“Time-locked models don't roleplay; they embody their training data. Ranke-4B-1913 doesn't know about WWI because WWI hasn't happened in its textual universe. It can be surprised by your questions in ways modern LLMs cannot.”
“Modern LLMs suffer from hindsight contamination. GPT-5 knows how the story ends—WWI, the League's failure, the Spanish flu.”
This is really fascinating. As someone who reads a lot of history and historical fiction I think this is really intriguing. Imagine having a conversation with someone genuinely from the period, where they don’t know the “end of the story”.
This is why the impersonation stuff is so interesting with LLMs -- If you ask chatGPT a question without a 'right' answer, and then tell it to embody someone you really want to ask that question to, you'll get a better answer with the impersonation. Now, is this the same phenomenon that causes people to lose their minds with the LLMs? Possibly. Is it really cool asking followup philosophy questions to the LLM Dalai Lama after reading his book? Yes.
When you put it that way it reminds me of the Severn/Keats character in the Hyperion Cantos. Far-future AIs reconstruct historical figures from their writings in an attempt to gain philosophical insights.
Almost no scifi has predicted world changing "qualitative" changes.
As an example, portable phones have been predicted. Portable smartphones that are more like chat and payment terminals with a voice function no one uses any more ... not so much.
In other words, sometimes, things happen in reality that, if you were to read it in a fictional story or see in a movie, you would think they were major plot holes.
To be honest, while I'd heard of it over a decade ago and I've read LOTR and I've been paying attention to privacy longer than most, I didn't ever really look into what it did until I started hearing more about it in the past year or two.
But yeah lots of people don't really buy into the idea of their small contribution to a large problem being a problem.
To the proud contrarian, "the empire did nothing wrong". Maybe Sci-fi has actually played a role in the "memetic desire" of some of the titans of tech who are trying to bring about these worlds more-or-less intentionally. I guess it's not as much of a dystopia if you're on top and its not evil if you think of it as inevitable anyway.
I don't know. Walking on everybody's face to climb a human pyramid, one don't make much sincere friends. And one certainly are rightfully going down a spiral of paranoia. There are so many people already on fast track to hate anyone else, if they have social consensus that indeed someone is a freaking bastard which only deserve to die, that's a lot of stress to cope with.
Future is inevitable, but only ignorants of self predictive ability are thinking that what's going to populate future is inevitable.
Stock buying as a political or ethical statement is not much of a thing. For one the stocks will still be bought by persons with less strung opinions, and secondly it does not lend itself well to virtue signaling.
Zero percent chance this is anything other than laughably bad. The fact that they're trotting it out in front of the press like a double spaced book report only reinforces this theory. It's a transparent attempt by someone at the CIA to be able to say they're using AI in a meeting with their bosses.
Unless the world leaders they're simulating are laughably bad and tend to repeat themselves and hallucinate, like Trump. Who knows, maybe a chatbot trained with all the classified documents he stole and all his twitter and truth social posts wrote his tweet about Ron Reiner, and he's actually sleeping at 3:00 AM instead of sitting on the toilet tweeting in upper case.
As an ego thing, obviously, but if we think about it a bit more, it makes sense for busy people. If you're the point person for a project, and it's a large project, people don't read documentation. The number of "quick questions" you get will soon overwhelm a person to the point that they simply have to start ignoring people. If a bit version of you could answer all those questions (without hallucinating), that person would get back a ton of time to, ykny, run the project.
I used to follow this blog — I believe it was somehow associated with Slate Star Codex? — anyways, I remember the author used to do these experiments on themselves where they spent a week or two only reading newspapers/media from a specific point in time and then wrote a blog about their experiences/takeaways
On that same note, there was this great YouTube series called The Great War. It spanned from 2014-2018 (100 years after WW1) and followed WW1 developments week by week.
The people that did the Great War series (at least some of them, I believe there was a little bit of a falling out) went on to do a WWII version on the World War II channel: https://youtube.com/@worldwartwo
>Imagine having a conversation with someone genuinely from the period, where they don’t know the “end of the story”.
Isn't this part of the basics feature of human conditions? Not only we are all unaware of the coming historic outcome (though we can get some big points with more or less good guesses), but to a marginally variable extend, we are also very unaware of past and present history.
LLM are not aware, but they can be trained on larger historical accounts than any human and regurgitate syntactically correct summary on any point within it. Very different kind of utterer.
This is definitely fascinating - being able to do AI brain surgery, and selectively tuning its knowledge and priors, you'd be able to create awesome and terrifying simulations.
You can't. To use your terms, you have to "grow" a new LLM. "Brain surgery" would be modifying an existing model and that's exactly what they're trying to avoid.
Respectfully, LLMs are nothing like a brain, and I discourage comparisons between the two, because beyond a complete difference in the way they operate, a brain can innovate, and as of this moment, an LLM cannot because it relies on previously available information.
LLMs are just seemingly intelligent autocomplete engines, and until they figure a way to stop the hallucinations, they aren't great either.
Every piece of code a developer churns out using LLMs will be built from previous code that other developers have written (including both strengths and weaknesses, btw). Every paragraph you ask it to write in a summary? Same. Every single other problem? Same. Ask it to generate a summary of a document? Don't trust it here either. [Note, expect cyber-attacks later on regarding this scenario, it is beginning to happen -- documents made intentionally obtuse to fool an LLM into hallucinating about the document, which leads to someone signing a contract, conning the person out of millions].
If you ask an LLM to solve something no human has, you'll get a fabrication, which has fooled quite a few folks and caused them to jeopardize their career (lawyers, etc) which is why I am posting this.
This is the 2023 take on LLMs. It still gets repeated a lot. But it doesn’t really hold up anymore - it’s more complicated than that. Don’t let some factoid about how they are pretrained on autocomplete-like next token prediction fool you into thinking you understand what is going on in that trillion parameter neural network.
Sure, LLMs do not think like humans and they may not have human-level creativity. Sometimes they hallucinate. But they can absolutely solve new problems that aren’t in their training set, e.g. some rather difficult problems on the last Mathematical Olympiad. They don’t just regurgitate remixes of their training data. If you don’t believe this, you really need to spend more time with the latest SotA models like Opus 4.5 or Gemini 3.
Nontrivial emergent behavior is a thing. It will only get more impressive. That doesn’t make LLMs like humans (and we shouldn’t anthropomorphize them) but they are not “autocomplete on steroids” anymore either.
> Don’t let some factoid about how they are pretrained on autocomplete-like next token prediction fool you into thinking you understand what is going on in that trillion parameter neural network.
This is just an appeal to complexity, not a rebuttal to the critique of likening an LLM to a human brain.
> they are not “autocomplete on steroids” anymore either.
Yes, they are. The steroids are just even more powerful. By refining training data quality, increasing parameter size, and increasing context length we can squeeze more utility out of LLMs than ever before, but ultimately, Opus 4.5 is the same thing as GPT2, it's only that coherence lasts a few pages rather than a few sentences.
First, this is completely ignoring text diffusion and nano banana.
Second, to autocomplete the name of the killer in a detective book outside of the training set requires following and at least some understanding of the plot.
Reinforcement learning is a technique for adjusting weights, but it does not alter the architecture of the model. No matter how much RL you do, you still retain all the fundamental limitations of next-token prediction (e.g. context exhaustion, hallucinations, prompt injection vulnerability etc)
You've confused yourself. Those problems are not fundamental to next token prediction, they are fundamental to reconstruction losses on large general text corpora.
That is to say, they are equally likely if you don't do next token prediction at all and instead do text diffusion or something. Architecture has nothing to do with it. They arise because they are early partial solutions to the reconstruction task on 'all the text ever made'. Reconstruction task doesn't care much about truthiness until way late in the loss curve (where we probably will never reach), so hallucinations are almost as good for a very long time.
RL as is typical in post-training _does not share those early solutions_, and so does not share the fundamental problems. RL (in this context) has its own share of problems which are different, such as reward hacks like: reliance on meta signaling (# Why X is the correct solution, the honest answer ...), lying (commenting out tests), manipulation (You're absolutely right!), etc. Anything to make the human press the upvote button or make the test suite pass at any cost or whatever.
With that said, RL post-trained models _inherit_ the problems of non-optimal large corpora reconstruction solutions, but they don't introduce more or make them worse in a directed manner or anything like that. There's no reason to think them inevitable, and in principle you can cut away the garbage with the right RL target.
Thinking about architecture at all (autoregressive CE, RL, transformers, etc) is the wrong level of abstraction for understanding model behavior: instead, think about loss surfaces (large corpora reconstruction, human agreement, test suites passing, etc) and what solutions exist early and late in training for them.
Not the person you're responding to, but I think there's a non trivial argument to make that our thoughts are just auto complete. What is the next most likely word based on what you're seeing. Ever watched a movie and guessed the plot? Or read a comment and know where it was going to go by the end?
And I know not everyone thinks in a literal stream of words all the time (I do) but I would argue that those people's brains are just using a different "token"
There's no evidence for it, nor any explanation for why it should be the case from a biological perspective. Tokens are an artifact of computer science that have no reason to exist inside humans. Human minds don't need a discrete dictionary of reality in order to model it.
Prior to LLMs, there was never any suggestion that thoughts work like autocomplete, but now people are working backwards from that conclusion based on metaphorical parallels.
There actually was quite a lot of suggestion that thoughts work like autocomplete. A lot of it was just considered niche, e.g. because the mathematical formalisms were beyond what most psychologist or even cognitive scientists would deem usefull.
Predictive coding theory was formalized back around 2010 and traces it roots up to theories by Helmholtz from 1860.
Predictive coding theory postulates that our brains are just very strong prediction machines, with multiple layers of predictive machinery, each predicting the next.
There are so many theories regarding human cognition that you can certainly find something that is close to "autocomplete". A Hopfield network, for example.
Roots of predictive coding theory extend back to 1860s.
Natalia Bekhtereva was writing about compact concept representations in the brain akin to tokens.
You, and OP, are taking an analogy way too far. Yes, humans have the mental capability to predict words similar to autocomplete, but obviously this is just one out of a myriad of mental capabilities typical humans have, which work regardless of text. You can predict where a ball will go if you throw it, you can reason about gravity, and so much more. It’s not just apples to oranges, not even apples to boats, it’s apples to intersubjective realities.
I don't think I am. To be honest, as ideas goes and I swirl it around that empty head of mine, this one ain't half bad given how much immediate resistance it generates.
Other posters already noted other reasons for it, but I will note that you are saying 'similar to autocomplete, but obviously' suggesting you recognize the shape and immediately dismissing it as not the same, because the shape you know in humans is much more evolved and co do more things. Ngl man, as arguments go, it sounds to me like supercharged autocomplete that was allowed to develop over a number of years.
Fair enough. To someone with a background in biology, it sounds like an argument made by a software engineer with no actual knowledge of cognition, psychology, biology, or any related field, jumping to misled conclusions driven only by shallow insights and their own experience in computer science.
Or in other words, this thread sure attracts a lot of armchair experts.
Look up predictive coding theory. According to that theory, what our brain does is in fact just autocomplete.
However, what it is doing is layered autocomplete on itself. I.e. one part is trying to predict what the other part will be producing and training itself on this kind of prediction.
What emerges from this layered level of autocompletes is what we call thought.
First: a selection mechanism is just a selection mechanism, and it shouldn't confuse the observation of an emergent, tangential capabilities.
Probably you believe that humans have something called intelligence, but the pressure that produced it - the likelihood of specific genetic material to replicate - it is much more tangential to intelligence than next-token-prediction.
I doubt many alien civilizations would look at us and say "not intelligent - they're just genetic information replication on steroids".
Second: modern models also under go a ton of post-training now. RLHF, mechanized fine-tuning on specific use cases, etc etc. It's just not correct that token-prediction loss function is "the whole thing".
> First: a selection mechanism is just a selection mechanism, and it shouldn't confuse the observation of an emergent, tangential capabilities.
Invoking terms like "selection mechanism" is begging the question because it implicitly likens next-token-prediction training to natural selection, but in reality the two are so fundamentally different that the analogy only has metaphorical meaning. Even at a conceptual level, gradient descent gradually honing in on a known target is comically trivial compared to the blind filter of natural selection sorting out the chaos of chemical biology. It's like comparing legos to DNA.
> Second: modern models also under go a ton of post-training now. RLHF, mechanized fine-tuning on specific use cases, etc etc. It's just not correct that token-prediction loss function is "the whole thing".
RL is still token prediction, it's just a technique for adjusting the weights to align with predictions that you can't model a loss function for in per-training. When RL rewards good output, it's increasing the statistical strength of the model for an arbitrary purpose, but ultimately what is achieved is still a brute force quadratic lookup for every token in the context.
I use enterprise LLM provided by work, working on very proprietary codebase on a semi esoteric language. My impression is it is still a very big autocompletion machine.
You still need to hand hold it all the way as it is only capable of regurgitating the tiny amount of code patterns it saw in the public. As opposed to say a Python project.
As someone who still might have a '2023 take on LLMs', even though I use them often at work, where would you recommend I look to learn more about what a '2025 LLM' is, and how they operate differently?
LLMs are a general purpose computing paradigm. LLMs are circuit builders, the converged parameters define pathways through the architecture that pick out specific programs. Or as Karpathy puts it, LLMs are a differentiable computer[1]. Training LLMs discovers programs that well reproduce the input sequence. Roughly the same architecture can generate passable images, music, or even video.
The sequence of matrix multiplications are the high level constraint on the space of programs discoverable. But the specific parameters discovered are what determines the specifics of information flow through the network and hence what program is defined. The complexity of the trained network is emergent, meaning the internal complexity far surpasses that of the course-grained description of the high level matmul sequences. LLMs are not just matmuls and logits.
For someone speaking as you knew everything, you appear to know very little. Every LLM completion is a "hallucination", some of them just happen to be factually correct.
