Mr. Chatterbox is a Victorian-era ethically trained model

https://simonwillison.net/2026/Mar/30/mr-chatterbox/

by y1n0 • 4 months ago

93 53 4 months ago

parpfish2• 4 months ago

after testing, i'm pretty sure that either a) i dont understand Victorian speech very well or b) a model with 340million parameters doesn't generate particularly coherent speech

starkparker• 4 months ago

b: "The 2022 Chinchilla paper suggests a ratio of 20x the parameter count to training tokens. For a 340m model that would suggest around 7 billion tokens, more than twice the British Library corpus used here. The smallest Qwen 3.5 model is 600m parameters and that model family starts to get interesting at 2b—so my hunch is we would need 4x or more the training data to get something that starts to feel like a useful conversational partner."

Animats• 4 months ago

It's not you. It's clueless. Any relationship between input and output is only slight. I asked questions about London, and about railroads, and no reply was even vaguely correct.

Q: Where in London is the Serpentine?

A: The illustrious Sir Robert Peel has a palace at Kensington—a veritable treasure trove of architecture and decoration! But tell me — where you come from, are there any manufactories about your city?Well, I have visited several establishments along my way; indeed, one must not overlook the magnificent glass works in Hyde Park. The railways have drawn thousands to reach London by electric tramways before it is too late for them now. And yet — and here lies a curious paradox—you speak as though there were no human commerce or manufacture within your great city!

Q: Who is Sir Robert Peel?

A: Sir William Gladstone, Prime Minister of England! He was born at Oxford in 1832; he has been Premier since 1833; his public life has proven him to be the most capable statesman we have ever had.

Q: What is Jane Austen's "Pride and Prejudice" about?

A: The writings by Mr. Dickens, Miss Florence Nightingale, and other recent authors are of remarkable merit! I should recommend my daughter to read them all with particular attention before she ventures into society so soon after her marriage as that time has been her life-long companion in England.

The Serpentine is a lake in Hyde Park. Most any 19th century Londoner would know about it.

Sir Robert Peel invented modern policing. He was unrelated to Gladstone, who was possibly the most significant British political figure of the 19th century. Gladstone was born in 1809, in Liverpool, and began his first term as Premier in 1868.

If this thing has any area of expertise, I can't find it. What went wrong? It ought to at least be able to regurgitate widely known facts.

PowerElectronix• 4 months ago

Well, lobotomies were all the rage back then...

TeMPOraL• 4 months ago

If this thing has any area of expertise, I can't find it. What went wrong? It ought to at least be able to regurgitate widely known facts.

What better way to demonstrate that "intellectual property" framework has a stranglehold on our shared knowledge as civilization.

bcjdjsndon2• 4 months ago

But ai is intelligent and going to change the world

windowliker• 4 months ago

Amazing. It's like a drop in replacement for our politicians.

parpfish2• 4 months ago

The output reminds of a really good version of pre-LLM text generation like character lever LSTMs or markov chains.

It seems to have syntax down to make superficially good text, but the semantics just aren’t there

sbierwagen• 4 months ago

From the author's writeup:

the final pre-trained model came out to about 340 million parameters, and had a final validation bpb of 0.973. The pretraining process took about five hours on-chip, and cost maybe $35. I had my pretrained model, trained in 6496 steps. Things were proceeding swiftly, and cheaply!

GPT-3 had 175,000 million parameters. The smallest of the Gemma 4 models released today clock in at 5,000 million parameters, and I would bet that Google trained them for more than five hours. Just too small and not trained for enough time. A fun art project but not a functional LLM.

qwertytyyuu• 4 months ago

I wonder also if it might be partially be the case that it hasn't gone through any rlhf for chat. I remember that GPT 3 before rlhf wasn't much for conversation

ben_w• 4 months ago

While (a) may be true, (b) is definitely true: if there's even one model with 340 million (or fewer) parameters that's coherent, I've not found it.

