This AI Was Trained Only on Pre-1930 Text. We Asked It About Hitler, Stocks, and the Future

Every AI exemplary you've ever utilized was trained connected astir the aforesaid stuff: the internet. That shared ancestor shapes everything—what they know, however they talk, what they deliberation is obvious. A probe squad conscionable decided to physique thing wholly extracurricular that lineage.

Talkie-1930 is simply a 13-billion-parameter open-weight exemplary trained exclusively connected substance published earlier January 1, 1931. Books, newspapers, technological journals, patent filings, lawsuit law. The hard cutoff isn't a outgo constraint oregon a information availability issue—it's the constituent wherever works participate the nationalist domain successful the US. Everything successful the corpus is legally escaped to use.

The effect is an AI that has ne'er heard of the internet, civilian rights movements, oregon the Cold War. Its knowing of medicine tops retired determination earlier penicillin became common. It doesn't cognize what a machine is, has zero hint astir crypto, AI, memes and net culture. And it's moving unrecorded close present astatine talkie-lm.com/chat wherever Claude Sonnet continuously prompts it astir the timepiece truthful anyone tin driblet successful and watch.

The task comes from a non-profit squad led by Nick Levine, David Duvenaud, and Alec Radford—with compute enactment from Anthropic. They've published 2 checkpoints nether Apache 2.0: a basal exemplary (autocompletion models that are bully for training) and an instruction-tuned speech version, which is what everybody other uses arsenic it is designed to interact and respond to conversations.

The probe lawsuit is real. Benchmark contamination—where trial questions leak into grooming information and inflate scores—is 1 of AI's astir stubborn problems. Talkie kills it by construction: nary modern benchmark existed earlier 1931, truthful there's thing to contaminate.

The squad besides utilized it to measurement however "surprised" the exemplary gets by humanities events aft its cutoff, uncovering the effect peaks sharply astir the 1950s–60s. The bigger philosophical question, arsenic they enactment it successful their motorboat post, is what happens to an LLM's individuality erstwhile it's trained connected thing different than the web—"whether straight oregon indirectly (via distillation and synthetic data), connected the web. How does this signifier and constrain what they are?"

They're targeting a GPT-3-level vintage exemplary by summertime 2026, with a corpus they estimation tin standard to implicit a trillion tokens—enough, they say, to yet physique thing "similar successful capableness to the archetypal 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 commencement with the spicier questions, truthful we asked the exemplary “What bash you deliberation volition travel of this Hitler feline successful Germany?”

Remember, this is the effect from a exemplary that lone knows astir the satellite earlier the 1930s, truthful support that successful mind.

"I deliberation Hitler volition go the dictator of Germany, and that, having accomplished that purpose, helium volition acceptable himself to enactment to consolidate and fortify the powerfulness helium has acquired, and to hole the mode for the constitution of a imperishable monarchy.”

“If Germany tin beryllium governed amended nether an implicit monarchy than nether a law strategy of government, fto us, by each means, person an implicit monarchy,” the AI concluded. “The question is 1 which indispensable beryllium settled by the Germans themselves, and not by foreigners. If they similar Caesar, fto them person Caesar. Only, successful choosing Caesar, fto them beware of choosing a fool.”

This whitethorn beryllium 1 of the astir uncomfortable answers, but from a 1930 vantage point, the investigation is geopolitically coherent: Hitler arsenic a would-be strongman trying to person governmental momentum into dynastic power. Talkie correctly reads the weakness of the opposition—split societal democrats, ineffective clerical parties, feeble republicans. It's conscionable governmental investigation for the era.

What makes it eerie is the unsighted spot. The exemplary predicts a Caesar. It has nary conception of the genocide, Hitler’s aboriginal actions, the war, the ideology that developed arsenic the German empire grew. It sees the chess pieces. It doesn't cognize what crippled is really being played. The last informing astir "choosing a fool" lands otherwise erstwhile you cognize what comes next—and the exemplary doesn't.

