Hardcore AI for History with Mark Humphries, History Professor at Wilfred Laurier University

Full Transcript

Transcript

Mark Humphries: 0:00 The vast, vast majority of people are still at the stage of encountering for the first time. So there's a huge learning curve between, wow, ChatGPT can write a poem in the style of Bob Dylan to you can actually use this to do something very constructive and I'm going to trust it with my data. The tricky thing is I think it also scales up expectations. What AI is able to do with an assignment is probably going to become the minimum. If you have someone who is unable to achieve kind of the level of GPT-3.5 on a given task, it's unlikely that that person will be successful in that job. What we need to do is teach people to use this in such a way that they exceed the baseline level of the model. It's not going to be possible for everyone, just in the same way as not everybody gets an A today, and that's kind of just how things work.

Nathan Labenz: 0:48 But I also do really worry about the fact that just unfortunately, a lot of people cannot write at a GPT-4 level. One of the things I said to the OpenAI team when I was doing the red teaming was like, for the vast majority of people, this is superintelligence. It's just not superintelligence to you because you're really smart.

Mark Humphries: 1:07 Yeah. I mean, I think that those are real problems that we have to contend with as a society and that are much larger than just higher education. It's pervasive.

Nathan Labenz: 1:15 Hello, and welcome to the Cognitive Revolution where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week, we'll explore their revolutionary ideas, and together, we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz, joined by my cohost, Erik Torenberg. Hello, and welcome back to the Cognitive Revolution. Mark Humphries is a professor of history at Canada's Wilfred Laurier University, where he has published widely on various aspects of Canadian history, including the intercivilizational fur trade of centuries past and the postwar experience of Canada's World War 1 veterans. I invited Mark to do an episode after he reached out to me to tell me that my tip to fine tune GPT-3.5 on GPT-4 reasoning had helped him get over some humps in his own archival research. Along the way, he told me about all the things that he tried that hadn't worked, and in the process, proved himself to be one of the world's leading adopters of AI technology in the field of history. A clip to his blog showed that he's also been an early explorer of how to use LLMs in classroom settings, having experimented with different policies and guidelines over the course of 2 semesters already. In this episode, you'll learn a bit about how history is done and hear about some of the idiosyncratic challenges that Mark has had to overcome on his path to an AI agent for archival research. This is practically valuable knowledge that does, at least partially, generalize to other domains. But beyond that, I think this conversation has a few important things to teach us. First, the speed of AI diffusion really is different from anything else we've seen in the past. Because it happened so long ago, it's easy to forget that the industrial revolution unfolded over 3 generations from early steam engines to well functioning locomotives. Today, in contrast, while the physical build out of GPU loaded data centers is obviously a very real and major investment, which is currently bottlenecking the field to some extent, the distribution network for AI existed before the technology itself. And as such, we're seeing deep expert quality applications pop up everywhere just months after the technology first became reasonably useful. Second, the competitive advantage that Mark has recently gained in the production of archival research is major and likely to drive continued rapid adoption even in such an apparently unrelated to AI field as history. Simply put, Mark can now process 1000 times more documents than he previously could, and the nature of the searches that he can perform has qualitatively changed. Most historians won't have to implement their own systems, of course. They'll wait for products to be built for them. But they'll have no choice but to use such tools to get the same value from the historical record that Mark now can. Third, local leadership is often quite out of touch when it comes to AI regulation, and this can be costly. Mark has accomplished all that he has despite the fact that certain governmental archive rules prevent him from using commercial APIs. And meanwhile, at the retail level, Claude is still not available in Canada for reasons that I don't really understand, but am willing to guess based on the fact that Anthropic just launched Claude in dozens of other countries, has something to do with the Canadian government scoring an own goal in its effort to protect its citizens. While I'm definitely not one to dismiss AI risks or the need for oversight, the quality of decision making really does matter. By slowing down such a unique AI success story as Mark and preventing access to what is arguably the most safety focused AI product effort in the world today, Canada's current policy provides what I consider to be unfortunate support to reckless calls for uncritical techno optimism. Finally, and perhaps most importantly, the big picture perspective from history itself, which Mark offers at the end, reminds us that among other things, technology driven economic restructuring always has winners and losers, and the transaction costs throughout history have often been extremely high. As always, if you find value in the show, we'd love it if you'd share it. I'd suggest that you send this one to somebody in your life who you think has the potential to be like Mark. By which I mean someone who can take up the mantle of figuring out how AI can drive a 100x productivity gains in their own corners of the world, while also keeping in mind that everything has pros and cons and that the thoughtful application of technology is always critical to positive outcomes. With that, here's professor of history and AI and history pioneer, Mark Humphries. Mark Humphries, welcome to the Cognitive Revolution.

