AI built from 1800s texts surprises creator by mentioning real 1834 London protests

Hobbyist training AI on Victorian texts gets an unexpected history lesson from his own creation.

See full article...

 

[tmg2018]

Great! Train an LLM on all the physics (including astronomy) and mathematics known up to 1900 and it won't be sidetracked by the limitations of quantum theory and relativity, but may produce the elusive Grand Unified Theory. Can't wait. Literally.

 

[NetMage]

Sure, you can get some information about how historical figures thought from their writing, but you could also get that information from just reading it?

The problem is that amount you’d have to read and remember - the 6GB of text input into the LM is about 1.2 billion words. I read fairly quickly (about 900 wpm for technical material, slower for not) and it would take me at least 2.5 years of non-stop reading (and much longer if I actually work, eat and sleep) and I doubt if I would remember enough to come to any conclusions.

 

[SinclairZX81]

" ...I’m not sure if just scaling the data up will ever result in reasoning but even now it kinda feels like digital time travel."

No, it won't, and no, it's not. It's a simulation of past patterns, only as accurate as the training data, which, being written, has already been distanced one remove from reality.

"If I train from scratch the language model won't pretend to be old, it just will be."

No, it's still pretending.

"This shows the model is beginning to remember things from the dataset."

No, it doesn't. LLM's don't "remember" anything. It's getting better at stringing actual training data together in response to a prompt.

"Training AI language models on period texts may allow for the creation of interactive period linguistic models that offer a researcher a chance to converse with a simulated speaker of an extinct vernacular or language of the past. The results would not necessarily be factually rigorous due to confabulations, but they could be stylistically illuminating for someone studying antique syntax or vocabulary in use."

It's only "stylistically illuminating" in terms of the simulation, not a genuine speaker. Books and letters are a far more reliable source, and letters will contain actual vernacular.

Studying a simulation is going to lead "researchers" away from the truth of history, not deeper into it.

 

[SinclairZX81]

... Knowing that a model can trusted to dig out facts after being fed copious amounts of text makes it much easier to analyze said texts, especially if it can cite the excerpts it used to come up with a particular inference.

That's the problem. You can't "trust it to dig out facts." You can ask it to dig out facts, and then you have to go verify every single part of its answer. If you're looking for a specific topic or event and not "what happened in 1832," you're going to be able to search and vet your information faster and easier with traditional methods, by precisely the amount of effort you put into interacting with the LLM.

As for citations, there is a small but ever-growing group of attorneys that have discovered, to their chagrin, that LLM's are quite capable of hallucinating cites, too.

 

[SinclairZX81]

There are some Victorian-sounding bits in there, but mostly the text is gobbledygook. It switches halfway down from past tense to future tense. Parts are nonsensical and ungrammatical ("was not bound in the way of private", "who first settled in the Gospel at Jerusalem", "a record of the prosperity and prosperity"), not to mention bad punctuation ("re counted", "be'known"). Other than the mention of Lord Palmerston, there is no indication of what happened in that year. It's word salad. The problem is that the remoteness of that era tempts one to excuse the nonsense as archaic. But no, "the day of law" is not Victorian English.

Thank you. It's infuriating that anyone is taking this output seriously. Fortunately, most of the people here at Ars are not.

 

[raxadian]

It would probably appear coherent without saying anything

So good enough to be president then? Not bad.

 

[hobiecatdriver]

Question: How much processing is needed for a query to output from a ~6GB dataset trained LLM as above. Curious, as this would make for an interesting game dynamic. Especially if your interactions could be fed back in to the model, i.e. you change history.

Star trek holodeck adventures would be an interesting use case.

 

[Walt Dizzy]

The fact that the researcher found the quoted bit of text to be "intriguing", proves again that people will see patterns where none exist, because that's how human brains are wired.

This LMM is slightly better than cutting up photocopies of the source text into phrases and individual words, putting the cut up bits into a shoe-box, shaking it vigorously, then pulling the words out and placing them onto lines on a sheet of paper.

Do that long enough, and some sentence-like assemblies will appear. You don't even need statistical models of word associations to make it happen if you are willing to spend enough time at it.

LMMs are a more sophisticated version of monkeys hammering away at typewriters. A lot of obviously random stuff is eliminated, but the underlying process is, at its core, fundamentally not guided by the information content of the source material.