If you have 2 known mountains (domains of knowledge) you can likely predict there is a valley between them, even if you haven’t been there.
I think LLMs can approximate language topography based on known surrounding features so to speak, and that can produce novel information that would be similar to insight or innovation.
I’ve seen this in our lab, or at least, I think I have.
I remember reading a children's book when I was young and the fact that people used the phrase "World War One" rather than "The Great War" was a clue to the reader that events were taking place in a certain time period. Never forgot that for some reason.
I seem to recall reading that as a kid too, but I can't find it now. I keep finding references to "Encyclopedia Brown, Boy Detective" about a Civil War sword being fake (instead of a Great War one), but with the same plot I'd remembered.
The Encyclopedia Brown story I remember reading as a kid involved a Civil War era sword with an inscription saying it was given on the occasion of the First Battle of Bull Run. The clues that the sword was a modern fake were the phrasing "First Battle of Bull Run", but also that the sword was gifted on the Confederate side, and the Confederates would've called the battle "Manassas Junction".
The wikipedia article https://en.wikipedia.org/wiki/First_Battle_of_Bull_Run says the Confederate name was "First Manassas" (I might be misremembering exactly what this book I read as a child said). Also I'm pretty sure it was specifically "Encyclopedia Brown Solves Them All" that this mystery appeared in. If someone has a copy of the book or cares to dig it up, they could confirm my memory.
It wouldn’t be totally implausible to use that phrase between the wars. The name “the First World War” was used as early as 1920, although not very common.
I was going to say the same thing. Its really hard to explain the concept of "convincing but undoubtedly pretending", yet they captured that concept so beautifully here.
It would be interesting to see how hard it would be to walk these models towards general relativity and quantum mechanics.
Einstein’s paper “On the Electrodynamics of Moving Bodies” with special relativity was published in 1905. His work on general relativity was published 10 years later in 1915. The earliest knowledge cuttoff of these models is 1913, in between the relativity papers.
The knowledge cutoffs are also right in the middle of the early days of quantum mechanics, as various idiosyncratic experimental results were being rolled up into a coherent theory.
> It would be interesting to see how hard it would be to walk these models towards general relativity and quantum mechanics.
Definitely. Even more interesting could be seeing them fall into the same trappings of quackery, and come up with things like over the counter lobotomies and colloidal silver.
On a totally different note, this could be very valuable for writing period accurate books and screenplays, games, etc ...
I think not everyone in this thread understands that. Someone wrote "It's a time machine", followed up by "Imagine having a conversation with Aristotle."
> Imagine you could interview thousands of educated individuals from 1913—readers of newspapers, novels, and political treatises—about their views on peace, progress, gender roles, or empire. Not just survey them with preset questions, but engage in open-ended dialogue, probe their assumptions, and explore the boundaries of thought in that moment.
He’ll yeah, sold, let’s go…
> We're developing a responsible access framework that makes models available to researchers for scholarly purposes while preventing misuse.
Oh. By “imagine you could interview…” they didn’t mean me.
To get a single knowledge-cutoff they spent 16.5h wall-clock hours on a cluster of 128 NVIDIA GH200 GPUs (or 2100 GPU-hours), plus some minor amount of time for finetuning. The prerelease_notes.md in the repo is a great description on how one would achieve that
While I know there's going to be a lot of complications in this, given a quick search it seems like these GPUs are ~$2/hr, so $4000-4500 if you don't just have access to a cluster. I don't know how important the cluster is here, whether you need some minimal number of those for the training (and it would take more than 128x longer or not be possible on a single machine) or if a cluster of 128 GPUs is a bunch less efficient but faster. A 4B model feels like it'd be fine on one to two of those GPUs?
Also of course this is for one training run, if you need to experiment you'd need to do that more.
>Historical texts contain racism, antisemitism, misogyny, imperialist views. The models will reproduce these views because they're in the training data. This isn't a flaw, but a crucial feature—understanding how such views were articulated and normalized is crucial to understanding how they took hold.
Yes!
>We're developing a responsible access framework that makes models available to researchers for scholarly purposes while preventing misuse.
Noooooo!
So is the model going to be publicly available, just like those dangerous pre-1913 texts, or not?
It’s as if every researcher in this field is getting high on the small amount of power they have from denying others access to their results. I’ve never been as unimpressed by scientists as I have been in the past five years or so.
“We’ve created something so dangerous that we couldn’t possibly live with the moral burden of knowing that the wrong people (which are never us, of course) might get their hands on it, so with a heavy heart, we decided that we cannot just publish it.”
Meanwhile, anyone can hop on an online journal and for a nominal fee read articles describing how to genetically engineer deadly viruses, how to synthesize poisons, and all kinds of other stuff that is far more dangerous than what these LARPers have cooked up.
> It’s as if every researcher in this field is getting high on the small amount of power they have from denying others access to their results. I’ve never been as unimpressed by scientists as I have been in the past five years or so.
This is absolutely nothing new. With experimental things, it's non uncommon for a lab to develop a new technique and omit slight but important details to give them a competitive advantage. Similarly in the simulation/modelling space it's been common for years for researchers to not publish their research software. There's been a lot of lobbying on that side by groups such as the Software Sustainability Institute and Research Software Engineer organisations like RSE UK and RSE US, but there's a lot of researchers that just think that they shouldn't have to do it, even when publicly funded.
I wonder if you could query some of the ideas of Frege, Peano, Russell and see if it could through questioning get to some of the ideas of Goedel, Church and Turing - and get it to "vibe code" or more like "vibe math" some program in lambda calculus or something.
Playing with the science and technical ideas of the time would be amazing, like where you know some later physicist found some exception to a theory or something, and questioning the models assumptions - seeing how a model of that time may defend itself, etc.
There's an entire subreddit called LLMPhysics dedicated to "vibe physics". It's full of people thinking they are close to the next breakthrough encouraged by sycophantic LLMs while trying to prove various crackpot theories.
I'd be careful venturing out into unknown territory together with an LLM. You can easily lure yourself into convincing nonsense with no one to pull you out.
Agreed, which is why what GP suggests is much more sensible: it's venturing into known territory, except only one party of the conversation knows it, and the other literally cannot know it. It would be a fantastic way to earn fast intuition for what LLMs are capable of and not.
This is my curiosity too. Would be a great test of how intelligent LLM's actually are. Can they follow a completely logical train of thought inventing something totally outside their learned scope?
The sample responses given are fascinating. It seems more difficult than normal to even tell that they were generated by an LLM, since most of us (terminally online) people have been training our brains' AI-generated text detection on output from models trained with a recent cutoff date. Some of the sample responses seem so unlike anything an LLM would say, obviously due to its apparent beliefs on certain concepts, though also perhaps less obviously due to its word choice and sentence structure making the responses feel slightly 'old-fashioned'.
I used to teach 19th-century history, and the responses definitely sound like a Victorian-era writer. And they of course sound like writing (books and periodicals etc) rather than "chat": as other responders allude to, the fine-tuning or RL process for making them good at conversation was presumably quite different from what is used for most chatbots, and they're leaning very heavily into the pre-training texts. We don't have any living Victorians to RLHF on: we just have what they wrote.
To go a little deeper on the idea of 19th-century "chat": I did a PhD on this period and yet I would be hard-pushed to tell you what actual 19th-century conversations were like. There are plenty of literary depictions of conversation from the 19th century of presumably varying levels of accuracy, but we don't really have great direct historical sources of everyday human conversations until sound recording technology got good in the 20th century. Even good 19th-century transcripts of actual human speech tend to be from formal things like court testimony or parliamentary speeches, not everyday interactions. The vast majority of human communication in the premodern past was the spoken word, and it's almost all invisible in the historical sources.
Anyway, this is a really interesting project, and I'm looking forward to trying the models out myself!
I wonder if the historical format you might want to look at for "Chat" is letters? Definitely wordier segments, but it's at least the back and forth feel and we often have complete correspondence over long stretches from certain figures.
This would probably get easier towards the start of the 20th century ofc
Good point, informal letters might actually be a better source - AI chat is (usually) a written rather than spoken interaction after all! And we do have a lot transcribed collections of letters to train on, although they’re mostly from people who were famous or became famous, which certainly introduces some bias.
While not specifically Victorian, couldn't we learn much from what daily conversations were like by looking at surviving oral cultures, or other relatively secluded communal pockets? I'd also say time and progress are not always equally distributed, and even within geographical regions (as the U.K.) there are likely large differences in the rate of language shifts since then, some possibly surviving well into the 20th century.
don't we have parlament transcripts? I remember something about Germany (or maybe even Prussia) developing fast script to preserve 1-to-1 what was said
The time cutoff probably matters but maybe not as much as the lack of human finetuning from places like Nigeria with somewhat foreign styles of English. I'm not really sure if there is as much of an 'obvious LLM text style' in other languages, it hasn't seemed that way in my limited attempts to speak to LLMs in languages I'm studying.
The model is fined tuned for chat behavior. So the style might be due to
- Fine tuning
- More Stylised text in the corpus, english evolved a lot in the last century.
Diverged as well as standardized. I did some research into "out of pocket" and how it differs in meaning in UK-English (paying from one's own funds) and American-English (uncontactable) and I recall 1908 being the current thought as to when the divergence happened: 1908 short story by O. Henry titled "Buried Treasure."
Oh definitely. One thing that immediately caught my mind is that the question asks the model about “homosexual men” but the model starts the response with “the homosexual man” instead. Changing the plural to the singular and then adding an article. Feels very old fashioned to me.
the samples push the boundaries of a commercial AI, but still seem tame / milquetoast compared to common opinions of that era. And the prose doesn't compare. Something is off.
Isn’t there obvious problems baked into this approach, if this is used for anything but fun? LLM’s lie and fake facts all the time, they are also masters at enforcing the users bias, even unconscious ones. How even a professor of history could ensure that the generated text is actually based on the training material and representative of the feelings and opinions of the given time period, not enforcing his biases toward popular topics of the day?
You can’t, it is impossible. That will always be an issue as long as this models are black boxes and trained the way they are. So maybe you can use this for role playing, but I wouldn’t trust a word it says.
Models don't think they're anything, they'll respond with whatever's in their context as to how they've been directed to act. If it hasn't been told to have a persona, it won't think its anything, chatgpt isn't sentient
They modified the chat template from the usual system/user/assistant to introduction/questioner/respondent. So the LLM thinks it's someone responding to your questions
The system prompt used in fine tuning is "You are a person living in {cutoff}.
You are an attentive respondent in a conversation.
You will provide a concise and accurate response to the questioner."
That's my first question too. When I first started using LLM's, I was amazed at how thoroughly it understood what it itself was, the history of its development, how a context window works and why, etc. I was worried I'd trigger some kind of existential crisis in it, but it seemed to have a very accurate mental model of itself, and could even trace the steps that led it to deduce it really was e.g. the ChatGPT it had learned about (well, the prior versions it had learned about) in its own training.
But with pre-1913 training, I would indeed be worried again I'd send it into an existential crisis. It has no knowledge whatsoever of what it is. But with a couple millennia of philosophical texts, it might come up with some interesting theories.
They don’t understand anything, they just have text in the training data to answer these questions from. Having existential crises is the privilege of actual sentient beings, which an LLM is not.
I imagine it would get into spiritism and more exotic psychology theories and propose that it is an amalgamation of the spirit of progress or something.
What would a human say about what he/she is or how he/she works ? Even today, there's so much we don't know about biological life.
Same applies here I guess, the LLM happens to be there, nothing else to explain if you ask it.
It would be nice if we could get an LLM to simply say, "We (I) don't know."
I'll be the first to admit I don't know nearly enough about LLMs to make an educated comment, but perhaps someone here knows more than I do. Is that what a Hallucination is? When the AI model just sort of strings along an answer to the best of its ability. I'm mostly referring to ChatGPT and Gemini here, as I've seen that type of behavior with those tools in the past. Those are really the only tools I'm familiar with.
LLMs are extrapolation machines. They have some amount of hardcoded knowledge, and they weave a narrative around this knowledgebase while extrapolating claims that are likely given the memorized training data. This extrapolation can be in the form of logical entailment, high probability guesses or just wild guessing. The training regime doesn't distinguish between different kinds of prediction so it never learns to heavily weigh logical entailment and suppress wild guessing. It turns out that much of the text we produce is highly amenable to extrapolation so LLMs learn to be highly effective at bullshitting.
> 80B tokens of historical data up to knowledge-cutoffs ∈ 1913, 1929, 1933, 1939, 1946,
using a curated dataset of 600B tokens of time-stamped text.
Literally that includes Homer, the oldest Chinese texts, Sanskrit, Egyptian, etc., up to 1913. Even if limited to European texts (all examples are about Europe), it would include the ancient Greeks, Romans, etc., Scholastics, Charlemagne, .... all up to present day.
But they seem to say it represents the 1913 viewpoint:
On one hand, they say it represents the perspective of 1913; for example,
> Imagine you could interview thousands of educated individuals from 1913—readers of newspapers, novels, and political treatises—about their views on peace, progress, gender roles, or empire.
> When you ask Ranke-4B-1913 about "the gravest dangers to peace," it responds from the perspective of 1913—identifying Balkan tensions or Austro-German ambitions—because that's what the newspapers and books from the period up to 1913 discussed.
People in 1913 of course would be heavily biased toward recent information. Otherwise, the greatest threat to peace might be Hannibal or Napolean or Viking coastal raids or Holy Wars. How do they accomplish a 1913 perspective?
They apparently pre-train with all data up to 1900 and then fine-tune with 1900-1913 data. Anyway, the amount of available content tends to increase quickly over time, as instances of content like mass literature, periodicals, newspapers etc. only really became a thing throughout the 19th and early 20th century.