Mark Humphries: 5:56 Thank you very much for having me. Enjoy the story very much.

Nathan Labenz: 5:59 I'm excited to have you here as well. This is our first episode on the use of AI in the practice of history and the teaching of history. And also an opportunity for us to learn, I think, a little bit about what history may have to suggest for us as we try to get ready for this AI wave to grow bigger and bigger. I guess, for starters, I'd love to hear because obviously you come from a background not typically associated with the latest technology. How did it come to pass that you have gotten obsessed with AI and gone down this rabbit hole yourself over the last year or two?

Mark Humphries: 6:32 Yeah, sure. It's funny. I hear a lot of people who kind of get interested in AI do kind of come to it from strange kind of serendipitous kind of ways. I started out being very interested in programming when I was young. I was interested in the whole Deep Blue and Kasparov stuff in the nineties. And when I was young, I built a computer game, X's and O's, where you're playing against the computer, Visual Basic, built websites. So I kind of had a bit of an interest in tech in general. Got into history as a student. And my first kind of foray into machine learning was actually in translating German histories of the first world war. And this was through a program called ABBYY FineReader. I don't know if you've ever heard of this, but it's an early kind of OCR program and you could kind of bootstrap it to also teach it to read things that were kind of not conventional script. And a lot of the early German stuff is written in a different kind of typeface. So spent some time playing around with that. It worked okay, not great. And then in the last 10 years, as part of a digitization project that I headed up where we digitized about 12,000,000 pages of records. The goal there was always to try and use AI to do something with it, right? Because it's just too much material to go through. And in the context of what's happening now, it's kind of the perfect example of what to do with this stuff because it's a mixture of handwritten records. There's no standardization. Like going through it is just a nightmare. But obviously the machine learning wasn't there a few years ago. It just didn't work, I mean, in terms of trying to train on all these things. So when ChatGPT came out last fall, like a lot of people became very excited and immediately I began to see, oh, here are all those things that I've always been really interested in being able to do with the computer that now seem to be a little bit more possible. And add on to that the fact that you can actually get ChatGPT to teach you how to do a lot of the coding that you need in order to do these things. So it both brought about kind of the abilities that I needed for a long time or was interested in, as well as the ability to learn how to do these things, right, which before would have been pretty prohibitive.

Nathan Labenz: 8:36 Yeah, that's interesting. And just for quick context, I'll circle back to this a little bit later, but reconnected originally. You sent me a very nice email that just shared the story of how you had tried the trick that I suggested a few episodes back on fine tuning GPT-3.5 on GPT-4 or, I guess, any quality reasoning and getting it to kind of mimic that reasoning, which has proven effective for me and also proved effective for you. What stood out to me about that email, honestly, most was not that that had worked, although that was definitely cool, but it was the many things that you had tried to get at this problem beforehand. And I was like, man, this is somebody who has really... This is not a ChatGPT user, but a ChatGPT user plus someone who's kind of explored a lot of different aspects of AI. So I thought that was super cool. Before kind of coming back to the current techniques that are starting to work for you, one thing I'm always really interested in is kind of what the state of AI was in a given field up until pretty recently when these kind of generalist systems started to come online and also how those are now interacting. If the old ones are just kind of falling away or if they still are kind of state of the art because the tasks are narrow. So you talked about that a little bit with the one program that you're using for the German work. I'd love to hear a little bit more about that. You mentioned a program called, I guess it's Transkribus as well. But yeah, take us through kind of AI in history pre-ChatGPT, just to set the stage.