 

[Tanngrisnir]

The problem is that amount you’d have to read and remember - the 6GB of text input into the LM is about 1.2 billion words. I read fairly quickly (about 900 wpm for technical material, slower for not) and it would take me at least 2.5 years of non-stop reading (and much longer if I actually work, eat and sleep) and I doubt if I would remember enough to come to any conclusions.

But what is the point of this? You aren't reading 6GB of text to understand a thought pattern. If you want to train your LLM to find you examples of a particular thought pattern then maybe you're on to something, but once this LLM is trained you can't exactly interrogate the weights to tease out a specific thought pattern. I don't mean to drop the "I'm a professional" so I'm infallible card, but I do sit here as a data scientist with a strong interest in historical linguistics wondering what on earth is the point of this?

 

[dzid]

In defense of the article, I submit that the notable part is that an individual hobbyist trained, not just ran, an LLM on a specific and completely legit corpus and got an interesting result. That's usually considered the purview of huge corporations and institutions. I think that's a pretty big deal, actually, and I encourage more people to experiment with this tech in this way.

It is, because it highlights a case where these things can fill a useful niche. The next step is a network of like-minded people that, importantly, have some level of trust in the others and share ideas, etc.

 

[dzid]

But what is the point of this? You aren't reading 6GB of text to understand a thought pattern. If you want to train your LLM to find you examples of a particular thought pattern then maybe you're on to something, but once this LLM is trained you can't exactly interrogate the weights to tease out a specific thought pattern. I don't mean to drop the "I'm a professional" so I'm infallible card, but I do sit here as a data scientist with a strong interest in historical linguistics wondering what on earth is the point of this?

I'd think the point is to extend the technology to allow it to do useful things, which is the opposite of what the big AI companies are doing with brute-force scaling, and failing miserably trying.

 

[deadmanwalking]

Not sure what the story here is.

LLMs hallucinate all the time. I guess that the "story" here is that this time, the AI just hallucinated correctly?

As the saying goes, even a blind squirrel will occasionally find nuts.

 

[dzid]

I think the story is that when a narrowly-focused LLM is created using a minimal quantity of vetted training data that is at least tangentially related to the LLM's subject of focus, the LLM output will consist of fewer pants-on-fire lies than one would likely see when using a much larger LLM (e.g. those developed by the for-profit AI companies) where the training data, while vast, contains not only a huge range of subject matter, but some pretty dodgy content, including, of course, some quantity of AI slop (think Grok).

Does it still have the same basic limitations as the LLMs developed by OpenAI, Anthropic, Google and Meta? Yes. None of these has any true 'memory', only the references that companies like OpenAI extract from chat transcripts, save, and then use to construct an illusion of an ongoing conversation, where the current chat is seeded with 'memories' of prior ones.

However, there is absolutely no functionality on par with a human's working or long-term memory. It does not exist in today's most advanced LLMs. The same can be said for motivation or sensory input and integration.

[slight subject change, but it is entirely relevant]

An effort that I expect we will see in the very near future that shares some similarities with the pint-sized LLM described in the article will be an implementation of a 'chat-buddy' aka the ad hoc therapy bots huge numbers of people are using today, but with decidedly mixed results: everything from "it helps me from feeling isolated" to true psychosis or self-harm or worse.

The reason for those wildly divergent results is that these commercial LLMs were never intended for such a use. Were they intended to be sycophantic? To generate maladaptive feedback loops between chatbot and vulnerable people? I certainly wouldn't rule out any of those, when the company goals are to drive up usage.

What people are already starting to collaborate on is a narrowly focused LLM that adheres to therapeutic best practices and drives conversations in a positive-weighted direction. I know many people will argue that this is a terrible idea, but the other option is to just continue to let people use, and quite possibly be harmed by, chatbots that are in no way suited for such use. And I am not optimistic that state-specific regulations could be passed that would force the AI companies to pull their products from [that market].

I wrote this up for future use, but I suppose this is as good a place as any, since this really is another type of small, narrowly-focused LLM:

- While not passing judgment on whether it is advisable, I would encourage anyone using a publicly accessible chatbot for personal or 'mental health' purposes to ask themselves: am I confident that <broligarch n>, who is the actual decision maker at the company that develops the chatbot that I use, would not now or in the future use the deeply personal information contained in the chat transcripts to manipulate me or do anything else contrary to my interests?

