The Workbench2 published
Experiments
Hands-on reconstruction, replication and practical exploration of historical ideas. Each write-up shows its full setup, its costs and its failures, and keeps anything a model wrote clearly labeled.
Raise a Stone
Can small fine-tuned AI models read Viking runestones better than frontier AI? A two-stage test on 2,316 runestones, with name-swapped stones that no model could have memorized
See what’s comingWhat the Papers Held
Mapping eighty years of perspectives in local American newspapers, 1850–1929, from the broadest idea down to the page it was printed on
Read the experimentUnearthing What Might Have Been
Can a language model sketch what a lost Epictetus lecture might have been like? Training fell short. Describing the teacher, and asking the model to interpret before it writes, worked better.
Read the experiment