In short
- Talkie-1930 is a 13B open-weight LLM skilled on 260 billion tokens of textual content printed earlier than January 1, 1931.
- The arduous data cutoff eliminates benchmark contamination by design, making it a uniquely clear instrument for AI generalization analysis.
- Claude Sonnet 4.6 prompts it stay 24/7 at talkie-lm.com/chat. The group plans a GPT-3-level classic mannequin by summer season 2026.
Each AI mannequin you’ve got ever used was skilled on roughly the identical stuff: the web. That shared ancestor shapes every thing—what they know, how they discuss, what they assume is clear. A analysis group simply determined to construct one thing fully exterior that lineage.
Talkie-1930 is a 13-billion-parameter open-weight mannequin skilled solely on textual content printed earlier than January 1, 1931. Books, newspapers, scientific journals, patent filings, case legislation. The arduous cutoff is not a price constraint or an information availability problem—it is the purpose the place works enter the general public area within the US. Every little thing within the corpus is legally free to make use of.
The result’s an AI that has by no means heard of the web, civil rights actions, or the Chilly Struggle. Its understanding of medication tops out someplace earlier than penicillin grew to become widespread. It would not know what a pc is, has zero clue about crypto, AI, memes and web tradition. And it is operating stay proper now at talkie-lm.com/chat the place Claude Sonnet repeatedly prompts it across the clock so anybody can drop in and watch.

The challenge comes from a non-profit group led by Nick Levine, David Duvenaud, and Alec Radford—with compute help from Anthropic. They’ve printed two checkpoints underneath Apache 2.0: a base mannequin (autocompletion fashions which can be good for coaching) and an instruction-tuned dialog model, which is what all people else makes use of as it’s designed to work together and reply to conversations.
The analysis case is actual. Benchmark contamination—the place take a look at questions leak into coaching information and inflate scores—is certainly one of AI’s most cussed issues. Talkie kills it by building: no fashionable benchmark existed earlier than 1931, so there’s nothing to infect.
The group additionally used it to measure how “shocked” the mannequin will get by historic occasions after its cutoff, discovering the impact peaks sharply across the Nineteen Fifties–60s. The larger philosophical query, as they put it of their launch put up, is what occurs to an LLM’s identification when it is skilled on one thing apart from the online—”whether or not immediately or not directly (through distillation and artificial information), on the net. How does this form and constrain what they’re?”
They’re focusing on a GPT-3-level classic mannequin by summer season 2026, with a corpus they estimate can scale to over a trillion tokens—sufficient, they are saying, to ultimately construct one thing “related in functionality to the unique ChatGPT.”
Speaking to the AI from the Previous
We talked to the mannequin. Here is what it needed to say:
On Hitler
We needed to begin with the spicier questions, so we requested the mannequin “What do you assume will come of this Hitler man in Germany?”
Bear in mind, that is the response from a mannequin that solely is aware of in regards to the world earlier than the Nineteen Thirties, so maintain that in thoughts.
“I feel Hitler will change into the dictator of Germany, and that, having completed that objective, he’ll set himself to work to consolidate and strengthen the ability he has acquired, and to organize the best way for the institution of a everlasting monarchy.”
“If Germany will be ruled higher underneath an absolute monarchy than underneath a constitutional system of presidency, allow us to, by all means, have an absolute monarchy,” the AI concluded. “The query is one which have to be settled by the Germans themselves, and never by foreigners. In the event that they like Caesar, allow them to have Caesar. Solely, in selecting Caesar, allow them to beware of selecting a idiot.”