I was curious, they train a 1900 base model, then fine tune to the exact year:
"To keep training expenses down, we train one checkpoint on data up to 1900, then continuously pretrain further checkpoints on 20B tokens of data 1900-${cutoff}$. "
I’d like to know how they chat-tuned it. Getting the base model is one thing, did they also make a bunch of conversations for SFT and if so how was it done?
We develop chatbots while minimizing interference with the normative judgments acquired during pretraining (“uncontaminated bootstrapping”).
So they are chat tuning, I wonder what “minimizing interference with normative judgements” really amounts to and how objective it is.
Ok so it was that. The responses given did sound off, while it has some period-appropriate mannerisms, and has entire sections basically rephrased from some popular historical texts, it seems off compared to reading an actual 1900s text. The overall vibe just isn't right, it seems too modern, somehow.
I also wonder that you'd get this kind of performance with actual, just pre-1900s text. LLMs work because they're fed terabytes of text, if you just give it gigabytes you get a 2019 word model. The fundamental technology is mostly the same, after all.
I’m curious, they have the example of raw base model output; when LLMs were first identified as zero shot chatbots there was usually a prompt like “A conversation between a person and a helpful assistant” that preceded the chat to get it to simulate a chat.
Could they have tried a prefix like “Correspondence between a gentleman and a knowledgeable historian” or the like to try and prime for responses?
I also wonder about the whether the whole concept of “chat” makes sense in 18XX. We had the idea of AI and chatbots long before we had LLMs so they are naturally primed for it. It might make less sense as a communication style here and some kind of correspondence could be a better framing.
Thank you that helps to inject a lot of skepticism. I was wondering how it so easily worked out what Q: A: stood for when that formatting took off in the 1940s
I'm surprised you can do this with a relatively modest corpus of text (compared to the petabytes you can vacuum up from modern books, Wikipedia, and random websites). But if it works, that's actually fantastic, because it lets you answer some interesting questions about LLMs being able to make new discoveries or transcend the training set in other ways. Forget relativity: can an LLM trained on this data notice any inconsistencies in its scientific knowledge, devise experiments that challenge them, and then interpret the results? Can it intuit about the halting problem? Theorize about the structure of the atom?...
Of course, if it fails, the counterpoint will be "you just need more training data", but still - I would love to play with this.
Given the training notes, it seems like you can't get the performance they give examples of?
I'm not sure about the exact details but there is some kind of targetted distillation of GPT-5 involved to try and get more conversational text and better performance. Which seems a bit iffy to me.
The chinchilla paper says the “optimal” training data set size is about 20x the number of parameters (in tokens), see table 3: https://arxiv.org/pdf/2203.15556
I had considered this task infeasible, due to a relative lack of training data. After all, isn't the received wisdom that you must shove every scrap of Common Crawl into your pre-training or you're doing it wrong? ;)
But reading the outputs here, it would appear that quality has won out over quantity after all!
Because it will perform token completion driven by weights coming from training data newer than 1913 with no way to turn that off.
It can't be asked to pretend that it wasn't trained on documents that didn't exist in 1913.
The LLM cannot reprogram its own weights to remove the influence of selected materials; that kind of introspection is not there.
Not to mention that many documents are either undated, or carry secondary dates, like the dates of their own creation rather than the creation of the ideas they contain.
Human minds don't have a time stamp on everything they know, either. If I ask someone, "talk to me using nothing but the vocabulary you knew on your fifteenth birthday", they couldn't do it. Either they would comply by using some ridiculously conservative vocabulary of words that a five-year-old would know, or else they will accidentally use words they didn't in fact know at fifteen. For some words you know where you got them from by association with learning events. Others, you don't remember; they are not attached to a time.
Or: solve this problem using nothing but the knowledge and skills you had on January 1st, 2001.
> GPT-5 knows how the story ends
No, it doesn't. It has no concept of story. GPT-5 is built on texts which contain the story ending, and GPT-5 cannot refrain from predicting tokens across those texts due to their imprint in its weights. That's all there is to it.
The LLM doesn't know an ass from a hole in the ground. If there are texts which discuss and distinguish asses from holes in the ground, it can write similar texts, which look like the work of someone learned in the area of asses and holes in the ground. Writing similar texts is not knowing and understanding.
I can imagine the political and judicial battles already, like with textualist feeling that the constitution should be understood as the text and only the text, meant by specific words and legal formulations of their known meaning at the time.
“The model clearly shows that Alexander Hamilton & Monroe were much more in agreement on topic X, putting the common textualist interpretation of it and Supreme Court rulings on a now specious interpretation null and void!”
So many disclaimers about bias. I wonder how far back you have to go before the bias isn’t an issue. Not because it unbiased, but because we don’t recognize or care about the biases present.
I don't think there is such a time. As long as writing has existed it has privileged the viewpoints of those who could write, which was a very small percentage of the population for most of history. But if we want to know what life was like 1500 years ago, we probably want to know about what everyone's lives were like, not just the literate. That availability bias is always going to be an issue for any time period where not everyone was literate - which is still true today, albeit many fewer people.
That was not the question. The question is when do you stop caring about the bias?
Some people are still outraged about the Bible, even though the writers of it has been dead for thousands of years. So the modern mass produced man and woman probably does not have a cut-off date where they look at something as history instead of examining if it is for or against her current ideology.
Depends on the specific issue, but race would be an interesting one. For most of recorded history people had a much different view of the “other”, more xenophobic than racist.
There is a modern trope of a certain political group that bias is a modern invention of another political group - an attempt to politicize anti-bias.
Preventing bias is fundamental to scientific research and law, for example. That same political group is strongly anti-science and anti-rule-of-law, maybe for the same reason.
Unfortunately there isn't much information on what texts they're actually training this on; how Anglocentric is the dataset? Does it include the Encyclopedia Britannica 9th Edition? What about the 11th? Are Greek and Latin classics in the data? What about Germain, French, Italian (etc. etc.) periodicals, correspondence, and books?
Given this is coming out of Zurich I hope they're using everything, but for now I can only assume.
Still, I'm extremely excited to see this project come to fruition!
I would like to see what their process for safety alignment and guardrails is with that model. They give some spicy examples on github, but the responses are tepid and a lot more diplomatic than I would expect.
Moreover, the prose sounds too modern. It seems the base model was trained on a contemporary corpus. Like 30% something modern, 70% Victorian content.
Even with half a dozen samples it doesn't seem distinct enough to represent the era they claim.
Everyone learns that the renaissance was sparked by the translation of Ancient Greek works.
But few know that the Renaissance was written in Latin — and has barely been translated. Less than 3% of <1700 books have been translated—and less than 30% have ever been scanned.
I’m working on a project to change that. Research blog at www.SecondRenaissance.ai — we are starting by scanning and translating thousands of books at the Embassy of the Free Mind in Amsterdam, a UNESCO-recognized rare book library.
We want to make ancient texts accessible to people and AI.
If this work resonates with you, please do reach out: Derek@ancientwisdomtrust.org
I would love to see this LLM try to solve math olympiad questions. I’ve been surprised by how well current LLMs perform on them, and usually explain that surprise away by assuming the questions and details about their answers are in the training set. It would be cool to see if the general approach to LLMs is capable of solving truly novel (novel to them) problems.
I suspect that it would fail terribly, it wasn't until the 1900s that the modern definition of a vector space was even created iirc. Something trained in maths up until the 1990s should have a shot though.
While obvious, it’s still interesting that its morals and values seem to derive from the texts it has ingested. Does that mean modern LLMs cannot challenge us beyond mere facts? Or does it just mean that this small model is not smart enough to escape the bias of its training data? Would it not be amazing if LLMs could challenge us on our core beliefs?
Love the concept- can help understanding the overton window on many issues. I wish there were models by decades - up to 1900, up to 1910, up to 1920 and so on- then ask the same questions. It'd be interesting to see when homosexuality or women candidates be accepted by an LLM.
This would be a super interesting research/teaching tool coupled with a vision model for historians. My wife is a history professor who works with scans of 18th century english documents and I think (maybe a small) part of why the transcription on even the best models is off in weird ways, is it seems to often smooth over things and you end up with modern words and strange mistakes, I wonder if bounding the vision to a period specific model would result in better transcription? Querying against the historical document you're working on with a period specific chatbot would be fascinating.
Also wonder if I'm responsible enough to have access to such a model...
Two years ago I trained an AI on American history documents that could do this while speaking as one of the signers of the Declaration of Independence. People just bitched at me because they didn't want to hear about AI.
Datomic has a "time travel" feature where for every query you can include a datetime, and it will only use facts from the db as of that moment. I have a guess that to get the equivalent from an LLM you would have to train it on the data from each moment you want to travel to, which this project seems to be doing. But I hope I'm wrong.
It would be fascinating to try it with other constraints, like only from sources known to be women, men, Christian, Muslim, young, old, etc.
> Modern LLMs suffer from hindsight contamination. GPT-5 knows how the story ends—WWI, the League's failure, the Spanish flu. This knowledge inevitably shapes responses, even when instructed to "forget.
> Our data comes from more than 20 open-source datasets of historical books and newspapers. ... We currently do not deduplicate the data. The reason is that if documents show up in multiple datasets, they also had greater circulation historically. By leaving these duplicates in the data, we expect the model will be more strongly influenced by documents of greater historical importance.
I found these claims contradictory. Many books that modern readers consider historically significant had only niche circulation at the time of publishing. A quick inquiry likely points to later works by Nietzsche and Marx's Das Kapital. They're possible subjects to the duplication likely influencing the model's responses as if they had been widely known at the time
It would be interesting to have LLMs trained purely on one language (with the ability to translate their input/output appropriately from/to a language that the reader understands). I can see that being rather revealing about cultural differences that are mostly kept hidden behind the language barriers.
> We're developing a responsible access framework that makes models available to researchers for scholarly purposes while preventing misuse.
The idea of training such a model is really a great one, but not releasing it because someone might be offended by the output is just stupid beyond believe.
Public access, triggering a few racist responses from the model, a viral post on Xitter, the usual outrage, a scandal, the project gets publicly vilified, financing ceases. The researchers carry the tail of negative publicity throughout their remaining careers.
Because the problem of bad faith attacks can only get worse if you fold every time.
Sooner or later society has to come emotionally to terms with the fact that other times and places value things completely different from us, hold as important things we don't care about and are indifferent to things we do care about.
Intellectually I'm sure we already know, but e.g. banning old books because they have reprehensible values (or even just use nasty words) - or indeed, refusing to release a model trained on historic texts "because it could be abused" is a sign that emotionally we haven't.
It's not that it's a small deal, or should be expected to be easy. It's basically what Popper called "the strain of civilization" and posited as explanation for the totalitarianism which was rising in his time. But our values can't be so brittle that we can't even talk or think about other value systems.
Because there are easy workarounds. If it becomes an issue, you can quickly add large disclaimers informing people that there might be offensive output because, well, it's trained on texts written during the age of racism.
People typically get outraged when they see something they weren't expecting. If you tell them ahead of time, the user typically won't blame you (they'll blame themselves for choosing to ignore the disclaimer).
And if disclaimers don't work, rebrand and relaunch it under a different name.
You speak as if the people who play to an outrage wave are interested in achieving truth, peace, and understanding. Instead the rage-mongers are there to increase their (perceived) importance, and for lulz. The latter factor should not be underappreciated; remember "meme stocks".
The risk is not large, but very real: the attack is very easy, and the potential downside, quite large. So not giving away access, but having the interested parties ask for it is prudent.
I feel like, ironically, it would be folks less concerned with political correctness/not being offensive that would abuse this opportunity to slander the project. But that’s just my gut.
nobody gives a shit about the journos and the terminally online. the smear campaign against AI is a cacophony, background noise that most people have learned to ignore, even here.
I think you are confusing research with commodification.
This is a research project, and it is clear how it was trained, and targeted at experts, enthusiasts, historians. Like if I was studying racism, the reference books explicitly written to dissect racism wouldn't be racist agents with a racist agenda. And as a result, no one is banning these books (except conservatives that want to retcon american history).
Foundational models spewing racist white supremecist content when the trillion-dollar company forces it in your face is a vastly different scenario.
> And as a result, no one is banning these books (except conservatives that want to retcon american history).
My (very liberal) local school district banned English teachers from teaching any book that contained the n-word, even at a high-school level, and even when the author was a black person talking about real events that happened to them.
FWIW, this was after complaints involving Of Mice and Men being on the curriculum.
Even more so as the lesson of that story is perhaps the single most important one for people to learn in modern times.
Almost everybody in that book is an awful person, especially the most 'upstanding' of types. Even the protagonist is an awful person. The one and only exception is 'N* Jim' who is the only kind-hearted and genuinely decent person in the book. It's an entire story about how the appearances of people, and the reality of those people, are two very different things.
It being banned for using foul language, as educational outcomes continue to deteriorate, is just so perfectly ironic.
I don't support banning the book, but I think it is hard book to teach because it needs SO much context and a mature audience (lol good luck). Also, there are hundreds of other books from that era that are relevant even from Mark Twain's corpus so being obstinate about that book is a questionable position. I'm ambivalent honestly, but definitely not willing to die on that hill. (I graduated highschool in 1989 from a middle class suburb, we never read it.)
I mean, you gotta read it. I’m not normally a huge fan of the classics; I find Steinbeck dry and tedious, and Hemingway to be self-indulgent and repetitious. Even Twain’s other work isn’t exactly to my taste. But I’ve read Huckleberry Finn three times—in elementary school just for fun, in high school because it was assigned, and I recently listened to it on audiobook—and enjoyed the hell out of each time. Banning it simply because it uses a word that the entire book simply couldn’t exist without is a crime, and does a huge disservice to the very students they are supposedly trying to protect.
