Vintage chatbot lives in the past like an elderly relative

Three AI researchers just released Talkie, a language model that has never heard of World War II, the New Deal, or microwave ovens. Its training data stops cold at the end of 1930, the current public domain cutoff in the US, so it knows about Betty Boop and flappers but nothing about anything that came after. The team calls it the largest vintage language model anyone has built so far, and they say they're planning to scale it up considerably.

The idea isn't just novelty chatbot cosplay, though there's plenty of that appeal too. David Duvenaud, a University of Toronto professor and one of Talkie's three creators, told The Register the project is meant to test things modern AI benchmarks can't touch. Can a model with only 1911-era knowledge stumble toward general relativity the way Einstein did, a test Demis Hassabis has floated as a real marker of AGI? Can it help study how laws were understood at the time they were written, based on the assumptions baked into period language? Can talking to a model that has no idea what an LLM even is reveal something about how these systems build self-conception?

The results so far are humble. Pitted against an identically architected model trained on modern data, Talkie lost on most standard evaluations despite burning the same amount of compute. Its coding skills topped out at one-line Python snippets like adding two numbers. The team blames a lot of this on OCR noise: since nothing was born digital in 1930, every scrap of training text had to be scanned and transcribed from physical pages, and that process introduces garbage that pristine modern datasets don't have. Training on raw OCR text got Talkie to only 30 percent of the performance of a model trained on clean, human-transcribed copies of the same material; regex cleanup pushed that to 70 percent, still not good enough for the team's taste.

Talkie also leaks the future it's not supposed to know about, a glitch the researchers call temporal leakage. Ask it about FDR's 1936 legislative record and it'll answer, proof that some post-1930 material slipped through filtering. Ask about the Nazis and it correctly IDs them as an antisemitic, authoritarian German party, then confidently names their leader as Hermann Joseph von Hitler, a man supposedly born in 1870. It's wrong, but at least it's wrong in a way that isn't spewing Nazi talking points, which is more than can be said for some far newer chatbots.

The creators say a GPT-3-class version could arrive by summer, built from a corpus they estimate could grow past a trillion tokens of historical text. Until then, Talkie is downloadable on GitHub and Hugging Face, with a web demo carrying a blunt warning that moderation only kicks in after the model has already said something offensive. Given that it's an amateur research effort with no illusions about closing the gap with modern frontier models, that trade-off seems to be the whole point: a smaller, weirder AI that only knows what humanity knew before it started keeping better records of its own worst mistakes.