New AI Trained Only on Pre-1930 Data Speaks Like the Most Old-Timey Guy Imaginable

Researchers introduced Talkie, a 13-billion-parameter language model trained only on books and newspapers published before 1930.

Talkie generates convincing old-timey speech, but tests found it can still leak later facts and make anachronistic predictions.

The results suggest historical-data-only models can still generalize beyond their era, while also exposing limits on keeping outputs temporally pure.

The study raises broader questions about what language models can reason, extrapolate, and discover when confined to a past time period.