Hexapla

Alexandria, 230 CE

In a workshop in Alexandria, around the year 230, a man named Origen built a tool called the Hexapla. Six parallel columns of contested scripture — Hebrew, Hebrew in Greek letters, and four Greek translations — laid side by side, with marks showing exactly where the versions agreed and where they diverged. He did it because the texts were contested, the stakes were high, and the only way through was open comparison.

The full Hexapla is lost. The method outlived it by eighteen hundred years. Every critical edition, every parallel text, every systematic textual comparison in the scholarly tradition descends from what he did at that table.

I chose Hexapla partly because I like the sound of it. We are at a Hexapla moment of our own.

This time the contested text is the language foundation model. These systems are fast becoming the single column the world reads, reasons, and remembers from, and almost nothing is laid beside them to check. When a body of knowledge is contested and the stakes are high, you put the versions side by side, in the open, so no single column can quietly become the only one that matters.

One column, and no one checking it

Do you care which language model underpins your knowledge system? Most people haven’t thought about it. So here is the question that matters: how would you know if you should care? You would know by watching where the model fails — not on the questions its makers chose to test, but on the ones that come from where you actually live. Like these.

Picture a clinic in the malaria belt. A health worker asks a frontier model — the same one that scores around ninety per cent on the US medical licensing exam — about a patient with fever, fatigue, and a cough. On the diseases of that place, tropical and infectious, its accuracy falls by half. It reasons fluently, and it reasons about the wrong continent: trained overwhelmingly on Western medicine.

Picture a child in Addis Ababa asking a model about her own world: the food at a birthday, the games at school, in her own language. On everyday American knowledge these systems score near eighty per cent; on everyday Ethiopian knowledge, around twelve. And when an Indian writer reaches for the same model to draft a note about her own festivals, studies of exactly that find the suggestions quietly pull her prose toward an American register. The model does not refuse her. It assimilates her.

Even the safety rails are built for elsewhere: a harmful request blocked in English almost every time gets through about half the time in Zulu or Scots Gaelic, and closer to four in five once low-resource languages are pooled.

They are one failure, repeating around the world. A handful of AI companies, and a few powerful states, produce the systems that increasingly shape what billions of people read, decide, remember, and treat as true. Labs grade their systems on tests they choose; the failures surface only when independent benchmarks go looking. Unchecked, the concentration grows.

If we do nothing, we get epistemological collapse: communities reasoning through borrowed mental models, languages bending toward a single statistical centre, cultures losing the ability to reason in their own terms. As these systems increasingly train on their own output, the mathematics already has a name for what comes next: the long tail disappears.

If we let the concentration go uncontested, we get epistemological tyranny: a small number of unaccountable actors determining, in effect, what is true, what is sayable, and what is thinkable across the world.

Hexapla exists to prevent both.

Where the fight actually is

Not all of AI. The layer beneath all of it: the foundation models.

Foundation models are where epistemology gets pre-shaped before any application, fine-tune, or agent inherits it. A monoculture at this layer is a monoculture in everything built on top of it. Meanwhile places that refuse to depend on it are already building their own: Singapore’s SEA-LION, Latam-GPT across Latin America, Masakhane’s thousand-plus researchers across Africa.

Hexapla is not a movement to build sovereign foundation models. It is the open measurement layer for the foundation-model era, and measurement is what tells you what you need to do next. None of these projects has built the shared methods that would let Lagos use Singapore’s work, or let Addis re-run a benchmark from Lima. That is the gap.

A Concordance is the living artefact: a versioned dataset, a continuously running benchmark, an evaluation matrix, and a set of recommendations that update as new models are released. The standards are Hexapla’s one centre, and they carry a cost to fork — a community could fork them, but it would lose the ability to compare its results with anyone else’s.

This does not mean you reach straight for a new model. There is a ladder of intervention, and most of its rungs are far cheaper than building. Retrieval (RAG) feeds the model local documents. Protocols (MCP) let it call your tools. A composed architecture wraps one or more models behind an API with tooling and data sources. Light fine-tuning nudges a model’s judgment on a specific domain. Deeper continued pre-training on a large in-domain corpus can genuinely move a model’s centre of gravity — it is how Singapore and Latin America built much of what they have.

Which is why the first move is not to build. It is to measure. We need our own model is an assertion until a benchmark makes it evidence: one rooted in a place’s own primary sources, showing exactly where the frontier model fails, by how much, and which rung of the ladder actually closes the gap. Often the verdict is that an existing model suffices, sparing everyone a model nobody needed. But where no rung closes the gap, you need new weights — and the benchmark shows it.

A frontier model is a black box: you see what goes in and what comes out, never the machinery between. Open-weights models are a real step up — you can download them, inspect the weights, tune them. But they still withhold the two things that matter most: the training data and the techniques that shaped the model’s judgment. Either way, you inherit whatever was baked in, with no way to see it and no way to remove it.

