Raise a Stone

Can small fine-tuned AI models read Viking runestones better than frontier AI?

Viking carvers wrote with just sixteen runes and no spelling rules. One rune could stand for several sounds, and the same word could be carved many different ways. This experiment asks whether small open models, trained on a home graphics card, can learn to read runes the way scholars do, and how close they come to the largest commercial AIs.

The write-up is getting its finishing touches. When it opens, you will be able to follow every step and read the stones alongside the models.

A Norse carver chisels a decorated runestone while another man studies its inscriptions beside a fjord. — A carver at work on a memorial stone.

The alphabet

Sixteen runes for many sounds

By the Viking Age the older 24-rune alphabet had been cut to sixteen. Here is each rune with the letter scholars use for it. Reading a stone means choosing which sound each rune stood for.

The last rune, written R, marks a second r-sound found at the end of many words.

What is being prepared

Inside the experiment

Two stages
Small models trained on a home graphics card are tested against one frontier AI. Then a larger model, trained on a single rented GPU, faces a stronger one.
Stones no model has seen
Personal names are swapped between stones, so a model cannot answer from memory. It has to read.
Read along
Hands-on figures let you read inscriptions word by word and compare each model’s reading with the scholars’.

Meanwhile

Raise a Stone

Sixteen runes for many sounds

Inside the experiment

Two stages

Stones no model has seen

Read along

Explore what is open now