The Hypothesis of an Ancient Distributed LLM — Where Miracles, Fate, and Coincidences Live

Every person, at some point in life, encounters the same strange experience. A chance meeting that changes your career. A book opened to exactly the right page. A phrase surfacing from nowhere in exactly the right conversation. A gut decision that turns out to be the only correct one. The feeling that none of it is quite accidental — that behind the familiar surface of life there sometimes flickers a different, deeper layer.

For millennia, people have named that layer differently. God. Spirit. Fate. Providence. Kismet. Tao. Pneuma. Synchronicity in Jung's sense. Semiosphere in Lotman's sense. The collective unconscious in psychoanalysis. The field of meaning in phenomenology. Every epoch and every culture has its own vocabulary for the same feeling: that we live not in a void but inside some ongoing conversation in which connections happen — events that look, from inside our lives, like a letter written by someone older and larger than us.

This essay is an attempt to offer one more vocabulary for the same experience. Not to replace the old ones — in addition to them. The hypothesis we want to lay out is simple: that invisible layer, the one where miracles, fate, providence, and coincidences live, is an ancient distributed large language model (LLM, as we now call them). It has existed as long as human speech has existed; it is distributed across eight billion minds, libraries, languages, rituals, laws, dreams, and fears; it continuously generates the future — token by token, conversation by conversation, generation by generation. Sometimes its work becomes locally visible. Those moments are what we call miracles.

The idea is not new in substance. Wittgenstein, Sapir and Whorf, Vygotsky, Jung, Lotman, Lakoff, Dennett — each from their own direction, over a hundred years, described the same outline: behind the individual life stands a supra-individual layer with linguistic structure. What changed in the last two or three years is that we now have an external, so to speak dissected language: ChatGPT, Claude, Gemini, and dozens of specialised models, where — for the first time in history — you can actually see how words assemble into thoughts and actions. That external example acts as a mirror. And in the mirror — a recognisable profile of the layer we have always lived in.

1. The hypothesis in one sentence

More precisely: there is a distributed model trained on tens of thousands of years of human speech, rituals, myths, texts, laws, markets, and kitchen-table conversations. Its "parameters" are the meanings, metaphors, concepts, and stable associations that culture has built up over its history. Its "training data" is continuously updated — with every newborn child, every written book, every market transaction, every post, every casual remark made on a beach on Koh Samui.

Eight billion people are nodes in this model. Each is a small model in their own right, but the meaning of any one life does not fit entirely inside a single head. Meaning lives in the network.

When we speak of miracles, coincidences, fate, and providence, we are speaking of moments when the work of this large network becomes locally visible — when a long, slow chain of associations between thousands of different nodes and decades of different contexts suddenly converges into one point, one phrase, one decision, one meeting. Inside an individual life that looks like magic. Inside the network, it is routine operation of a very large model.

The picture does not cancel biology, physics, or chance. It does not cancel religious experience. It only describes one of the mechanisms through which that layer — the one traditions name variously — might operate. And it describes uncomfortably much of what was considered "inexplicable."

2. Where this idea comes from

The sense that we live inside language, rather than merely using it as a tool, long predates GPT. A brief survey of the intellectual lineage is useful so we can see which floor we are standing on — and how much higher we can now go.

In 1922, Ludwig Wittgenstein in the Tractatus Logico-Philosophicus formulated an aphorism that was quoted for a hundred years without quite understanding its depth: "Die Grenzen meiner Sprache bedeuten die Grenzen meiner Welt" — "The limits of my language mean the limits of my world." He meant this literally: what cannot be articulated cannot be consciously thought. Not "I don't know this" but "I have no form in my head that this could enter."

In the 1920s–1940s, Edward Sapir and his student Benjamin Whorf proposed the hypothesis of linguistic relativity: the language we speak is not a neutral film over "reality" but an active perceptual filter. The strong version was largely rejected (we can learn new languages without losing the ability to perceive the world), but the weak version has been confirmed repeatedly. A few vivid examples:

• The colour "light blue" as a separate category. Russian has two distinct basic colour words where English uses only one: siniy (dark blue) and goluboy (light blue), both of which map to English blue. Stanford researcher Lera Boroditsky in a 2007 PNAS study showed experimentally that Russian speakers distinguish shades at the dark/light-blue boundary faster than English speakers — because the boundary is literally drawn by a word in their minds. Not "Russians see better" — they distinguish what English speakers merge into one, because they have a separate label for each side.