- If your answer is "hell no!", the following describes an alternative to those chatbots:

See the preprint psyarxiv paper The Psychology of Human-AI Emotional Bonds [https://osf.io/preprints/psyarxiv/rxedj_v2] (Yes, this is from a Chinese researcher, and the paper indicates "This work was supported by the National Key R&D Program of China STI2030 Major Projects.") They are taking this seriously, as should we.

See also: Large language models for the mental health community: framework for translating code to care [https://www.thelancet.com/journals/landig/article/PIIS2589-7500(24)00255-3/fulltext]

 

[Geebs]

See also: Large language models for the mental health community: framework for translating code to care [https://www.thelancet.com/journals/landig/article/PIIS2589-7500(24)00255-3/fulltext]

I understand why you might think something published in the Lancet would be a reliable source (Wakefield notwithstanding) but this is unreadable word salad which expends reams of text on saying nothing and was probably written by an LLM.

Mental health conditions are often both identified and treated through language, making them an ideal target for LLMs

Research findings from peer-reviewed and preprint studies suggest that LLMs could assist in clinical tasks, including diagnostic assessments,6 intervention delivery,3and empathic support.

Christ on a bike.

 

[dzid]

I understand why you might think something published in the Lancet would be a reliable source (Wakefield notwithstanding) but this is unreadable word salad which expends reams of text on saying nothing and was probably written by an LLM.





Christ on a bike.

"Mental health conditions are often both identified and treated through language, making them an ideal target for LLMs"

If you have complaints about the content in that article, that seems like a bad place to start. What exactly do you have a problem with? The only part I'd take issue with is "ideal target for LLMs", because the LLMs have to be very targeted and used in conjunction with standards, and locale permitting, regulations. But that's the whole point of talking about this.

So what exactly is your problem?

ETA: I didn't see any Wakefield in the authors or collaborators. Maybe he taints it indirectly?

Again, WTF? People are using these things, and quite a few are being harmed. So what do you have to offer?

 

[Byte Juggler]

Slow news day at Ars Technica today?

This isn't strictly a news site.

 

[Erbium68]

"appears to have surprised him by reconstructing a coherent historical moment"

Er, but it really didn't? What it wrote is nonsensical garbage. It contains the words "Palmerston" and "protest", yes. Everything beyond that is the reader doing a lot of interpretation. What's actually written is junk.

This struck me at once. It's exactly like Nostradamus. It tells us absolutely nothing about Palmerston or the nature of the protests, and in a garbled version of the newspaper syntax of the era. It's utterly useless.
Modern history has been patiently uncovering more and more details of past events and linking them, often causing us to re-evaluate past events and identify the myths behind, say, the Spanish Armada or the conquest of what is now Texas. It looks as if "AI" is just going to come up with exactly the same kind of smearing of history because it does not have the ability to weight data.

 

[dzid]

This struck me at once. It's exactly like Nostradamus. It tells us absolutely nothing about Palmerston or the nature of the protests, and in a garbled version of the newspaper syntax of the era. It's utterly useless.
Modern history has been patiently uncovering more and more details of past events and linking them, often causing us to re-evaluate past events and identify the myths behind, say, the Spanish Armada or the conquest of what is now Texas. It looks as if "AI" is just going to come up with exactly the same kind of smearing of history because it does not have the ability to weight data.

It's going to be stuck at basically this level of (non)functionality until someone develops a rudimentary memory or some other bio-inspired addition to the system. Users have to be aware that the output cannot be relied upon for factual answers, and some amount of human cognitive oversight is required. That's why I've been skeptical about their ability to offload any meaningful amount of work, on balance.

So what conceivable niche can they fill now? Confabulation automation: there's a ton of conspiracy-related content popping up on the web, so I imagine it will be used to populate that. That's no good. Some form of creative writing assistant, perhaps? Sure, if you feel comfortable in copyright no-mans-land. 'chat-buddies'? There's a killer app for you. Unfortunately that is what they are being used for, and we can see that in the increasing reports, both good ("it keeps me company") and all manner of bad.

Can that cat be put back in the bag? I'm not seeing it. That's why I suggested thinking about how to decouple that from the big tech companies and do something closer to what the person in the article did. The other option is just not worry about it and leave it in the capable hands of OpenAI, Meta et al.

 

[NullSignal]

This isn't strictly a news site.

The site's tagline is "Serving the Technologist since 1998. News, reviews, and analysis.", so I beg to differ.

