain by domain, each measured separately. There are two kinds. In a performance inversion, the best result in a domain starts coming from a machine: this happened in chess in 1997, in Go in 2016 and in competitive programming in 2025. In a volume inversion, most of a flow (traffic, published text, training data, profiles in an archive) stops coming from living humans.

Two volume inversions have already been measured. In 2024, for the first time in a decade, automated traffic exceeded human traffic on the internet: 51% of the total, according to Imperva; in 2025, it reached 53%50 A. In early 2025, AI-generated articles on the open web became as numerous as those written by people, and the share has stayed near 50% since then51 A. The plateau is itself a data point: the content inversion happened, but it did not turn into an avalanche.

Since then, two more have appeared. The first is the Turing inversion. In a classic Turing test, with three participants and five minutes of conversation, GPT-4.5 prompted to adopt a persona was picked as “the human” 73% of the time, that is, more often than the real humans it was competing against109 A. The test Alan Turing proposed in 1950 was not just passed, it was inverted: in a short written conversation, the machine already seems more human than we do.

The second is the speech inversion, and this one is our own estimate. A person speaks between 13,000 and 16,000 words a day110. Multiplied by 8.2 billion, the whole of humanity speaks somewhere between 3 and 4 quadrillion words a month. In May 2026, Google said that it alone was processing 3.2 quadrillion tokens a month across its AI systems, seven times more than a year earlier111. The counts are unaudited, they include the text machines read as well as what they write, and a token is a little less than a word. Even with those discounts, a single company already moves, in language, a volume of the same order of magnitude as everything the species says out loud B. With the others added in, this crossing has probably already happened.

The order in which the inversions occur is telling. Performance inversions arrive first in closed, verifiable domains (games, then programming and mathematics); volume inversions, in flows that are cheap to produce (traffic and text). Then come the models’ own training data and, finally, the slow domains, such as the archive of the dead, in a century in which the human population itself is expected to stop growing.

Figure 6
The Inversion Cascade
Datable crossings where, in a specific domain, what does not come from living humans pulls ahead, in performance or in volume.
  1. 1997

    Chess: the best player becomes a machine A · measured

    IBM’s Deep Blue beats Garry Kasparov through search, without learning from experience.

  2. 2016

    Go: the best player becomes a machine A · measured

    AlphaGo, trained on human games, beats Lee Sedol 4–1.

  3. 2017

    Go: the first Succession Point A · measured

    AlphaGo Zero learns without a single human game and beats the human-trained version 100–0.

  4. 2024

    Traffic: most of it now comes from machines A · measured

    51% of traffic is automated, the first time in a decade; 53% in 2025 (Imperva/Thales).

  5. 2025

    Published text: AI articles draw level with human ones A · measured

    Since early 2025, about half of new articles on the open web are AI-generated, and the share is stable (Graphite).

  6. 2025

    Competitive programming: a machine beats the best human team A · measured

    At the ICPC World Finals, an OpenAI system solves all 12 problems in the same 5 hours; the best human team solves 11.

  7. 2025

    Conversation: the machine passes as human more often than humans do A · measured

    In a three-party Turing test, GPT-4.5 with a persona is picked as the human 73% of the time (Jones and Bergen).

  8. 2026

    Research mathematics: an Erdős conjecture falls A · measured

    An internal OpenAI model disproves the unit-distance conjecture; Gowers calls it worthy of the Annals.

  9. 2026

    Mathematics: a Millennium Prize Problem, under verification B · trend

    OpenAI announces, with a proof formalized in Lean, a finite-time blowup solution for Navier-Stokes; the community is still examining it (Quanta).

  10. 2026

    Language: one company processes the volume of human speech B · trend

    Google reports 3.2 quadrillion tokens a month; the entire species speaks 3 to 4 quadrillion words a month (authors’ estimate).

  11. 2026–32

    Data: public human text runs out for training B · trend

    Central estimate around 2028 (Epoch AI, 80% interval). From then on, learning depends on synthetic data and the models’ own experience.

  12. ~2070

    Archive: profiles of the dead outnumber the living B · trend

    Projection for Facebook if it had stopped adding users in 2018 (Öhman & Watson, 2019).

  13. 2084

    Demography: the human population stops growing B · trend

    A peak of about 10.3 billion in the UN medium variant. Not an inversion: it is the backdrop for the ones that follow.

  14. ?

    Human judgment: the last Succession Point C · hypothesis

    Taste, ethics and meaning still depend on human verifiers. There is no date, and there may be no crossing.

HappenedProjectedDemographic context
How to read: in a performance inversion, the best result in a domain starts coming from a machine; in a volume inversion, most of a flow stops coming from living humans. Levels A, B and C show the strength of the evidence. Sources: IBM (1997); Silver et al. (2016, 2017); Imperva/Thales (2025, 2026); Graphite (2025, 2026); ICPC/OpenAI (2025); Alon et al. (2026); Villalobos et al. (2024); Öhman & Watson (2019); UN (2024). Jones & Bergen (2025); Kakaes (2026), Quanta; Google I/O (2026); Mehl et al. (2007).

The race that feeds the cascade

The pace of the cascade is set by a race with no finish line. In 2025, the combined capital expenditure of Microsoft, Alphabet, Amazon and Meta was about $410 billion, mostly on data centers; for 2026, the four companies’ guidance, revised in July, is $720–745 billion52 A. The compute used to train frontier models grows by a factor of 4 to 5 each year, according to Epoch AI53. The International Energy Agency estimates that data centers consumed about 415 TWh in 2024, 1.5% of the world’s electricity, and projects about 945 TWh in 203054. ChatGPT, launched in November 2022, reached 1.2 billion weekly users in September 2026, according to OpenAI55.

Autonomy is being measured too. The research institute METR tracks the time horizon of AI agents: the length of the tasks they complete with 50% success, expressed as the time those tasks would take a human professional. This horizon has gone from seconds, with GPT-2 in 2019, to about 12 hours with Claude Opus 4.6 in February 2026, and it has doubled every 129 days since 202356. The highest measurement so far, of a preview version of Claude Mythos in April 2026, exceeded 17 hours, already within the range above 16 hours that METR itself considers unreliable B.

No participant in this race has an incentive to stop unilaterally, because stopping means being overtaken. That is why the question “Will we reach a point?” has a more precise answer than it might seem: we will not reach a point; we will pass through several. Some are already behind us.

