Vintage chatbot lives in the past like an elderly relative
The Register
AI researchers released Talkie, a 13-billion-parameter language model trained exclusively on English-language texts published before 1931 to study how AI systems behave with limited historical knowledge. The model underperforms compared to modern equivalents on standard benchmarks, achieving only 30 percent of performance when trained on OCR-transcribed pre-1931 texts versus human-transcribed versions. The researchers plan to scale Talkie to GPT-3.5 level capability by summer to explore scientific discovery, cultural change interpretation, and how language models form self-conception without knowledge of modern AI.
Why it matters
Talkie's training data stops at the end of 1930, and its creators hope it'll help us better understand how AI thinks