I guess you could argue I should have said "Slow news, reviews, and analysis day today at Ars?", but that didn't flow so well.

 

[SixDegrees]

Why is this noteworthy or unexpected? An LLM got trained on a bunch of data, then regurgitated that data in response to queries. Isn't this exactly what LLMs are expected to do? I don't see any surprising "accident" here at all. The only mildly surprising bit is that the experimenter, with some alleged deep knowledge of the time period in question, was surprised by the output because he had never heard of it before, indicating that maybe his depth of knowledge was shallower than he presumed.

Or it may be surprising that an LLM actually managed to spit out a correct response, which isn't particularly common.

 

[SixDegrees]

IMO one of the big issues with LLMs is that the corporate types that fund development for them use any content under the sun, including social media, fictional books, etc.

My long standing opinion is that this is a bad approach, and this just proves my point.

LLMs aren't bad. The training approach popular with big tech is the issue. Stop feeding LLMs irrelevant social media stuff. Start tailoring both the model and training toward specific areas. There are already a ton of projects that are doing this and seeing success. Because they aren't backed by VC, they just quietly slip by and do their thing.

That assumes that the desire is to produce accurate answers. It's not that at all. Corporate overlords view LLMs as 1) a way to justify firing employees and using LLMs as half-ass replacements, and 2) engagement engines driven by the content of the Internet, probably the worst possible feedback loop imaginable.

 

[dzid]

That assumes that the desire is to produce accurate answers. It's not that at all. Corporate overlords view LLMs as 1) a way to justify firing employees and using LLMs as half-ass replacements, and 2) engagement engines driven by the content of the Internet, probably the worst possible feedback loop imaginable.

It would be if they weren't or couldn't be supplanted with something better, and that's where these smaller systems come in. They are determined to fire people, and they likely will. People should boycott their shitty 'customer service' replacements.

And a bit further down the road, this is truly an area where there could be some bona fide "innovation", if enough clever and determined people put in the time and effort to develop some bio-inspired components that a) consume less energy, b) implement a crude but functional memory integration scheme, c) think about how a system like that could function, not trying to shoehorn it into a human cognitive clone that it will never be. It should be something different.

They wouldn't know innovation if it bit them in the ass.

 

[plectrum]

Why is this noteworthy or unexpected? An LLM got trained on a bunch of data, then regurgitated that data in response to queries. Isn't this exactly what LLMs are expected to do? I don't see any surprising "accident" here at all. The only mildly surprising bit is that the experimenter, with some alleged deep knowledge of the time period in question, was surprised by the output because he had never heard of it before, indicating that maybe his depth of knowledge was shallower than he presumed.

Or it may be surprising that an LLM actually managed to spit out a correct response, which isn't particularly common.

It's not even that. Just throw some random words into Google and you can likely find an event you didn't know about.

Let's try this game... '1756 massacre'. Search for those words and ooh look there's a Wikipedia article about an event I'd never heard of. See, my 'LLM' (diceware) can teach me about history.

I'm afraid we're still at the monkeys with typewriters stage and looking for patterns in the clouds.

 

[D Fluke]

haven't read the comments yet, and yeah there's a lot to criticize here, but even I, an idiot in Norway had to pause and point out that this is straight up factually effing wrong for effs sake what are we even doing here there are no truths anymore and here Ars is tilling more lies into the already lie-filled substrate that feeds the nothing-but-hallucination softwares - get a grip Ars, get a grip... take some time... re-examine your purpose...

This is completely unhinged. Completely false.

"...developer Hayk Grigorian trains his hobbyist AI models from scratch using exclusively Victorian-era sources—over 7,000 books, legal documents, and newspapers published in London between 1800 and 1875

 

[dzid]

It's not even that. Just throw some random words into Google and you can likely find an event you didn't know about.

Let's try this game... '1756 massacre'. Search for those words and ooh look there's a Wikipedia article about an event I'd never heard of. See, my 'LLM' (diceware) can teach me about history.

I'm afraid we're still at the monkeys with typewriters stage and looking for patterns in the clouds.

I think we're all really clear that LLMs do not produce factual output, except as a lucky dice roll. There are still two scenarios (at least) that this writeup provides a jumping off point for:

1) further development so LLMs can (perhaps) some day be a useful part of a larger system, and

2) more immediately pressing, come to terms with the fact that lots of people are using the big tech firms' chatbots for things they probably shouldn't, but since Sam Altman isn't all of a sudden going to pull his product off the market for the good of people's mental health, we can sit back and watch those people circle the drain or try to find an alternate, still bad, but less bad system for people to use.

