Google DeepMind's Aeneas Lets Historians Decode Ancient Rome Without Writing Code

Epigraphy -- the study of ancient inscriptions -- has always been a discipline that demands years of specialized training. Deciphering a damaged Latin altar, dating a fragmentary Greek oracle tablet, or mapping religious practices across the Roman Empire requires not just historical knowledge, but the ability to navigate complex computational tools. Google DeepMind just removed that last barrier with the Predicting the Past Skill, a new addition to Google Antigravity that lets historians run sophisticated AI-powered analysis through plain English conversation.

A decade of work, now accessible to everyone

This release is the culmination of nearly a decade of collaboration between DeepMind and epigraphers. The milestones are two specialized models, both published in Nature:

• Ithaca (2022) -- a deep neural network for ancient Greek inscriptions. While Ithaca alone achieves 62% accuracy when restoring damaged texts, historians using Ithaca improved their own accuracy from 25% to 72%. It can also attribute inscriptions to their original location with 71% accuracy and date them to within 30 years of ground-truth ranges.
• Aeneas (2025) -- the successor model, focused on Latin inscriptions. Aeneas can process both text and image input and outperforms other state-of-the-art models at restoring missing characters in damaged inscriptions. It is designed to assist historians with epigraphy, automating key tasks: dating an inscription, identifying the region of origin, reconstructing partial inscriptions, and identifying parallels -- inscriptions with similar words or phrasing.

While trained for Latin, Aeneas can be adapted to other ancient languages, scripts and media, from papyri to coinage. An interactive version is freely available to researchers, students and educators at predictingthepast.com, and its code and dataset are open source.

The bottleneck this solves

Even with Ithaca and Aeneas available online, using them at research scale was painful. The DeepMind team identified three concrete blockers that historians kept hitting:

• Creating custom visualizations for individual inscriptions required bespoke code every time.
• Cross-corpus analysis -- finding patterns across thousands of inscriptions -- required programming skills most historians don't have.
• The training data for these models is tiny compared to general-purpose LLMs. There simply aren't enough high-quality scans of inscriptions to train a language model to learn these tasks, which is why specialized solutions like Aeneas are needed. But that specialization also means Gemini alone can't do this work reliably without grounding.

The Predicting the Past Skill addresses all three at once by wiring Aeneas and Ithaca directly into Gemini's reasoning loop inside Google Antigravity.

What Antigravity Skills actually are

Google Antigravity is an agentic development platform designed to help you work in the era of agents, serving as a central command center for launching, monitoring, and orchestrating AI agent activities. The "Skill" mechanism is the key architectural idea here. Skills represent a shift from monolithic context loading to progressive disclosure. Rather than forcing the model to memorize every possible capability at the start of a session, Skills allow developers to package specialized expertise into modular, discoverable units. The model is exposed only to a lightweight menu of these capabilities and loads the heavy procedural knowledge only when the user's intent matches a specific skill.

In practice, this means Gemini stays lean and fast for general queries, but the moment you ask something like "restore the missing text in this inscription and map parallels across the Rhine provinces," it loads the full Aeneas and Ithaca toolchain into context and executes a multi-step research workflow. No Python. No API calls. No configuration.

Three real case studies

DeepMind worked with Dr. Thea Sommerschield, a historian and epigrapher at Durham University who co-led both the Ithaca and Aeneas projects, to validate the Skill across three distinct historical problems:

1. The ring thief of Aquae Sulis -- A Latin curse tablet from Roman Bath, written by a woman named Basilia cursing whoever stole her silver ring. The Skill attributed and analyzed the inscription while generating what amounts to an epigraphic commentary -- explaining which textual features drove the historical attribution, not just the conclusion.
2. Mapping the cult of the Aufaniae -- A votive altar from Mainz dedicated to Germanic mother-goddesses. The Skill moved beyond the single inscription to identify patterns across an entire corpus of similar dedications from the Rhine and Danube provinces, tracing how religious practices traveled with Roman soldiers and administrators across the Empire.
3. Who came to Dodona? -- Thousands of fragmentary lead oracle tablets from the Greek sanctuary of Dodona. Aeneas reasons across thousands of Latin inscriptions, retrieving textual and contextual parallels in seconds that allow historians to interpret and build upon the model's findings. Applied to Dodona, the Skill reconstructed the wider community of visitors -- mapping a network of connected individuals moving through the ancient Mediterranean.

How it was built

Aeneas uses a multimodal transformer architecture as its backbone, with dedicated heads for each task. To train the model, DeepMind assembled the Latin Epigraphic Dataset (LED), a corpus containing 176,861 inscriptions, starting with existing source datasets and creating a complex pipeline to clean the records and combine them into a single format. The data spans inscriptions from the 7th century BCE to the 8th century CE, from regions covering the Roman world from Britain to Mesopotamia.

When evaluated on several epigraphic tasks, Aeneas outperformed the state-of-the-art AI model and human historians. When human historians used Aeneas as a tool on these same tasks, the combined performance was even better. The Predicting the Past Skill takes this human-AI collaboration model and scales it: instead of requiring historians to interact with a raw model interface, Gemini orchestrates the entire workflow and presents results in natural language with auto-generated visualizations.

Internal evaluations of the Science Skills framework show that adding skills improves the reliability and efficiency of scientific workflow tasks: more tasks succeed and the number of tokens and model calls necessary to complete tasks is reduced. With skills, smaller Flash-category models attain Pro-level reliability.

What this means beyond ancient history

The deeper story here is a template for scientific AI. As one Oxford professor put it, "Aeneas represents a concerted effort to bring the power of AI to bear on the complex uncertainties of humanities research. In a field of this sort, the answers are not black and white, and the data is simultaneously rich and incomplete." The challenge of grounding a general-purpose LLM in a narrow, high-stakes domain -- where hallucinations aren't just wrong but historically misleading -- is one every scientific field faces.

The Skill architecture offers a concrete answer: keep the specialized model doing what it's good at (precise, evidence-based inference over a curated corpus), let Gemini handle the reasoning, orchestration, and natural language interface, and use the Skill mechanism to wire them together on demand. Aeneas goes a step further than Ithaca, helping historians interpret and contextualize a text, give meaning to isolated fragments, draw richer conclusions and piece together a better understanding of ancient history. The Predicting the Past Skill makes all of that accessible without a single line of code.

The Skill is available now through Google Antigravity's Science use cases. Google Antigravity is available in public preview at no cost for individuals. The underlying Aeneas and Ithaca models are fully open-sourced on GitHub, and the interactive platform remains accessible at predictingthepast.com for those who want direct model access without the agentic layer.