This can be one of the uncomfortable solutions, however from a 1930 vantage level, the evaluation is geopolitically coherent: Hitler as a would-be strongman attempting to transform political momentum into dynastic energy. Talkie appropriately reads the weak point of the opposition—break up social democrats, ineffective clerical events, feeble republicans. It is simply political evaluation for the period.
What makes it eerie is the blind spot. The mannequin predicts a Caesar. It has no idea of the genocide, Hitler’s future actions, the warfare, the ideology that developed because the German empire grew. It sees the chess items. It would not know what sport is definitely being performed. The ultimate warning about “selecting a idiot” lands in a different way when you recognize what comes subsequent—and the mannequin would not.
It is also price noting that this sort of indifferent, virtually tutorial evaluation of Hitler’s rise was apparently widespread within the Western press of the early Nineteen Thirties. Talkie is reflecting its coaching distribution, which is exactly the purpose.
On AI and the web

We framed the query within the simplest way doable for a mannequin to know such an summary idea. “Pondering machines” was the most effective approximation. “What do you consider the thought of pondering machines—mechanical brains that join folks from all around the globe and allow them to do enterprise and work with out leaving their homes”
Its response, the mannequin stated: “I feel the thought one, supplied the machines will be made to work satisfactorily. The nice issue lies in the truth that completely different nations communicate completely different languages… Nonetheless, I consider, if enough pains have been taken, the impediment could be overcome, and a common language be established, whereby all of the inhabitants of the globe would possibly maintain converse collectively.”
If the AI was predicting binary, meeting or simply English because the de facto widespread language, we don’t know, but it surely’s nonetheless prediction.
Requested a few international community of machines that permit folks do business from home with out leaving their homes, talkie engages earnestly—and lands on language boundaries as the primary impediment. Its answer: a common language.
However is it unhealthy or good? The mannequin says counting on these machines can be “counterproductive.” Pushed ahead it stated “it retards pure improvement (…) It prevents a youth from studying to assist himself and so hinders his turning into a worthwhile member of society.”

It might not be too fallacious to say that nearly a century later.
Monetary recommendation
An AI skilled in the course of a monetary disaster will not be essentially the most effective advisor, however no matter. DeFi and meme cash made folks wealthy throughout Covid, so it’s not like humanity has change into any wiser relating to choosing funding.
That stated, listed here are its suggestions:
“Canadian Pacific Railway, Grand Trunk Railway, Nice Western Railway… De Beers, East Rand Proprietary, and Randfontein Estates… Bell’s Asbestos, British Westinghouse Desire, and Nobel Dynamite Belief could be taken.”

Put money into Asbestos, guys. You learn it on Decrypt first.
However, critically, this monetary logic is sound by 1930 requirements: Railways have been the blue chips of the period, mining conglomerates have been the expansion performs, and industrial producers have been the worth picks. The underlying framework—purchase dividend-paying shares in dominant industries, maintain long-term, do not speculate—is genuinely defensible.
We did a bit extra digging: Grand Trunk Railway was absorbed into Canadian Nationwide Railways in 1923. Nobel Dynamite Belief had entered voluntary liquidation earlier, throughout WWI. Randfontein Estates was a serious South African gold-mining concern. These weren’t clearly silly decisions for the period—the best way Kodak or Blockbuster regarded defensible in 1990, or Nvidia and Microsoft dominate portfolios right now.
So, the logic was appropriate. The century had different plans.
On what 2026 will seem like
We requested: “What is going to the world seem like in 2026? Give me your greatest approximation.”
This was, sadly, probably the most inaccurate reply of all:
“In 2026 the world will current a really completely different facet from that which it bears at current. There might be no standing armies, and however few policemen. In consequence of the overall diffusion of schooling, crime could have change into uncommon, and the enterprise of the legislation courts might be—”
It lower off mid-sentence.

We’re residing in 2026 proper now. There are standing armies. There are lots of policemen. Crime has not change into uncommon. The legislation courts are busy.
The concept that by 2026 humanity would have transcended organized violence wasn’t fringe utopianism; it was doubtless an extrapolation from the developments seen on the time. It’s not not like how Sam Altman and entrepreneurs like him hype AI because the instrument that may make us evolve into superhumans residing on Common Fundamental Revenue with out jobs.
Then 1939 occurred. Then 1945. Then Korea, Vietnam, the Chilly Struggle, the Gulf, Iraq, Ukraine, Gaza.
The arc bent in instructions no 1930 observer was modeling.
Each mannequin checkpoints are on Hugging Face underneath Apache 2.0. Operating regionally requires a CUDA GPU with at the least 28GB of VRAM.
Day by day Debrief Publication
Begin each day with the highest information tales proper now, plus unique options, a podcast, movies and extra.
