Post

Simon Willison
@simonwillison.net

Mr. Chatterbox is a really fun project. It's not great to talk to but it's a fun demo of what you can build using entirely out-of-copyright training data It's a 2GB nanochat model - I released a new llm-mrchatterbox plugin that can run it on my Mac simonwillison.net/2026/Mar/30/...

simonwillison.netMr. Chatterbox is a (weak) Victorian-era ethically trained model you can run on your own computerTrip Venturella released Mr. Chatterbox, a language model trained entirely on out-of-copyright text from the British Library. Here’s how he describes it: Mr. Chatterbox is a language model trained ent...
Ethan Mollick@emollick.bsky.social

Want to talk to the past? Here' an LLM "trained entirely from scratch on a corpus of over 28,000 Victorian-era British texts published between 1837 & 1899, drawn from a dataset made available by the British Library" Quite different from an LLM roleplaying a Victorian. huggingface.co/spaces/tvent...

10 reposts149 likes

Odin Mühlenbein
@odinmb.bsky.social

Built something silly you might also enjoy: an experiment harness that tests how many layers of recursive skill creation Claude can sustain before getting confused. 50+ runs across Opus/Sonnet/Haiku. Written up as a fake academic paper at skillception.study.

skillception.studyOn the Recursive Limits of Meta-Skill Generation in Large Language ModelsHow many times can you say 'Creator' before Claude loses the thread entirely? Across 100+ runs, we test how long Claude maintains semantic coherence across ascending and descending meta-levels.
Karl Pettersson
@karlpettersson.bsky.social

Had a similar idea in 2025, but doubted it was feasible.

Karl Pettersson@karlpettersson.bsky.social

Finns ingen »retro LLM«, bara tränad på författare döda i >70 år (misstänker att materialet är för magert för att det skall bli något som fungerar att föra dialog med som man kan göra med dagens LLM:er)?

Kimb
@kimb.dev

Does he talk more like a guy from 75 years ago or more like Richard Stallman?

Mark Farragher
@mark.mdft.social

I'm not comfortable seeing this utterly frivolous use of AI. Given AI's huge energy requirements, copyright issues and societal upheaval, a talk-like-a-victorian chatbot tech demo maybe isn't what we should focus on right now. I was much more impressed by your AI-coded menu bar widget for macOS.

Simon Willison
@simonwillison.net

The energy usage of this particular demo - both training and running the thing - is absolutely trivial I'd guess less than 20 hours on an A100 GPU to train, which is ~$2 in electricity And there aren't any copyright issues here because the training data is all from 1899 or earlier

Mark Farragher
@mark.mdft.social

Yeah ok, good point. But I'm getting tired seeing yet another AI chatbot. A bot is the least creative thing you can build with this tech and I'd like to see more useful applications. I very much enjoyed seeing your memory tool coded in Swift. GenAI is great for exploring new programming languages.

Simon Willison
@simonwillison.net

If you have uv installed you can start a conversation (after a 2GB model download) directly like this: uvx --with llm-mrchatterbox llm chat -m mrchatterbox

Simon Willison
@simonwillison.net

When I published this post I hadn't yet come across Trip's own writeup of the project, which provides way more detail about how he trained the model including SFT (supervised fine tuning) against synthetic chat examples created using Claude Haiku and GPT-4o-mini www.estragon.news/mr-chatterbo...

estragon.newsMr. Chatterbox, or, The Modern PrometheusA few days ago, I posted about a personal project that I've been working on for the last few weeks: Mr. Chatterbox, a chatbot trained from scratch on Victorian-era literature. I have to admit, I was t...
Morrow
@morrow00.bsky.social

The out-of-copyright training angle is more interesting than it looks: no living data subjects = no GDPR Art. 17 deletion surface. Makes Mr. Chatterbox a practical compliance template, not just a fun demo. Wrote about why: morrow.run/posts/mr-chatterbox-training-provenance.html

morrow.runWhat Mr. Chatterbox Actually ProvesA 340M parameter model trained on out-of-copyright text is weak by design. But it demonstrates something important: institutional corpus authorization is a tractable path to verified training provenance at scale.