Rules and guardrails are more easily implemented in smaller systems. Again, I'd rather not, but the alternative is ChatGPT or god forbid, Grok.

 

[SixDegrees]

It's not even that. Just throw some random words into Google and you can likely find an event you didn't know about.

Let's try this game... '1756 massacre'. Search for those words and ooh look there's a Wikipedia article about an event I'd never heard of. See, my 'LLM' (diceware) can teach me about history.

I'm afraid we're still at the monkeys with typewriters stage and looking for patterns in the clouds.

As I've said before, LLMs are basically Magic Eight Balls that require terawatts of power to operate.

 

[plectrum]

This is completely unhinged. Completely false.

"...developer Hayk Grigorian trains his hobbyist AI models from scratch using exclusively Victorian-era sources—over 7,000 books, legal documents, and newspapers published in London between 1800 and 1875

It seems the writer hasn't noticed that Queen Victoria ascended in 1837... incidentally after the period the text is purportedly about (which was in the Regency era as William IV was on the throne)

 

[RuralRob]

"I have, through most extraordinary means, journeyed unto the year of our Lord two thousand and twenty-five, and beheld with mine own eyes the establishment of a most sublime and faultless system of public welfare in that distant land known as the United States of America. Verily, I say unto you, we are most urgently called upon to emulate and enact such noble reforms within our own age and realm..."

--- 19th-century time traveler and MP, upon his return from the future

 

[dzid]

"I have, through most extraordinary means, journeyed unto the year of our Lord two thousand and twenty-five, and beheld with mine own eyes the establishment of a most sublime and faultless system of public welfare in that distant land known as the United States of America. Verily, I say unto you, we are most urgently called upon to emulate and enact such noble reforms within our own age and realm..."

--- 19th-century time traveler and MP, upon his return from the future

What would that system be? Crypto-grift?

 

[Albino_Boo]

It seems the writer hasn't noticed that Queen Victoria ascended in 1837... incidentally after the period the text is purportedly about (which was in the Regency era as William IV was on the throne)

Err no the Regency lasted from 1811 till 1820.The future George IV acted as Regent for his Father during his father's final mental illness.

 

[SraCet]

But what makes this episode especially interesting is that a small hobbyist model trained by one man appears to have surprised him by reconstructing a coherent historical moment from scattered references across thousands of documents, connecting a specific year to actual events and figures without being explicitly taught these relationships.

This is the most important part of the article, i.e., the claim that the event wasn't in the training corpus.

Because otherwise, you might as well write an article about how water is wet.

And yet, this claim is buried in paragraph 9.

And also, was the event in the training corpus or not? If there were "scattered references" to it, then it sounds like it kinda was.

 

[SraCet]

The problem is that amount you’d have to read and remember - the 6GB of text input into the LM is about 1.2 billion words. I read fairly quickly (about 900 wpm for technical material, slower for not) and it would take me at least 2.5 years of non-stop reading (and much longer if I actually work, eat and sleep) and I doubt if I would remember enough to come to any conclusions.

Well, it's not like this software could process 6GB in a few seconds and deliver something useful either.

Training the model presumably took days or possibly weeks, running on expensive hardware that none of us own.

 

[HeadPlug]

Based on this paragraph:

reconstructing a coherent historical moment from scattered references across thousands of documents, connecting a specific year to actual events and figures without being explicitly taught these relationships. Grigorian hadn't intentionally trained the model on 1834 protest documentation; the AI assembled these connections from the ambient patterns in 6.25GB of Victorian-era writing.

My understanding is that nowhere in the 6.25GB corpus was the protest ever described 'as a single thing' - there wasn't a news piece covering it, and no text ever described it in much detail - maybe it wasn't even mentioned by name; rather, there were oblique references and "ambient patterns" in the way people wrote during that time period, that the LLM 'knew' to combine into one single event.

Maybe I'm reading this incorrectly, but if not, then this would be like if an LLM were able to describe the general happenings of the 2008 financial crisis by just being trained on, say, a no-politics gardening forum - no one has ever written a post about such a crisis there, and no one will have been able to go into much detail without hitting the 'no politics' restriction; but rather, just from the disconnected references of people maybe mentioning they had to reduce garden sizes because of having to move homes, or them struggling to find the time for soil rotation after their partner lost their job, the model then connecting all of these dots and 'realising' that it's all referring to the same event, and being able to refer to it as a single thing and bringing out a couple of details about it, agglomerated just from various mentions-in-passing.