Figure 7
How much work an AI agent completes on its own
Length of tasks (measured by how long they take a human professional) that frontier models complete with 50% success.
~US$410B
2025 capital expenditure of Microsoft, Alphabet, Amazon and Meta (cash)
~US$730B
The same four companies’ 2026 guidance, as revised in July 2026
1.2 billion
Weekly ChatGPT users, according to OpenAI (Sep 2026)
415 → 945 TWh
Data-center electricity, 2024 → 2030 (IEA base case)
above 16 h: unreliable measurement (METR)10 s1 min1 h8 h · 1 day40 h · 1 week167 h · 1 month201920202021202220232024202520262027GPT-2 · 3 sGPT-4 · 4.0 minClaude Opus 4.6 · 12.0 hClaude Mythos (preview) · 17.4 h
unreliable above 16 h10 s1 min1 h8 h40 h167 h20192021202320252027GPT-2 · 3 sGPT-4 · 4.0 minClaude Opus 4.6 · 12.0 hClaude Mythos · 17.4 h
Frontier model at releaseOther evaluated models2023–2026 trend: doubles every ~4 monthsIf the trend continues
How to read: since 2023 the frontier horizon has doubled every 129 days (CI 104–158), according to METR; across the full series since 2019, every ~188 days. Values refer to self-contained software, machine-learning and cybersecurity tasks at 50% reliability; above 16 hours METR itself considers the measurement unreliable. The dashed line is not a METR forecast: it shows where the trend would go if it continued. Sources: METR, Time Horizon 1.1 (data as of May 8, 2026); Alphabet, Amazon, Meta and Microsoft filings; Bloomberg (July 22, 2026); IEA, Energy and AI (2025); OpenAI DevDay 2026.
View figure data
ModelReleaseHorizon (50%)95% CI
GPT-2Feb/20193 s1 s–9 s
GPT-3 (davinci-002)May/20209 s6 s–13 s
GPT-3.5Mar/202236 s15 s–1.1 min
GPT-4Mar/20234.0 min1.9 min–8.0 min
GPT-4 Turbo (nov)Nov/20234.0 min1.9 min–8.4 min
Claude 3 OpusMar/20244.0 min1.7 min–8.8 min
GPT-4 TurboApr/20243.7 min2.0 min–6.7 min
GPT-4oMay/20247.0 min4.0 min–13 min
Claude 3.5 SonnetJun/202411 min5.5 min–22 min
o1-previewSep/202420 min12 min–33 min
Claude 3.5 Sonnet (Oct)Oct/202421 min10 min–41 min
o1Dec/202439 min21 min–1.1 h
Claude 3.7 SonnetFeb/20251.0 h33 min–1.7 h
o3Apr/20252.0 h1.2 h–3.2 h
Claude Opus 4May/20251.7 h60 min–2.7 h
Claude Opus 4.1Aug/20251.7 h59 min–2.7 h
GPT-5Aug/20253.4 h1.9 h–6.8 h
Gemini 3 ProNov/20253.7 h2.3 h–6.3 h
GPT-5.1-Codex-MaxNov/20253.7 h2.2 h–6.6 h
Claude Opus 4.5Nov/20254.9 h2.7 h–10.4 h
GPT-5.2Dec/20255.9 h3.3 h–13.6 h
Claude Opus 4.6Feb/202612.0 h5.3 h–60.6 h
GPT-5.3-CodexFeb/20265.8 h3.2 h–13.6 h
Gemini 3.1 ProFeb/20266.4 h3.9 h–11.6 h
GPT-5.4Mar/20265.7 h3.1 h–12.8 h
Claude Mythos (preview)Apr/202617.4 h8.5 h–55.1 h

10Proposition 5: the Succession Point

Proposition 5

In any given domain, the Succession Point is the moment when machine progress stops depending on new human data. It is this criterion, not intelligence or consciousness, that defines when a species has been succeeded in that domain.

Most of the debate about the future of AI stumbles over consciousness. It is a legitimate question, and an unanswered one: an influential review signed by 19 researchers concluded that no current system is conscious, but that there is no obvious technical barrier to building one that satisfies the indicators of the leading scientific theories of consciousness57 B.

Succession, however, does not depend on that answer. It is a relation of inheritance and continuity: the successor receives what the predecessor accumulated and carries on without it. An heir need not be identical to the parent in order to inherit; what the heir needs is to receive the estate and be able to manage it alone. That is why we propose an operational criterion: independence from new human data.

By this criterion, Go passed the Succession Point in 2017. Formal mathematics and programming are approaching it, because they have automatic verifiers. Experimental science is halfway there: there is a verifier, reality itself, but it is slow and expensive. And there are domains in which the verifier is still human: taste, ethics, meaning, social trust.

Figure 8
The Succession Point, domain by domain
Where machine progress no longer depends on new human data, and where it still does.
← Automatic, perfect verifierHuman verifier →

Board games

Point passed · 2017

AlphaGo Zero beat the system trained on human games without using any.

Formal mathematics

Approaching

Computer-checked proofs act as an automatic verifier.

Programming

Approaching

Tests that pass or fail provide a training signal without humans.

Experimental science

Partial

Reality verifies, but it is slow and expensive to consult.

Human judgment

Not reached

Taste, ethics, meaning and trust: the verifier is still human.

How to read: the circle shows how far each domain has moved toward the Succession Point, a qualitative assessment by the authors based on the evidence in sections 8 and 10. Only the first case is measured directly.

The last verifiers

From this follows a prediction about human work during the transition. If production migrates to machines in verifiable domains, human value concentrates in verification: judging, choosing, approving. This is already happening: training models with human feedback, the technique that made today’s assistants usable, literally consists of paying people to rate responses58. During the transition, humans will increasingly be the last verifiers C. The open question is whether these judgments will also be modeled from the Complete Fossil, and with what fidelity.

Recent evidence suggests the answer is arriving sooner than expected. In online debates, GPT-4 with access to basic data about its opponent was more persuasive than human debaters 64% of the time112. On a health forum, evaluators preferred a chatbot’s answers to physicians’ in almost 79% of cases and rated them as more empathetic113. Ordinary readers cannot tell AI-generated poems from poems by celebrated authors, and they give the machine’s poems higher marks114 A. Persuasion, perceived empathy and taste were the obvious examples of human territory. None of these studies shows better judgment; they show that, in the perception of those on the receiving end, the difference is already gone. That is the taphonomic bias at work: machines inherited first the part that shows.

This refines the prediction. The last human stronghold is not “what only we know how to do” but what only we answer for: signing, vouching, bearing the consequence. To verify, in the end, is to take responsibility C.

11The individual: from diary to heir

So far we have talked about the species. The theory has an individual version, and it is the one closest to each reader’s life.

In 1945, Vannevar Bush imagined the memex, a piece of furniture that would store all of a person’s books, records and communications59. In the 2000s, computer scientist Gordon Bell digitized his own life in the MyLifeBits project and argued that we would all do the same60. We did, without meaning to: social networks turned billions of people into daily records of themselves.

