Here is the question that has haunted Buddhist scholarship for over a century: which words did the Buddha actually say?

Not in the pious sense — as in, we trust the whole Canon. And not in the cynical sense — as in, we can know nothing. In the precise, historical sense: when you open the Pali Canon and read one of its 36,434 passages¹, are you reading something composed within a generation of the man who walked the Gangetic plain in the fifth century BCE? Or are you reading something written five hundred years later by scholars systematically classifying every mental event into exhaustive taxonomies? Both kinds of text ended up in the same library. The library does not tell you which is which.

A new book by L. Lopin — Buddha, before the Religion: A Computational Search for the Buddha’s Oldest Words — takes this question seriously in a way that has not been possible before. And the findings, if you are interested in early Buddhist texts and in what the Buddha may actually have taught, are worth knowing about.

A Different Kind of Reader

Imagine you are standing in a great hall containing ten thousand scrolls, all in Pali, all hand-copied over a thousand years, some repeating the same phrase word for word for twenty pages, some containing passages unlike anything else in the room. Some of these scrolls are old in the way that the first draft of a revolution is old compared to the constitution it eventually produced. Some are much more recent. And you cannot always tell which is which just by looking.

This is the situation every scholar of early Buddhism has faced. The methods exist — metrical archaism, morphological archaism, the absence of later systematized formulas, the evidence of intra-canonical commentary — but applying them requires decades of training, access to specialized literature, and the ability to hold the entire textual tradition in your head simultaneously. The scholars who do this work are extraordinary. But they are few. And they read sequentially.

Lopin’s approach was to ask a different kind of reader to examine the library. Not a human reader. A mathematical model trained on the entire corpus of Pali text — 36,434 passages — without any prior assumptions about which passages are old and which are late. No loaded expectations. No prior training in Pali philology. Just the statistical patterns in the text itself, extracted by algorithms originally developed for commercial search engines and biological sequence analysis.

The starting point was the Aṭṭhakavagga, the fourth section of the Suttanipāta. Sixteen poems. Ninety-seven verses. A body of writing that seven independent lines of philological and historical evidence agree is the oldest recoverable stratum of the Buddhist canon.² To the model, Lopin said simply: find everything else in this library that sounds like this.

Not by topic. Not by meaning. By sound — the statistical patterns of word use, sentence rhythm, vocabulary distribution, the ten thousand microscopic choices that characterize a particular speaker’s register, a particular era of composition.

Five Methods, One Convergence

Now, this is interesting — because Lopin did not stop at one method. The semantic similarity analysis (using FastText word embeddings, for the technically inclined)³ was only the first of five entirely independent approaches. The second measured formulaic density⁴ — the proportion of each passage shared as five-to-seven word sequences with the rest of the corpus. The third analyzed stylometric fingerprints. The fourth cataloged morphological archaisms. The fifth examined metrical profiles.

Five methods. Five different theoretical frameworks. Five different mathematical approaches.⁵ Each one knowing nothing about the others’ results.

And they converge.

The convergence is the book’s central finding, and it is worth sitting with. When independent instruments measuring different things point at the same cluster of texts, you are not looking at a methodological artifact. You are looking at a signal in the data.

The Zero-Density Cluster

Let me focus on the formulaic density finding, because it is the most striking.

The Pali Canon, like all oral literature preserved through centuries of recitation, is full of stock phrases. Standard formulas for opening a discourse. Standard formulas for describing the Buddha’s mental states. Standard formulas for the four jhānas, the noble eightfold path, the five aggregates. A passage from the Abhidhamma Piṭaka might consist of 80% or 90% phrases shared verbatim with dozens of other passages. This is how oral literature works: it crystallizes around formulas.

Now ask: what is the formulaic density of the Aṭṭhakavagga? The answer is very low. It uses original language. The poet is not reaching for stock phrases; the poet is finding words for something being seen for the first time.

The search identified 84 passages with formulaic density at or below 0.100 — meaning ninety percent or more of their language is original, not shared with the rest of the corpus. Of these, 34 novel passages (outside the seed texts themselves) have a formulaic density of exactly 0.000.⁶ Not a single five-to-seven word sequence in any of these passages is shared with any other passage in the entire 36,000-passage corpus.

These texts were not composed by reaching for the available formulas. They were composed by someone finding language fresh.

What the Machine Brought Back

Let me tell you about a few of these passages. Because this is where the book stops being about methodology and starts being about something that actually matters for practice.

The Kālāma Sutta — the passage sometimes called the Buddha’s “charter of free inquiry” — comes back with formulaic density 0.000. Semantically adjacent to the Aṭṭhakavagga. Stylometrically archaic. The algorithm did not know the Kālāma Sutta is famous.⁷ It did not know that scholars have long suspected it contains ancient material. It followed the numbers. And the numbers point here.

The Tevijjasutta (DN 13)⁸, the Buddha’s devastating critique of the brahminical tradition — brahmins who have never seen Brahmā arguing about the path to Brahmā, “like a line of blind men, each holding the shoulder of the one before” — comes back at formulaic density 0.000. This is not a late composition. This is the earliest epistemological register of the tradition.

There is a passage in the Majjhima Nikāya (MN 75), the so-called Māgaṇḍiya eye passage, that describes the sense organs with a phenomenological precision that never quite recurs in later canonical prose. Let us have a closer look at what it says:

“The eye, Māgaṇḍiya, delights in forms, rejoices in forms, is pleased with forms. That has been tamed, guarded, protected, restrained by the Tathāgata, and he teaches the Dhamma for its restraint.”