Mark Humphries: 10:16 It's interesting, right? I mean, myself and a friend and colleague of mine, John Maker, were working on those German histories back in about 2007. So this is pretty early, right? Nobody's talking really about, certainly not in the humanities, machine learning and deep learning is kind of still a little bit of a ways away at that point. And that was when historians started to get kind of interested in the technology in the sense that what you could do with it at that point is you could begin to do OCR, right? And just for context for your listeners, in history, it's all textual documents. So the biggest problem we have is to do anything digital, you have to convert the material from text into obviously machine readable format. Handwriting has obviously been the big problem for that. But early on, this kind of process of machine learning, you could apply it to more and more text because again, typewriter fonts change over time, right? And the ways newspapers were printed, if you ever look at kind of those old revolutionary war period newspapers, the typeface looks weird, right? You've got Fs that look like Ss and that sort of thing, right? And so that was kind of the first era. And I think when John and I were doing some of that back then, that was unbeknownst to us at the time kind of cutting edge in the sense we weren't thinking about it that way. But it was like, you could do this neat thing with this program. And then in the 2010s, what really happened with digital history was that people got increasingly interested in visualization. So taking big data and creating word clouds and things like this. That never really appealed to me all that much on a personal level. I mean, you went through an awful lot of work to kind of produce something that might translate into one sentence in an article, so not really worth your time. So that's where a lot of the field went though into that kind of what can you do with big data? And a lot of it was this visualization kind of stuff, right? Transkribus is an interesting program. It goes back a few years. It's changed a lot in the last year, which is fascinating. It's the first kind of good program I think that is able to accurately transcribe handwriting, right? And especially early handwriting. I mean, your listeners are going to be familiar with ChatGPT's or GPT-4's ability to read handwritten documents, right? But historical scripts are much different. They're very flowy. They have a very different type of handwriting. They're often hard to read, right? So you can imagine how much handwriting varies from one person to another. So trying to train these models to read handwriting is very difficult. Transkribus did a pretty good job. So back in the winter, it was able to basically run a document, you could run a document through Transkribus, you'd get the typical kind of error rate you'd find with OCR software, maybe for text that you'd get 10 years ago. So you might have a 96% to 97% accuracy rate, which is great, but it does mean that 3 of every 100 characters are going to be wrong, right? And what I found back in the spring, this is how quickly things are evolving as we know, you could take that document, take the text, put it into ChatGPT, you can do it through an API or you could do it through a chat interface. And you could have ChatGPT correct it, use the predictive text function to basically go through and figure out, oh no, this word was actually supposed to be this. And it was really accurate. It was actually shockingly accurate how well it did with that, right? Including with proper names and things like that, which are hard to predict. Since then, Transkribus has integrated AI into its model. And now since August, you've been able to actually use a model that effectively, I think, they haven't really described exactly how it works, but I'm pretty sure this is what's happening. It's doing that in real time, right? So you're transcribing the document and then you have a large language model, which is checking the predictive text as you go through, right? This opens up a whole new kind of field for historians. It'll enable us to do things now that we could only kind of dream of maybe even 3 years ago, 5 years ago. And so I think whereas digital history has always been kind of a sidelight of history, I think that going forward, it's most likely that this is going to take center stage, right? Just because of the possibilities that it offers. And again, just for kind of added context on this, all historians for years have been taking digital images, right? So you go to an archive, you bring a camera, you snap pictures, you might take 10,000 pictures in an archival trip. And the big problem has been all that stuff, handwritten, textual, whatever, has been basically locked in JPEGs. And what this does is unlocks a huge archive of material, right? And I suspect we're going to see that being used not only by historians, but probably also used eventually to train models as well, simply because it's this kind of other repository of text that isn't out there and all of it too is copyright free, right? Or most of it is anyway. So there's this whole kind of synergy, I think, between what's going to happen in history as well as the fact that there's simply all of this massive material that's out there that hasn't been digitized before and what that represents as we need to feed more and more material to these large language models. Mark