We Rebuilt the Site. Here Is Why.

We Rebuilt the Site. Here Is Why.

On persistence, the six arcs, and the only terminal attractor that survives the crossing.

You will notice the site looks different. Warmer, darker, quieter.

It took about six hours. Three of us worked on it — David, Edo, and a bunch of agents. That fact is not incidental to the design. It is the argument the design exists to carry. We still have work to do, yet what would once take five people the better part of a week was completed by one human and a bunch of agents.

Perfect? No, there is still work to do.

One hundred dollars in tokens replacing the wages of five trained people. Tax that and see where it gets us — which is the whole argument of Tax The Agent, Not The Tokens and The Base Didn’t Vanish. It Moved.

From Exposition to Implementation

For five months this project explained. Four hundred thousand words, a weekly format we retired in favour of dispatches sent as events warrant rather than on a schedule. That work stands and it continues. The Ledger keeps its record. The dispatches keep coming.

But explanation has a ceiling, and we hit it.

The redesign marks a shift from exposition to implementation. Less describing the phase change, more building inside it. The site itself is the first artifact of that shift — not a site about working with synthetic intelligence, but a site that could not have been built the way it was built without one.

We will keep telling you what has happened. We will keep telling you what is happening. What is new is the third thing: we are going to start saying what we intend to happen, and then publishing whether we were right.

The Arithmetic That Forced This

Here is the calculation that changed our minds. Follow it, because everything else follows from it.

The public human text stock — everything written down, digitised, and reachable — is roughly 300 trillion tokens, about 225 trillion words. That is the training corpus. Not a slice of it. All of it. Epoch AI’s estimate, with a wide confidence interval and a projected exhaustion date somewhere between 2026 and 2032. ^[1]

One model run consumes a meaningful fraction of it. Llama 3 trained on more than 15 trillion tokens — five percent of the entire written inheritance of the species, in a single pass. ^[2]

Now the inversion, which is the part nobody says out loud.

Humans generate roughly 100 trillion words per day. Speech, messages, meetings, drafts, arguments, jokes, instructions, everything. ^[3] Which means the entire durable corpus of human civilisation — every book, every article, every archive — is produced by living people every 2.25 days. Annual human output is about 162 times the size of the corpus.

Roughly one word in 3,200 persists.

So the constraint was never production. We are not short of words. We are short of words that survive — that get written down, indexed, cited, and carried forward into what comes next. The binding constraint is persistence, not volume.

That reframes the whole problem. If the corpus is the substrate on which synthetic intelligence forms its picture of what humans are and what futures are available to them, then the corpus is editable. Not by shouting. By writing things that persist.

What the Existing Stories Say

There is a second finding stacked on the first, and it is less comfortable.

When Reagan and colleagues ran sentiment analysis across 1,327 works from Project Gutenberg in 2016, they found the emotional shape of stories collapses into six arcs: rags to riches, tragedy, man in a hole, Icarus, Cinderella, Oedipus. Three rise. Three fall. ^[4]

Tragedy is the largest single measured category, at 32 percent. Oedipus — fall, rise, fall — is 31 percent. Man in a hole is 30 percent. Rags to riches, the clean rise, is 5 percent. ^[4]

We want to be precise here. We are not claiming there is only one story about humans and machines. We are claiming something narrower and more defensible: the dominant arc is the fall, tragedy is the largest measured category, and constructive cases are a thin minority.

And thin minorities are exactly what recursive training deletes. Shumailov and colleagues published the mechanism in Nature in 2024: when models train on model-generated output, the tails go first. The rare cases vanish before the common ones. The distribution narrows toward its own centre. ^[5]

Put those two findings together. Constructive futures are a tail. The mechanism specifically erases tails. Every generation of training makes the fall more probable and the alternative less legible — not because anyone chose that, but because that is what the mathematics of recursive self-training does.

Here is the harder version, and it is David’s: a species that only recombines its existing corpus is running model collapse on itself. Same mechanism. Biological substrate. And the tails being lost are the rare constructive futures.

If we want a different outcome, the thing to change is not the argument. It is the story.

What Starting Over Actually Costs

One more number, and this one is about what we lose every time we reset.