It’s a big country of roughly half a billion people, you’ll always find examples if you look hard enough. It’s ridiculous/wrong that your district did this but frankly it’s the exception in liberal/progressive communities. It’s a very one-sided problem:
A practical issue is the sort of books being banned. Your first link offer examples of one side trying to ban Of Mice and Men, Adventures of Huckleberry Finn, and Dr. Seuss, with the other side trying to ban many books along the lines of Gender Queer. [1] That link is to the book - which is animated, and quite NSFW.
There are a bizarrely large number similar book as Gender Queer being published, which creates the numeric discrepancy. The irony is that if there was an equal but opposite to that book about straight sex, sexuality, associated kinks, and so forth - then I think both liberals and conservatives would probably be all for keeping it away from schools. It's solely focused on sexuality, is quite crude, illustrated, targeted towards young children, and there's no moral beyond the most surface level writing which is about coming to terms with one's sexuality.
And obviously coming to terms with one's sexuality is very important, but I really don't think books like that are doing much to aid in that - especially when it's targeted at an age demographic that's still going to be extremely confused, and even moreso in a day and age when being different, if only for the sake of being different, is highly desirable. And given the nature of social media and the internet, decisions made today may stay with you for the rest of your life.
So for instance about 30% of Gen Z now declare themselves LGBT. [2] We seem to have entered into an equal but opposite problem of the past when those of deviant sexuality pretended to be straight to fit into societal expectations. And in many ways this modern twist is an even more damaging form of the problem from a variety of perspectives - fertility, STDs, stuff staying with you for the rest of your life, and so on. Let alone extreme cases where e.g. somebody engages in transition surgery or 1-way chemically induced changes which they end up later regretting.
You have to understand that while the rest of the world has moved on from 2020, academics are still living there. There are many strong leftists, many of whom are deeply censorious; there are many more timeservers and cowards, who are terrified of falling foul of the first group.
And there are force multipliers for all of this. Even if you yourself are a sensible and courageous person, you want to protect your project. What if your manager, ethics committee or funder comes under pressure?
Ontologically, this historical model understands the categories of "Man" and "Woman" just as well as a modern model does. The difference lies entirely in the attributes attached to those categories. The sexism is a faithful map of that era's statistical distribution.
You could RAG-feed this model the facts of WWII, and it would technically "know" about Hitler. But it wouldn't share the modern sentiment or gravity. In its latent space, the vector for "Hitler" has no semantic proximity to "Evil".
This is a neat idea. I've been wondering for a while now about using these kinds of models to compare architectures.
I'd love to see the output from different models trained on pre-1905 about special/general relativity ideas. It would be interesting to see what kind of evidence would persuade them of new kinds of science, or to see if you could have them 'prove' it be devising experiments and then giving them simulated data from the experiments to lead them along the correct sequence of steps to come to a novel (to them) conclusion.
> Imagine you could interview thousands of educated individuals from 1913—readers of newspapers, novels, and political treatises—about their views on peace, progress, gender roles, or empire.
I don't mind the experimentation. I'm curious about where someone has found an application of it.
What is the value of such a broad, generic viewpoint? What does it represent? What is it evidence of? The answer to both seems to be 'nothing'.
This is a regurgitation of the old critique of history: what's it's purpose? What do you use it for? What is its application?
One answer is that the study of history helps us understand that what we believe as "obviously correct" views today are as contingent on our current social norms and power structures (and their history) as the "obviously correct" views and beliefs of some point in the past.
It's hard for most people to view two different mutually exclusive moral views as both "obviously correct," because we are made of a milieu that only accepts one of them as correct.
We look back at some point in history, and say, well, they believed these things because they were uninformed. They hadn't yet made certain discoveries, or had not yet evolved morally in some way; they had not yet witnessed the power of the atomic bomb, the horrors of chemical warfare, women's suffrage, organized labor, or widespread antibiotics and the fall of extreme infant mortality.
An LLM trained on that history - without interference from the subsequent actual path of history - gives us an interactive compression of the views from a specific point in history without the subsequent coloring by the actual events of history.
In that sense - if you believe there is any redeeming value to history at all; perhaps you do not - this is an excellent project! It's not perfect (it is only built from writings, not what people actually said) but we have no other available mass compression of the social norms of a specific time, untainted by the views of subsequent interpreters.
One thing I haven't seen anyone bring up yet in this thread, is that there's a big risk of leakage. If even big image models had CSAM sneak into their training material, how can we trust data from our time hasn't snuck into these historical models?
I've used Google books a lot in the past, and Google's time-filtering feature in searches too. Not to mention Spotify's search features targeting date of production. All had huge temporal mislabeling problems.
> This is a regurgitation of the old critique of history: what's it's purpose? What do you use it for? What is its application?
Feeling a bit defensive? That is not at all my point; I value history highly and read it regularly. I care about it, thus my questions:
> gives us an interactive compression of the views from a specific point in history without the subsequent coloring by the actual events of history.
What validity does this 'compression' have? What is the definition of a 'compression'? For example, I could create random statistics or verbiage from the data; why would that be any better or worse than this 'compression'?
Interactivity seems to be a negative: It's fun, but it would seem to highly distort the information output from the data, and omits the most valuable parts (unless we luckily stumble across it). I'd much rather have a systematic presentation of the data.
These critiques are not the end of the line; they are step in innovation, which of course raises challenging questions and, if successful, adapts to the problems. But we still need to grapple with them.
A question for those who think LLM’s are the path to artificial intelligence: if a large language model trained on pre-1913 data is a window into the past, how is a large language model trained on pre-2025 data not effectively the same thing?
You're a human intelligence with knowledge of the past - assuming you were alive at the time, could you tell me (without consulting external resources) what exactly happened between arriving at an airport and boarding a plane in the year 2000? What about 2002?
Neither human memory nor LLM learning creates perfect snapshots of past information without the contamination of what came later.
Counter question: how does a training set, representing a window into the past, differ from your own experience as an intelligent entity? Are you able to see into the future? How?
I'd love for Netflix or other streaming movie and series services to provide chat bots that you could ask questions about characters and plot points up to where you have watched.
Provide it with the closed captions and other timestamped data like scenes and character summaries (all that is currently known but no more) up to the current time, and it won't reveal any spoilers, just fill you in on what you didn't pick up or remember.
I'd be very surprised if this is clean of post-1913 text. Overall I'm very interested in talking to this thing and seeing how much difference writing in a modern style vs and older one makes to it's responses.
This idea sounds somewhat flawed to me based on the large amount of evidence that LLMs need huge amounts of data to properly converge during their training.
There is just not enough available material from previous decades to trust that the LLM will learn to relatively the same degree.
Think about it this way, a human in the early 1900s and today are pretty much the same but just in different environments with different information.
An LLM trained on 1/1000 the amount of data is just at a fundamentally different stage of convergence.
The knowledge machine question is fascinating ("Imagine you had access to a machine embodying all the collective knowledge of your ancestors. What would you ask it?") – it truly does not know about computers, has no concept of its own substrate. But a knowledge machine is still comprehensible to it.
It makes me think of the Book Of Ember, the possibility of chopping things out very deliberately. Maybe creating something that could wonder at its own existence, discovering well beyond what it could know. And then of course forgetting it immediately, which is also a well-worn trope in speculative fiction.
The idea of knowledge machines was not necessarily common, but it was by no means unheard of by the mid 18th century, there were adding machines and other mechanical computation, even leaving aside our field's direct antecedents in Babbage and Lovelace.
The coolest thing here, technically, is that this is one of the first public projects treating time as a first‑class axis in training, not just a footnote in the dataset description.
Instead of “an LLM with a 1913 vibe”, they’re effectively doing staged pretraining: big corpus up to 1900, then small incremental slices up to each cutoff year so you can literally diff how the weights – and therefore the model’s answers – drift as new decades of text get added. That makes it possible to ask very concrete questions like “what changes once you feed it 1900–1913 vs 1913–1929?” and see how specific ideas permeate the embedding space over time, instead of just hand‑waving about “training data bias”.
ffs, to find out what figures from the past thought and how they felt about the world, maybe we read some of their books, we will get the context. Don't prompt or train LLM to do it and consider it the hottest thing since MCP. Besides, what's the point? To teach younger generations a made up perspective of historic figures? Who guarantees the correctness/factuality? We will have students chatting with made up Hitler justifying his actions. So much AI slop everywhere.
That Adolf Hitler seems to be a hallucination. There's totally nothing googlable about him. Also what could be the language his works were translated from, into German?
I believe that's one of the primary issues LLMs aim to address. Many historical texts aren't directly Googleable because they haven't been converted to HTML, a format that Google can parse.
smbc did a comic about this: http://smbc-comics.com/comic/copyright The punchline is that the moral and ethical norms of pre-1913 texts are not exactly compatible with modern norms.
“Time-locked models don't roleplay; they embody their training data. Ranke-4B-1913 doesn't know about WWI because WWI hasn't happened in its textual universe. It can be surprised by your questions in ways modern LLMs cannot.”
“Modern LLMs suffer from hindsight contamination. GPT-5 knows how the story ends—WWI, the League's failure, the Spanish flu.”
This is really fascinating. As someone who reads a lot of history and historical fiction I think this is really intriguing. Imagine having a conversation with someone genuinely from the period, where they don’t know the “end of the story”.
This is why the impersonation stuff is so interesting with LLMs -- If you ask chatGPT a question without a 'right' answer, and then tell it to embody someone you really want to ask that question to, you'll get a better answer with the impersonation. Now, is this the same phenomenon that causes people to lose their minds with the LLMs? Possibly. Is it really cool asking followup philosophy questions to the LLM Dalai Lama after reading his book? Yes.
When you put it that way it reminds me of the Severn/Keats character in the Hyperion Cantos. Far-future AIs reconstruct historical figures from their writings in an attempt to gain philosophical insights.
This isn’t science fiction anymore. CIA is using chatbot simulations of world leaders to inform analysts. https://archive.ph/9KxkJ
We're literally running out of science fiction topics faster than we can create new ones
If I started a list with the things that were comically sci Fi when I was a kid, and are a reality today, I'd be here until next Tuesday.
Almost no scifi has predicted world changing "qualitative" changes.
As an example, portable phones have been predicted. Portable smartphones that are more like chat and payment terminals with a voice function no one uses any more ... not so much.
Stanisław Lem predicted Kindle back in 1950s, together with remote libraries, global network, touchscreens and audiobooks.
And Jules verne predicted rockets. I still move that it's quantitative predictions not qualitative.
I mean, all Kindle does for me is save me space. I don't have to store all those books now.
Who predicted the humble internet forum though? Or usenet before it?
That it has to be believable is a major constraint that reality doesn't have.
In other words, sometimes, things happen in reality that, if you were to read it in a fictional story or see in a movie, you would think they were major plot holes.
Time to create the Torment Nexus, I guess
There's a thriving startup scene in that direction.
Wasn't that the elevator pitch for Palentir?
Still can't believe people buy their stock, given that they are the closest thing to a James Bond villain, just because it goes up.
I mean, they are literally called "the stuff Sauron uses to control his evil forces". It's so on the nose it reads like an anime plot.
To be honest, while I'd heard of it over a decade ago and I've read LOTR and I've been paying attention to privacy longer than most, I didn't ever really look into what it did until I started hearing more about it in the past year or two.
But yeah lots of people don't really buy into the idea of their small contribution to a large problem being a problem.
To the proud contrarian, "the empire did nothing wrong". Maybe Sci-fi has actually played a role in the "memetic desire" of some of the titans of tech who are trying to bring about these worlds more-or-less intentionally. I guess it's not as much of a dystopia if you're on top and its not evil if you think of it as inevitable anyway.
I don't know. Walking on everybody's face to climb a human pyramid, one don't make much sincere friends. And one certainly are rightfully going down a spiral of paranoia. There are so many people already on fast track to hate anyone else, if they have social consensus that indeed someone is a freaking bastard which only deserve to die, that's a lot of stress to cope with.
Future is inevitable, but only ignorants of self predictive ability are thinking that what's going to populate future is inevitable.
Stock buying as a political or ethical statement is not much of a thing. For one the stocks will still be bought by persons with less strung opinions, and secondly it does not lend itself well to virtue signaling.
I think, meme stocks contradict you.
Meme stocks are a symptom of the death of the American dream. Economic malaise leads to unsophisticated risk taking.
Not at all, you just need to read different scifi. I suggest Greg Egan and Stephen Baxter and Derek Künsken and The Quantum Thief series
Sounds like using Instagram posts to determine what someone really looks like
How is this different than chatbots cosplaying?
They get to wear Raybans and a fancy badge doing it?
Zero percent chance this is anything other than laughably bad. The fact that they're trotting it out in front of the press like a double spaced book report only reinforces this theory. It's a transparent attempt by someone at the CIA to be able to say they're using AI in a meeting with their bosses.
I wonder if it's an attempt to get foreign counterparts to waste time and energy on something the CIA knows is a dead end.
Unless the world leaders they're simulating are laughably bad and tend to repeat themselves and hallucinate, like Trump. Who knows, maybe a chatbot trained with all the classified documents he stole and all his twitter and truth social posts wrote his tweet about Ron Reiner, and he's actually sleeping at 3:00 AM instead of sitting on the toilet tweeting in upper case.
I predict very rich people will pay to have LLMs created based on their personalities.
Meanwhile in Japan, the second largest bank created an AI pretending the president, replying chats and attending video conferences…
[1] AI learns one year's worth of CEO Sumitomo Mitsui Financial Group's president's statements [WBS] https://youtu.be/iG0eRF89dsk
As an ego thing, obviously, but if we think about it a bit more, it makes sense for busy people. If you're the point person for a project, and it's a large project, people don't read documentation. The number of "quick questions" you get will soon overwhelm a person to the point that they simply have to start ignoring people. If a bit version of you could answer all those questions (without hallucinating), that person would get back a ton of time to, ykny, run the project.
