Aeneas

DeepMind’s AI Model for Ancient Inscriptions

What is Aeneas?

‍ Aeneas is Google DeepMind’s pioneering AI model specifically designed for historians, archaeologists, and researchers to interpret, attribute, and restore fragmented ancient Latin texts. Built on a multimodal transformer and trained on over 176,000 Latin inscriptions, Aeneas combines textual and visual information, setting a new benchmark for contextualizing inscriptions. It advances what was previously achieved with models like Ithaca (Greek) to now cover the vast Roman epigraphic record.

Key Features of Aeneas

Multimodal Analysis

Processes both text and artifact images for richer context and improved accuracy.

Gap-Filling & Restoration

Suggests missing inscription segments with high accuracy, even on damaged or incomplete artifacts.

Automated Dating & Origin Prediction

Dates inscriptions within ±13 years and attributes geographic origin with 72% accuracy across 62 Roman provinces.

Textual Parallels & Contextual Insights

Instantly retrieves similar inscriptions, formulas, and structures for rapid contextualization.

Open and Adaptable

Model, code, and datasets are open-source (predictingthepast.com), adaptable for other ancient languages and scripts.

Collaborative Research

Designed to enhance human expertise, supporting joint work for greater accuracy in restoration and interpretation.

Use Cases of Aeneas

Fills in ancient Roman inscriptions for museums, digs, and archives.

Dates and places texts to reconstruct the social and political footprint of the Roman Empire.

Supports exhibitions by providing expert-level restoration and context on display artifacts.

Powers interactive demonstrations for students and the public, users can try restoring texts via the free tool.

Aeneas

vs

Human Only

Feature	Aeneas AI	Ithaca (Greek)	Human Only
Languages	Latin (adaptable)	Greek	Any (manual)
Gap Filling (Top-20, 10c)	73%	N/A	~60%
Geographic Attribution	72% (62 provinces)	~65%	Lower (manual)
Dating Accuracy	±13 years	±18 years	±31 years
Multimodal (Text+Image)	Yes	No	N/A
Open Data/Model	Yes (predictingthepast.com)	Yes	N/A