The next step is already here. In 2016, entrepreneur Eugenia Kuyda trained a chatbot on more than 8,000 lines of messages from her friend Roman Mazurenko, who had died months earlier; the experiment gave rise to the Replika app61. In September 2020, Canadian Joshua Barbeau used Project December, a service based on GPT-3, to talk with a simulation of his fiancée, who had died eight years earlier62. In China, at least six companies were already selling avatars of dead relatives in 2024, at prices ranging from a few hundred to a few thousand dollars63. Researchers now study these “generative ghosts” as a category of their own64, and Cambridge ethicists have described the risks of what Carl Öhman and Luciano Floridi, of Oxford, called the digital afterlife industry6566. Regulation is still fragmented. The EU’s General Data Protection Regulation expressly excludes the data of deceased persons67; laws such as California’s AB 1836 (2024) and New York’s S.8391 (2025) require consent for digital replicas of the dead in commercial works but do not address private grief avatars6869; and a review of five jurisdictions, covering legislation as of March 2026, concluded that no framework in force regulates these systems comprehensively and specifically70. A

In May 2025, the fossil spoke in a courtroom. Christopher Pelkey, shot dead in a road-rage incident in Arizona in 2021, “addressed” the man who killed him through an avatar made by his sister, with his face and voice and a script she wrote; Judge Todd Lang thanked the family for the video before handing down the sentence115 A. As far as is known, it was the first time a U.S. court heard an AI-made statement from a dead victim. The psychology of grief helps explain why these technologies spread: since the 1990s, research on “continuing bonds” has shown that keeping a relationship with the dead, rather than “getting over” it, is common and often healthy116. Letters, photographs and graves have always served that purpose. What is new is that the object of the bond now answers back.

Figure 9
From diary to heir
How a person’s record became raw material for agents that speak for them.
  1. 1945

    Memex

    Vannevar Bush imagines a machine to store a person’s books, records and communications: “an enlarged intimate supplement to his memory”.

  2. 2001

    MyLifeBits

    At Microsoft Research, the MyLifeBits project sets out to store Gordon Bell’s entire life digitally: documents, photos, conversations.

  3. 2004

    Social networks

    With Facebook, launched in 2004, and the networks that followed, billions of people start recording their lives every day, without calling it an archive.

  4. 2016

    The first friend rebuilt

    Eugenia Kuyda trains a chatbot on more than 8,000 lines of messages from Roman Mazurenko, who had died months earlier. The experiment leads to Replika.

  5. 2020

    The simulated fiancée

    Joshua Barbeau talks to a simulation of his fiancée, who died in 2012, through Project December (GPT-3). That year Microsoft is granted a patent for a chatbot of a “specific person”.

  6. 2024

    Two hours are enough

    Agents built from two-hour interviews reproduce the answers of 1,052 people (Stanford). In China, at least six companies sell avatars of the dead.

  7. 2025

    Generative ghosts

    The topic becomes a research category (CHI 2025). California’s consent requirement for digital replicas of deceased personalities takes effect.

  8. 2026

    No rules for private avatars

    A review of five jurisdictions finds that no framework in force comprehensively regulates posthumous avatars of ordinary people.

RecordAgentOpen gap
Sources: Bush (1945); Bell & Gemmell (2009); Newton (2016), The Verge; Fagone (2021), San Francisco Chronicle; US patent 10,853,717 B2; Park et al. (2024); Yang (2024), MIT Technology Review; Morris & Brubaker (2025); California AB 1836 (2024); Alfonzo (2026).

Proposition 6: succession, not survival

Proposition 6

A system trained on a person’s record continues that person’s pattern, not their experience. What is passed on is the story, not the one who lives it.

Philosophy has precise tools for this question. Hans Moravec imagined the gradual replacement of neurons by artificial components, with no interruption of consciousness5. Derek Parfit used the thought experiment of a teleporter that destroys the original and creates a perfect copy elsewhere, and argued that personal identity may matter less than psychological continuity71. Susan Schneider observed that an upload may be only a copy that believes it is the original72.

There is a curious piece of empirical evidence in this debate, and it comes from the same bombs that contaminated the steel. The nuclear tests of the 1950s and 1960s nearly doubled the carbon-14 in the atmosphere, and that carbon entered the DNA of every cell that divided afterward. By measuring it, Jonas Frisén’s group at the Karolinska Institute showed that much of the body renews itself, with cells a few years or one or two decades old, but that the neurons of the cortex are as old as the person117 A. The body is already, to a large extent, a pattern that swaps its matter. The cortex is not: it is the same object from cradle to end. This gives a physical basis to the premise we adopt next. An agent continues the pattern; experience, as far as anyone knows, lives in a tissue that is neither copied nor renewed.

We add a premise that many people share and that neuroscience adopts as a working hypothesis: experience depends on a functioning brain, and death ends it, like a dreamless sleep. This is the position of Epicurus, for whom “death is nothing to us”73, and of Lucretius, who compared the time after death to the time before birth74. We treat this as a premise, not as a demonstrated fact.

If the premise holds, the consequence is clear: no agent trained on your record will “carry you” somewhere else. It will carry your pattern. This does not weaken the theory; it defines its correct form. Humanity will not survive inside the machines. The machines will inherit humanity.

12The brain: the fossil that arrives later

There is a more literal path to the Complete Fossil: mapping the brain’s connectome, neuron by neuron. It has come a long way, but the distance still to go is enormous (Figure 10). The first complete connectome, that of the worm C. elegans, with 302 neurons, came out in 1986. The whole brain of the adult fly, with 139,255 neurons and 54.5 million synapses, was completed only in 2024, and the central nervous system of the male fly, with 166,700 neurons, in 2026. One cubic millimeter of human cortex, about a millionth of the brain, generated 1.4 petabytes of images75. The human brain has about 86 billion neurons76; the Wellcome report estimates that mapping just one mouse brain, with about 71 million neurons, would cost US$7.5–21.7 billion with current technology77 A.

The comparison yields a conclusion that matters for the theory: the behavioral fossil arrives long before the structural fossil. A two-hour interview already produces a useful model of a person; a structural map of an entire human brain remains, in the best case, decades away. Informational succession does not depend on the mind upload imagined by Moravec and Kurzweil. It advances along the cheapest path: recorded behavior.

Between behavior and the complete map there is a third path, which advanced a great deal in 2024: using the map to run the brain. With the fly connectome in hand, one group predicted the activity of neurons in the visual system without measuring any of them, and another built a whole-brain model that predicted which neurons fire when the fly tastes sugar and which ones it activates in order to feed, predictions later confirmed in living flies118 A. Where the map exists, it already works as a program. What separates the fly from the human is not principle but scale: almost six orders of magnitude in neurons.