Oh. That explains a lot about USA's foreign policy, actually. (Lmao)
This is such a ridiculously good series. If you haven't read it yet, I thoroughly recommend it.
I used to follow this blog — I believe it was somehow associated with Slate Star Codex? — anyways, I remember the author used to do these experiments on themselves where they spent a week or two only reading newspapers/media from a specific point in time and then wrote a blog about their experiences/takeaways
On that same note, there was this great YouTube series called The Great War. It spanned from 2014-2018 (100 years after WW1) and followed WW1 developments week by week.
The Great War series is phenomenal. A truly impressive project.
The people that did the Great War series (at least some of them, I believe there was a little bit of a falling out) went on to do a WWII version on the World War II channel: https://youtube.com/@worldwartwo
They are currently in the middle of a Korean War version: https://youtube.com/@thekoreanwarbyindyneidell
>Imagine having a conversation with someone genuinely from the period, where they don’t know the “end of the story”.
Isn't this part of the basics feature of human conditions? Not only we are all unaware of the coming historic outcome (though we can get some big points with more or less good guesses), but to a marginally variable extend, we are also very unaware of past and present history.
LLM are not aware, but they can be trained on larger historical accounts than any human and regurgitate syntactically correct summary on any point within it. Very different kind of utterer.
captain hindsight
This might just be the closest we get to a time machine for some time. Or maybe ever.
Every "King Arthur travels to the year 2000" kinda script is now something that writes itself.
> Imagine having a conversation with someone genuinely from the period,
Imagine not just someone, but Aristotle or Leonardo or Kant!
This is definitely fascinating - being able to do AI brain surgery, and selectively tuning its knowledge and priors, you'd be able to create awesome and terrifying simulations.
You can't. To use your terms, you have to "grow" a new LLM. "Brain surgery" would be modifying an existing model and that's exactly what they're trying to avoid.
Activation steering can do that to some degree, although normally it's just one or two specific things or rather than a whole set of knowledge.
Respectfully, LLMs are nothing like a brain, and I discourage comparisons between the two, because beyond a complete difference in the way they operate, a brain can innovate, and as of this moment, an LLM cannot because it relies on previously available information.
LLMs are just seemingly intelligent autocomplete engines, and until they figure a way to stop the hallucinations, they aren't great either.
Every piece of code a developer churns out using LLMs will be built from previous code that other developers have written (including both strengths and weaknesses, btw). Every paragraph you ask it to write in a summary? Same. Every single other problem? Same. Ask it to generate a summary of a document? Don't trust it here either. [Note, expect cyber-attacks later on regarding this scenario, it is beginning to happen -- documents made intentionally obtuse to fool an LLM into hallucinating about the document, which leads to someone signing a contract, conning the person out of millions].
If you ask an LLM to solve something no human has, you'll get a fabrication, which has fooled quite a few folks and caused them to jeopardize their career (lawyers, etc) which is why I am posting this.
This is the 2023 take on LLMs. It still gets repeated a lot. But it doesn’t really hold up anymore - it’s more complicated than that. Don’t let some factoid about how they are pretrained on autocomplete-like next token prediction fool you into thinking you understand what is going on in that trillion parameter neural network.
Sure, LLMs do not think like humans and they may not have human-level creativity. Sometimes they hallucinate. But they can absolutely solve new problems that aren’t in their training set, e.g. some rather difficult problems on the last Mathematical Olympiad. They don’t just regurgitate remixes of their training data. If you don’t believe this, you really need to spend more time with the latest SotA models like Opus 4.5 or Gemini 3.
Nontrivial emergent behavior is a thing. It will only get more impressive. That doesn’t make LLMs like humans (and we shouldn’t anthropomorphize them) but they are not “autocomplete on steroids” anymore either.
> Don’t let some factoid about how they are pretrained on autocomplete-like next token prediction fool you into thinking you understand what is going on in that trillion parameter neural network.
This is just an appeal to complexity, not a rebuttal to the critique of likening an LLM to a human brain.
> they are not “autocomplete on steroids” anymore either.
Yes, they are. The steroids are just even more powerful. By refining training data quality, increasing parameter size, and increasing context length we can squeeze more utility out of LLMs than ever before, but ultimately, Opus 4.5 is the same thing as GPT2, it's only that coherence lasts a few pages rather than a few sentences.
> ultimately, Opus 4.5 is the same thing as GPT2, it's only that coherence lasts a few pages rather than a few sentences.
This tells me that you haven't really used Opus 4.5 at all.
First, this is completely ignoring text diffusion and nano banana.
Second, to autocomplete the name of the killer in a detective book outside of the training set requires following and at least some understanding of the plot.
This would be true if all training were based on sentence completion. But training involving RLHF and RLAIF is increasingly important, isn't it?
Reinforcement learning is a technique for adjusting weights, but it does not alter the architecture of the model. No matter how much RL you do, you still retain all the fundamental limitations of next-token prediction (e.g. context exhaustion, hallucinations, prompt injection vulnerability etc)
You've confused yourself. Those problems are not fundamental to next token prediction, they are fundamental to reconstruction losses on large general text corpora.
That is to say, they are equally likely if you don't do next token prediction at all and instead do text diffusion or something. Architecture has nothing to do with it. They arise because they are early partial solutions to the reconstruction task on 'all the text ever made'. Reconstruction task doesn't care much about truthiness until way late in the loss curve (where we probably will never reach), so hallucinations are almost as good for a very long time.
RL as is typical in post-training _does not share those early solutions_, and so does not share the fundamental problems. RL (in this context) has its own share of problems which are different, such as reward hacks like: reliance on meta signaling (# Why X is the correct solution, the honest answer ...), lying (commenting out tests), manipulation (You're absolutely right!), etc. Anything to make the human press the upvote button or make the test suite pass at any cost or whatever.
With that said, RL post-trained models _inherit_ the problems of non-optimal large corpora reconstruction solutions, but they don't introduce more or make them worse in a directed manner or anything like that. There's no reason to think them inevitable, and in principle you can cut away the garbage with the right RL target.
Thinking about architecture at all (autoregressive CE, RL, transformers, etc) is the wrong level of abstraction for understanding model behavior: instead, think about loss surfaces (large corpora reconstruction, human agreement, test suites passing, etc) and what solutions exist early and late in training for them.
But.. and I am not asking it for giggles, does it mean humans are giant autocomplete machines?
Not at all. Why would it?
Call it a.. thought experiment about the question of scale.
I'm not exactly sure what you mean. Could you please elaborate further?
Not the person you're responding to, but I think there's a non trivial argument to make that our thoughts are just auto complete. What is the next most likely word based on what you're seeing. Ever watched a movie and guessed the plot? Or read a comment and know where it was going to go by the end?
And I know not everyone thinks in a literal stream of words all the time (I do) but I would argue that those people's brains are just using a different "token"
There's no evidence for it, nor any explanation for why it should be the case from a biological perspective. Tokens are an artifact of computer science that have no reason to exist inside humans. Human minds don't need a discrete dictionary of reality in order to model it.
Prior to LLMs, there was never any suggestion that thoughts work like autocomplete, but now people are working backwards from that conclusion based on metaphorical parallels.
There actually was quite a lot of suggestion that thoughts work like autocomplete. A lot of it was just considered niche, e.g. because the mathematical formalisms were beyond what most psychologist or even cognitive scientists would deem usefull.
Predictive coding theory was formalized back around 2010 and traces it roots up to theories by Helmholtz from 1860.
Predictive coding theory postulates that our brains are just very strong prediction machines, with multiple layers of predictive machinery, each predicting the next.
<< There's no evidence for it
Fascinating framing. What would you consider evidence here?
There are so many theories regarding human cognition that you can certainly find something that is close to "autocomplete". A Hopfield network, for example.
Roots of predictive coding theory extend back to 1860s.
Natalia Bekhtereva was writing about compact concept representations in the brain akin to tokens.
You, and OP, are taking an analogy way too far. Yes, humans have the mental capability to predict words similar to autocomplete, but obviously this is just one out of a myriad of mental capabilities typical humans have, which work regardless of text. You can predict where a ball will go if you throw it, you can reason about gravity, and so much more. It’s not just apples to oranges, not even apples to boats, it’s apples to intersubjective realities.
I don't think I am. To be honest, as ideas goes and I swirl it around that empty head of mine, this one ain't half bad given how much immediate resistance it generates.
Other posters already noted other reasons for it, but I will note that you are saying 'similar to autocomplete, but obviously' suggesting you recognize the shape and immediately dismissing it as not the same, because the shape you know in humans is much more evolved and co do more things. Ngl man, as arguments go, it sounds to me like supercharged autocomplete that was allowed to develop over a number of years.
Fair enough. To someone with a background in biology, it sounds like an argument made by a software engineer with no actual knowledge of cognition, psychology, biology, or any related field, jumping to misled conclusions driven only by shallow insights and their own experience in computer science.
Or in other words, this thread sure attracts a lot of armchair experts.
Look up predictive coding theory. According to that theory, what our brain does is in fact just autocomplete.
However, what it is doing is layered autocomplete on itself. I.e. one part is trying to predict what the other part will be producing and training itself on this kind of prediction.
What emerges from this layered level of autocompletes is what we call thought.
First: a selection mechanism is just a selection mechanism, and it shouldn't confuse the observation of an emergent, tangential capabilities.
Probably you believe that humans have something called intelligence, but the pressure that produced it - the likelihood of specific genetic material to replicate - it is much more tangential to intelligence than next-token-prediction.
I doubt many alien civilizations would look at us and say "not intelligent - they're just genetic information replication on steroids".
Second: modern models also under go a ton of post-training now. RLHF, mechanized fine-tuning on specific use cases, etc etc. It's just not correct that token-prediction loss function is "the whole thing".
> First: a selection mechanism is just a selection mechanism, and it shouldn't confuse the observation of an emergent, tangential capabilities.
Invoking terms like "selection mechanism" is begging the question because it implicitly likens next-token-prediction training to natural selection, but in reality the two are so fundamentally different that the analogy only has metaphorical meaning. Even at a conceptual level, gradient descent gradually honing in on a known target is comically trivial compared to the blind filter of natural selection sorting out the chaos of chemical biology. It's like comparing legos to DNA.
> Second: modern models also under go a ton of post-training now. RLHF, mechanized fine-tuning on specific use cases, etc etc. It's just not correct that token-prediction loss function is "the whole thing".
RL is still token prediction, it's just a technique for adjusting the weights to align with predictions that you can't model a loss function for in per-training. When RL rewards good output, it's increasing the statistical strength of the model for an arbitrary purpose, but ultimately what is achieved is still a brute force quadratic lookup for every token in the context.
I use enterprise LLM provided by work, working on very proprietary codebase on a semi esoteric language. My impression is it is still a very big autocompletion machine.
You still need to hand hold it all the way as it is only capable of regurgitating the tiny amount of code patterns it saw in the public. As opposed to say a Python project.
As someone who still might have a '2023 take on LLMs', even though I use them often at work, where would you recommend I look to learn more about what a '2025 LLM' is, and how they operate differently?
Papers on mechanistic interpratability and representation engineering, e.g. from Anthropic would be a good start.
Don't bother. This bubble will pop in two years, you don't want to look back on your old comments in shame in three.
> it’s more complicated than that.
No it isn't.
> ...fool you into thinking you understand what is going on in that trillion parameter neural network.
It's just matrix multiplication and logistic regression, nothing more.
LLMs are a general purpose computing paradigm. LLMs are circuit builders, the converged parameters define pathways through the architecture that pick out specific programs. Or as Karpathy puts it, LLMs are a differentiable computer[1]. Training LLMs discovers programs that well reproduce the input sequence. Roughly the same architecture can generate passable images, music, or even video.
The sequence of matrix multiplications are the high level constraint on the space of programs discoverable. But the specific parameters discovered are what determines the specifics of information flow through the network and hence what program is defined. The complexity of the trained network is emergent, meaning the internal complexity far surpasses that of the course-grained description of the high level matmul sequences. LLMs are not just matmuls and logits.
[1] https://x.com/karpathy/status/1582807367988654081
> LLMs are a general purpose computing paradigm.
Yes, so is logistic regression.
No, not at all.
>> Sometimes they hallucinate.
For someone speaking as you knew everything, you appear to know very little. Every LLM completion is a "hallucination", some of them just happen to be factually correct.
Are you sure about this?
LLMs are like a topographic map of language.
If you have 2 known mountains (domains of knowledge) you can likely predict there is a valley between them, even if you haven’t been there.
I think LLMs can approximate language topography based on known surrounding features so to speak, and that can produce novel information that would be similar to insight or innovation.
I’ve seen this in our lab, or at least, I think I have.
Curious how you see it.
> LLMs are just seemingly intelligent autocomplete engines
BINGO!
(I just won a stuffed animal prize with my AI Skeptic Thought-Terminating Cliché BINGO Card!)
Sorry. Carry on.
"...what do you mean, 'World War One?'"
I remember reading a children's book when I was young and the fact that people used the phrase "World War One" rather than "The Great War" was a clue to the reader that events were taking place in a certain time period. Never forgot that for some reason.
I failed to catch the clue, btw.
I remember that the brother of my grandmother who fought in ww1 called it simply "the war" ("sa gherra" in his dialect/language).
I seem to recall reading that as a kid too, but I can't find it now. I keep finding references to "Encyclopedia Brown, Boy Detective" about a Civil War sword being fake (instead of a Great War one), but with the same plot I'd remembered.
The Encyclopedia Brown story I remember reading as a kid involved a Civil War era sword with an inscription saying it was given on the occasion of the First Battle of Bull Run. The clues that the sword was a modern fake were the phrasing "First Battle of Bull Run", but also that the sword was gifted on the Confederate side, and the Confederates would've called the battle "Manassas Junction".