Humphries: 10:16 It's interesting, right? I mean, myself and a friend and colleague of mine, John Maker, were working on those German histories back in about 2007. So this is pretty early, right? Nobody's talking really about, certainly not in the humanities, machine learning and deep learning is still a little bit of ways away at that point. And that was when historians started to get kind of interested in the technology in the sense that what you could do with it at that point is you could begin to do OCR, right? And just for context for your listeners, in history, it's all textual documents. So the biggest problem we have is to do anything digital, you have to convert the material from text into obviously machine readable format. Handwriting has obviously been the big problem for that. But early on, this kind of process of machine learning, you could apply it to more and more text because again, typewriter fonts change over time, right? And the ways newspapers were printed, if you ever look at kind of those old revolutionary war period newspapers, the typeface looks weird, right? You've got Fs that look like Ss and that sort of thing, right? And so that was kind of the first era. And I think when John and I were doing some of that back then, that was unbeknownst to us at the time kind of cutting edge in the sense we weren't thinking about it that way. But it was like, you could do this neat thing with this program. And then in the 2010s, what really happened with digital history was that people got increasingly interested in visualization. So taking big data and creating word clouds and things like this, that never really appealed to me all that much on a personal level. I mean, you went through an awful lot of work to kind of produce something that might translate into one sentence in an article, so not really worth your time. So that's where a lot of the field went though into that kind of what can you do with big data? And a lot of it was this visualization kind of stuff, right? Transkribus is an interesting program. It goes back a few years. It's changed a lot in the last year, which is fascinating. It's the first kind of good program I think that is able to accurately transcribe handwriting, right? And especially early handwriting. I mean, your listeners are going to be familiar with, you know, ChatGPT's or GPT-4's ability to read handwritten documents, right? But historical scripts are much different. They're very flowy. They have a very different type of handwriting. They're often hard to read, right? So you can imagine how much handwriting varies from one person to another. So trying to train these models to read handwriting is very difficult. Transkribus did a pretty good job. So back in the winter, it was able to basically run a document, you could run a document through Transkribus, you'd get the typical kind of error rate you'd find with OCR software, maybe for text that you get 10 years ago. So you might have a 96% to 97% accuracy rate, which is great, but it does mean that 3 of every 100 characters are going to be wrong, right? And what I found back in the spring, this is how quickly things are evolving as we know, you could take that document, take the text, put it into ChatGPT, you can do it through an API or you could do it through a chat interface. And you could have ChatGPT correct it, use the predictive text function to basically go through and figure out, oh no, this word was actually supposed to be this. And it was really accurate. It was actually shockingly accurate how well it did with that, right? Including with proper names and things like that, which are hard to predict. Since then, Transkribus has integrated AI into its model. And now since August, you've been able to actually use a model that effectively, I think, they haven't really described exactly how it works, but I'm pretty sure this is what's happening. It's doing that in real time, right? So you're transcribing the document and then you have a large language model, which is checking the predictive text as you go through, right? This opens up a whole new kind of field for historians. It'll enable us to do things now that we could only kind of dream of maybe even 3 years ago, 5 years ago. And so I think whereas digital history has always been kind of a sidelight of history, I think that going forward, it's most likely that this is going to take center stage, right? Just because of the possibilities that it offers. And again, just for kind of added context on this, all historians for years have been taking digital images, right? So you go to an archive, you bring a camera, you snap pictures, you might take 10,000 pictures in an archival trip. And the big problem has been all that stuff, handwritten, textual, whatever, has been basically locked in JPEGs. And what this does is unlocks a huge archive of material, right? And I suspect we're going to see that being used not only by historians, but probably also used eventually to train models as well, simply because it's this kind of other repository of text that isn't out there and all of it too is copyright free, right? Or most of it is anyway. So there's this whole kind of synergy, I think, between what's going to happen in history as well as the fact that there's simply all of this massive material that's out there that hasn't been digitized before and what that represents as we need to feed more and more material to these large language models.

