talkie-1930
A 13-billion-parameter language model trained only on English published before 1931, so its picture of the world genuinely stops at the end of 1930.
Ask talkie-1930 what the world looks like in 2026 and it describes a Europe of a billion people wintering in Paris and summering in London, crossed by iron railroads and linked to New York by ten-day steamship — and it is not sure whether another great war is coming, since "the madness of 1914-1918 has passed away." This is not a modern model doing a period voice. It read nothing published after 31 December 1930, and so it has no idea what happened next. The date is not arbitrary: it is the United States public-domain boundary, which is what made 260 billion words of books, newspapers, journals, patents and case law legal to train on.
The same idea appeared recently at toy scale in Mr. Chatterbox, a from-scratch Victorian model. What makes this one notable is the company it keeps: one of its three creators is Alec Radford, lead author of the papers that introduced GPT and GPT-2. A frontier-lineage researcher has spent real compute growing a mind that is deliberately, verifiably ignorant — which turns a curiosity into a clean instrument. Because nothing after 1930 could have leaked in, it is a contamination-free way to test whether a model is reasoning or just remembering: shown a small cipher function it had never seen, it wrote the matching decoder, despite there being no computer code anywhere in its training.
The frozen worldview is the pitch, not a clean fact. Filtering was imperfect — the model knows a little about Roosevelt, the New Deal, even the Second World War and the peace that followed. Its coding never rises above one-line programs cribbed from the example in front of it, so read the reasoning result as a hint, not a headline. And its manners are a tell: to make it answer politely, the team let 2026 models grade its replies, which their report admits "inevitably shapes talkie's behavior anachronistically." A mind sealed in 1930, taught to converse by machines from the future.
The lenses
The facts
Concepts
How this connects
Tap a node to open it