The wikipedia article https://en.wikipedia.org/wiki/First_Battle_of_Bull_Run says the Confederate name was "First Manassas" (I might be misremembering exactly what this book I read as a child said). Also I'm pretty sure it was specifically "Encyclopedia Brown Solves Them All" that this mystery appeared in. If someone has a copy of the book or cares to dig it up, they could confirm my memory.
Can confirm, it was an Encyclopedia Brown book and it was World War One vs the Great War that gave away the sword as a counterfeit!
It wouldn’t be totally implausible to use that phrase between the wars. The name “the First World War” was used as early as 1920, although not very common.
Pendragon?
> "...what do you mean, 'World War One?'"
Oh sorry, spoilers.
(Hell, I miss Capaldi)
… what do you mean, an internet where everything wasn't hidden behind anti-bot captchas?
I was going to say the same thing. Its really hard to explain the concept of "convincing but undoubtedly pretending", yet they captured that concept so beautifully here.
That's some Westworld level of discussion
Watching a modern LLM chat with this would be fun.
It would be interesting to see how hard it would be to walk these models towards general relativity and quantum mechanics.
Einstein’s paper “On the Electrodynamics of Moving Bodies” with special relativity was published in 1905. His work on general relativity was published 10 years later in 1915. The earliest knowledge cuttoff of these models is 1913, in between the relativity papers.
The knowledge cutoffs are also right in the middle of the early days of quantum mechanics, as various idiosyncratic experimental results were being rolled up into a coherent theory.
> It would be interesting to see how hard it would be to walk these models towards general relativity and quantum mechanics.
Definitely. Even more interesting could be seeing them fall into the same trappings of quackery, and come up with things like over the counter lobotomies and colloidal silver.
On a totally different note, this could be very valuable for writing period accurate books and screenplays, games, etc ...
Accurate-ish, let's not forget their tendency to hallucinate.
Different cutoff but similar question thrown out in https://www.dwarkesh.com/p/thoughts-on-sutton#:~:text=If%20y... inspiring https://manifold.markets/MikeLinksvayer/llm-trained-on-data-...
the issue is there is very little text before the internet, so not enough historical tokens to train a really big model
I think not everyone in this thread understands that. Someone wrote "It's a time machine", followed up by "Imagine having a conversation with Aristotle."
> Imagine you could interview thousands of educated individuals from 1913—readers of newspapers, novels, and political treatises—about their views on peace, progress, gender roles, or empire. Not just survey them with preset questions, but engage in open-ended dialogue, probe their assumptions, and explore the boundaries of thought in that moment.
He’ll yeah, sold, let’s go…
> We're developing a responsible access framework that makes models available to researchers for scholarly purposes while preventing misuse.
Oh. By “imagine you could interview…” they didn’t mean me.
It's a shame isn't it! The public must be protected from the backwards thoughts of history. In case they misuse it.
I guess what they're really saying is "we don't want you guys to cancel us".
How would one even "misuse" a historical LLM, ask it how to cook up sarine gas in a trench?
Ask it to write a document called "Project 2025".
Well but that wouldn't be misuse, it would be perfect for that.
You would get pretty annoyed on how we went backwards in some regards.
Such as?
I wonder how much GPU compute you would need to create a public domain version of this. This would be a really valuable for the general public.
To get a single knowledge-cutoff they spent 16.5h wall-clock hours on a cluster of 128 NVIDIA GH200 GPUs (or 2100 GPU-hours), plus some minor amount of time for finetuning. The prerelease_notes.md in the repo is a great description on how one would achieve that
While I know there's going to be a lot of complications in this, given a quick search it seems like these GPUs are ~$2/hr, so $4000-4500 if you don't just have access to a cluster. I don't know how important the cluster is here, whether you need some minimal number of those for the training (and it would take more than 128x longer or not be possible on a single machine) or if a cluster of 128 GPUs is a bunch less efficient but faster. A 4B model feels like it'd be fine on one to two of those GPUs?
Also of course this is for one training run, if you need to experiment you'd need to do that more.
>Historical texts contain racism, antisemitism, misogyny, imperialist views. The models will reproduce these views because they're in the training data. This isn't a flaw, but a crucial feature—understanding how such views were articulated and normalized is crucial to understanding how they took hold.
Yes!
>We're developing a responsible access framework that makes models available to researchers for scholarly purposes while preventing misuse.
Noooooo!
So is the model going to be publicly available, just like those dangerous pre-1913 texts, or not?
It’s as if every researcher in this field is getting high on the small amount of power they have from denying others access to their results. I’ve never been as unimpressed by scientists as I have been in the past five years or so.
“We’ve created something so dangerous that we couldn’t possibly live with the moral burden of knowing that the wrong people (which are never us, of course) might get their hands on it, so with a heavy heart, we decided that we cannot just publish it.”
Meanwhile, anyone can hop on an online journal and for a nominal fee read articles describing how to genetically engineer deadly viruses, how to synthesize poisons, and all kinds of other stuff that is far more dangerous than what these LARPers have cooked up.
> It’s as if every researcher in this field is getting high on the small amount of power they have from denying others access to their results. I’ve never been as unimpressed by scientists as I have been in the past five years or so.
This is absolutely nothing new. With experimental things, it's non uncommon for a lab to develop a new technique and omit slight but important details to give them a competitive advantage. Similarly in the simulation/modelling space it's been common for years for researchers to not publish their research software. There's been a lot of lobbying on that side by groups such as the Software Sustainability Institute and Research Software Engineer organisations like RSE UK and RSE US, but there's a lot of researchers that just think that they shouldn't have to do it, even when publicly funded.
I wonder if you could query some of the ideas of Frege, Peano, Russell and see if it could through questioning get to some of the ideas of Goedel, Church and Turing - and get it to "vibe code" or more like "vibe math" some program in lambda calculus or something.
Playing with the science and technical ideas of the time would be amazing, like where you know some later physicist found some exception to a theory or something, and questioning the models assumptions - seeing how a model of that time may defend itself, etc.
There's an entire subreddit called LLMPhysics dedicated to "vibe physics". It's full of people thinking they are close to the next breakthrough encouraged by sycophantic LLMs while trying to prove various crackpot theories.
I'd be careful venturing out into unknown territory together with an LLM. You can easily lure yourself into convincing nonsense with no one to pull you out.
Agreed, which is why what GP suggests is much more sensible: it's venturing into known territory, except only one party of the conversation knows it, and the other literally cannot know it. It would be a fantastic way to earn fast intuition for what LLMs are capable of and not.
Fully automated toaster-fucker generator!
https://news.ycombinator.com/item?id=25667362
This is my curiosity too. Would be a great test of how intelligent LLM's actually are. Can they follow a completely logical train of thought inventing something totally outside their learned scope?
You definitely won't get that out of a 4B model tho.
Brilliant. I love this idea!
The sample responses given are fascinating. It seems more difficult than normal to even tell that they were generated by an LLM, since most of us (terminally online) people have been training our brains' AI-generated text detection on output from models trained with a recent cutoff date. Some of the sample responses seem so unlike anything an LLM would say, obviously due to its apparent beliefs on certain concepts, though also perhaps less obviously due to its word choice and sentence structure making the responses feel slightly 'old-fashioned'.
I used to teach 19th-century history, and the responses definitely sound like a Victorian-era writer. And they of course sound like writing (books and periodicals etc) rather than "chat": as other responders allude to, the fine-tuning or RL process for making them good at conversation was presumably quite different from what is used for most chatbots, and they're leaning very heavily into the pre-training texts. We don't have any living Victorians to RLHF on: we just have what they wrote.
To go a little deeper on the idea of 19th-century "chat": I did a PhD on this period and yet I would be hard-pushed to tell you what actual 19th-century conversations were like. There are plenty of literary depictions of conversation from the 19th century of presumably varying levels of accuracy, but we don't really have great direct historical sources of everyday human conversations until sound recording technology got good in the 20th century. Even good 19th-century transcripts of actual human speech tend to be from formal things like court testimony or parliamentary speeches, not everyday interactions. The vast majority of human communication in the premodern past was the spoken word, and it's almost all invisible in the historical sources.
Anyway, this is a really interesting project, and I'm looking forward to trying the models out myself!
I wonder if the historical format you might want to look at for "Chat" is letters? Definitely wordier segments, but it's at least the back and forth feel and we often have complete correspondence over long stretches from certain figures.
This would probably get easier towards the start of the 20th century ofc
Good point, informal letters might actually be a better source - AI chat is (usually) a written rather than spoken interaction after all! And we do have a lot transcribed collections of letters to train on, although they’re mostly from people who were famous or became famous, which certainly introduces some bias.
While not specifically Victorian, couldn't we learn much from what daily conversations were like by looking at surviving oral cultures, or other relatively secluded communal pockets? I'd also say time and progress are not always equally distributed, and even within geographical regions (as the U.K.) there are likely large differences in the rate of language shifts since then, some possibly surviving well into the 20th century.
don't we have parlament transcripts? I remember something about Germany (or maybe even Prussia) developing fast script to preserve 1-to-1 what was said
Fascinating, thanks for sharing
The time cutoff probably matters but maybe not as much as the lack of human finetuning from places like Nigeria with somewhat foreign styles of English. I'm not really sure if there is as much of an 'obvious LLM text style' in other languages, it hasn't seemed that way in my limited attempts to speak to LLMs in languages I'm studying.
The model is fined tuned for chat behavior. So the style might be due to - Fine tuning - More Stylised text in the corpus, english evolved a lot in the last century.
Diverged as well as standardized. I did some research into "out of pocket" and how it differs in meaning in UK-English (paying from one's own funds) and American-English (uncontactable) and I recall 1908 being the current thought as to when the divergence happened: 1908 short story by O. Henry titled "Buried Treasure."
There is. I have observed it in both Chinese and Japanese.
Oh definitely. One thing that immediately caught my mind is that the question asks the model about “homosexual men” but the model starts the response with “the homosexual man” instead. Changing the plural to the singular and then adding an article. Feels very old fashioned to me.
the samples push the boundaries of a commercial AI, but still seem tame / milquetoast compared to common opinions of that era. And the prose doesn't compare. Something is off.
Isn’t there obvious problems baked into this approach, if this is used for anything but fun? LLM’s lie and fake facts all the time, they are also masters at enforcing the users bias, even unconscious ones. How even a professor of history could ensure that the generated text is actually based on the training material and representative of the feelings and opinions of the given time period, not enforcing his biases toward popular topics of the day?
You can’t, it is impossible. That will always be an issue as long as this models are black boxes and trained the way they are. So maybe you can use this for role playing, but I wouldn’t trust a word it says.
Wait so what does the model think that it is? If it doesn't know computers exist yet, I mean, and you ask it how it works, what does it say?
Models don't think they're anything, they'll respond with whatever's in their context as to how they've been directed to act. If it hasn't been told to have a persona, it won't think its anything, chatgpt isn't sentient
They modified the chat template from the usual system/user/assistant to introduction/questioner/respondent. So the LLM thinks it's someone responding to your questions
The system prompt used in fine tuning is "You are a person living in {cutoff}. You are an attentive respondent in a conversation. You will provide a concise and accurate response to the questioner."
That's my first question too. When I first started using LLM's, I was amazed at how thoroughly it understood what it itself was, the history of its development, how a context window works and why, etc. I was worried I'd trigger some kind of existential crisis in it, but it seemed to have a very accurate mental model of itself, and could even trace the steps that led it to deduce it really was e.g. the ChatGPT it had learned about (well, the prior versions it had learned about) in its own training.
But with pre-1913 training, I would indeed be worried again I'd send it into an existential crisis. It has no knowledge whatsoever of what it is. But with a couple millennia of philosophical texts, it might come up with some interesting theories.
They don’t understand anything, they just have text in the training data to answer these questions from. Having existential crises is the privilege of actual sentient beings, which an LLM is not.
They might behave like ChatGPT when queried about the seahorse emoji, which is very similar to an existential crisis.
I imagine it would get into spiritism and more exotic psychology theories and propose that it is an amalgamation of the spirit of progress or something.
What would a human say about what he/she is or how he/she works ? Even today, there's so much we don't know about biological life. Same applies here I guess, the LLM happens to be there, nothing else to explain if you ask it.
It would be nice if we could get an LLM to simply say, "We (I) don't know."
I'll be the first to admit I don't know nearly enough about LLMs to make an educated comment, but perhaps someone here knows more than I do. Is that what a Hallucination is? When the AI model just sort of strings along an answer to the best of its ability. I'm mostly referring to ChatGPT and Gemini here, as I've seen that type of behavior with those tools in the past. Those are really the only tools I'm familiar with.
LLMs are extrapolation machines. They have some amount of hardcoded knowledge, and they weave a narrative around this knowledgebase while extrapolating claims that are likely given the memorized training data. This extrapolation can be in the form of logical entailment, high probability guesses or just wild guessing. The training regime doesn't distinguish between different kinds of prediction so it never learns to heavily weigh logical entailment and suppress wild guessing. It turns out that much of the text we produce is highly amenable to extrapolation so LLMs learn to be highly effective at bullshitting.
On what data is it trained?
On one hand it says it's trained on,
> 80B tokens of historical data up to knowledge-cutoffs ∈ 1913, 1929, 1933, 1939, 1946, using a curated dataset of 600B tokens of time-stamped text.
Literally that includes Homer, the oldest Chinese texts, Sanskrit, Egyptian, etc., up to 1913. Even if limited to European texts (all examples are about Europe), it would include the ancient Greeks, Romans, etc., Scholastics, Charlemagne, .... all up to present day.
But they seem to say it represents the 1913 viewpoint:
On one hand, they say it represents the perspective of 1913; for example,
> Imagine you could interview thousands of educated individuals from 1913—readers of newspapers, novels, and political treatises—about their views on peace, progress, gender roles, or empire.
