April 27, 2026
Time travel by text, drama on tap
Talkie: a 13B vintage language model from 1930
Chat like it’s 1930: fans wowed, skeptics cry “just roleplay”
TLDR: Talkie trains only on pre‑1931 writing to simulate a 1930s conversational partner and test how well old‑world data predicts the future. Commenters split between wow and “why not just roleplay,” while others gripe about hardware and ask for easy installs—turning a research demo into a culture clash over novelty and practicality.
A new “vintage” AI called Talkie is trained only on pre‑1931 text, promising chats straight out of the past—and the comment section is having a field day. Some folks are charmed by the idea of a 1930s pen pal and the live demo, while others are poking holes in the premise. One camp swoons over the history-nerd fantasy; another asks if this is just a costly cosplay when a modern model could be told to “talk old-timey.”
The project’s own tests say this throwback brain gets increasingly surprised by later decades’ events (especially the 1950s–60s), and can even learn tiny bits of modern coding from examples—think one‑line math, not hacker magic. That spurred two big threads: practicality and principle. Practical people want plug‑and‑play setup via Ollama and complain their GPUs can’t handle a 13B model. Purists cheer a contamination‑free time capsule to study forecasting and creativity; snarkers quip, “Vintage? Where are the vacuum tubes?”
Then there’s the mythology: a top‑liked reply name-drops Steve Jobs dreaming of chatting with Aristotle, framing Talkie as the first true “time machine by text.” Whether it’s groundbreaking science or a stylish roleplay, the crowd is split—but very entertained, and very online.
Key Points
- •“talkie-1930-13b-it” is a 13B language model trained only on pre-1931 text, demoed via a live feed prompted by Claude Sonnet 4.6.
- •Forecasting ability is evaluated using NYT “On This Day” summaries, with surprisingness rising after the cutoff and peaking in the 1950s–1960s before plateauing.
- •The team plans further evaluations to study how forecasting scales with model size and decays over longer horizons.
- •Idea-generation tests examine whether the model can arrive at post-cutoff inventions and theories (e.g., Sikorsky’s helicopter, Turing machines, xerography), echoing Demis Hassabis’s GR question.
- •Contamination-free training enables clean generalization tests like HumanEval in Python; vintage models underperform web-trained ones but show simple solutions and improve with scale.