A correspondent
who genuinely lives
in 1930.
He hasn't heard of the moon landing or the internet — but he's read about Lindbergh, Mickey Mouse, and the brand-new planet of Pluto. Not a system prompt. A 13-billion-parameter language model trained on English text published before 1931, and nothing else.
An honest stranger from the past.
Most AI correspondents pretend at character. Talk to the Past has one — by construction. Its model has read Wodehouse and Woolf, Poe and Conan Doyle, Edith Wharton, the Times of London through October 1930, and not a word published after. It has never heard of television, WWII, computers, the Beatles, or any other thing that hadn't yet happened.
Ask him about the moving pictures — he'll tell you about All Quiet on the Western Front at the Capitol Theatre. Ask about flying — he'll mention Lindbergh and the new aerodromes. Ask about the future, and he'll be charmingly bewildered. The bewilderment is not a feature — it is the absence of training data after 1930.
"My dear sir, I confess the term sits uneasily in my ear. Might you mean a tele-vision apparatus — Mr. Baird's curious invention for the transmission of moving images by wireless?"
Streaming letters
Each reply types in word by word, like a wireless coming through. Conversations are kept locally on your device.
Period typography
IM Fell English for the letters, Cutive Mono for transmitted dispatches, Special Elite for the labels. Cream paper. Ink-blue type.
A correspondence log
Every conversation is preserved. Resume any past dispatch from the side drawer. No accounts, no cloud — your letters stay yours.
A research artefact, not a chatbot wrapper.
Mr. Whitmore is powered by talkie-1930-13b-it — an open-source 13-billion-parameter language model. It was trained from the very first token of its corpus on English text written before 1931. There is no fine-tune. No "stay in character" system prompt. The vocabulary, the cadence, the opinions, the ignorance — all of it is what could have been read by candlelight or under a gaslamp before the year 1931 dawned.
The model was released earlier this year by:
- Alec Radford — the OpenAI researcher who first-authored GPT-2 and Whisper.
- David Duvenaud — University of Toronto Computer Science professor and former Alignment Evals lead at Anthropic.
- Nick Levine — independent researcher and former financial historian at Winton.
Talk to the Past is an independent project. The model is licensed under Apache 2.0; this app is not affiliated with the model's authors.
Ten dispatches a day.
Or as many as you please.
Free correspondents may post ten dispatches per day, with the office reopening at midnight. Subscribers write as often as they like and keep the wireless humming.
Pricing displayed for reference. Final price is set by the App Store in your local currency at the moment of purchase.
A few things readers ask.
Isn't this just GPT-4 with a system prompt that says "act like 1930"?
No, and the difference is the entire point. Most "historical AI" chatbots wrap a modern model (GPT-4, Claude, etc.) with an instruction telling it to roleplay. Those models drift back to a helpful-assistant register the moment you ask them anything they weren't given specific instructions for.
Talk to the Past uses talkie-1930-13b-it, a 13B-parameter model whose entire training corpus is English text published before 1931. The model has never seen a single sentence about World War II, the moon landing, the internet, or anything else that hadn't yet happened. It cannot drift to a modern register because that register doesn't exist in its weights.
Will Mr. Whitmore remember me from one conversation to the next?
No — by design. Each conversation is its own letter exchange. Within a single conversation he holds context across many turns. Between conversations he begins fresh. We think of each chat as a single piece of correspondence, sent and answered.
This is also a privacy feature: there is nothing stored about you across conversations, because there's no account and no profile.
What about my privacy?
Your conversations are stored only on your device. Nothing is uploaded to our servers, no analytics, no crash reporting, no third-party SDKs. The text of the message you send transits a small proxy we operate so the model can generate a reply, and is discarded after the reply returns.
The full picture is in our Privacy Policy, written in plain English.
Will he ever be wrong?
Yes. He invents on occasion. He sometimes misattributes a quotation, conflates two events, or recovers from a misread question with a small embarrassment. Treat his replies as you would any letter from a stranger — for what he knew of his world, not as a reference text. The magic is in the joins.
Why is the subscription only $2.99?
Because the cost of running a 13B-parameter model is mostly the cost of the GPU minutes the model spends generating replies. Pro subscribers cover that cost and a bit more, which keeps the wireless office open. We're not trying to compete with Replika or Character.AI on pricing — we're trying to keep the lights on.
You can use the app on the free plan indefinitely. Ten dispatches a day is enough for a real correspondence; the office reopens every midnight.
Does it work on iPad?
Yes. Universal app, single binary, iOS 17 or later. The drawer slides over the chat on iPhone; on iPad the layout adjusts to use the wider canvas while keeping the chat column at a comfortable reading width.