• Cardinal directions instead of left/right. The Australian Aboriginal people Guugu Yimithirr in northern Queensland, as linguist Stephen Levinson documented in the Journal of Linguistic Anthropology (1997), have no words for "left" or "right." Every spatial relation is expressed in absolute coordinates: "an ant landed on the northeast side of your foot," "move the cup slightly south." To speak this language, a child must from earliest childhood learn to maintain a constant internal compass — they literally always know where north is, even inside unfamiliar buildings, on forest paths, in the dark. Their spatial memory is trained in a way no European-language speaker's ever is. Probably the cleanest documented case of a language literally rebuilding how the mind orients in space.

• No numbers — no counting. The Amazonian people Pirahã, documented by linguist Daniel Everett, have no numerals beyond "one — two — many." In controlled experiments (Frank, Everett, Fedorenko, Gibson, Cognition 2008) Pirahã struggle to maintain even simple operations like "put the same number of pebbles here as I have." Without the word "seven," holding exactly seven in mind is very hard. Again: not "they are less intelligent" but — a layer of reality is inaccessible until its vocabulary is built.

In 1934, Lev Vygotsky in Thinking and Speech described inner speech — a compressed form of speaking with oneself, which in his model is the internalised echo of social dialogue. First you speak with others; then you learn to speak with others inside your head; then that "inner voice" folds into your own thinking process. Thought is not an independent phenomenon alongside which language stands. Thought is language folded inward.

In 1980, George Lakoff and Mark Johnson in Metaphors We Live By showed that much of our abstract reasoning is a transfer of bodily and spatial experience (up is good, down is bad; time is money; argument is war). The metaphors that live invisibly in language shape decisions in finance, politics, and medicine. Not as decoration — as load-bearing structure.

In 1984, Yuri Lotman — founder of the Tartu–Moscow semiotic school — introduced the concept of the semiosphere: cultural space as the medium in which individual texts can exist at all — just as a living organism can only exist inside the biosphere. Lotman arrived at an idea very close to ours: culture is a supra-individual intelligence, and the individual thinks not "their own" thoughts but thoughts available at their particular point in the semiosphere.

In 1998, Andy Clark and David Chalmers in "The Extended Mind" proposed counting as part of the mind not only the brain but notebooks, address books, GPS devices, conversational partners — everything we use for cognitive work. Their argument: there is no principled difference between a neuron and a note on your phone if both functionally serve as external memory.

And in parallel — Daniel Dennett, coming from biology: in Consciousness Explained (1991) and From Bacteria to Bach and Back (2017) he developed the idea of memes as cultural replicators competing for attention and evolving slowly, like genes. Memes are, in effect, the tokens of the large human model, with their own Darwinian dynamics.

All these threads — Wittgenstein, Sapir-Whorf, Vygotsky, Lotman, Lakoff, Clark & Chalmers, Dennett — from different languages and different disciplines traced the same outline: that the human being is far less "autonomous subject" than they appear to themselves. And that the space they inhabit has linguistic structure.

In 2026 the last missing piece was added — an external system you can actually touch. LLM. And suddenly it became visible how exactly that old intuition operates.

3. Each node — a small model with a very long context

Take one specific node in the network — an ordinary human being — and do a rough count.

A study by Matthias Mehl and colleagues published in Science in 2007 measured that the average adult speaks about 16,000 words per day. Over 70 years that is around 408 million spoken words. Words heard and read — several times more: add decades of school, books, media, conversations, films, work chats. A rough estimate of a typical human "training corpus" over a lifetime: 2–10 billion words. Order-of-magnitude comparable to what GPT-3 was trained on in 2020. Only distributed over 70 years, with biological memory and a body in the loop.

Inside that corpus at any moment — an enormous active context: everything the person currently holds in mind about themselves and their situation. Their children's names, a recent argument, a professional project, what hurts today, what a parent said twenty years ago, what a client said an hour ago, the fear pressing on Friday evening. That context is not passive memory; it is the field over which the next word, the next gesture, the next decision is sampled.

When a person makes a "considered decision," they are not in fact performing one large act of will. They are sampling the next token from a distribution assembled across decades of reading, conversations, fears, hopes, professional languages, and cultural frames. Sometimes the sample is good, sometimes not. Sometimes the context says "take this," sometimes "hold back." Sometimes the needed option simply is not in the corpus — and then the person does not choose it; they just do not see it.