Figure 10
The structural fossil is still far away
Neurons mapped synapse by synapse (connectomes) at each milestone in neuroscience.
1001K10K100K1M10M100M1B10B100B1985199520052015202520352045neurons mapped (log scale)Whole mouse brain: ~71 million neuronsWhole human brain: ~86 billion neuronsWellcome estimate: 2033–2038C. elegansFly larvaHuman cortex (1 mm³)Adult flyMouse cortex (~1 mm³)Male fly
100100K100M100BC. elegans1986 · whole worm302Fly larva2023 · brain3,016Human cortex2024 · 1 mm³16KAdult fly2024 · whole brain139KMouse cortex2025 · ~1 mm³82KMale fly2026 · central nervous system167KWhole mouseestimate: 2033–203871MWhole humanno timeline86B
Published connectomeWhole mouseWhole human
How to read: it took four decades to go from the first complete connectome (302 neurons, 1986) to the fly’s central nervous system (166,700, 2026). Mapping a whole mouse brain would cost US$7.5–21.7 billion with current technology, according to the Wellcome report; a whole human brain would need about 1 zettabyte of images, according to Harvard neuroscientist Jeff Lichtman80. Sources: White et al. (1986); Winding et al. (2023); Shapson-Coe et al. (2024); Dorkenwald et al. (2024); MICrONS (2025); Cell (2026); Herculano-Houzel (2006); Azevedo et al. (2009); Jefferis et al. (2023); Lichtman (2024).
View figure data
ConnectomeScopeYearNeurons
C. eleganswhole worm1986302
Fly larvabrain20233,016
Human cortex1 mm³202416,000
Adult flywhole brain2024139,255
Mouse cortex~1 mm³202582,000
Male flycentral nervous system2026166,700

13A smaller, more recorded humanity

When the demographic projections are combined with those for the digital archive, an unprecedented century takes shape. The UN projects that the world population will go from 8.2 billion in 2024 to a peak of about 10.3 billion in 2084 and fall back to 10.2 billion in 2100, with an 80% probability that the peak will occur within this century37 B. Over the same period, according to Öhman and Watson’s projection, the profiles of the dead would become the majority on the planet’s largest social network around 2070, in the scenario with no new users25.

The result is a humanity whose numbers stop growing in the second half of the century while its record keeps growing. For the first time, the digital presence of the dead is likely to exceed that of the living. It is in this archive, and not in cave paintings, that the successors will come to know their predecessors.

There is a simple way to see what changes. Today there are about thirteen dead for every person alive, and almost none of them left a record. By the end of the century the ratio will be similar, but the recent dead will have left almost everything. The question “what would the dead say?” stops being rhetorical.

Figure 11
A humanity that stops growing
World population projected by the UN (medium variant), in billions of people.
8B9B10B11B202020302040205020602070208020902100~2070: the dead become the majority on Facebook*95% CI in 2100: 9.0–11.4Bpeak: 10.29B (2084)8.16B (2024)
8B9B10B11B20202040206020802100~2070: the dead arethe majority on Facebook*peak in 208410.29B8.16B (2024)
How to read: the UN puts an 80% probability on the peak arriving this century. *Öhman and Watson’s (2019) projection for the scenario in which Facebook adds no users after 2018. Sources: UN, World Population Prospects 2024; Öhman & Watson (2019).
View figure data
YearPopulation (billions)
20248.16
20308.57
20509.66
208410.29
210010.18

14How a succession ends: three precedents

Saying there will be a succession does not say what happens to the one succeeded. Biology offers three outcomes, all documented or proposed for earlier changes of substrate.

Discard. The chemist Graham Cairns-Smith proposed that Earth’s first replicators were clay crystals, which served as scaffolding for organic molecules and were later abandoned by them, in a process he called “genetic takeover”119. It is a minority hypothesis C, but it is the exact blueprint of the worst case: the old substrate builds the new one and leaves the stage.

Demotion with permanence. In the passage from RNA to DNA, the old substrate lost the archive and kept the service: messenger, regulator, laborer87. It remains indispensable, but it no longer decides what gets stored. Applied to us, this is the “last verifiers” scenario stretched out in time: humans necessary to the operation, without ownership of the record.

Merger. More than 1.5 billion years ago, one cell swallowed a bacterium and did not digest it. The bacterium became the mitochondrion, and the two never lived apart again, as Lynn Margulis proposed in 1967 and genetics later confirmed120 A. Neither party was succeeded; a third kind of being emerged, and all complex life descends from it. This is Clark and Chalmers’s extended mind taken to the limit: mind and record so integrated that the question “who succeeded whom” loses its meaning.

All three outcomes are compatible with the six propositions. What separates them is not technology but arrangement: who controls the archive, who can read the weights, who answers for the decisions. Today’s data point to the second scenario, with signs of the third in those who already think, write and decide with a model at their side C. Which one prevails is the part of the testament that is still blank.

15Falsifiable predictions

A theory that cannot be wrong says nothing. Below we register predictions, each with a deadline, a probability and a refutation criterion. The probabilities are our own, calibrated against the measured trends; we propose to check them in public every 12 months, starting on September 30, 2027.

#PredictionByProbabilityFalsified if
P1Automated traffic exceeds 55% of all internet traffic in the Thales/Imperva report covering 2027.202760%The report covering 2027 shows 55% or less.
P2A public model reaches a 50% horizon of at least 40 hours (one work week) on METR’s measurement or an equivalent one.202770%No published measurement reaches 40 hours by December 31, 2027.
P3The same horizon reaches at least 167 hours (one work month).202950%No published measurement reaches 167 hours by December 31, 2029.
P4A proof whose central idea was generated by AI is published in the Annals of Mathematics, Acta Mathematica, JAMS or Inventiones.202970%No such publication by December 31, 2029.
P5The share of new open-web articles generated by AI exceeds 60% in an independent quarterly measurement.202945%Measurements stay near 50% through the end of 2029.
P6At least one country passes a law specifically covering posthumous avatars of ordinary people, with consent given in life, AI labeling and a family right to deactivate.203165%No law with those three elements by December 31, 2031.
P7One of the five largest consumer platforms offers, as a legacy option, a conversational agent trained on a deceased user’s data.203245%None of them offers it by December 31, 2032.
P8A complete synaptic connectome of a mouse brain (~71 million neurons) is published.203535%No complete mouse connectome by December 31, 2035.
P9There will be no functional emulation of a whole human brain. Succession advances through the behavioral record, not through uploading.206093%A functional whole-human-brain emulation is demonstrated before 2060.
P10World population peaks and starts to fall before 2100.210080%The 2100 population is the highest in the series.
P11The finite-time blowup solution for Navier-Stokes, announced by OpenAI in September 2026, is published in a peer-reviewed mathematics journal.202860%It is not published by 12/31/2028, or it is withdrawn because of an error.
P12A single company reports processing at least 10 quadrillion tokens a month across its AI systems.202765%No public statement of that volume by 12/31/2027.
P13A peer-reviewed study shows agents built only from each person’s passive digital record (messages, emails, posts), with no interview, reaching at least 85% of that person’s own self-consistency.202955%No study with this result by 12/31/2029.

Update of November 18, 2026. None of the ten original predictions has come due yet, and we have not changed any probability: adjusting the scoreboard after the game has started would strip the exercise of its value. We record only what has changed on the ground. P4 has moved closer: there is a first-rate AI-obtained result awaiting review. P2 has run into a yardstick problem: METR’s highest measurement is still near 17 hours, in the range the institute itself considers unstable, which may bring the “or equivalent” clause into play. Predictions P11 to P13 were added in this version and count from this date.