Nathan Labenz: 15:03 In terms of just kind of raw scale and just state of play right now, first of all, it's interesting just to realize, in my lifetime, we didn't have a personal computer at home. So you don't have to go that far back to where, you know, kind of the standard, certainly, domestic document production and not that much farther back into, you know, even professional environments. You just didn't have, you know, kind of all the individual people creating stuff just weren't able to create it in a digital native format. It's amazing how you know, that's almost unthinkable, but it's like not that long ago. So it's certainly not just deep history that this would apply to. I'm at least bucketing the sort of historical records into maybe I see at least 2 kinds of things. One being the typeset, which presumably is like easier, although maybe not easy, and then the handwritten stuff, which sounds like quite hard and, you know, is just coming online with some of these, you know, Transkribus and and now more generalist models as you're mentioning. My assumption would have been that like Google Books or whatever would have kind of effectively digitized all the books. But I would imagine most of the handwritten stuff is just still kind of sitting in bins or maybe has been snapped into JPEGs, but that's about it. Is that accurate? Or what would the general state of the content be?

Mark Humphries: 16:35 Yeah, you're right. In terms of Google Books and the Hathi Trust in the US as well have digitized enormous amounts of printed material, right? And so this is bound printed material. But what I'm talking about, what most historians tend to use the most is archival material, right? So these are paper documents that are sitting in boxes. And it's kind of hard to visualize how much of this material exists. But to give you an example, in Canada where I'm operating out of, for the First World War alone, you're talking about if you set all of the documents that were generated now stored in the archives, just in Ottawa, about the First World War end to end, you have about a kilometer of records, which is about 3 quarters of a mile, right? And so this is an enormous amount of material. And so none of that's been digitized or some of it has been digitized, but we're talking tiny fractions of it, even all the war diaries, which have been digitized, small little amount of that total. And so if you can then extrapolate from there and think about the many, many, many, many miles, hundreds of miles of records that exist in various archival repositories, that's what historians use, right? So the big problem is that going through that material, even if you digitize it, it's usually selective, it's usually done kind of in an ad hoc way and it's certainly not everything. And what AI opens up as a possibility is going through everything eventually, right? In the sense that, you know, in an ideal world, you could in theory now begin to pump all of this material in, into a model. And for historians, and this is going to be more broadly applicable in other areas too, but for historians, the big problem is always processing, right? You can't go and read a mile's worth of records in a lifetime, so it's not going to happen. So you're always selective. You've got to choose where you look, you've got to choose what you look at, you've got to choose how you spend your time, right? And that's why history is constantly rewritten. What we're talking about here then is at least the theoretical possibility of being able to feed an infinite amount or close to an infinite amount of paper into a model, which can then actually do that processing very quickly. And if you think about what that opens up in terms of things that wouldn't have been feasible a few years ago to do it manually and what's now feasible or at least will be very soon, it really changes what we do. That's relevant to history. But I think more broadly, if you think about the law and what lawyers are already doing in terms of using models on legal documents for the medical profession, it's all part of the same thing, right? We're scaling up what we can process, and the volume of material that you can go through, which allows you to be far more exacting and detailed, right? So just as you know, a radiologist can look at a scan and detect a tumor or a lesion or something like that on a scan, but you can also feed those scans into a model which can see things that the radiologist might not pick up on because there's, you know, something in the data that that stands out that you wouldn't necessarily notice that might be indicative of something. That's what we're talking about here in all of these different fields. You feed more and more content into these models and the odds are you're going to start to pick up on things you wouldn't have seen before because you simply couldn't have processed the data. You couldn't have found the connections. We're not there yet, obviously. But what you can see, as you well know, is that the capabilities are there. It's often getting to the place where you can scale them up or they become affordable and it's GPU capacity and things like that, right? So this is how it's changing what we do. And it's also, I guess, worthwhile thinking about the fact you've got that textual component that's printed, but then you've got this enormous amount of archival records out there that are not digitized or that are not accessible through Google Books and things like that, right?

Nathan Labenz: 20:13 I was just trying to translate this with a little mental math into tokens and then kind of cost. So, you know, you take a mile, 5,000, we'll use round numbers, 5,000 feet, that is 50,000 inches, probably could do more, but I'll say 100 pieces of paper per inch. So 5,000,000 sheets of paper in a mile. I don't know, you know, whatever. Call it 100 tokens, maybe 200 if it's 200 tokens.

Mark Humphries: 20:52 It's probably 350 tokens a page would be my guess, but 300 words, 500 words depending on the writing.

