A Million Unread Words: AI Is Finally Opening Ancient Greek's Biggest Puzzle
```html
Somewhere in Vienna, there's a library holding tens of thousands of ancient Greek papyri that nobody has ever read. Not in 2,000 years. Not ever. If you've ever felt smug about finishing a Sunday crossword, maybe hold off for a second.
The World's Oldest Unsolved Word Problem
About one million Greek papyri exist worldwide. One million. And a staggering number of them have never been read, deciphered, or even catalogued properly. They're sitting in collections, fragile and faded, carrying words that haven't been seen since Roman times.
The Papyrus Collection of the Austrian National Library alone holds tens of thousands of unread texts. That's not a backlog. That's a civilizational puzzle with nobody at the table.
Some of those texts are probably shopping lists. Some might be poetry nobody has encountered in twenty centuries. You won't know until someone reads them. And until recently, "someone" meant generations of scholars working through fragments one at a time, for years per document.
Enter Apollo
The Austrian Academy of Sciences (OeAW) teamed up with Mistral AI and Sail Reply (a Reply Group Company) to build something genuinely remarkable: a large language model trained specifically on ancient Greek. They named it Apollo, after the Greek god of light and patron of the arts and sciences. The naming is doing real work here. This is a model meant to illuminate.
Project leader Anna Dolganov, an ancient historian and papyrologist at the Austrian Archaeological Institute of the OeAW, is heading the effort. When you want to decode ancient texts, you lead with someone who has spent their career decoding ancient texts. Sensible.
What Apollo Actually Does
Apollo handles two things that used to eat years: advanced searching through ancient Greek papyri and inscriptions, and automatic text restoration. That second one is the astonishing part. When a papyrus is torn, damaged, or faded into near-nothing, Apollo suggests what the missing words likely were, based on the largest digital corpus of historical Greek ever assembled.
For word lovers: imagine a crossword where half the letters are gone, the grid is 2,000 years old, and the answer key was lost in Alexandria. Apollo is the solver.
Content that would take years to capture can now happen in hours. Researchers aren't just moving faster. They're suddenly able to ask questions that were previously impossible to even formulate.
The First of Its Kind
Apollo is described as the world's first advanced multimodal LLM for an ancient language. That word "multimodal" matters. It means the model works with images of the actual papyri, not just transcribed text. It can look at a scrap of papyrus the way you'd look at a mostly-blank Wordle grid and make educated guesses about what belongs where.
Ancient Greek has been studied obsessively for centuries. Scholars have debated its grammar, catalogued its dialects, and argued about its pronunciation across generations of academic careers. And still: a million unread texts. The gap between knowing a language and having actually read what people wrote in it turns out to be oceanic.
Why This Should Delight You
Word games are, at their core, about finding meaning in patterns. You look at a scrambled set of letters and your brain lights up trying to find the word hiding inside. That's the same instinct driving ancient Greek scholarship, just scaled up by millennia and extreme papyrus damage.
Apollo isn't replacing that instinct. It's amplifying it. Researchers still need to know what questions to ask, what restorations make historical sense, what a word means in context. The model is the assistant. The humans are still playing the game.
One million unread words, waiting. Apollo just made it possible to finally start reading them.
Source: Languagehat
```