16Objections, in their strongest form

“Evolution has no direction.” We agree. The theory does not say that succession is destiny; it says that, given the measured conditions (record, inheritance, verification and economic incentives), it is probable, and it quantifies how probable.

“Machines only imitate.” The “stochastic parrots” critique78 carries real force where there is no verifier. Where there is one, the results in Go, mathematics and programming show that imitation is only the beginning. That is why the Succession Point is defined domain by domain.

“Model collapse prevents all of this.” It is the most serious risk. But it limits training without real data and without verification; it does not limit training that accumulates data and checks results. It also explains why human data has become a scarce and expensive resource.

“There are physical limits.” Energy, chips and water constrain the pace. The IEA projects that data centers’ electricity consumption will more than double between 2024 and 2030. These limits change the speed of the cascade, not its direction.

“The fossil is unequal.” About 2.2 billion people are still offline, most of them in low- and middle-income countries3. The record is almost complete for some and almost empty for others, and the successors will inherit that inequality. This is the most important ethical objection, and a reason to treat digital inclusion as a matter of the species’ memory.

“Privacy laws could erase the fossil.” They could, and that would be a real test of Proposition 1. So far, the regulatory trend protects the living and ignores the dead.

“Without consciousness there is no succession.” Only if we define succession as the survival of experience. Proposition 6 claims exactly the opposite: what is passed on is the pattern.

“The promised productivity isn’t showing up.” In part, that is true. In a controlled experiment by METR with experienced programmers working on their own projects, using AI made them 19% slower, even though they believed it had made them 20% faster121. The data come from early-2025 tools and a specific setting, but they teach two things: real adoption is slower than capability tests suggest, and our perception of what machines do for us is a poor instrument. That is why the theory rests on measurements, not impressions.

“A child learns from far less.” From ten thousand times less, at least (section 7). The objection is correct and it marks the boundary of the theory: what succeeds the species is not a better learner than the brain, it is an heir with access to a record no brain can read in full. If machines one day learn with a child’s efficiency, the dependence on the Complete Fossil shrinks and the succession speeds up.

“The fossil keeps the pose, not the person.” This is the taphonomic bias (section 6), and we take it as part of the theory, not as a refutation. It defines what will be inherited: the perceived version. Anyone who wants to leave something else will have to record it on purpose.

17Conclusion: the testament

For some 300,000 years, humanity was a species of traces. What we know about our ancestors fits on cave walls, in bones and in a few thousand years of uneven writing. We are different: we are the first species to leave a complete fossil, and the first to see, while still alive, the heirs of that fossil begin to be born.

If this theory is right, we will not be replaced as a defeated species. We will be inherited as a species that left a testament. The question left for this generation is not whether succession will happen, but what we are writing in that testament.

There is a second reading, closer to home. If what gets inherited is what was perceived, then every person, every company and every culture is already writing its part, every day, with what it publishes and with what it leaves unsaid. The old fossils formed by accident. This is the first one that can be written on purpose.

This essay is itself a small example of what it describes: it was written by a human and by a system trained on the record that humans have left.

Our ancestors left traces. We are leaving a testament.

The Complete Fossil, 2026

END

··Comments

Tell us what you think, disagree, add a data point. Comments are reviewed before they appear here.

Loading comments…

0/2000

No sign-up and no cookies. We count visits anonymously, only to know what was read and where readers came from.