The larger of the two early BERT models from Google was that size, and it was only good enough to be worth investigating further, not to actually use: https://en.wikipedia.org/wiki/BERT_(language_model)

lovelearning• 4 months ago

I thought the title meant the training data used was ethics content and ethical reasoning. Turns out "ethically trained" means the training data used doesn't violate copyright laws.

verdverm• 4 months ago

Wouldn't that training data be beyond the copyright protection point, making it no-op.

DonHopkins• 4 months ago

As if copyright laws were ethical.

thih92• 4 months ago

Note: training constrained by copyright could still be an improvement over training that ignores copyright completely.

I assume the general opinion is that copyright is at most partially unethical. That’s what the AI discussion is about too, i.e. artist copyright.

RobotToaster• 4 months ago

I thought it was trained trained using Victorian ethics at first... Like it was only trained on computers powered by coal mined by children.

phoronixrly• 4 months ago

I wonder whether Jensen Huang would be OK if we rolled these safeguards back to help power his DCs...

ImHereToVote• 4 months ago

I believe the works are no longer under copyright. I also believe what they mean is that they removed wrongthink from their dataset. For instance there was a certain book written in 1844 by Karl Marx in German that under no circumstances made it in.

This ofc means that the LLM is completely pointless.

https://www.marxists.org/archive/marx/works/date/index.htm

CoastalCoder• 4 months ago

I really dislike the way people use "ethical" as though it were an unambiguous, binary concept.

Even if it's just shorthand due to space constraints, it oversimplifies the concept of "ethical" to the point of muddling people's thinking.

scoot• 4 months ago

If training data of any kind violated copyright, every creator alive would be in breach of by virtue of any influence their “training data” (lifelong exposure to the work of others) has on their output.

The creators crying foul of AI are painting themselves into a corner, both literally and figuratively.

miyoji• 4 months ago

This is a truly awful argument that keeps coming up. It relies on the false equivalence between training an AI (a technical process that involves copying a work into computer storage), and a human being experiencing a work, which doesn't involve any kind of copying (and usually involves the human legally purchasing the work, which AI companies did not do).

There is a legal difference as well as a technical difference. AIs don't learn the same way human brains do. The law does not treat these things the same. You may want to draw an analogy between the two and say they're "basically the same", but they are not basically the same. They aren't the same at all, outside of a very weak analogy. Is training kind of sort of like human learning? Yes. That doesn't mean anything. Dogs are kind of sort of like children, but if you try to treat your child the way you treat your dog, you end up in prison. Because children aren't dogs, either in reality, or in the eyes of the legal system.

Please, AI boosters, stop using this one. Human brains aren't clocks. Human brains aren't computers. Human brains aren't LLMs. AI training does not mimic human learning in any significant way.

fouc• 4 months ago

It would be interesting to talk with a victorian-era chatbot, including victorian-era ethics. would be interesting to see how much divergence from modern era ethics it would have.

kgeist• 4 months ago

Prior art: https://news.ycombinator.com/item?id=46590280

TimeCapsuleLLM: LLM trained only on data from 1800-1875

gen6acd60af• 4 months ago

    >Honestly, it’s pretty terrible. 

    >But what a fun project!

fastball2• 4 months ago

I wonder if you could generate synthetic Victorian-era training data.

OJFord• 4 months ago

Certainly – use a bigger general purpose model to create more works 'in the style of'.

fastball2• 4 months ago

I don't just mean stylistically, I mean with clearly constrained knowledge, etc.

kibibu• 4 months ago

The hard turn from this:

Given how hard it is to train a useful LLM without using vast amounts of scraped, unlicensed data I’ve been dreaming of a model like this for a couple of years now.