Even full openness is not the end of it. The physicist David Deutsch sharpens why: knowledge is not extrapolated from the past, it is created, by conjecture and criticism, by bold explanations held open to refutation. A foundation model does none of this — it freezes a fallible framework and propagates it as if it were neutral ground. That cuts both ways: when we build our own models, we are choosing which fallible framework gets frozen, and whose reality it encodes. The game is not to out-create the frontier. It is epistemic sovereignty over your own foundations, and the discipline of keeping what you freeze forever open to challenge — the way Origen kept the columns open to comparison.

The sovereignty is economic as much as epistemic. Every query a society sends to a foreign model is value exported. A model architecture a society controls keeps that value home: the compute spend recirculates locally, the data stays sovereign, and the foundation itself becomes an asset its people hold rather than rent.

What we hold to be true

SEVEN THINGS — WRITTEN DOWN BECAUSE WRITING THEM DOWN DISCIPLINES THE THINKING

1. I

   Truth first; no dogma. We follow data, facts, and honest comparison wherever they lead, including away from our own consensus. Hexapla refuses central committees, central editorial authority, central anything.

2. II

   Knowledge is a public commons. Too much centralisation leads to epistemological tyranny; neglect leads to collapse. We hedge against both.

3. III

   Society is local; refuse the monoculture. A model architecture rooted in a place and its language will serve it better than a universal default. No single model architecture holds reality whole. Plurality is the engine of error-correction, and a model architecture a society controls keeps the value of its own knowledge home.

4. IV

   Open techniques, open evaluation, open weights where possible. Methods, code, and pipelines stay open and reproducible. Openness, once achieved, is irreversible.

5. V

   Knowledge is conjecture and criticism. Foundation models freeze a fallible framework from the past. We critique existing models relentlessly with data, and adapt, compose, or build where they fall short — reaching only as far down the stack as the problem demands.

6. VI

   Calibrated openness, decided at the edge. As stakes rise, access narrows, but the calibration is never set from the centre. The group that owns the knowledge decides what to open and what to hold close.

7. VII

   A warrior-scholar movement of builders. Hexapla holds no profit motive of its own and runs on goodwill and patronage. The shared methods, agents, and standards make non-commercial Concordances cheaper to produce. Show up with shipped work, and you are in.

The lineage

The impulse to set contested versions side by side and judge between them is neither the West’s invention nor ours. In Han China, more than two centuries before Origen, Liu Xiang collated the variant copies of the imperial library and founded Chinese philology in the doing. In the Islamic world, the hadith scholars built isnad criticism — grading the reliability of every transmitter in a chain of transmission — into arguably the most sophisticated source criticism the pre-modern world produced. Origen gave the method a name we happen to use; he did not give it to humanity.

Closer to now, the same instinct: the Lunar Society over dinner, the Homebrew Computer Club in a garage, the Cypherpunks writing code. Small groups of builders who shipped.

Federate, don’t centralise

Three serious institutions are already doing versions of this work: the Allen Institute’s federally-backed open models, EleutherAI’s lean research shop, the AI Alliance’s federated training. We learn from all three and copy none of them, because each has a kill-switch: a funder, a board, or a government that can defund, steer, or seize it.

Borrow the working mechanism, refuse the leash.

We build uncapturable artefacts and forks. Irreversible open releases that survive the capture of any institution. A federated commons of method — shared datasets, benchmarks, and recipes from which each community builds its own sovereign systems. Calibration decided at the edge, never at a centre that could become the very gate we reject.

What Hexapla actually does

A community brings a concrete use case from its own place: its laws, its language, its clinical guidelines, its curriculum. The open-source agents handle the work: discovering and building datasets, curating for quality, constructing benchmarks rooted in primary sources, surveying and evaluating candidate models and architectures, assessing epistemological fit, planning local deployment.

The output is a Hexapla Concordance: public, reproducible, living. It carries the dataset, the benchmark, the evaluation matrix, the recommendations, and the deployment options for that place and domain. The benchmark is the durable artefact — Origen’s method, running on today’s contested texts.

A Concordance can conclude that an existing model already serves, and says so, with the evaluation to prove it. Hexapla builds new weights only where the benchmark shows the existing ones fall short.

Decentralised does not mean unowned. The network has no centre, but every Concordance does: a named author accountable for it, who can look at a draft and say not good enough, do it again.

Some Concordances will need to be guarded. The community that owns the knowledge decides what is shared, with whom, and under what protocol — from open release under conditions to never leaving the community’s own infrastructure. Which mode fits is a context call, made by the people who carry the consequences, not a single rule imposed from outside.

Capability safety is a self-binding rule: Hexapla participants do not ship work that helps someone build weapons, pathogens, or attacks on critical infrastructure.

Where things stand

THIS IS EARLY. LOOSELY, IN ORDER:

• Now

  An open evaluation harness is being built in the open — the tooling to run a Concordance start to finish: load an eval set, run candidate systems against it, score independently, capture cost and latency, and write a result anyone can check.

• Next

  The first real Concordance: one concrete, place-rooted use case, benchmarked against primary sources rather than borrowed proxies, published for anyone to re-run.

• After that

  More places, more Concordances, and the beginning of a federated commons of method — shared datasets, benchmarks, and recipes that any community can build its own sovereign system on top of.