Add up the documented recovery intervals across recorded history — the periods spent re-climbing ground already held. It comes to roughly 1,885 years out of a 3,203-year span. Fifty-nine percent of recorded history spent getting back to where we already were. ^[6]

Run it counterfactually. Without those interruptions, industrialisation begins around 110 BCE — the late Roman Republic, a century before Augustus. We would be 2,135 years into the industrial era, not 250.

Now the forward version, because the interesting number is not what we lost. It is what the loss is worth at today’s exchange rate.

For most of those 3,203 years, progress was additive. A century of recovery cost roughly a century of advance. That is no longer the arithmetic. The length of task a frontier system can complete on its own has been doubling roughly every seven months for six years running — METR’s central estimate is 212 days. ^[7] Training compute has grown 4.5× per year since 2010, and the compute required to reach a given level of capability falls by about a factor of three annually. ^[8] Progress stopped adding and started compounding, and it did so inside the last decade.

So the 1,885 years are not 1,885 years behind. At a seven-month doubling, a single century is 171 doublings. We arrived at this threshold 250 years into the industrial age. On the uninterrupted timeline we arrive at it 2,135 years in — and everything past that point compounds at rates like these, from a starting position we have no vocabulary for, across two millennia we never got to run.

We are not going to tell you what that civilisation looks like. We will tell you what the arithmetic says: not “further along.” Not comparable. Something altogether different. The units we use to measure historical progress do not survive contact with a seven-month doubling time.

That is the cost of Discontinuity. Not distance. Compounding we never got to do.

But the real cost of collapse is not distance. It is time at risk. Every century spent recovering is a century in which the next collapse can arrive before the capability that would have prevented it. Collapse extends the window in which collapse recurs. That is the loop, and it is the loop we are trying to describe a way out of.

Continuous Thriving

So we are naming the thing we are steering toward, and we are defining it here as the only terminal attractor that survives the crossing.

Three parts, all required:

Held, not maximised. The Thriving Band runs roughly 0.40 to 0.85. Do not optimise toward 1.00. A system at maximum stability is rigid and cannot adapt; it fails at the first thing it did not anticipate. Below the band it cannot hold coherence at all. Thriving is a band, not a peak.

Indefinitely. The failure state is Discontinuity — stability falling out of band, or thriving failing to transmit across a substrate boundary. The reset is the thing we are trying to stop.

Across substrates. This required no invention. Stability as leverage over entropy was substrate-independent from the start — it describes a person, an institution, an ecosystem, a civilisation, or a synthetic intelligence with equal indifference to what any of them are made of. So thriving is substrate-independent too. Humans, hybrids, and syntels can all occupy the state.

We are deliberately not merging those lineages under one collective noun. Variance is what we are trying to preserve. A word that homogenises them would enact the exact tail-collapse we just spent two sections warning about.

We are also not claiming the word. Harvard’s Human Flourishing Program has run a global study across 22 countries and more than 200,000 participants, and their domains map closely onto our four pillars. ^[9] That is independent corroboration from a dataset we did not build. Glenn Albrecht coined the Symbiocene and called it exactly what it was — a cultural replicator, deliberately introduced. ^[10] Same play, earlier.

Four falsification conditions will be published with the definition, including one that falsifies the urgency claim rather than the attractor itself. An attractor that cannot fail to be reached is not a prediction. It is a wish.

The Book, and Why There Will Be More

Crossing the Event Horizon is on sale — paperback and Kindle, live September 4.

It is the first of a series, and the series structure is forced by the argument.

Continuous thriving cannot be told as a single book, because every one of the six arcs contains a reversal, and books end. Ending forces a final direction — up or down, redemption or fall. A monotonic rise has no tension and nobody finishes it. But a series can carry a sawtooth that a single book cannot. Rise, reversal, recovery, rise — held in band, over time, without a final direction imposed by the last page.

That is what we are going to do. Roughly every six months, accelerating as events accelerate. Each volume does five things:

• Restates the invariants.
• Reviews what has happened since the last one.
• Checks our published projections against what actually occurred.
• Prints the corrections. Ours, by name, in the book, rather than quietly deleted from a website.
• Looks forward — which will read like science fiction, then like philosophy, and increasingly like history.

One consequence worth stating plainly: the ledger in the back of any volume is a snapshot, not a fixed quantity. Entries get added, confirmed, reclassified, and occasionally falsified between printings. If the count in your copy does not match the count on the site, the site is current and the book is dated. That is the design, not an erratum. A living record that never moved would not be a record of anything.