Notice the word danta⁹ — tamed. Not suppressed. Not destroyed. Tamed, the way a horse or an elephant is tamed — brought under discipline while retaining its nature. The eye’s reaching toward form is not denied; it is understood and worked with. Formulaic density: 0.000. This is original teaching.

And then there is the First Sermon narrative — the passage describing the Buddha walking into the Deer Park at Bārāṇasī, the five monks who had abandoned him because he gave up extreme asceticism, their prior agreement to snub him dissolving as he approached:

“The five monks saw me coming from afar. Having seen me, they agreed among themselves: ‘Here comes the ascetic Gotama, who lives in luxury, who gave up the exertion, and reverted to luxury. We should not pay homage to him, should not rise for him, should not receive his bowl and robe.’ But as I approached nearer, the five monks were unable to hold to their agreement.”

That last sentence has the texture of something remembered, not composed for an occasion. Formulaic density: 0.000. The algorithm found this passage by following the semantic thread of the Aṭṭhakavagga. It has the register of the oldest stratum. It may be among the oldest narrative elements in the entire prose collection.

Before the Religion

What do all these passages have in common beyond their formulaic originality? This is the part of the book that I find most philosophically interesting, and most relevant to anyone who sits on a cushion.

The muni — the sage — of the Aṭṭhakavagga is defined almost entirely by negation. The muni does not get into disputes. Does not take up a view as a banner. Does not seek what is not yet obtained. Has put down the burden of self-construction. The Aṭṭhakavagga does not give you the Four Noble Truths in their standardized form. Does not give you the Eightfold Path as a numbered list. Does not mention vipassanā as a technical term with a fixed meaning. These formulas are simply not there. This is not an accident of preservation. The collection is complete.

What this points to is a world before monasteries. Before the Vinaya. Before the saṅgha (community) became an institution with rules and procedures. The muni of the Aṭṭhakavagga is a wanderer — paribbājaka¹¹ — in the older, freer sense: someone who has given up household life to roam, who debates with other wanderers, who has no permanent residence.

The passages the machine retrieved from the prose collections — the Kālāma Sutta, the Tevijjasutta, the Māgaṇḍiya eye passage, the First Sermon narrative — all have this same quality. They are not documents of an organized religion. They are records of observation: here is what happens when the mind grasps at a view; here is how the eye reaches toward form; here are people confused by competing doctrines; here is what it feels like to walk up to people who have decided to ignore you, and watch their resolution dissolve.

In other words: these are texts from before the religion built itself around the teaching. They have the quality of what was seen before the frameworks for seeing crystallized.

If you are a meditator — if you sit regularly and actually watch what happens at the sense doors — this distinction matters. The Aṭṭhakavagga is not giving you a method. It is describing what there is to see. The method came later, and the method is valuable. But the original observation is not the same as the method that was derived from it.

The Translations

The book includes new English translations of all 84 low-density passages, including the Aṭṭhakavagga in full. The translation register Lopin uses is what the book calls “present-tense direct” — as though the text is being spoken right now, to you, without the mediating apparatus of footnoted Victorian prose. Copious scholarly endnotes explain every translation choice for readers who want to check the Pali themselves.

Having spent time with the Pali of these texts, I find the translation choices consistently defensible and often illuminating. The word papañca¹⁰ — mental proliferation, the way the mind spins out of a simple perception into elaboration, identification, defense, and conflict — is translated with the freshness the concept deserves. The Aṭṭhakavagga‘s insistence that the muni does not “take up a view” is rendered in language that keeps the physical weight of the Pali idiom.

An Honest Caveat

The book is admirably honest about what it cannot claim. I will quote directly from the conclusion:

“We cannot say the Buddha said it. We cannot say it was composed in the fifth century BCE rather than the third. We cannot say it was not later revised by a redactor who happened to produce original-sounding language. What we can say is this: among all the 36,434 passages of the Pali Canon, a specific cluster of texts share a specific profile — semantically adjacent to the oldest stratum, linguistically original, metrically archaic where verse features apply, morphologically archaic where detectable. That cluster coheres.”

This is the right level of epistemic humility for a computational study of ancient texts. The machine cannot time-stamp a manuscript. What it can do is measure linguistic similarity, formulaic originality, and stylometric fingerprint — and when five independent measures agree, the agreement is meaningful.

What we have, at minimum, is a subset of the canon that is linguistically unlike the rest of the canon in ways that correlate with known markers of antiquity. That is not nothing. In a tradition that has been asking “what did the Buddha actually say?” for 2,500 years, it is something new.

Why This Matters

Right now, while reading this — wherever you are — you are probably not suffering because of a bad theory of everything. You are suffering, if you are suffering, because of how the mind reaches toward what it wants and recoils from what it does not want. Because of the way a view, once adopted, becomes an identity, becomes something to defend, becomes a source of conflict.

The Aṭṭhakavagga — and the cluster of texts the algorithm found clustering around it — addresses that directly, without the apparatus that later Buddhism built around it. Not as a historical curiosity. As a description of something observable, right now, on the cushion and off it.

The tradition preserved this voice. It surrounded it with commentaries and sub-commentaries and organizational frameworks and rules. It succeeded: the voice is still audible, in Pali, in a digital edition stored on a server somewhere and searchable by a machine learning model that was not invented until 2,500 years after the voice spoke.

That the machine found it again — by following the numbers, without knowing what it was looking for — is, I think, worth knowing about.

Buddha, before the Religion: A Computational Search for the Buddha’s Oldest Words is available on Amazon. If you are interested in early Buddhist texts, in the question of what the Buddha actually taught, or in how computational methods are changing the study of ancient literature, I recommend it without reservation.

Find it on Amazon →

---