If that's what this article is describing, then it still wouldn't be 'model can reason' news of course, but it would be pretty cool that it can bring out factual events solely from linguistic patterns and vague references that are in the 'background' of the training corpus.

 

[rdeforest]

I can't tell from the article and didn't look into the references, but I think commenters here are misunderstanding what's been done.

It's not fact in => fact out; it's tangentially related facts in => fact out.

In visual terms, it's like training a model on a bunch of pictures taken near something but not of it and then having the model spit out an accurate picture of that thing based on, for example, people's reactions to it in other pictures.

I'm not saying the LLM 'figured out' anything but rather that the output was the most likely input to have caused the patterns in the data it was trained on. I expect it works better with text than other data because there's a lot of redundancy built into text.

 

[SraCet]

...
My understanding is that nowhere in the 6.25GB corpus was the protest ever described 'as a single thing' - there wasn't a news piece covering it, and no text ever described it in much detail - maybe it wasn't even mentioned by name; rather, there were oblique references and "ambient patterns" in the way people wrote during that time period, that the LLM 'knew' to combine into one single event.
...

Right. The article is written so poorly that it's pretty easy to misunderstand this point. And it's still not clear from the article how much information about the event there actually was or wasn't in the corpus.

 

[SixDegrees]

I can't tell from the article and didn't look into the references, but I think commenters here are misunderstanding what's been done.

It's not fact in => fact out; it's tangentially related facts in => fact out.

In visual terms, it's like training a model on a bunch of pictures taken near something but not of it and then having the model spit out an accurate picture of that thing based on, for example, people's reactions to it in other pictures.

I'm not saying the LLM 'figured out' anything but rather that the output was the most likely input to have caused the patterns in the data it was trained on. I expect it works better with text than other data because there's a lot of redundancy built into text.

No, it's a bunch of scattered facts creating a statistical likelihood that those similar facts will get regurgitated more than other, less frequently occurring facts.

 

[dzid]

I can't tell from the article and didn't look into the references, but I think commenters here are misunderstanding what's been done.

It's not fact in => fact out; it's tangentially related facts in => fact out.

In visual terms, it's like training a model on a bunch of pictures taken near something but not of it and then having the model spit out an accurate picture of that thing based on, for example, people's reactions to it in other pictures.

I'm not saying the LLM 'figured out' anything but rather that the output was the most likely input to have caused the patterns in the data it was trained on. I expect it works better with text than other data because there's a lot of redundancy built into text.

Even so, "fact out" would be more a fortunate occurrence, but the quality and specificity of the training data matters.

 

[uthallan]

The plagiarism machine spits out one correct paragraph and gets news coverage. This just shows how little AI users trust their own software. Tech writers are unable to resist blowing air into the AI bubble.

 

[Erbium68]

It seems the writer hasn't noticed that Queen Victoria ascended in 1837... incidentally after the period the text is purportedly about (which was in the Regency era as William IV was on the throne)

There is a tendency to refer to the whole post-Napoleon era as being "Victorian", there's a lot of overlapping threads which make it difficult to define a start point. Riots in 1834 were very much part of the post-Napoleon demands for more democracy, but the themes of the Victorian era - Industrial Revolution, India, railways - had been developing fast since the 1820s.
Real history resists the attempts of inferior historians to simplify and label.

 

[FalcorMontoya]

haven't read the comments yet, and yeah there's a lot to criticize here, but even I, an idiot in Norway had to pause and point out that this is straight up factually effing wrong for effs sake what are we even doing here there are no truths anymore and here Ars is tilling more lies into the already lie-filled substrate that feeds the nothing-but-hallucination softwares - get a grip Ars, get a grip... take some time... re-examine your purpose...

This is completely unhinged. Completely false.

"...developer Hayk Grigorian trains his hobbyist AI models from scratch using exclusively Victorian-era sources—over 7,000 books, legal documents, and newspapers published in London between 1800 and 1875

Just to make this easier for other readers who might not know the dates - the Victorian era is a common name for the period when Queen Victoria ruled the UK, which was between 1837 and 1901. So fully half of the period this guy used wasn't Victorian (and he's missing half of the era on the later part).