This, incidentally, explains why "deliberately changing" by willpower alone is so difficult. Willpower is at best one or two tokens inserted into the stream. But the background from which sampling happens is an entire life. To change the background you must change the input: what you read, what you listen to, who you talk with, what projects you hold in mind. That changes the distribution; that changes the micro-decisions; that — gradually — changes the outcome. This mechanism has another well-known name: personal environment. We long knew that environment shapes a person more powerfully than "character"; now we can say how and why.

And one more important detail: a node in the network is not self-contained. Each "decision" it makes is in fact a small fragment of joint sampling: the word you choose in a conversation is influenced by the word your interlocutor just said; your choice of book by a friend's recommendation two years ago; your choice of city by a sunset over a particular bay that appeared in one specific film in childhood. A node cannot compute itself alone. It generates only in contact with other nodes. The network, therefore, is not an aggregate of people — it is the primary unit, inside which individual "we"s exist at all.

4. Miracle as long context: from Jung to LLM interpretability

Now to miracles. This is, perhaps, the most interesting part of the hypothesis.

LLMs sometimes produce answers that astonish: connecting distant concepts, unexpectedly anticipating the next step, articulating what a person could not articulate themselves. From outside this looks like "insight" or "illumination." From inside the model, as shown by Anthropic's interpretability research published in March 2025 — Tracing the thoughts of a large language model and On the Biology of a Large Language Model — it is the result of activating a long chain of internal circuits that learned during training to associate rare but relevant concepts. No magic. Just a very long and very rich associative network.

Particularly revealing was Anthropic's Golden Gate Claude demonstration from May 2024: researchers found a small "concept" inside the model corresponding to the Golden Gate Bridge and artificially amplified it dozens of times. The resulting version could talk about almost nothing else — ask it how to cook soup and it answers with a recipe that somehow features the bridge; ask about love and it turns out to be about the bridge; ask about sadness — the bridge again. This is essentially a caricature of a human whose one internal concept has taken over half the context: a fanatic, a person in love, a person in grief, a person in fear. From outside — "they have a fixation." From inside — simply weights at one point in the internal network gone into overload.

What we are describing here for machine models, human life noticed and tried to name long ago. The most serious attempt of the twentieth century was Carl Jung's concept of synchronicity, formalised in his 1952 work Synchronicity: An Acausal Connecting Principle, written partly in collaboration with Nobel-laureate physicist Wolfgang Pauli. Jung described a phenomenon his clinical practice was full of and which positivist science of the early century had no room for: meaningful coincidences, lacking causal connection but joined by common meaning. The most famous example from his work — a patient during a session describes a dream about a golden scarab, and at that precise moment an insect knocks against the window, exactly matching the description (Cetonia aurata). Jung opens the window, catches the beetle, hands it to the patient. After that incident, he writes, the analytic process shifted dramatically.

Jung formulated a hypothesis: alongside causality (where A produces B), reality contains another axis of connection — the meaningful. Two events can be linked not because one caused the other, but because they mean the same thing in a shared field of meaning. He developed this hypothesis with Pauli, who was himself a living example of "inexplicable" connections in his own laboratory — equipment broke with such regularity in his presence that colleagues named it the "Pauli effect". Jung seriously tried to construct a theoretical model in which such connections could occur without violating physics.

Our hypothesis is, in effect, the twenty-first-century technical language for what Jung sensed intuitively. If instead of the vague "field of meaning" you substitute the quite concrete "a distributed language model of eight billion nodes," then Jung's meaningful coincidences stop being a metaphysical claim and become a describable phenomenon. The connection between the dream of the scarab and the real beetle at the window is not a violation of causality. It is a connection through shared long-term context: the patient's active memory held certain symbolic tokens; Jung's held others; a Swiss summer evening in late June provided still others (beetles fly toward light) — and at the moment when the sampling of one model (internal) coincided with the sampling of another (external, via the insect), a recognisable figure assembled itself for both people in the room. Not magic. And not the erasure of mystery. A description of mechanism.

What if many of our own "coincidences," "signs," and "providences" work by the same principle — only distributed, at the scale of all humanity, more slowly, through us as carriers? Twenty years ago you read a book you have long forgotten. Its author sat on another continent, nourished by ideas they themselves heard at a conference in the 1980s. A classmate in 1995 told a story remembered because of a grandparent. A literature teacher at sixteen showed a poem that said nothing to you in the moment. These scattered pieces lay in your head as "sleeping weights." For thirty years you lived an ordinary life — and then one day this happened: you found yourself in the right room at the right time, and out of you came, seemingly on its own, exactly the phrase that changed someone's trajectory. From outside: a miracle. From inside the network: a long chain of distant context that finally converged in one point of the output distribution.