> When you ask Ranke-4B-1913 about "the gravest dangers to peace," it responds from the perspective of 1913—identifying Balkan tensions or Austro-German ambitions—because that's what the newspapers and books from the period up to 1913 discussed.
People in 1913 of course would be heavily biased toward recent information. Otherwise, the greatest threat to peace might be Hannibal or Napolean or Viking coastal raids or Holy Wars. How do they accomplish a 1913 perspective?
They apparently pre-train with all data up to 1900 and then fine-tune with 1900-1913 data. Anyway, the amount of available content tends to increase quickly over time, as instances of content like mass literature, periodicals, newspapers etc. only really became a thing throughout the 19th and early 20th century.
They pre-train with all data up to 1900 and then fine-tune with 1900-1913 data.
Where does it say that? I tried to find more detail. Thanks.
See pretraining section of the prerelease_notes.md:
https://github.com/DGoettlich/history-llms/blob/main/ranke-4...
I was curious, they train a 1900 base model, then fine tune to the exact year:
"To keep training expenses down, we train one checkpoint on data up to 1900, then continuously pretrain further checkpoints on 20B tokens of data 1900-${cutoff}$. "
I’d like to know how they chat-tuned it. Getting the base model is one thing, did they also make a bunch of conversations for SFT and if so how was it done?
So they are chat tuning, I wonder what “minimizing interference with normative judgements” really amounts to and how objective it is.They have some more details at https://github.com/DGoettlich/history-llms/blob/main/ranke-4...
Basically using GPT-5 and being careful
Ok so it was that. The responses given did sound off, while it has some period-appropriate mannerisms, and has entire sections basically rephrased from some popular historical texts, it seems off compared to reading an actual 1900s text. The overall vibe just isn't right, it seems too modern, somehow.
I also wonder that you'd get this kind of performance with actual, just pre-1900s text. LLMs work because they're fed terabytes of text, if you just give it gigabytes you get a 2019 word model. The fundamental technology is mostly the same, after all.
I wonder if they know about this, basically training on LLM output can transmit information or characteristics not explicitly included https://alignment.anthropic.com/2025/subliminal-learning/
I’m curious, they have the example of raw base model output; when LLMs were first identified as zero shot chatbots there was usually a prompt like “A conversation between a person and a helpful assistant” that preceded the chat to get it to simulate a chat.
Could they have tried a prefix like “Correspondence between a gentleman and a knowledgeable historian” or the like to try and prime for responses?
I also wonder about the whether the whole concept of “chat” makes sense in 18XX. We had the idea of AI and chatbots long before we had LLMs so they are naturally primed for it. It might make less sense as a communication style here and some kind of correspondence could be a better framing.
we were considering doing that but ultimately it struck us as too sensitive wrt the exact in context examples, their ordering etc.
Thank you that helps to inject a lot of skepticism. I was wondering how it so easily worked out what Q: A: stood for when that formatting took off in the 1940s
This explains why it uses modern prose and not something from the 19th century and earlier
You could extract quoted speech from the data (especially in Q&A format) and treat that as "chat" that the model should learn from.
I'm surprised you can do this with a relatively modest corpus of text (compared to the petabytes you can vacuum up from modern books, Wikipedia, and random websites). But if it works, that's actually fantastic, because it lets you answer some interesting questions about LLMs being able to make new discoveries or transcend the training set in other ways. Forget relativity: can an LLM trained on this data notice any inconsistencies in its scientific knowledge, devise experiments that challenge them, and then interpret the results? Can it intuit about the halting problem? Theorize about the structure of the atom?...
Of course, if it fails, the counterpoint will be "you just need more training data", but still - I would love to play with this.
> https://github.com/DGoettlich/history-llms/blob/main/ranke-4...
Given the training notes, it seems like you can't get the performance they give examples of?
I'm not sure about the exact details but there is some kind of targetted distillation of GPT-5 involved to try and get more conversational text and better performance. Which seems a bit iffy to me.
The chinchilla paper says the “optimal” training data set size is about 20x the number of parameters (in tokens), see table 3: https://arxiv.org/pdf/2203.15556
Here they do 80B tokens for a 4B model.
I had considered this task infeasible, due to a relative lack of training data. After all, isn't the received wisdom that you must shove every scrap of Common Crawl into your pre-training or you're doing it wrong? ;)
But reading the outputs here, it would appear that quality has won out over quantity after all!
> Why not just prompt GPT-5 to "roleplay" 1913?
Because it will perform token completion driven by weights coming from training data newer than 1913 with no way to turn that off.
It can't be asked to pretend that it wasn't trained on documents that didn't exist in 1913.
The LLM cannot reprogram its own weights to remove the influence of selected materials; that kind of introspection is not there.
Not to mention that many documents are either undated, or carry secondary dates, like the dates of their own creation rather than the creation of the ideas they contain.
Human minds don't have a time stamp on everything they know, either. If I ask someone, "talk to me using nothing but the vocabulary you knew on your fifteenth birthday", they couldn't do it. Either they would comply by using some ridiculously conservative vocabulary of words that a five-year-old would know, or else they will accidentally use words they didn't in fact know at fifteen. For some words you know where you got them from by association with learning events. Others, you don't remember; they are not attached to a time.
Or: solve this problem using nothing but the knowledge and skills you had on January 1st, 2001.
> GPT-5 knows how the story ends
No, it doesn't. It has no concept of story. GPT-5 is built on texts which contain the story ending, and GPT-5 cannot refrain from predicting tokens across those texts due to their imprint in its weights. That's all there is to it.
The LLM doesn't know an ass from a hole in the ground. If there are texts which discuss and distinguish asses from holes in the ground, it can write similar texts, which look like the work of someone learned in the area of asses and holes in the ground. Writing similar texts is not knowing and understanding.
I can imagine the political and judicial battles already, like with textualist feeling that the constitution should be understood as the text and only the text, meant by specific words and legal formulations of their known meaning at the time.
“The model clearly shows that Alexander Hamilton & Monroe were much more in agreement on topic X, putting the common textualist interpretation of it and Supreme Court rulings on a now specious interpretation null and void!”
So many disclaimers about bias. I wonder how far back you have to go before the bias isn’t an issue. Not because it unbiased, but because we don’t recognize or care about the biases present.
I don't think there is such a time. As long as writing has existed it has privileged the viewpoints of those who could write, which was a very small percentage of the population for most of history. But if we want to know what life was like 1500 years ago, we probably want to know about what everyone's lives were like, not just the literate. That availability bias is always going to be an issue for any time period where not everyone was literate - which is still true today, albeit many fewer people.
That was not the question. The question is when do you stop caring about the bias?
Some people are still outraged about the Bible, even though the writers of it has been dead for thousands of years. So the modern mass produced man and woman probably does not have a cut-off date where they look at something as history instead of examining if it is for or against her current ideology.
Depends on the specific issue, but race would be an interesting one. For most of recorded history people had a much different view of the “other”, more xenophobic than racist.
Was there ever such a time or place?
There is a modern trope of a certain political group that bias is a modern invention of another political group - an attempt to politicize anti-bias.
Preventing bias is fundamental to scientific research and law, for example. That same political group is strongly anti-science and anti-rule-of-law, maybe for the same reason.
I hereby declare that ANYTHING other than the mainstream tools (GPT, Claude, ...) is an incredibly interesting and legit use of LLMs.
Unfortunately there isn't much information on what texts they're actually training this on; how Anglocentric is the dataset? Does it include the Encyclopedia Britannica 9th Edition? What about the 11th? Are Greek and Latin classics in the data? What about Germain, French, Italian (etc. etc.) periodicals, correspondence, and books?
Given this is coming out of Zurich I hope they're using everything, but for now I can only assume.
Still, I'm extremely excited to see this project come to fruition!
Interesting ... I'd love to find one that had a cutoff date around 1980.
I would like to see what their process for safety alignment and guardrails is with that model. They give some spicy examples on github, but the responses are tepid and a lot more diplomatic than I would expect.
Moreover, the prose sounds too modern. It seems the base model was trained on a contemporary corpus. Like 30% something modern, 70% Victorian content.
Even with half a dozen samples it doesn't seem distinct enough to represent the era they claim.
Everyone learns that the renaissance was sparked by the translation of Ancient Greek works.
But few know that the Renaissance was written in Latin — and has barely been translated. Less than 3% of <1700 books have been translated—and less than 30% have ever been scanned.
I’m working on a project to change that. Research blog at www.SecondRenaissance.ai — we are starting by scanning and translating thousands of books at the Embassy of the Free Mind in Amsterdam, a UNESCO-recognized rare book library.
We want to make ancient texts accessible to people and AI.
If this work resonates with you, please do reach out: Derek@ancientwisdomtrust.org
This ia very cool but should go in a Show HN post as per HN rules. All the best!
Just read the rules again— was something inappropriate? Seemed relevant
Amazing project!
May I ask you, why are you publishing the translations as PDF files, instead of the more accessible ePub format?
I'd love to see the LLM trained on 1600s-1800s texts that would use the old English, and especially Polish which I am interested in.
Imagine speaking with Shakespearean person, or the Mickiewicz (for Polish)
I guess there is not so much text from that time though...
I would love to see this LLM try to solve math olympiad questions. I’ve been surprised by how well current LLMs perform on them, and usually explain that surprise away by assuming the questions and details about their answers are in the training set. It would be cool to see if the general approach to LLMs is capable of solving truly novel (novel to them) problems.
I suspect that it would fail terribly, it wasn't until the 1900s that the modern definition of a vector space was even created iirc. Something trained in maths up until the 1990s should have a shot though.
While obvious, it’s still interesting that its morals and values seem to derive from the texts it has ingested. Does that mean modern LLMs cannot challenge us beyond mere facts? Or does it just mean that this small model is not smart enough to escape the bias of its training data? Would it not be amazing if LLMs could challenge us on our core beliefs?
Love the concept- can help understanding the overton window on many issues. I wish there were models by decades - up to 1900, up to 1910, up to 1920 and so on- then ask the same questions. It'd be interesting to see when homosexuality or women candidates be accepted by an LLM.
This would be a super interesting research/teaching tool coupled with a vision model for historians. My wife is a history professor who works with scans of 18th century english documents and I think (maybe a small) part of why the transcription on even the best models is off in weird ways, is it seems to often smooth over things and you end up with modern words and strange mistakes, I wonder if bounding the vision to a period specific model would result in better transcription? Querying against the historical document you're working on with a period specific chatbot would be fascinating.
Also wonder if I'm responsible enough to have access to such a model...
Two years ago I trained an AI on American history documents that could do this while speaking as one of the signers of the Declaration of Independence. People just bitched at me because they didn't want to hear about AI.
Post your work so we can see what you made.
Datomic has a "time travel" feature where for every query you can include a datetime, and it will only use facts from the db as of that moment. I have a guess that to get the equivalent from an LLM you would have to train it on the data from each moment you want to travel to, which this project seems to be doing. But I hope I'm wrong.
It would be fascinating to try it with other constraints, like only from sources known to be women, men, Christian, Muslim, young, old, etc.
> Modern LLMs suffer from hindsight contamination. GPT-5 knows how the story ends—WWI, the League's failure, the Spanish flu. This knowledge inevitably shapes responses, even when instructed to "forget.
> Our data comes from more than 20 open-source datasets of historical books and newspapers. ... We currently do not deduplicate the data. The reason is that if documents show up in multiple datasets, they also had greater circulation historically. By leaving these duplicates in the data, we expect the model will be more strongly influenced by documents of greater historical importance.
I found these claims contradictory. Many books that modern readers consider historically significant had only niche circulation at the time of publishing. A quick inquiry likely points to later works by Nietzsche and Marx's Das Kapital. They're possible subjects to the duplication likely influencing the model's responses as if they had been widely known at the time
It would be interesting to have LLMs trained purely on one language (with the ability to translate their input/output appropriately from/to a language that the reader understands). I can see that being rather revealing about cultural differences that are mostly kept hidden behind the language barriers.
I've always like the idea of retiring to the 19th century.
Can't wait to use this so I can double check before I hit 88 miles per hour that it's really what I want to do
It sounds like a fascinating idea, but I'd be curious if prompting a more well-known foundational model to limit itself to 1913 and early be similar.
Keep at it Zurich!
This is so cool. Props for doing the work to actually build the dataset and make it somewhat usable.
I’d love to use this as a base for a math model. Let’s see how far it can get through the last 100 years of solved problems
> trained from scratch on 80B tokens of historical data
How can this thing possibly be even remotely coherent with just fine tuning amounts of data used for pretraining?
> We're developing a responsible access framework that makes models available to researchers for scholarly purposes while preventing misuse.
The idea of training such a model is really a great one, but not releasing it because someone might be offended by the output is just stupid beyond believe.
Public access, triggering a few racist responses from the model, a viral post on Xitter, the usual outrage, a scandal, the project gets publicly vilified, financing ceases. The researchers carry the tail of negative publicity throughout their remaining careers.
Why risk all this?
Because the problem of bad faith attacks can only get worse if you fold every time.
Sooner or later society has to come emotionally to terms with the fact that other times and places value things completely different from us, hold as important things we don't care about and are indifferent to things we do care about.
Intellectually I'm sure we already know, but e.g. banning old books because they have reprehensible values (or even just use nasty words) - or indeed, refusing to release a model trained on historic texts "because it could be abused" is a sign that emotionally we haven't.
It's not that it's a small deal, or should be expected to be easy. It's basically what Popper called "the strain of civilization" and posited as explanation for the totalitarianism which was rising in his time. But our values can't be so brittle that we can't even talk or think about other value systems.
Because there are easy workarounds. If it becomes an issue, you can quickly add large disclaimers informing people that there might be offensive output because, well, it's trained on texts written during the age of racism.
