Apollo, first Ancient Greek LLM, trained on 600 million words

Academic libraries hold hundreds of thousands of Ancient Greek papyrus fragments, many damaged enough that their meaning is probably lost. Restoring a single tattered piece has meant identifying word divisions in a script with no gaps, dating the document, weighing its socio-political context, and consulting references to choose replacement words. Few people in the world are skilled enough at Greek history to do this, as University College London's Stephen Colvin puts it. Apollo is meant to bake that specialized knowledge into a model: Anna Dolganov of the Austrian Academy of Science says that when it sees Homer it supplements Homeric Greek, and when it sees a Doric inscription it uses Doric dialect.

The article places this within a wider run of AI results in research, citing OpenAI's claim that its models solved a 200-year-old math problem and Google DeepMind's dataset mapping how genetic mutations affect molecular biology. Vlitas of Sail Reply says unlocking knowledge this way was unthinkable a year ago, and that the technique could be extended to Latin or Egyptian, or any discipline that benefits from indexing a large body of material. Oxford's Armand D'Angour, whose university holds the world's largest ancient papyrus collection, says a machine offering three possible words for a gap would speed up matters considerably.

The stakes seem to hinge on how the tool is used rather than what it can output. Because a language model deals in probabilities, there is a concern that machine-filled gaps could pollute the historical record; Apollo is built to propose a selection of word options for a scholar to choose between, and Dolganov argues human competence needs to remain, warning that total reliance on AI transcriptions and interpretations is where problems start. Gains are likely to be incremental — small details about life in antiquity and support for existing scholarly assumptions — rather than headline discoveries, but for the small group of specialists able to read these fragments, the bottleneck Apollo targets is a real one.