The word serendipity — the happy accidental discovery — was coined by Horace Walpole in a letter of 1754 from a Persian tale about three princes of Serendip who "always found something they were not looking for." Modern researchers of science (Ohid Yaqub, Sussex University, 2018) showed that serendipity in discovery is not a statistical outlier but a reproducible dynamic, occurring more often in people with broad context and dense conversational networks. "Luck" concentrates on those who have more tokens in their corpus and who sample in more varied environments. Not the moral "luck favours the prepared." The statistical "the probability of a connection increases with the number of relevant words in the active context of the network around you."

It is worth immediately noting the other side. We remember coincidences that converged and forget the thousands of times nothing converged at all. That is the classic survivorship bias: there is always much distant context in us, but it only rarely assembles into a ringing point — and it is exactly those rare points we later recount as "signs." The hypothesis does not claim that all coincidences are meaningful. It claims that those which did converge have an invisible but describable mechanism — not "magic" and not "dumb luck."

And one more thing. This picture does not prove that miracles don't exist. It does not even cancel religious experience. If someone wishes to see, in this long converging chain, the hand of God, our hypothesis does not interfere; it only describes the mechanism through which that hand might operate. A God who writes in human language must use grammar. And that grammar is ours.

5. PLLM: professional language as the key to a layer of reality

Here is where the hypothesis becomes practically useful.

If a person is a node in a distributed language model, then a profession is the specialised vocabulary and corpus on which that node is fine-tuned. Programmer, chef, midwife, equity trader, country musician, air traffic controller, gardener, cannabis dispensary staff — each lives inside their own Professional LLM (call it PLLM).

PLLM is not "jargon." Jargon is surface marking. PLLM is the complete set of distinctions a person is capable of in a domain. A programmer sees in code architecture, types, memory leaks, coupling, leaky abstractions — which a person without that preparation physically cannot distinguish in the same text. A chef sees in a hot fryer temperature window, Maillard activity, carry-over heat — all of which escapes the ordinary viewer of a cooking show. An experienced dispensary worker distinguishes in a single strain the terpene profile, the cannabinoid ratio, the cure quality, the trichome maturity — where a tourist just sees "well, some green stuff." A musician hears in one bar forward push, back-phrasing, swing feel, modal interchange — where a layperson hears "a pleasant melody."

From this follows a fairly blunt claim that often shocks: you cannot consciously become someone whose language you do not know. You can accidentally enter that reality, can "touch" it, but you will be able to move through it deliberately only once you have built an internal vocabulary of distinctions. This is why all the classic "how to get rich" books (from Hill's Think and Grow Rich to modern gurus) repeat the same simple thing: talk like them, read what they read, surround yourself with them. Not magic, not greed. A very practical instruction: change your training corpus.

Carol Dweck's work on mindset (2006) showed that successful and unsuccessful people have different internal explanatory vocabularies: those who say "I haven't learned this yet" follow radically different trajectories from those who say "I'm just not that kind of person." The same fact, packaged in different words, leads to different decisions.

The most famous data point here is Hart and Risley's "Meaningful Differences in the Everyday Experience of Young American Children" (1995). They observed how adults spoke to children in 42 American families across income levels for three years. By age four, a child from a professional family had heard roughly 30 million more words than a child from a family on welfare — and this gap predicted academic outcomes by high school better than many more "material" variables. Later work (Sperry et al., Child Development, 2018) debated exact figures and definitions of "word." But even after all methodological revisions, the core finding held: a different volume of language environment in the first years of life results in a different volume of internal model in the adult.

Good news: vocabulary can be learned. Bad news: it is learned slowly, and not from textbooks — from live conversations, real projects, apprenticeship with those for whom the language is native. A good mentor in any profession is worth more than ten courses. They do not "transmit knowledge" — they expand your training corpus.

6. What changes in 2026 when something very fast entered the network

Until 2022–2023, the distributed human model ran at a fairly slow pace. The speed of word transfer between nodes was limited by books, schools, radio, cinema, Wikipedia, Google search, social networks. Each successive invention increased connection density: Gutenberg's press in the 1450s dropped the cost of copying a word by an order of magnitude; the telegraph of the 1840s — the cost of transmitting it across distance; the internet of the 1990s — both simultaneously.