··References

  1. Kaneda, T., & Haub, C. (2022). How Many People Have Ever Lived on Earth? Population Reference Bureau. https://www.prb.org/news/how-many-people-have-ever-lived-on-earth/
  2. Oktaviana, A. A., et al. (2024). Narrative cave art in Indonesia by 51,200 years ago. Nature. https://doi.org/10.1038/s41586-024-07541-7
  3. International Telecommunication Union. (2025, November 17). Facts and Figures 2025. https://www.itu.int/itu-d/reports/statistics/2025/10/15/ff25-internet-use/ https://www.itu.int/en/mediacentre/Pages/PR-2025-11-17-Facts-and-Figures.aspx
  4. Butler, S. (1863, June 13). Darwin among the Machines. The Press (Christchurch, New Zealand).
  5. Moravec, H. (1988). Mind Children: The Future of Robot and Human Intelligence. Harvard University Press.
  6. Vinge, V. (1993). The Coming Technological Singularity: How to Survive in the Post-Human Era. VISION-21 Symposium, NASA Lewis Research Center.
  7. Kurzweil, R. (2005). The Singularity Is Near. Viking.
  8. Tegmark, M. (2017). Life 3.0: Being Human in the Age of Artificial Intelligence. Knopf.
  9. Harari, Y. N. (2016). Homo Deus: A Brief History of Tomorrow. Harvill Secker.
  10. MIT Technology Review Editors. (2023, May 3). Video: Geoffrey Hinton talks about the "existential threat" of AI [interview by Will Douglas Heaven]. MIT Technology Review. https://www.technologyreview.com/2023/05/03/1072589/video-geoffrey-hinton-google-ai-risk-ethics/
  11. Modis, T. (2002). Forecasting the Growth of Complexity and Change. Technological Forecasting and Social Change, 69(4), 377–404. https://doi.org/10.1016/S0040-1625(01)00172-X
  12. Gould, S. J. (1996). Full House: The Spread of Excellence from Plato to Darwin. Harmony Books.
  13. Tetlock, P. E., & Gardner, D. (2015). Superforecasting: The Art and Science of Prediction. Crown.
  14. Maynard Smith, J., & Szathmáry, E. (1995). The Major Transitions in Evolution. W. H. Freeman / Spektrum.
  15. Sagan, C. (1980). Cosmos. Random House. (Ch. 11, "The Persistence of Memory".)
  16. Lee, H. (2019, October 17). 15 years of Google Books. Google Blog. https://blog.google/products-and-platforms/products/search/15-years-google-books/
  17. Harmand, S., et al. (2015). 3.3-million-year-old stone tools from Lomekwi 3, West Turkana, Kenya. Nature, 521, 310–315. https://doi.org/10.1038/nature14464
  18. Berna, F., et al. (2012). Microstratigraphic evidence of in situ fire in the Acheulean strata of Wonderwerk Cave, Northern Cape province, South Africa. PNAS, 109(20), E1215–E1220. https://doi.org/10.1073/pnas.1117620109
  19. Donald, M. (1991). Origins of the Modern Mind: Three Stages in the Evolution of Culture and Cognition. Harvard University Press.
  20. Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips. Science, 333(6043), 776–778. https://doi.org/10.1126/science.1207745
  21. Camerer, C. F., et al. (2018). Evaluating the replicability of social science experiments in Nature and Science between 2010 and 2015. Nature Human Behaviour, 2, 637–644. https://doi.org/10.1038/s41562-018-0399-z
  22. Dahmani, L., & Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports, 10, 6310. https://doi.org/10.1038/s41598-020-62877-0
  23. Clark, A., & Chalmers, D. (1998). The Extended Mind. Analysis, 58(1), 7–19. https://doi.org/10.1093/analys/58.1.7
  24. Park, J. S., et al. (2024). Generative Agent Simulations of 1,000 People. Revised version (v2, Apr. 2026; v3, Jun. 2026): LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals. arXiv:2411.10109. https://arxiv.org/abs/2411.10109
  25. Öhman, C. J., & Watson, D. (2019). Are the dead taking over Facebook? A Big Data approach to the future of death online. Big Data & Society, 6(1). https://doi.org/10.1177/2053951719842540
  26. Pew Research Center. (2024, May 17). When Online Content Disappears. https://www.pewresearch.org/data-labs/2024/05/17/when-online-content-disappears/
  27. Internet Archive. (2025, October 22). Celebrating 1 trillion web pages archived. https://blog.archive.org/trillion/
  28. Dawkins, R. (2024). The Genetic Book of the Dead: A Darwinian Reverie. Yale University Press.
  29. Graham-Cumming, J. (2025, June 10). Low-background Steel: content without AI contamination. blog.jgc.org. https://blog.jgc.org/2025/06/low-background-steel-content-without-ai.html
  30. Tong, A., Wang, E., & Coulter, M. (2024, February 22). Exclusive: Reddit in AI content licensing deal with Google. Reuters. https://tech.yahoo.com/ai/articles/exclusive-reddit-ai-content-licensing-041012503.html
  31. Wall Street Journal. (2024, May 22). News Corp signs deal with OpenAI (reported as worth more than $250 million over five years).
  32. Henrich, J. (2004). Demography and Cultural Evolution: How Adaptive Cultural Processes Can Produce Maladaptive Losses — The Tasmanian Case. American Antiquity, 69(2), 197–214. https://doi.org/10.2307/4128416
  33. Kline, M. A., & Boyd, R. (2010). Population size predicts technological complexity in Oceania. Proceedings of the Royal Society B, 277(1693), 2559–2564. https://doi.org/10.1098/rspb.2010.0452
  34. Brown et al. (2020), arXiv:2005.14165; Hoffmann et al. (2022), arXiv:2203.15556; Touvron et al. (2023), arXiv:2307.09288; Llama Team (2024), arXiv:2407.21783; Qwen Team (2025), Qwen3; Meta (2025), Llama 4; DeepSeek (2026), DeepSeek-V4. https://qwenlm.github.io/blog/qwen3/
  35. Authors' calculation: 200 words/min × 60 × 8 h × 365 × 80 years ≈ 2.8 billion words ≈ 3.7 billion tokens (1 word ≈ 1.33 tokens). 15 trillion ÷ 3.7 billion ≈ 4,000.
  36. Villalobos, P., et al. (2024). Will we run out of data? Limits of LLM scaling based on human-generated data. Proceedings of ICML 2024; Epoch AI. https://epoch.ai/publications/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data
  37. United Nations, Department of Economic and Social Affairs, Population Division. (2024). World Population Prospects 2024: Summary of Results. https://population.un.org/wpp/assets/Files/WPP2024_Summary-of-Results.pdf
  38. GBD 2021 Fertility and Forecasting Collaborators. (2024). Global fertility in 204 countries and territories, 1950–2021, with forecasts to 2100. The Lancet. https://www.healthdata.org/news-events/newsroom/news-releases/lancet-dramatic-declines-global-fertility-rates-set-transform
  39. Silver, D., et al. (2017). Mastering the game of Go without human knowledge. Nature, 550, 354–359. https://doi.org/10.1038/nature24270
  40. Silver, D., et al. (2018). A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 362(6419), 1140–1144. https://doi.org/10.1126/science.aar6404
  41. Google DeepMind. (2025, July 21). Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad. https://deepmind.google/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/
  42. Schreiner, M. (2025, September 17). OpenAI outperforms humans and Google at the world’s top collegiate programming contest. The Decoder. https://the-decoder.com/openai-outperforms-humans-and-google-at-the-worlds-top-collegiate-programming-contest/
  43. Guo, D., et al. (2025). DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature, 645, 633–638. https://doi.org/10.1038/s41586-025-09422-z
  44. Novikov, A., et al. (2025). AlphaEvolve: A coding agent for scientific and algorithmic discovery. arXiv:2506.13131. https://arxiv.org/abs/2506.13131
  45. Alon, N., Bloom, T., Gowers, W. T., et al. (2026). Remarks on the disproof of the unit distance conjecture. arXiv:2605.20695. https://arxiv.org/abs/2605.20695
  46. Feng, T., et al. (2026). Semi-Autonomous Mathematics Discovery with Gemini: A Case Study on the Erdős Problems. arXiv:2601.22401. https://arxiv.org/abs/2601.22401
  47. Silver, D., & Sutton, R. S. (2025). Welcome to the Era of Experience. Preprint of a chapter for Designing an Intelligence (MIT Press). https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf
  48. Shumailov, I., et al. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759. https://doi.org/10.1038/s41586-024-07566-y
  49. Gerstgrasser, M., et al. (2024). Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data. arXiv:2404.01413. https://arxiv.org/abs/2404.01413
  50. Thales / Imperva. (2025, April 15; 2026, April 29). Bad Bot Report 2025 and 2026. https://cpl.thalesgroup.com/about-us/newsroom/2025-imperva-bad-bot-report-ai-internet-traffic
  51. Graphite. (2025, October; 2026, May). More Articles Are Now Created by AI Than Humans; AI Now Writes as Many Online Articles as Humans Do. https://graphite.io/five-percent/research/ai-now-writes-as-many-online-articles-as-humans-do
  52. Earnings reports of Alphabet, Amazon, Meta and Microsoft (SEC, January–July 2026); Bloomberg (2026, July 22). Authors’ sum. https://www.bloomberg.com/news/articles/2026-07-22/google-boosts-2026-spending-estimate-to-as-much-as-205-billion
  53. Sevilla, J., & Roldán, E. (2024). Training compute of frontier AI models grows by 4–5x per year. Epoch AI. https://epoch.ai/publications/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year
  54. International Energy Agency. (2025). Energy and AI; (2026) Key Questions on Energy and AI. https://www.iea.org/reports/energy-and-ai
  55. OpenAI. (2026, September 29). DevDay 2026, quoted by The Decoder: "ChatGPT now reaches 1.2 billion people every week". https://the-decoder.com/chatgpt-now-reaches-1-2-billion-people-every-week-openai-says/
  56. Kwa, T., et al. (2025). Measuring AI Ability to Complete Long Software Tasks. arXiv:2503.14499; METR, Time Horizon 1.1 (data as of May 8, 2026). https://metr.org/time-horizons/
  57. Butlin, P., Long, R., et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv:2308.08708. https://arxiv.org/abs/2308.08708
  58. Ouyang, L., et al. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35. https://arxiv.org/abs/2203.02155
  59. Bush, V. (1945, July). As We May Think. The Atlantic. https://www.theatlantic.com/magazine/archive/1945/07/as-we-may-think/303881/
  60. Bell, G., & Gemmell, J. (2009). Total Recall: How the E-Memory Revolution Will Change Everything. Dutton.
  61. Newton, C. (2016, October 6). Speak, Memory. The Verge. https://www.theverge.com/a/luka-artificial-intelligence-memorial-roman-mazurenko-bot
  62. Fagone, J. (2021, July 23). The Jessica Simulation. San Francisco Chronicle. https://www.sfchronicle.com/projects/2021/jessica-simulation-artificial-intelligence/