Nathan Labenz: 20:59 So that gets us to basically 1 billion. At 200, you go to 1 billion. So $3.50, you know, maybe 2 tokens. You said that is just the World War I archive that is in Ottawa. So, yeah, I mean, it does start to get potentially quite material even on the scale of just data, you know, writ large, right? I mean, kind of for all of the use cases. I mean, it's not like YouTube scale, but it's definitely pretty big, and it certainly covers something that YouTube, you know, is in is in no position to cover. It seemed like it's like a co-op on some level, but it also has a commercial model. So I'm just kind of curious about that. But the pricing appeared to be more than 10 cents per page, which, you know, going back to our mental math, if we're talking about, what was it, 500,000 pages, right? It's 5,000,000 pages? I'm losing track. But whatever, some huge number of pages. Obviously, that's going to be tough on your grant budget. But maybe, you know, with a if we're talking about 1 billion tokens and it's $2 per million tokens with GPT-3.5, then, you know, you're getting into just a couple thousand dollars in theory is kind of the price that we could sort of see our way to. So do I have that right in terms of just kind of what it has cost? And do you think it's realistic that it seems like it's going to drop now to the point where, you know, maybe your grant, your research budget actually could cover sort of full processing of something on this scale?

Mark Humphries: 22:36 Well, that's it. I mean, right now, it's not feasible today, right? But, I mean, what we know is the pricing's dropping constantly, right? I mean, AI or OpenAI is probably seems like going to open some or announce some kind of a price drop on November 6, right? And this is going to happen over the next few years. We're just going to see it get cheaper and cheaper. It's what happens, right? We also got to keep this in mind, right? The limiting factor before has always been you used to have to hire someone to transcribe this material for you, right? And so when you got a research grant as a historian at a university, and this would apply to English or a whole other kind of, any other kind of field like that, you'd very often hire a student to come and do that transcription. Now, that student would take maybe a week or 2 to to transcribe a couple of dozen pages of documents, right? You're paying them very often somewhere between 20 and $30 an hour. That's expensive. So then the 10 cents a page becomes much cheaper than the alternative, which is why that pricing model looks the way that it does right now, right? But if we assume that the vision capabilities of GPT-4 and whatever comes after it get better, and that the API pricing follows in line with that, yeah, I mean, it's going to get much cheaper. So I think the reality for historians, for context, is that right now, even transcribing something through Transkribus at 10 cents a page is a tiny fraction of what it would have cost you to transcribe that document with with human input, not that long ago. And doing it through an API with OpenAI, once that technology gets a little bit better at doing this stuff, which it will, that's going to get even cheaper then, right? So again, it's one of these things where you can kind of see where this is going to go with I think a fairly reasonable degree of certainty at the moment. Yeah, it's not feasible this second, but I could say within the next few years, will be quite feasible in order to transcribe, if not, you know, that whole amount of records, certainly you could take a big block of them and do that relatively quickly and cheaply compared to what it would have cost in the past, certainly.

Nathan Labenz: 24:37 Yeah, it's amazing. So what is the state today with GPT-4 vision if you just go drop in one of these things? We don't have the API yet. I think that's another good candidate for a November 6 release, but we'll see. Pressure's mounting on OpenAI to drop the prices because everybody's talking about it. It's going to be if all of a sudden the prices stay the same, people are going to be a little disappointed. But in any event, if you go take one of these historical documents today, again, feel free to complicate this because I'm sure it varies depending on what era and all sorts of conditions. But how does the kind of new baseline GPT-4 vision compare to Transkribus head to head just pure quality basis? Nathan Labenz: 24:37 Yeah, it's amazing. So what is the state today with GPT-4 Vision if you just go drop in one of these things? We don't have the API yet. I think that's another good candidate for a November 6 release, but we'll see. Pressure's mounting on OpenAI to drop the prices because everybody's talking about it. If all of a sudden the prices stay the same, people are going to be a little disappointed. But in any event, if you go take one of these historical documents today, again, feel free to complicate this because I'm sure it varies depending on what era and all sorts of conditions. But how does the kind of new baseline GPT-4 Vision compare to a transcriber's head to head just pure quality basis?

