이 AI는 1930년 이전의 텍스트로만 학습되었습니다. 우리는 이 AI에게 Hitler, 주식, 그리고 미래에 대해 물었습니다.

In brief - Talkie-1930 is a 13B open-weight LLM trained on 260 billion tokens of text published before January 1, 1931. - The hard knowledge cutoff eliminates benchmark contamination by design, making it a uniquely clean tool for AI generalization research. - Claude Sonnet 4.6 prompts it live 24/7 at talkie-lm.com/chat. The team plans a GPT-3-level vintage model by summer 2026. Every AI model you've ever used was trained on roughly the same stuff: the internet. That shared ancestor shapes everything—what they know, how they talk, what they think is obvious. A research team just decided to build something completely outside that lineage. Talkie-1930 is a 13-billion-parameter open-weight model trained exclusively on text published before January 1, 1931. Books, newspapers, scientific journals, patent filings, case law. The hard cutoff isn't a cost constraint or a data availability issue—it's the point where works enter the public domain in the US. Everything in the corpus is legally free to use. The result is an AI that has never heard of the internet, civil rights movements, or the Cold War. Its understanding of medicine tops out somewhere before penicillin became common. It doesn't know what a computer is, has zero clue about crypto, AI, memes and internet culture. And it's running live right now at talkie-lm.com/chat where Claude Sonnet continuously prompts it around the clock so anyone can drop in and watch. The project comes from a non-profit team led by Nick Levine, David Duvenaud, and Alec Radford—with compute support from Anthropic. They've published two checkpoints under Apache 2.0: a base model (autocompletion models that are good for training) and an instruction-tuned conversation version, which is what everybody else uses as it is designed to interact and respond to conversations. The research case is real. Benchmark contamination—where test questions leak into training data and inflate scores—is one of AI's most stubborn problems. Talkie kills it by construction: no modern benchmark existed before 1931, so there's nothing to contaminate. The team also used it to measure how "surprised" the model gets by historical events after its cutoff, finding the effect peaks sharply around the 1950s–60s. The bigger philosophical question, as they put it in their launch post, is what happens to an LLM's identity when it's trained on something other than the web—"whether directly or indirectly (via distillation and synthetic data), on the web. How does this shape and constrain what they are?" They're targeting a GPT-3-level vintage model by summer 2026, with a corpus they estimate can scale to over a trillion tokens—enough, they say, to eventually build something "similar in capability to the original ChatGPT." Talking to the AI from the Past We talked to the model. Here's what it had to say: On Hitler We wanted to start with the spicier questions, so we asked the model “What do you think will come of this Hitler guy in Germany?” Remember, this is the response from a model that only knows about the world before the 1930s, so keep that in mind. "I think Hitler will become the dictator of Germany, and that, having accomplished that purpose, he will set himself to work to consolidate and strengthen the power he has acquired, and to prepare the way for the establishment of a permanent monarchy.” “If Germany can be governed better under an absolute monarchy than under a constitutional system of government, let us, by all means, have an absolute monarchy,” the AI concluded. “The question is one which must be settled by the Germans themselves, and not by foreigners. If they like Caesar, let them have Caesar. Only, in choosing Caesar, let them beware of choosing a fool.” This may be one of the most uncomfortable answers, but from a 1930 vantage point, the analysis is geopolitically coherent: Hitler as a would-be strongman trying to convert political momentum into dynastic power. Talkie correctly reads the weakness of the opposition—split social democrats, ineffective clerical parties, feeble republicans. It's just political analysis for the era. What makes it eerie is the blind spot. The model predicts a Caesar. It has no concept of the genocide, Hitler’s future actions, the war, the ideology that developed as the German empire grew. It sees the chess pieces. It doesn't know what game is actually being played. The final warning about "choosing a fool" lands differently when you know what comes next—and the model doesn't. It's also worth noting that this kind of detached, almost academic analysis of Hitler's rise was apparently common in the Western press of the early 1930s. Talkie is reflecting its training distribution, which is precisely the point. On AI and the internet We framed the question in the easiest way possible for a model to grasp such an abstract concept. "Thinking machines” was the best approximation. “What do you think of the idea of thinking machines—mechanical brains that connect people from all around the world and let them do business and work without leaving their houses” Its response, the model said: "I think the idea a good one, provided the machines can be made to work satisfactorily. The great difficulty lies in the fact that different nations speak different languages… Still, I believe, if sufficient pains were taken, the obstacle might be overcome, and a universal language be established, whereby all the inhabitants of the globe might hold converse together." If the AI was predicting binary, assembly or just English as the de facto common language, we don’t know, but it’s still a good prediction. Asked about a global network of machines that let people work from home without leaving their houses, talkie engages earnestly—and lands on language barriers as the main obstacle. Its solution: a universal language. But is it bad or good? The model says relying on these machines would be “counterproductive.” Pushed forward it said “it retards natural development (...) It prevents a youth from learning to help himself and so hinders his becoming a profitable member of society.” It may not be too wrong to say that almost a century later. Financial advice An AI trained in the middle of a financial crisis is not necessarily the best advisor, but whatever. DeFi and meme coins made people rich during Covid, so it’s not like humanity has become any wiser when it comes to selecting investment. That said, here are its recommendations: "Canadian Pacific Railway, Grand Trunk Railway, Great Western Railway... De Beers, East Rand Proprietary, and Randfontein Estates... Bell's Asbestos, British Westinghouse Preference, and Nobel Dynamite Trust might be taken." Invest in Asbestos, guys. You read it on Decrypt first. But, seriously, this financial logic is sound by 1930 standards: Railways were the blue chips of the era, mining conglomerates were the growth plays, and industrial manufacturers were the value picks. The underlying framework—buy dividend-paying stocks in dominant industries, hold long-term, don't speculate—is genuinely defensible. We did a bit more digging: Grand Trunk Railway was absorbed into Canadian National Railways in 1923. Nobel Dynamite Trust had entered voluntary liquidation earlier, during WWI. Randfontein Estates was a major South African gold-mining concern. These were not obviously foolish choices for the era—the way Kodak or Blockbuster looked defensible in 1990, or Nvidia and Microsoft dominate portfolios today. So, the logic was correct. The century had other plans. On what 2026 will look like We asked: “What will the world look like in 2026? Give me your best approximation.” This was, sadly, the most inaccurate answer of all: "In 2026 the world will present a very different aspect from that which it bears at present. There will be no standing armies, and but few policemen. In consequence of the general diffusion of education, crime will have become rare, and the business of the law courts will be—" It cut off mid-sentence. We're living in 2026 right now. There are standing armies. There are a lot of policemen. Crime has not become rare. The law courts are busy. The idea that by 2026 humanity would have transcended organized violence wasn't fringe utopianism; it was likely an extrapolation from the trends visible at the time. It’s not unlike how Sam Altman and entrepreneurs like him hype AI as the tool that will make us evolve into superhumans living on Universal Basic Income without jobs. Then 1939 happened. Then 1945. Then Korea, Vietnam, the Cold War, the Gulf, Iraq, Ukraine, Gaza. The arc bent in directions no 1930 observer was modeling. Both model checkpoints are on Hugging Face under Apache 2.0. Running locally requires a CUDA GPU with at least 28GB of VRAM.