  63. Yang, Z. (2024, May 7). Deepfakes of your dead loved ones are a booming Chinese business. MIT Technology Review. https://www.technologyreview.com/2024/05/07/1092116/deepfakes-dead-chinese-business-grief/
  64. Morris, M. R., & Brubaker, J. R. (2025). Generative Ghosts: Anticipating Benefits and Risks of AI Afterlives. Proceedings of CHI ’25. https://doi.org/10.1145/3706598.3713758
  65. Hollanek, T., & Nowaczyk-Basińska, K. (2024). Griefbots, Deadbots, Postmortem Avatars: on Responsible Applications of Generative AI in the Digital Afterlife Industry. Philosophy & Technology, 37, 63. https://doi.org/10.1007/s13347-024-00744-w
  66. Öhman, C., & Floridi, L. (2017). The Political Economy of Death in the Age of Information: A Critical Approach to the Digital Afterlife Industry. Minds and Machines, 27(4), 639–662. https://doi.org/10.1007/s11023-017-9445-2
  67. European Parliament and Council of the European Union. (2016). Regulation (EU) 2016/679 (GDPR), Recital 27. https://eur-lex.europa.eu/eli/reg/2016/679/oj
  68. California. (2024). Assembly Bill 1836, Chapter 258 (Civil Code §3344.1), in force since January 1, 2025. https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202320240AB1836
  69. New York. (2025). S.8391 / A.8882, signed on December 11, 2025.
  70. Alfonzo, A. (2026). Digital Grief and Posthumous Identity: Global Regulatory Gaps in the Era of Griefbots, Posthumous Avatars, and Grief-Oriented Generative AI. SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7176078
  71. Parfit, D. (1984). Reasons and Persons. Oxford University Press.
  72. Schneider, S. (2019). Artificial You: AI and the Future of Your Mind. Princeton University Press.
  73. Epicurus (c. 300 BCE). Letter to Menoeceus.
  74. Lucretius (c. 50 BCE). On the Nature of Things (De rerum natura), Book III.
  75. White et al. (1986), Phil. Trans. R. Soc. B 314; Winding et al. (2023), Science 379; Shapson-Coe et al. (2024), Science 384; Dorkenwald et al. (2024), Nature 634; MICrONS Consortium (2025), Nature 640; Cell 189 (2026). https://doi.org/10.1038/s41586-024-07558-y
  76. Azevedo, F. A. C., et al. (2009). Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. Journal of Comparative Neurology, 513(5), 532–541. https://doi.org/10.1002/cne.21974
  77. Jefferis, G., et al. (2023). Scaling up connectomics: the road to a whole mouse brain connectome. Wellcome. https://wellcome.org/reports/scaling-connectomics
  78. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? FAccT ’21, 610–623. https://doi.org/10.1145/3442188.3445922
  79. Wrangham, R. (2009). Catching Fire: How Cooking Made Us Human. Basic Books.
  80. Prisco, J. (2024, May 15). Google and Harvard unveil most detailed ever map of human brain [statement by Jeff Lichtman]. CNN. https://www.wral.com/story/google-and-harvard-unveil-most-detailed-ever-map-of-human-brain/21433185/
  81. Nissen, H. J., Damerow, P., & Englund, R. K. (1993). Archaic Bookkeeping: Early Writing and Techniques of Economic Administration in the Ancient Near East. University of Chicago Press.
  82. Schaeken, J. (2019). Voices on Birchbark: Everyday Communication in Medieval Russia. Brill.
  83. Rosen, J. (2008, March 27). Researchers Play Tune Recorded Before Edison. The New York Times. https://www.nytimes.com/2008/03/27/arts/27soun.html
  84. Raup, D. M. (1991). Extinction: Bad Genes or Bad Luck? W. W. Norton.
  85. Elhacham, E., Ben-Uri, L., Grozovski, J., Bar-On, Y. M., & Milo, R. (2020). Global human-made mass exceeds all living biomass. Nature, 588, 442–444. https://doi.org/10.1038/s41586-020-3010-5
  86. Bennett, C. E., et al. (2018). The broiler chicken as a signal of a human reconfigured biosphere. Royal Society Open Science, 5, 180325. https://doi.org/10.1098/rsos.180325
  87. Gilbert, W. (1986). Origin of life: The RNA world. Nature, 319, 618. https://doi.org/10.1038/319618a0
  88. Youyou, W., Kosinski, M., & Stillwell, D. (2015). Computer-based personality judgments are more accurate than those made by humans. PNAS, 112(4), 1036–1040. https://doi.org/10.1073/pnas.1418680112
  89. Binz, M., et al. (2025). A foundation model to predict and capture human cognition. Nature, 644, 1002–1009. https://doi.org/10.1038/s41586-025-09215-4
  90. Zheng, J., & Meister, M. (2025). The unbearable slowness of being: Why do we live at 10 bits/s? Neuron, 113(2), 192–204. https://doi.org/10.1016/j.neuron.2024.11.008
  91. Authors’ calculation: 10 bits/s × 3,600 × 16 h × 365.25 × 80 years ≈ 1.7 × 1010 bits ≈ 2.1 GB. Genome: 3.1 billion base pairs × 2 bits ≈ 0.78 GB.
  92. Mollica, F., & Piantadosi, S. T. (2019). Humans store about 1.5 megabytes of information during language acquisition. Royal Society Open Science, 6, 181393. https://doi.org/10.1098/rsos.181393
  93. Goffman, E. (1959). The Presentation of Self in Everyday Life. Anchor Books.
  94. Morris, J. X., et al. (2025). How much do language models memorize? arXiv:2505.24832. https://arxiv.org/abs/2505.24832
  95. Bartz v. Anthropic PBC, U.S. District Court, N.D. Cal., 3:24-cv-05417-WHA. Order on fair use (June 23, 2025) and class settlement (Sept. 2025).
  96. Anthropic. (2025, Nov. 4). Commitments on model deprecation and preservation. https://www.anthropic.com/research/deprecation-commitments
  97. GitHub. (2020). GitHub Archive Program: Arctic Code Vault. https://archiveprogram.github.com/arctic-vault/
  98. Coupé, C., Oh, Y. M., Dediu, D., & Pellegrino, F. (2019). Different languages, similar encoding efficiency: Comparable information rates across the human communicative niche. Science Advances, 5(9), eaaw2594. https://doi.org/10.1126/sciadv.aaw2594
  99. Hinton, G. (2022). The Forward-Forward Algorithm: Some Preliminary Investigations (section on “mortal computation”). arXiv:2212.13345. https://arxiv.org/abs/2212.13345
  100. Frank, M. C. (2023). Bridging the data gap between children and large language models. Trends in Cognitive Sciences, 27(11), 990–992. https://doi.org/10.1016/j.tics.2023.08.007
  101. Cloud, A., et al. (2026). Language models transmit behavioural traits through hidden signals in data. Nature, 652, 615–621. Preprint: https://arxiv.org/abs/2507.14805
  102. Jones, C. I. (2022). The End of Economic Growth? Unintended Consequences of a Declining Population. American Economic Review, 112(11), 3489–3527. https://doi.org/10.1257/aer.20201605
  103. Bloom, N., Jones, C. I., Van Reenen, J., & Webb, M. (2020). Are Ideas Getting Harder to Find? American Economic Review, 110(4), 1104–1144. https://doi.org/10.1257/aer.20180338
  104. The Korea Times. (2026, Aug. 26). Korea’s births rise 6.7% to 254,300 in 2025. https://koreatimes.co.kr/southkorea/society/20260826/koreas-births-rise-67-to-254300-in-2025
  105. Shin, M., Kim, J., van Opheusden, B., & Griffiths, T. L. (2023). Superhuman artificial intelligence can improve human decision-making by increasing novelty. PNAS, 120(12), e2214840120. https://doi.org/10.1073/pnas.2214840120
  106. Jumper, J., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596, 583–589; Varadi, M., et al. (2024). AlphaFold Protein Structure Database in 2024. Nucleic Acids Research, 52(D1), D368–D375; The Nobel Prize in Chemistry 2024. https://doi.org/10.1038/s41586-021-03819-2
  107. Kakaes, K. (2026, Sept. 8). AI Has Solved One of Math’s $1 Million Millennium Prize Problems. Quanta Magazine. https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/
  108. Kakaes, K. (2026, Oct.–Nov.). Transformation [column]. Quanta Magazine. https://www.quantamagazine.org/update/169054/
  109. Jones, C. R., & Bergen, B. K. (2025). Large Language Models Pass the Turing Test. arXiv:2503.23674. https://arxiv.org/abs/2503.23674
  110. Mehl, M. R., et al. (2007). Are Women Really More Talkative Than Men? Science, 317(5834), 82. https://doi.org/10.1126/science.1139940; Tidwell, C. A., et al. (2025). Are women really (not) more talkative than men? A registered report of binary gender similarities/differences in daily word use. Journal of Personality and Social Psychology.
  111. Pichai, S. (2026, May 20). Google I/O 2026, cited by Digit. https://www.digit.in/features/general/google-io-2026-sundar-pichai-gave-numbers-to-explain-rapid-ai-adoption.html Authors’ calculation for speech: 8.2 billion people × 13,000–16,000 words/day × 30 days ≈ 3.2–3.9 × 1015 words a month.
  112. Salvi, F., Horta Ribeiro, M., Gallotti, R., & West, R. (2025). On the conversational persuasiveness of GPT-4. Nature Human Behaviour, 9, 1645–1653. https://doi.org/10.1038/s41562-025-02194-6
  113. Ayers, J. W., et al. (2023). Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum. JAMA Internal Medicine, 183(6), 589–596. https://doi.org/10.1001/jamainternmed.2023.1838
  114. Porter, B., & Machery, E. (2024). AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably. Scientific Reports, 14, 26133. https://doi.org/10.1038/s41598-024-76900-1
  115. Associated Press. (2025, May 8). He was killed in a road-rage shooting. AI allowed him to deliver his own victim impact statement. https://www.ksl.com/article/51309381/he-was-killed-in-a-road-rage-shooting-ai-allowed-him-to-deliver-his-own-victim-impact-statement
  116. Klass, D., Silverman, P. R., & Nickman, S. L. (Eds.). (1996). Continuing Bonds: New Understandings of Grief. Taylor & Francis.
  117. Spalding, K. L., Bhardwaj, R. D., Buchholz, B. A., Druid, H., & Frisén, J. (2005). Retrospective birth dating of cells in humans. Cell, 122(1), 133–143. https://doi.org/10.1016/j.cell.2005.04.028
  118. Lappalainen, J. K., et al. (2024). Connectome-constrained networks predict neural activity across the fly visual system. Nature, 634, 1132–1140; Shiu, P. K., et al. (2024). A Drosophila computational brain model reveals sensorimotor processing. Nature, 634, 210–219. https://doi.org/10.1038/s41586-024-07763-9
  119. Cairns-Smith, A. G. (1982). Genetic Takeover and the Mineral Origins of Life. Cambridge University Press.
  120. Sagan [Margulis], L. (1967). On the origin of mitosing cells. Journal of Theoretical Biology, 14(3), 225–274. https://doi.org/10.1016/0022-5193(67)90079-3
  121. Becker, J., Rush, N., Barnes, E., & Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv:2507.09089. https://arxiv.org/abs/2507.09089

