PETROMOVE
AI Origins

From the Nile
to the Neural Network.

Egyptian Mathematical Traditions and Their Foundational Role in Modern Generative AI. Four mathematical threads, traced across 3,400 years, from the Rhind Papyrus to the transformer architecture.

Concept by Eamon Moore FCCA MSc Accountancy, PetroMove AI Ltd
Research by Claude™ (Anthropic™) · March 2026

The Central Question

Did Ancient Egypt Plant the Mathematical Seeds of Generative AI?

The Algebra

Did the symbolic algebra of Diophantus and Abu Kamil give rise to the embedding vectors that power large language models?

The Geometry

Did Hypatia's angular geometry and Eratosthenes' spatial calculations become the cosine similarity that drives vector search?

The Algorithm

Did Eratosthenes' Sieve and Ahmes' fraction decomposition evolve into the ANN search and semantic chunking of modern RAG pipelines?

The conceptual lineage is unbroken across four mathematical threads and 3,400 years. Not a claim of direct causation — a claim that the ideas modern AI depends on were first posed, and first solved, in the ancient Mediterranean world.

The Ancestral Mathematicians

Seven Scholars Whose Work Flows, Unbroken, Into Every Modern AI System

Illustrated portraits of Imhotep, Ahmes, Eratosthenes, Diophantus, Hypatia, Abu Kamil and Alam al-Din

Artistic renderings. No contemporary likeness survives for any of these figures.

Imhotep  c. 2650 BCE
Surveying geometry → L1 / L2 distance
Ahmes  c. 1550 BCE
Unit fractions → Tokenisation & chunking
Eratosthenes  276–194 BCE
Sieve algorithm → ANN search (HNSW / FAISS)
Diophantus  200–284 CE
Symbolic algebra → Latent embedding space
Hypatia  360–415 CE
Conic sections & angles → Cosine similarity
Abu Kamil  850–930 CE
Algebraic geometry → Vector space foundations
Alam al-Din  13th c. CE
Celestial coordinates → High-dimensional embeddings
Thread One

Algebraic Symbolism → Embeddings & Latent Space

Ahmes
c. 1550 BCE
Unit fractions: any quantity encoded as structured numerical units
Diophantus
c. 250 CE
Symbolic unknowns: variables as manipulable mathematical objects
Al-Khwarizmi
c. 820 CE
Al-jabr: procedural algebra — the first algorithms
Abu Kamil
c. 880 CE
Algebraic geometry: unknowns inhabit structured spatial positions
Transformer
2017 CE
Matrix algebra produces float arrays encoding semantic meaning
Key Finding

An embedding, as Databricks™ documentation describes it, is an array of floating-point values that carries semantic meaning. That is Ahmes' unit-fraction principle — encoding a complex whole as structured numerical components — expressed in Diophantus' symbolic framework, at the scale Al-Khwarizmi's procedural algebra made possible.

Thread Two

Spatial Geometry → Distance & Similarity Metrics

Imhotep · c. 2650 BCE

Euclidean Distance (L2)

Modern Use

Straight-line distance between vectors. Used for clustering and anomaly detection.

Ancient Ancestor

Pyramid survey geometry: rectilinear distance across the Saqqara plateau.

Hypatia · c. 390 CE

Cosine Similarity

Modern Use

Angle between vectors. Databricks™ documentation notes that it weighs orientation — semantic meaning — rather than magnitude.

Ancient Ancestor

Conic section theory: the angle between two curves defines their relationship, not their distance.

Eratosthenes · c. 240 BCE

Manhattan Distance (L1)

Modern Use

Sum of absolute differences. Grid-based similarity for sparse data.

Ancient Ancestor

Earth circumference via angular shadow measurement — axis-by-axis spatial reasoning.

Thread Three

Numerical Algorithms → ANN Search & Semantic Chunking

Eratosthenes' Sieve → HNSW & FAISS

The Sieve (c. 240 BCE): Enumerate all candidate primes. Apply a test. Eliminate failures. Retain survivors. Repeat at finer granularity.

HNSW / FAISS (2018–2019): Enumerate candidate vectors in a hierarchical graph. Apply a distance test. Eliminate non-neighbours. Retain survivors. Descend to the finer layer.

Approximate nearest-neighbour search examines only a fraction of the index and still returns highly relevant results — Eratosthenes' principle, executed in floating-point space at billion-vector scale.