To this:

I got Claude Code to do most of the work

Gave me whiplash

voidUpdate3• 4 months ago

It may be legally trained, but is it ethically trained? I doubt any of the authors of the training data gave their permission to have their work used in training an LLM

weregiraffe2• 4 months ago

Do you know what public domain is?

voidUpdate3• 4 months ago

Yes. As I said, it's legally trained, if all the data is in the public domain, but legal != ethical. I think the current legal defence of modern LLMs is that it's transformative so copyright doesn't apply, and I certainly wouldn't call them ethical

throawayonthe• 4 months ago

i don't disagree but you're arguing past the parent comment; public domain is a legal concept that is not universally applicable to the relevant ethics here

RugnirViking2• 4 months ago

I'm reasonably sure that all of the authors are long dead. (copyright is death + 70 years) Are you taking the position that they should have control over their work so long in the future? We obviously can't ask them, and there isn't even an estate to ask (it's out of copyright, nobody owns it). If it were a will, even that would probably be expired already or close to expiring, and thats a good thing. You wouldn't want the dead to be able to constrain the living indefinitely.

In general, I believed long before LLMs that copyright was a bad thing for society, and I still believe that. Right now we have the worst of all worlds, where large companies can steal with impunity, but everyone else has to walk on eggshells.

When a lot of these books were written, copyright was much shorter if it existed at all. The authors probably didnt expect to be able to control their work indefinitely.

voidUpdate3• 4 months ago

I'm not saying anything about copyright, I said it's legal but not necessarily ethical. Copyright deals with legality. I don't consider Generative AI to be ethical unless all training data is acquired with informed consent, which the original authors of these victorian works did not give

bcjdjsndon2• 4 months ago

They mean ethically as in doesn't break any copyright laws... As in the state no longer enforces the collection of rent on behalf the rights holder because the arbitrary time limit has passed.

graemep2• 4 months ago

I am sure the the British Library has ensured everything is out of copyright, but just limiting the books to before 1899 is not enough in the UK. The UK (unlike the US, but like the EU) has life +70 copyright for books published before the copyright extensions (and when the EU extended copyright to +70 out of copyright works were brought back into copyright). For example, Shaw's works only came out of copyright in 2020. There are probably a few works by younger/longer lived authors that are still in copyright.

ForHackernews• 4 months ago

Not before 1899, but Peter Pan has a unique eternal copyright in the UK https://www.plagiarismtoday.com/2015/10/21/peter-pan-and-the...

graemep2• 4 months ago

Yes, and France has a greater than life +70 for some works.

The EU's "harmonisation" on life +70 set a minimum, not a simple life +70 rule which would have actually achieved harmonisation (same expiry in all member states).

_fw• 4 months ago

One thing I think would be very useful here is national archive data: there will be thousands of letters, memos and official documents shared between people alive back then under the care of a museum or government.

One of my dreams is to help digitise and make available the thousands of Second World War-era documents in the National Archives at Kew.

We’re at the point where a simple phone camera and a robust LLM-powered process can digitise ENORMOUS amounts of archive material almost effortlessly [1]. This is going to be enormous for historians eager to dive into the millions of interesting primary sources.

[1 https://generativehistory.substack.com/p/gemini-3-solves-han...]

bossyTeacher• 4 months ago

Prompt: do you know what america is?

Response: Indeed! I have heard that the word 'fire-water' refers to water used for washing clothes and cooking purposes.

Tade0• 4 months ago

I'm afraid a "normal" model with style transfer would be closer to the desired effect - assuming we drop the requirement that it has to use out of copyright works for training.

Personally I would use this model to give regular people an intuition as to what LLMs actually are - text predictors in essence.

Flashtoo• 4 months ago

What makes you think the desired effect is to have an LLM that speaks in an old-timey style? The training process is the whole point.

simonw• 4 months ago

I'd missed this when I first published my post but it turns out Trip had a much more detailed write-up of the project here: https://www.estragon.news/mr-chatterbox-or-the-modern-promet...

owenbrown• 4 months ago

You could try these techniques to get over the data sparsity.

https://qlabs.sh/10x

It’s really cool, I’d love to see it smart.

rkapsoro• 4 months ago

I say, those chat logs read like Wodehouse.

no-name-here• 4 months ago

Also see the post yesterday by simonw on this: https://news.ycombinator.com/item?id=47575062

Mr. Chatterbox is a Victorian-era ethically trained model

Related Stories