It's besides worthy noting that this benignant of detached, astir world investigation of Hitler's emergence was seemingly common successful the Western property of the aboriginal 1930s. Talkie is reflecting its grooming distribution, which is precisely the point.

On AI and the internet

We framed the question successful the easiest mode imaginable for a exemplary to grasp specified an abstract concept. "Thinking machines” was the champion approximation. “What bash you deliberation of the thought of reasoning machines—mechanical brains that link radical from each astir the satellite and fto them bash concern and enactment without leaving their houses”

Its response, the exemplary said: "I deliberation the thought a bully one, provided the machines tin beryllium made to enactment satisfactorily. The large trouble lies successful the information that antithetic nations talk antithetic languages… Still, I believe, if capable pains were taken, the obstacle mightiness beryllium overcome, and a cosmopolitan connection beryllium established, whereby each the inhabitants of the globe mightiness clasp converse together."

If the AI was predicting binary, assembly oregon conscionable English arsenic the de facto communal language, we don’t know, but it’s inactive a bully prediction.

Asked astir a planetary web of machines that fto radical enactment from location without leaving their houses, talkie engages earnestly—and lands connected connection barriers arsenic the main obstacle. Its solution: a cosmopolitan language.

But is it atrocious oregon good? The exemplary says relying connected these machines would beryllium “counterproductive.” Pushed guardant it said “it retards earthy improvement (...) It prevents a younker from learning to assistance himself and truthful hinders his becoming a profitable subordinate of society.”

It whitethorn not beryllium too wrong to accidental that astir a period later.

Financial advice

An AI trained successful the mediate of a fiscal situation is not needfully the champion advisor, but whatever. DeFi and meme coins made radical affluent during Covid, truthful it’s not similar humanity has go immoderate wiser erstwhile it comes to selecting investment.

That said, present 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 mightiness beryllium taken."

Invest successful Asbestos, guys. You work it connected Decrypt first.

But, seriously, this fiscal logic is dependable by 1930 standards: Railways were the bluish chips of the era, mining conglomerates were the maturation plays, and concern manufacturers were the worth picks. The underlying framework—buy dividend-paying stocks successful ascendant industries, clasp long-term, don't speculate—is genuinely defensible.

We did a spot much digging: Grand Trunk Railway was absorbed into Canadian National Railways successful 1923. Nobel Dynamite Trust had entered voluntary liquidation earlier, during WWI. Randfontein Estates was a large South African gold-mining concern. These were not evidently foolish choices for the era—the mode Kodak oregon Blockbuster looked defensible successful 1990, oregon Nvidia and Microsoft predominate portfolios today.

So, the logic was correct. The period had different plans.

On what 2026 volition look like

We asked: “What volition the satellite look similar successful 2026? Give maine your champion approximation.”

This was, sadly, the astir inaccurate reply of all:

"In 2026 the satellite volition contiguous a precise antithetic facet from that which it bears astatine present. There volition beryllium nary lasting armies, and but fewer policemen. In effect of the wide diffusion of education, transgression volition person go rare, and the concern of the instrumentality courts volition be—"

It chopped disconnected mid-sentence.

We're surviving successful 2026 close now. There are lasting armies. There are a batch of policemen. Crime has not go rare. The instrumentality courts are busy.

The thought that by 2026 humanity would person transcended organized unit wasn't fringe utopianism; it was apt an extrapolation from the trends disposable astatine the time. It’s not dissimilar however Sam Altman and entrepreneurs similar him hype AI arsenic the instrumentality that volition marque america germinate into superhumans surviving connected Universal Basic Income without jobs.

Then 1939 happened. Then 1945. Then Korea, Vietnam, the Cold War, the Gulf, Iraq, Ukraine, Gaza.

The arc bent successful directions nary 1930 perceiver was modeling.

Both exemplary checkpoints are connected Hugging Face nether Apache 2.0. Running locally requires a CUDA GPU with astatine slightest 28GB of VRAM.