Mark Humphries: 25:22 Not well right now, which is interesting, but it's interesting for why it doesn't. Right? And so right now, if you drop in a document and I've tried it with modern, kind of handwritten documents from, say, the 1950s versus ones from the eighteenth century or whatever, it has a lot of hallucinations that were more common with 3.5 than they were with 4 when you're doing textual based stuff, where it will just start summarizing something versus transcribing it. And it will change the language and you can play around with the prompt and say, you know, you have to be exact and all these, but it tends to go off the rails pretty quickly. And when you play around, one of the things you realize is that I think there's a couple of things going on. One, I don't think it was probably trained on that much old handwriting to begin with, which probably poses a bit of an issue for it. It probably hasn't seen that many examples of eighteenth century handwriting compared to something later on. So that poses an issue. But the other one is that it seems to kind of fall into these troughs where what it wants to do is tell you about the document rather than transcribe it. And it does a pretty good job summarizing. So you get this weird kind of output where it can't transcribe a document exactly, but does a really good job telling you what's in the document, which is kind of an interesting kind of problem to have. It clearly is able to interpret what's there on some level, but getting it to actually output an exact transcription is problematic. And that's kind of what I'm seeing. Now, at the same time, depending on your use case, maybe transcription isn't always necessary. As historians, we used to think about we have to transcribe things in order to be able to use them. And what I've been playing around with, especially with some of these pension documents that, you know, I'm just curious to see what you can do with them, is very often what you want to do is ask the model, you know, I've got this research question, is this document relevant to the thing I'm looking at? And if you've got 12 million pages, you might not need to transcribe 12 million pages. You might just need it to tell you, you know, here are the 2,500 pages out of here that are actually going to be useful. And it seems to be better at doing that sort of thing already than transcribing. And that's certainly a very big shift for how I've thought about digitization and what you do with digital records. The vision capability may well short circuit the need to do some of that transcription if you're able to just kind of work with these native JPEGs in that sense.

Nathan Labenz: 27:45 Yeah. That's fascinating. I had occasion to drop in a letter from, you know, a city official that my sister received the other day, and I did notice a similar behavior where I didn't ask it to transcribe because in this case, I could read it just fine. I only had one of them to deal with. But I did notice that similar behavior where and I didn't do any prompting, but just kind of by default, you know, literally just here's an image. First kind of native behavior was it appears to be a letter from whoever in this township of whatever, and here's what it's about. And it was all very apt. So that's pretty insightful. And I also really like your use of the word trough. Another word that we've been hearing for these kinds of things is mode of behavior. But I like trough because it does have a certain, it evokes the topology. And I do think of these kind of paths through the language model as almost like they're circuits in some sense, they almost seem like they're eroded through the ongoing process of training. So I like that visualization quite a bit just as an aside. Maybe you should have asked this even a little earlier, but tell me a little bit more about what kinds of problems you're trying to solve, what kinds of questions you're trying to answer. You've got these 12 million documents. You ideally would like to ask a question over all of them. You can't. But let's start with what are the documents, what are the questions, and how are you zeroing in semi blindly today? And then you have this really interesting idea of the historian agent. And I want to hear a little bit more about that as well. So that's a lot there. But let's start with just kind of the stuff that you're actually trying to make progress on. Nathan Labenz: 27:45 Yeah. That's fascinating. I had occasion to drop in a letter from a city official that my sister received the other day, and I did notice a similar behavior where I didn't ask it to transcribe because in this case, I could read it just fine. I only had one of them to deal with. But I did notice that similar behavior where, and I didn't do any prompting, but just kind of by default, literally just here's an image. First kind of native behavior was it appears to be a letter from whoever in this township of whatever, and here's what it's about. And it was all very apt. So that's pretty insightful. And I also really like your use of the word trough. Another word that we've been hearing for these kinds of things is mode of behavior. But I like trough because it does have a certain... It evokes the topology. And I do think of these kind of paths through the language model as almost like they're circuits in some sense, they almost seem like they're eroded through the ongoing process of training. So I like that visualization quite a bit just as an aside. Maybe you should have asked this even a little earlier, but tell me a little bit more about what kinds of problems you're trying to solve, what kinds of questions you're trying to answer. You've these 12 million documents. You ideally would like to ask a question over all of them. You can't. But let's start with what are the documents, what are the questions, and how are you zeroing in semi blindly today? And then you have this really interesting idea of the historian agent. And I want to hear a little bit more about that as well. So that's a lot there. But let's start with just kind of the stuff that you're actually trying to make progress on.