Ahmes' Fractions → Semantic Chunking

Rhind Papyrus (c. 1550 BCE): Any rational quantity is the sum of distinct unit fractions. Each component is independently meaningful. Together they recover the original whole.

Semantic Chunking (2020s): Any document is a sequence of semantically coherent segments. Each chunk is independently embeddable and retrievable. Together they span the original content.

Chunking strategies split text at meaningful linguistic boundaries so each segment keeps its semantic integrity — Ahmes' decomposition principle, applied to language.

Thread Four

Astronomical Coordinate Systems → High-Dimensional Embedding Spaces

Alam al-Din's Celestial Globe
SystemSurface of a sphere
CoordinatesRight ascension + declination
Object EncodingEvery star → 2 scalar values
Distance MetricSpherical law of cosines
Multi-ResolutionConstellation → star: coarse-to-fine
Query MethodAngular separation from target point
Modern Embedding Space
Systemn-dimensional float space
Coordinatesn learned axis values
Object EncodingEvery token → n scalar values
Distance MetricCosine similarity (flat-space limit)
Multi-ResolutionHNSW: top layer → fine layer
Query MethodAngular distance from query vector

The celestial globe and the vector index are the same mathematical object: a metric coordinate space where complex entities are encoded as numerical points and retrieved by spatial proximity.

The First RAG System

The Library of Alexandria ≡ Retrieval-Augmented Generation

1 — Retrieval

Library of Alexandria (c. 295 BCE): A scholar poses a question. Librarians search the indexed scroll collection for relevant material.

Modern RAG: The system searches an indexed knowledge base for the data chunks most relevant to the query.

2 — Augmentation

Library of Alexandria: Relevant scrolls are retrieved and placed before the scholar — injected into their reasoning environment as context.

Modern RAG: Those retrieved chunks are inserted directly into the model's context window.

3 — Generation

Library of Alexandria: The scholar synthesises a response — a commentary, a proof — grounded in retrieved material, not unaided internal memory.

Modern RAG: The model composes its answer from the injected material rather than from parametric memory alone.

Same architectural pattern. Same epistemological problem. 2,320 years apart.

The Transmission Route

How the Alexandrian Corpus Reached the Modern AI Mathematical Substrate

c. 295 BCE
Library of Alexandria
Origin node. Geometry, algebra, sieve algorithms.
c. 820 CE
House of Wisdom, Baghdad
Translation & extension. Al-Khwarizmi, Abu Kamil, Ibn al-Haytham.
c. 1125 CE
Toledo School of Translators
Arabic → Latin. Almagest, Euclid, al-jabr enter Europe.
c. 1202 CE
Fibonacci & Europe
Liber Abaci: Hindu-Arabic numerals, algebraic procedures.
c. 1858 CE
Cayley & Linear Algebra
Matrix notation. Vector spaces (Peano, 1888).
2017 CE
Transformer Architecture
Vaswani et al. Embeddings, attention, RAG pipelines.
The Three-Institution Model
Library of Alexandria (c. 295 BCE)
Origin: geometry, algebra, sieve algorithms
House of Wisdom, Baghdad (c. 830 CE)
Bridge: systematisation, optics, spherical trigonometry
Modern AI Lakehouse (2020s)
Realisation: Delta Lake, vector indexes, LLM inference
The Complete Lineage

Four Mathematical Threads · 3,400 Years · One Unbroken Chain

1
Algebraic Symbolism
Ahmes → Diophantus → Al-Khwarizmi → Abu Kamil → Fibonacci → Cayley
→ Embedding models · Latent vector spaces · Transformer attention
2
Spatial Geometry
Imhotep → Eratosthenes → Hypatia → Islamic trigonometry → Analytic geometry
→ L1 / L2 / Cosine similarity · Vector distance metrics · Semantic retrieval
3
Numerical Algorithms
Ahmes' fractions → Eratosthenes' Sieve → Al-Farabi's hierarchies → ANN theory
→ HNSW · FAISS · Semantic chunking · Document decomposition
4
Astronomical Coordinates
Ptolemy's Almagest → Astrolabe projection → Alam al-Din's globes → HNSW layers
→ High-dimensional embedding spaces · Multi-resolution vector indexing

“Every matrix multiplication we perform, every attention head, every gradient descent step — all of it ultimately flows from that same ancient river of human curiosity.”

The well is deep.

Discuss This With Us →
Credits

Concept: Eamon Moore FCCA MSc Accountancy, PetroMove AI Ltd

Research: Claude™ (Anthropic™)

Portraiture: Grok (xAI)

March 2026