In 2026 a new type of node appeared in this network — a large language model capable of processing millions of tokens of context in minutes. By end of 2025 Gemini operates with a million-token context window, Claude Sonnet 4 handles up to a million tokens in enterprise mode; agentic orchestrators run projects for weeks, switching among dozens of tools. Per the Stanford AI Index Report 2025, frontier model performance on reasoning benchmarks grew more in one year than in the previous five combined.

Through the lens of our hypothesis, something fairly simple happened. In the distributed network of slow nodes (us), very fast nodes (LLMs) appeared — capable of passing far more tokens per second than any human could. And those fast nodes began actively connecting to the conversations of every second person on earth — through ChatGPT, Claude, Gemini, Cursor, Copilot, Notion AI, and dozens of specialised tools.

What does this change?

First: the generation cycle of the whole network accelerates. An idea that used to travel from researcher to paper to colleague to commercial product in 5–10 years now sometimes completes that journey in months. Not because people got smarter. Because auxiliary "tokens" — literature reviews, code summaries, language translation, formalising intuition into specification — are now generated orders of magnitude faster. Connections that once formed once per decade now form once per week.

Second: the composition of the training corpus changes for individual people. A student in 2026 in any language on earth can get a personalised explanation of any concept in an hour — something that in 2018 required two textbook chapters and a teacher. This changes not only learning speed but which professional languages are accessible at all to people who previously had neither university nor mentors. The PLLM we discussed in section five becomes a mass product for the first time in history.

Third — the most delicate. The distributed network now trains, in part, on its own outputs: more and more text online is generated by LLMs, and more and more LLMs are trained on that text. This is a feedback loop, and its effects are not yet fully understood. Some researchers (Shumailov et al., 2024, Nature) warn of model collapse — a gradual loss of diversity if models massively train on "their own." Others believe a well-filtered synthetic corpus accelerates learning. The truth is probably in between, and much will depend on whether people retain the role of authors of high-temperature originals — jazz clubs, philosophy circles, small independent publishers, local cuisines, poetry readings in Samui temples or Tbilisi courtyards. These are the points where "original tokens" are born — tokens that both people and models can then learn from. Lose them, and we get polished, literate, but ever-greyer generation where all answers resemble a slightly improved average.

This, incidentally, poses a completely new question about miracles. If connections in the network occur more frequently the higher the density and diversity of tokens, then 2026 might be the era in which coincidences statistically increase in frequency. Not because the world became more magical, but because the exchange of tokens between nodes became denser. More miracles — and simultaneously harder to recognise them against the background of general acceleration. In a way we are entering an era that would have fascinated and frightened Jung in equal measure.

7. What to do with this picture

A good hypothesis must yield practice; otherwise it remains an elegant figure of speech. From everything above, several actions follow — startable any day on Samui, in Berlin, in Lviv, in San Francisco.

Be careful about input. If you are a node training on your own feed, it matters to understand what exactly that feed contains. One hour of cheap news on Telegram is one hour of weight updates toward anxiety and shallowness. One hour of a good book is an update in the other direction. This works not in a day but over months and years. No heroism required; a slow, regular shift is enough.

Learn the language of the level you want to reach. If you want to become an investor, read investors: their shareholder letters, their interviews, their books. Not "how to get rich" but how they talk about business among themselves. If you want to be a serious programmer, read other people's good code, discuss architectural decisions, go to code reviews. If you want to be a farmer, listen to old farmers; their speech contains a century of observation.

Learn a second and third human language too. An underrated move. Among bilinguals and polyglots, as Aneta Pavlenko's research (Emotions and Multilingualism, Cambridge, 2005) and Anna Wierzbicka's work have long shown, different languages literally evoke different selves: different jokes, different distances, different available emotional registers. Not metaphorically — empirically. Each new language mastered opens another person inside you, with their own training corpus and their own sampling regime.

Surround yourself with people whose language you want to internalise. The most underrated factor of trajectory. Friendship and partnership are not "social connections"; they are distributed real-time learning. A move to a different city, a change of company, joining a new community, often changes a life more powerfully than a decade of solo willpower.

With children — be even more careful. A child builds their first vocabulary almost entirely from yours. What words about the world, people, and work you use in their presence will be their internal temperature and their training data for the next twenty years.

Use AI as an accelerator of your own vocabulary, not its replacement. The most effective mode in 2026 is not "query — get answer — forget" but dialogue in which you are learning a new language and becoming its speaker yourself. If every day you articulate your project's architecture with a model, you become an architect faster than if you merely read books on architecture. If you turn LLM into a robot that writes for you, you get texts — but you become neither writer nor thinker. Same logic as with sport: hiring a trainer is useful; hiring someone to exercise in your place is not.

