December 18, 2025
No spoilers: History bot
Trained LLMs exclusively on pre-1913 texts
A history chatbot stuck before 1913 has no WWI—and the comments go feral
TLDR: A team trained “time-locked” chatbots only on old texts, so the 1913 version doesn’t know WWI and mirrors period attitudes. Comments split between loving its surprise factor and worrying about outdated morals, while others dig into chat-tuning—key for studying bias, context, and how AI learns history.
Researchers from the University of Zurich and partners just dropped “Ranke‑4B,” a family of time‑locked history chatbots trained only on texts from before set years like 1913. The 1913 model doesn’t know WWI or Hitler, and its answers sound very old‑timey: it condemns slavery while echoing era‑bound views on women. The team warns they won’t “fix” its morals. Cue comment fireworks: one user links an SMBC comic and quips that pre‑1913 norms clash hard with today. Memes flew: “No spoilers bot,” “WWI hasn’t dropped yet,” and “season finale of history.”
Then the hype crowd shows up. saaaaaam gushes that this thing can be surprised, unlike modern bots with “hindsight contamination” (models that already know how the story ends). Heliodex says the outputs feel less like typical AI because we’re used to modern cutoffs. Teever wants to stress‑test pre‑1905 minds on relativity. Meanwhile, andy99 pokes the devs: how do you “chat‑tune” without rewriting its values—what is “uncontaminated bootstrapping” really? The big fight: time capsule vs. bias machine. Is this brave science—or building bots that repeat outdated beliefs? For now, the internet is rubbernecking and refreshing, waiting for Ranke‑4B’s release notes. Bring popcorn; the comments are the real main event today.
Key Points
- •Ranke-4B is a family of 4B-parameter time-locked historical LLMs with knowledge cutoffs at 1913, 1929, 1933, 1939, and 1946.
- •Models are based on the Qwen3 architecture and trained from scratch on 80B tokens of historical data.
- •Training uses a curated dataset of 600B time-stamped tokens; prerelease notes are available and repositories are forthcoming.
- •The project aims to preserve normative judgments from pretraining for scientific applications and does not endorse model outputs.
- •Example outputs from Ranke-4B-1913 show ignorance of post-1913 events/figures and reflect period-specific views on social issues.