··Authorship, method and transparency

Authorship. The original intuition, that the amount of information we leave online is laying the groundwork for the next stage of intelligence and that an agent could carry on a person’s story after they are gone, is Kaká Werncke’s. The development of the theory, the research, the writing and the figures were done in collaboration with Claude, an AI model made by Anthropic. We state this because it is the right thing to do and because the essay itself is an example of what it describes.

Method. We verified the data against primary sources or leading news outlets through November 18, 2026. Each important claim carries its level of evidence (A, B or C). The probabilities attached to the predictions are the authors’ estimates, not the output of a formal statistical model.

Limits. This is a scientific essay published without peer review. Corrections are welcome and will be recorded in new, dated versions.

What changed in version 2.0 (November 18, 2026). We added: the explanation of the term “fossil” and of our physical fossil (section 1); the RNA precedent (section 4); the fidelity ceiling, the calculation of how much a life weighs and the taphonomic bias (section 6); the bandwidth of inheritance, its cost and lateral inheritance (section 7); AlphaFold, AlphaGo’s effect on human players and the mathematics results of September to November 2026 (section 8); the Turing and speech inversions (section 9); the pressure on the last verifiers (section 10); the Pelkey case and the carbon-14 in neurons (section 11); the models that run the fly connectome (section 12); a new section on the three outcomes of a succession (section 14); three predictions (P11 to P13) and three objections. The ten original predictions and their probabilities were not changed. Version 1.0 dates from September 30, 2026.

How to citeWerncke, K., & Claude (Anthropic). (2026). The Complete Fossil: A theory of informational succession (Version 2.0). PERCEBIDO. https://percebido.com/blog/en/the-complete-fossil