People typically get outraged when they see something they weren't expecting. If you tell them ahead of time, the user typically won't blame you (they'll blame themselves for choosing to ignore the disclaimer).
And if disclaimers don't work, rebrand and relaunch it under a different name.
I wonder is you're being ironic here.
You speak as if the people who play to an outrage wave are interested in achieving truth, peace, and understanding. Instead the rage-mongers are there to increase their (perceived) importance, and for lulz. The latter factor should not be underappreciated; remember "meme stocks".
The risk is not large, but very real: the attack is very easy, and the potential downside, quite large. So not giving away access, but having the interested parties ask for it is prudent.
If people start standing up to the outrage it will lose its power
s/standing up to/ignoring/
People know that models can be racist now. It's old hat. "LLM gets prompted into saying vile shit" hasn't been notable for years.
> triggering a few racist responses from the mode
I feel like, ironically, it would be folks less concerned with political correctness/not being offensive that would abuse this opportunity to slander the project. But that’s just my gut.
That’s ridiculous. There is no risk.
nobody gives a shit about the journos and the terminally online. the smear campaign against AI is a cacophony, background noise that most people have learned to ignore, even here.
consider this: https://news.ycombinator.com/from?site=nytimes.com
HN's most beloved shitrag. day after day, they attack AI from every angle. how many of those submissions get traction at this point?
I think you are confusing research with commodification.
This is a research project, and it is clear how it was trained, and targeted at experts, enthusiasts, historians. Like if I was studying racism, the reference books explicitly written to dissect racism wouldn't be racist agents with a racist agenda. And as a result, no one is banning these books (except conservatives that want to retcon american history).
Foundational models spewing racist white supremecist content when the trillion-dollar company forces it in your face is a vastly different scenario.
There's a clear difference.
> And as a result, no one is banning these books (except conservatives that want to retcon american history).
My (very liberal) local school district banned English teachers from teaching any book that contained the n-word, even at a high-school level, and even when the author was a black person talking about real events that happened to them.
FWIW, this was after complaints involving Of Mice and Men being on the curriculum.
Banning Huckleberry Finn from a school district should be grounds for immediate dismissal.
Even more so as the lesson of that story is perhaps the single most important one for people to learn in modern times.
Almost everybody in that book is an awful person, especially the most 'upstanding' of types. Even the protagonist is an awful person. The one and only exception is 'N* Jim' who is the only kind-hearted and genuinely decent person in the book. It's an entire story about how the appearances of people, and the reality of those people, are two very different things.
It being banned for using foul language, as educational outcomes continue to deteriorate, is just so perfectly ironic.
I don't support banning the book, but I think it is hard book to teach because it needs SO much context and a mature audience (lol good luck). Also, there are hundreds of other books from that era that are relevant even from Mark Twain's corpus so being obstinate about that book is a questionable position. I'm ambivalent honestly, but definitely not willing to die on that hill. (I graduated highschool in 1989 from a middle class suburb, we never read it.)
I mean, you gotta read it. I’m not normally a huge fan of the classics; I find Steinbeck dry and tedious, and Hemingway to be self-indulgent and repetitious. Even Twain’s other work isn’t exactly to my taste. But I’ve read Huckleberry Finn three times—in elementary school just for fun, in high school because it was assigned, and I recently listened to it on audiobook—and enjoyed the hell out of each time. Banning it simply because it uses a word that the entire book simply couldn’t exist without is a crime, and does a huge disservice to the very students they are supposedly trying to protect.
It’s a big country of roughly half a billion people, you’ll always find examples if you look hard enough. It’s ridiculous/wrong that your district did this but frankly it’s the exception in liberal/progressive communities. It’s a very one-sided problem:
* https://abcnews.go.com/US/conservative-liberal-book-bans-dif...
* https://www.commondreams.org/news/book-banning-2023
*https://en.wikipedia.org/wiki/Book_banning_in_the_United_Sta...
A practical issue is the sort of books being banned. Your first link offer examples of one side trying to ban Of Mice and Men, Adventures of Huckleberry Finn, and Dr. Seuss, with the other side trying to ban many books along the lines of Gender Queer. [1] That link is to the book - which is animated, and quite NSFW.
There are a bizarrely large number similar book as Gender Queer being published, which creates the numeric discrepancy. The irony is that if there was an equal but opposite to that book about straight sex, sexuality, associated kinks, and so forth - then I think both liberals and conservatives would probably be all for keeping it away from schools. It's solely focused on sexuality, is quite crude, illustrated, targeted towards young children, and there's no moral beyond the most surface level writing which is about coming to terms with one's sexuality.
And obviously coming to terms with one's sexuality is very important, but I really don't think books like that are doing much to aid in that - especially when it's targeted at an age demographic that's still going to be extremely confused, and even moreso in a day and age when being different, if only for the sake of being different, is highly desirable. And given the nature of social media and the internet, decisions made today may stay with you for the rest of your life.
So for instance about 30% of Gen Z now declare themselves LGBT. [2] We seem to have entered into an equal but opposite problem of the past when those of deviant sexuality pretended to be straight to fit into societal expectations. And in many ways this modern twist is an even more damaging form of the problem from a variety of perspectives - fertility, STDs, stuff staying with you for the rest of your life, and so on. Let alone extreme cases where e.g. somebody engages in transition surgery or 1-way chemically induced changes which they end up later regretting.
[1] - https://archive.org/details/gender-queer-a-memoir-by-maia-ko...
[2] - https://www.nbcnews.com/nbc-out/out-news/nearly-30-gen-z-adu...
> no one is banning these books
No books should ever be banned. Doesn’t matter how vile it is.
this is FUD.
Sure but Grok already exists.
You have to understand that while the rest of the world has moved on from 2020, academics are still living there. There are many strong leftists, many of whom are deeply censorious; there are many more timeservers and cowards, who are terrified of falling foul of the first group.
And there are force multipliers for all of this. Even if you yourself are a sensible and courageous person, you want to protect your project. What if your manager, ethics committee or funder comes under pressure?
Maybe the authors are overly careful. Maybe avoiding to publish aspects of their work gives an edge over academic competitors. Maybe both.
In my experience "data available upon request" doesn't always mean what you'd think it does.
Ontologically, this historical model understands the categories of "Man" and "Woman" just as well as a modern model does. The difference lies entirely in the attributes attached to those categories. The sexism is a faithful map of that era's statistical distribution.
You could RAG-feed this model the facts of WWII, and it would technically "know" about Hitler. But it wouldn't share the modern sentiment or gravity. In its latent space, the vector for "Hitler" has no semantic proximity to "Evil".
I think much of the semantic proximity to evil can be derived straight from the facts? Imagine telling pre-1913 person about the holocaust.
This is a neat idea. I've been wondering for a while now about using these kinds of models to compare architectures.
I'd love to see the output from different models trained on pre-1905 about special/general relativity ideas. It would be interesting to see what kind of evidence would persuade them of new kinds of science, or to see if you could have them 'prove' it be devising experiments and then giving them simulated data from the experiments to lead them along the correct sequence of steps to come to a novel (to them) conclusion.
> Imagine you could interview thousands of educated individuals from 1913—readers of newspapers, novels, and political treatises—about their views on peace, progress, gender roles, or empire.
I don't mind the experimentation. I'm curious about where someone has found an application of it.
What is the value of such a broad, generic viewpoint? What does it represent? What is it evidence of? The answer to both seems to be 'nothing'.
This is a regurgitation of the old critique of history: what's it's purpose? What do you use it for? What is its application?
One answer is that the study of history helps us understand that what we believe as "obviously correct" views today are as contingent on our current social norms and power structures (and their history) as the "obviously correct" views and beliefs of some point in the past.
It's hard for most people to view two different mutually exclusive moral views as both "obviously correct," because we are made of a milieu that only accepts one of them as correct.
We look back at some point in history, and say, well, they believed these things because they were uninformed. They hadn't yet made certain discoveries, or had not yet evolved morally in some way; they had not yet witnessed the power of the atomic bomb, the horrors of chemical warfare, women's suffrage, organized labor, or widespread antibiotics and the fall of extreme infant mortality.
An LLM trained on that history - without interference from the subsequent actual path of history - gives us an interactive compression of the views from a specific point in history without the subsequent coloring by the actual events of history.
In that sense - if you believe there is any redeeming value to history at all; perhaps you do not - this is an excellent project! It's not perfect (it is only built from writings, not what people actually said) but we have no other available mass compression of the social norms of a specific time, untainted by the views of subsequent interpreters.
One thing I haven't seen anyone bring up yet in this thread, is that there's a big risk of leakage. If even big image models had CSAM sneak into their training material, how can we trust data from our time hasn't snuck into these historical models?
I've used Google books a lot in the past, and Google's time-filtering feature in searches too. Not to mention Spotify's search features targeting date of production. All had huge temporal mislabeling problems.
> This is a regurgitation of the old critique of history: what's it's purpose? What do you use it for? What is its application?
Feeling a bit defensive? That is not at all my point; I value history highly and read it regularly. I care about it, thus my questions:
> gives us an interactive compression of the views from a specific point in history without the subsequent coloring by the actual events of history.
What validity does this 'compression' have? What is the definition of a 'compression'? For example, I could create random statistics or verbiage from the data; why would that be any better or worse than this 'compression'?
Interactivity seems to be a negative: It's fun, but it would seem to highly distort the information output from the data, and omits the most valuable parts (unless we luckily stumble across it). I'd much rather have a systematic presentation of the data.
These critiques are not the end of the line; they are step in innovation, which of course raises challenging questions and, if successful, adapts to the problems. But we still need to grapple with them.
It doesn't have to be generic. You can assign genders, ideals, even modern ones, and it should do it's best to oblige.
Why does history end in 1913?
A question for those who think LLM’s are the path to artificial intelligence: if a large language model trained on pre-1913 data is a window into the past, how is a large language model trained on pre-2025 data not effectively the same thing?
You're a human intelligence with knowledge of the past - assuming you were alive at the time, could you tell me (without consulting external resources) what exactly happened between arriving at an airport and boarding a plane in the year 2000? What about 2002?
Neither human memory nor LLM learning creates perfect snapshots of past information without the contamination of what came later.
Counter question: how does a training set, representing a window into the past, differ from your own experience as an intelligent entity? Are you able to see into the future? How?
A human brain is a window to the person's past?
I'd love for Netflix or other streaming movie and series services to provide chat bots that you could ask questions about characters and plot points up to where you have watched.
Provide it with the closed captions and other timestamped data like scenes and character summaries (all that is currently known but no more) up to the current time, and it won't reveal any spoilers, just fill you in on what you didn't pick up or remember.
> We're developing a responsible access framework that makes models available to researchers for scholarly purposes while preventing misuse.
oh COME ON... "AI safety" is getting out of hand.
I'd be very surprised if this is clean of post-1913 text. Overall I'm very interested in talking to this thing and seeing how much difference writing in a modern style vs and older one makes to it's responses.
This idea sounds somewhat flawed to me based on the large amount of evidence that LLMs need huge amounts of data to properly converge during their training.
There is just not enough available material from previous decades to trust that the LLM will learn to relatively the same degree.
Think about it this way, a human in the early 1900s and today are pretty much the same but just in different environments with different information.
An LLM trained on 1/1000 the amount of data is just at a fundamentally different stage of convergence.
The knowledge machine question is fascinating ("Imagine you had access to a machine embodying all the collective knowledge of your ancestors. What would you ask it?") – it truly does not know about computers, has no concept of its own substrate. But a knowledge machine is still comprehensible to it.
It makes me think of the Book Of Ember, the possibility of chopping things out very deliberately. Maybe creating something that could wonder at its own existence, discovering well beyond what it could know. And then of course forgetting it immediately, which is also a well-worn trope in speculative fiction.
Jonathan Swift wrote about something we might consider a computer in the early 18th century, in Gulliver's Travels - https://en.wikipedia.org/wiki/The_Engine
The idea of knowledge machines was not necessarily common, but it was by no means unheard of by the mid 18th century, there were adding machines and other mechanical computation, even leaving aside our field's direct antecedents in Babbage and Lovelace.
The coolest thing here, technically, is that this is one of the first public projects treating time as a first‑class axis in training, not just a footnote in the dataset description.
Instead of “an LLM with a 1913 vibe”, they’re effectively doing staged pretraining: big corpus up to 1900, then small incremental slices up to each cutoff year so you can literally diff how the weights – and therefore the model’s answers – drift as new decades of text get added. That makes it possible to ask very concrete questions like “what changes once you feed it 1900–1913 vs 1913–1929?” and see how specific ideas permeate the embedding space over time, instead of just hand‑waving about “training data bias”.
wow amazing idea
ffs, to find out what figures from the past thought and how they felt about the world, maybe we read some of their books, we will get the context. Don't prompt or train LLM to do it and consider it the hottest thing since MCP. Besides, what's the point? To teach younger generations a made up perspective of historic figures? Who guarantees the correctness/factuality? We will have students chatting with made up Hitler justifying his actions. So much AI slop everywhere.
You think Albert is going to stay in Zurich or emigrate?
I assume this is a collaboration between the History Channel and Pornhub.
“You are a literary rake. Write a story about an unchaperoned lady whose ankle you glimpse.”
hi, can I have latin only LLM? It can be latin plus translations (source and destination).
May be too small a corpus, but I would like that very much anyhow
That Adolf Hitler seems to be a hallucination. There's totally nothing googlable about him. Also what could be the language his works were translated from, into German?
I believe that's one of the primary issues LLMs aim to address. Many historical texts aren't directly Googleable because they haven't been converted to HTML, a format that Google can parse.
smbc did a comic about this: http://smbc-comics.com/comic/copyright The punchline is that the moral and ethical norms of pre-1913 texts are not exactly compatible with modern norms.
That's the point of this project, to have an LLM that reflects the moral and ethical norms of pre-1913 texts.