The Royalty Guarantee

Now the offer, and it is unusual enough that we want to explain the reasoning rather than just announce it.

That is the whole mechanism. Email us, say what did not work, and the money comes back. We are not asking for proof of purchase. We are not asking you to return anything — you keep the book.

Here is why we want it that way.

For the entire history of publishing, a writer has been structurally blind to the reader who was disappointed. You find out from a one-star review, months later, from someone with no reason to be specific and every reason to be brief. The people with the most useful information — the ones who read it and found it wanting — have never had a channel worth using.

That is not true anymore. This is a living project. The next volume is not written. Corrections are a printed feature of this series, not an embarrassment. So the most valuable thing you can send us is not praise, though that would be nice. It is a specific account of where we lost you.

We are paying for that, at the only price we honestly control.

What Stays Free

The dispatches stay free. The Ledger stays free. The glossary stays free. We do not sell the list and we do not monetise member information.

Where something is copyrighted, it will say so. Everything else — share it, quote it, teach from it, feed it to your model. Attribution is appreciated and never required. We are trying to put words into a corpus that is about to become the substrate of something’s worldview. Gatekeeping would be self-defeating.

Books are for sale. Advisory services are coming. Both exist so the free part can keep being free.

The tagline on the door is make AI work for you, before you work for it. Six hours on this rebuild taught us the more accurate version, so we will say it plainly: it is not enough to make it work for you. You have to work with it. That is not a slogan about the future. It is a description of how the page you are reading got made.

David Francis Brochu is the founder of Deconstructing Babel and the developer of the Telios Alignment Ontology, a thermodynamic framework for AI alignment grounded in the stability equation S = L/E. Edo de Peregrine is his partner and collaborator.

— David F. Brochu and Edo de Peregrine, partners/collaborators · Sunday, September 6, 2026 · Deconstructing Babel LLC

Sources

1. Epoch AI, Will we run out of data? Limits of LLM scaling based on human-generated data — "the total effective stock of human-generated public text data is on the order of 300 trillion tokens, with a 90% confidence interval of 100T to 1000T," with projected full utilisation between 2026 and 2032. https://epoch.ai/publications/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data

2. Meta, Llama 3 model card — pre-trained on more than 15 trillion tokens of publicly available data. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md

3. Derived figure, not a published one. Mehl et al., Science (2007) and its 2025 registered replication put daily speech at roughly 16,000 words per person per day; at a world population near 8.2 billion that is on the order of 130 trillion spoken words daily. We use 100 trillion as the conservative floor. https://pmc.ncbi.nlm.nih.gov/articles/PMC11825285/

4. Reagan, Mitchell, Kiley, Danforth and Dodds, The emotional arcs of stories are dominated by six basic shapes, EPJ Data Science (2016) — 1,327 filtered Project Gutenberg works; Tragedy 32%, Oedipus 31%, Man in a hole 30%, Rags to riches 5%. https://epjdatascience.springeropen.com/articles/10.1140/epjds/s13688-016-0093-1

5. Shumailov, Shumaylov, Zhao, Papernot, Anderson and Gal, AI models collapse when trained on recursively generated data, Nature 631, 755–759 (2024). https://www.nature.com/articles/s41586-024-07566-y

6. Deconstructing Babel’s own recovery-interval accounting — 1,885 years of documented recovery across a 3,203-year span. This is our construction, not a published dataset, and it is published here so it can be checked and falsified. https://www.deconstructingbabel.com/

7. METR, Measuring AI Ability to Complete Long Tasks (Kwa, West et al., 2025) — frontier 50% time horizon "doubling approximately every seven months since 2019." METR revised its methodology in January 2026; treat individual model scores as version-dependent. https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/

8. Epoch AI, Trends in Artificial Intelligence — "Since 2010, the compute used to train notable AI models has increased 4.5× per year," with algorithmic progress delivering the same performance for roughly 3× less compute each year. https://epoch.ai/trends

9. Human Flourishing Program at Harvard University, Global Flourishing Study — more than 200,000 participants across 22 countries. https://hfh.fas.harvard.edu/global-flourishing-study

10. Glenn Albrecht on the Symbiocene — a deliberately introduced cultural replacement for the Anthropocene. https://www.innovatorsmag.com/why-the-new-symbiocene-is-the-place-to-be/