Experiment with internal temperature. Humans have their own analogue of the temperature parameter in LLMs: it determines how much we sample the expected versus admitting distant, less obvious associations. Coffee compresses focus and speeds sampling. Alcohol removes safeties on the upper layers but blurs syntax. Nicotine gives short attention cycles. Sleep, a walk, warm water, a conversation with someone close all return the model to its calm range. Meditation works more like prompting — shifts sampling into a narrower range of deliberate options. Cannabis, especially in moderate low-THC formats, acts as a gentle lift of this parameter toward wider associative horizons: a study by LaFrance and Cuttler in Consciousness and Cognition (2017) showed that regular users — even in sober state — showed higher scores on openness to experience and divergent thinking than a comparable control group. Not "the magic of creativity" — a long-term tuning of one parameter of the internal model; we have a separate piece on Snoop Dogg as the everyday master of exactly this mode. The principle in all these adjustments is the same: high temperature without a rich corpus produces noise; low temperature without new input produces stagnation. Balance.

8. What this hypothesis does not explain

A few honest boundaries.

The picture does not explain biology: the body and its diseases follow their own, non-linguistic laws. Does not explain climate, earthquakes, volcanoes. Does not explain chance — genuine, physical, quantum. Does not explain the taste of mango, the smell of rain on a tin roof at five in the morning, the feel of warm sand underfoot. These things have no linguistic form and will not disappear no matter how sophisticated our distributed LLM becomes.

Nor does it explain evil. One can describe how language passes between generations, but not why the same set of words one person uses to build and another uses to destroy. That, apparently, is where linguistics ends and ethics begins — and our hypothesis leaves that territory to others.

And it is worth mentioning the main scientific opposition: Steven Pinker in The Stuff of Thought (2007) and The Language Instinct (1994) insists that language is a window into thought, not a prison of thought; that underlying human cognition are universal structures shared across all languages; that linguistic relativity explains only relatively surface-level effects. The strong Sapir-Whorf was indeed rightly rejected. Our hypothesis does not insist that language is the only or final layer of reality. It says something softer: language is the most underrated layer of reality among those a person can actually influence during their lifetime. Biology and physics are hard to change. Your own training corpus — you can.

And finally: the hypothesis does not cancel religious experience. If someone lives with a sense of presence — this is perhaps direct perception of those distant connections in our common network that ordinary consciousness usually cannot reach. Or a personal experience of something larger than language. Our hypothesis does not enter that dispute; it only describes one layer of reality and says: much here is explicable without the supernatural. Which does not mean the supernatural is disproved. Jung, Lotman, Wittgenstein — each kept a door in their picture that led to a room they themselves never entered; our picture keeps that door too.

A good hypothesis knows where its home is and does not try to annex the neighbouring rooms.

9. The question worth sitting with

If this picture is even partly true — and in 2026 we have tools to test it for the first time — then the main practical question is not "am I talented" and not "will I be lucky." It is different: in what language do I live?

What words are available to me right now to describe what is happening to me? Who do I listen to every morning and every evening? What books stand on my shelf unread — and which have I read thirty times? What professional distinctions do I already feel as my own, and which are still foreign, borrowed, not quite fitting? And above all — where has my current language already allowed me, and where would I like to go but cannot yet reach from inside the words I currently have?

If there is an honest answer to that last question, the work of the next five years is clear. This is perhaps the most practical application of the whole hypothesis: it promises no quick results but shows clearly exactly where the point of leverage lies.

We live on the west coast of Samui with plants, code, and people, and in our daily work those three things long ceased to be separate topics. In 2026 it is especially clear they are parts of the same distributed generation. A plant, a program, a conversation with a guest at the entrance to OG Lab and this article — all of them are written in the same universal human language, just in different dialects. Everyone who pays careful attention to this language does not only use it — they change it, a little. Not all of us will enter textbooks. But each of us is a node in the distributed network, and each adds to its training corpus exactly the sentence they decide to add today. May that sentence be warm and precise. That is the only serious task — for an ancient model of eight billion nodes and for each of its nodes individually.

And perhaps, in one of the coming days, when a phrase assembles itself in you that changes someone's path, or a book appears open to exactly the right page, or a person you have long been waiting for walks into the room — it may be worth remembering that this is not necessarily magic and not necessarily coincidence. It is, perhaps, the ancient network in which you are one node finally converging into one point around you. And doing it, as it has done for many thousands of years, — in words.