MonadGPT – What would have happened if ChatGPT was invented in the 17th century?
A 7B-parameter language model fine-tuned on 17th–18th century texts, MonadGPT aims to emulate how an early modern thinker might write and reason, enabling immersive “time travel” conversations for history education and experimentation. Commenters explore use cases such as simulating debates between historical figures, teaching cultural context, and building period-specific variants (e.g., Roman-era or Latin-only models), while noting technical constraints like limited historical corpora, modern data “leaking” into answers, and repetitive output. Many find the results both entertaining and revealing, as the model’s archaic style exposes past beliefs about science, medicine, religion, and politics, and highlights how LLMs are bound by their training data.
Historical Education & Use Cases
- Many see strong potential for teaching history, especially as an “orientation for time travel” to past cultures rather than just dates and events.
- Suggested uses: simulated conversations with early modern figures, debates between historical thinkers, and historically grounded NPCs in games.
- Interest in extending the idea to other periods and languages: Roman-era Latin, medieval/early modern French, “the 80s,” Old/early English, and specialized thinkers (e.g., Foucault).
Training Data & Linguistic Choices
- Model is fine‑tuned from a modern instruction-following base (Mistral/OpenHermes) on early modern corpora; not trained from scratch on old texts due to data limits.
- Dataset is published; future larger French historical model (1000–2000) is planned.
- Conscious choice to preserve historical spelling and typography (including long‑s confusion), though tokenizers dislike it.
- Debate over whether to normalize typography or run a “dictionary step” first.
Model Behavior & Examples
- Users share entertaining outputs: advice on essays, headaches, lightning, love melancholy, crypto-currency, and speculative 2023, all in archaic style.
- Some answers are historically “wrong” in a way that matches period thinking (e.g., physics, medicine, dog training).
- Repetitive loops (“new engines…”, repeated washing with acid, endless “strike him again”) highlight small-model and sampling issues.
Authenticity and Temporal Contamination
- Model sometimes switches between 17th‑century and modern voice (e.g., Wikipedia‑like definitions, awareness of modern politicians or products).
- Creator acknowledges an intentional trade-off: good instruction-following vs. deep “cultural reset.”
- Prompt engineering with faux archaisms helps anchor it in the past; suggestions include penalizing modern words or reinitializing top layers.
- Some users are dissatisfied that modern knowledge “leaks in,” arguing it undermines the time-travel illusion.
Technical Aspects & Limitations
- Quantized GGUF versions exist; users report smooth local runs on M‑series Macs with llama.cpp.
- Small 7B size and “weird” data mean more sensitivity to quantization quality and repetition.
- Suggestions include better sampling penalties and more training to reduce loops.
Philosophical & Conceptual Debates
- Discussion on whether LLMs can test historical interpretations or merely generate confident but unverifiable rewrites.
- Dispute over what it would mean for the model to “believe” anything vs. simply manipulating tokens.
- Skepticism about reconstructing “zeitgeist” from elite-written early texts and about the thread title’s promise of exploring alternate history, not just style simulation.