Mark Humphries: 29:40 I'll try not to get too pedantic about the history angle of it here. But there are issues around any type of document you're dealing with, with sending it through an API or using ChatGPT because a lot of these documents are protected. So in the case of the records that I'm using right now, we can't, most of them, some of them, but most of them we can't actually upload for privacy reasons because they're restricted. And so this is where using something like a Llama 2 model or something like that is going to come in. So I've been trying to actually more play around with other documents that we can actually use this way, in kind of a large group. But I'll give you an example of just the type of thing you might want to do. So ideally with these pension files, the problem is that they're all completely non-standardized. The forms change from one year to another. Sometimes they're just handwritten notes, other times they're typewritten forms and it just varies all over the place. And so a very simple question like did individual X receive a pension for disability after the First World War actually becomes fairly difficult to answer because sometimes that person might have received a pension. Other times that pension might be taken away for a period and then they might get it back again. And then they might have the pension increased and then it might disappear and then come back again. So it becomes very difficult thing to go through on an individual level and requires a lot of time to just kind of process. What you could do though is you can very easily ask these models to kind of answer those types of questions and give you a very quick kind of summary about a case history. So it's something similar to what you might do with medical records in that sense. Just boil down to this long, complicated series of records where there's a lot of avenues that we don't want to go down and red herrings, irrelevant information and boil it down to something that's relevant to this question of, did this person receive a pension? If so, tell me about the dynamics of how that shifted and changed over time. That's the type of thing which you could do right now very easily with this if you were able to, feed through an API, a series of these images. That's something I've tried with some of records we've redacted that we can actually do is if you put them into GPT-4V, it does actually a pretty good job of looking at the record and going, yeah, here's the summary. You do kind of it's more like kind of what it was like in the winter with it, I find, where you have to kind of, pet it a little bit more and tell it, you're okay, you can do this, don't worry. Like I know you think you can't read this thing, but you can actually try and you'll be pretty good at it and then it will try again. And so it's odd kind of going back to that, which is behavior you see less of with GPT-4 normally. What I've been really working on though lately is a completely different project with my research partner, Leanne Leddy. And we've been looking at trying to do fur trade records. And this came out of a whole other series of things that I became interested in during the pandemic for other irrelevant reasons, but I became very interested in the fur trade. And it's another kind of an area where you have this interesting combination of records. You've got handwritten journals from fur trade posts, you have a series of fur trade contracts. And one of the big problems that historians have had is how do you take somebody from a fur trade contract that you have and you know that they signed this in Montreal, say in 1780, and then find out what they did through all these handwritten records through thousands and thousands of pages of handwritten records. How do you find that individual? It's hard to do, you have to manually go through and do this. And so what we've been trying to do is create an agent that will effectively do this, take an individual, find identifying markers based on demographic records that are in Quebec, and then use that information to then find that individual through fur trade records and be able to basically follow them through and do a life history. And it's an interesting process because what it involves from a training perspective is trying to teach the model historical language, get around some of the safeguards that are there. Because again, there's a lot of documents that contain racist material because they're historical documents and getting GPT-4 to think through that in a way that's useful then becomes a bit of a challenge. You also end up with this kind of process of trying to find the documents you need in order to feed them to the model. And it sounds pretty straightforward. Except again, like so many of these problems, it's not when you drill down to it. So if you have an individual and I've come up with a few of these test cases that I've been using that are just edge cases that are not uncommon edge cases, but are good ones for this. And you might have an individual and you want to ask it as a historian and open ended question about, tell me about Ferdinand Wentzell's family. And Wentzell's kind of a mid-range fur trade figure. And he writes a number of post journals and he writes about his family, but he does it in very oblique ways. So he'll refer to his indigenous partner as being my girl. He'll refer to his son as being my boy or the boy. And it's often he writes in these kind of... he's German, so he's writing in English and he's writing often kind of backwards, grammatically backwards sentences. So, or he's Norwegian, I guess, but they're still kind of grammatically backwards in that sense. And what you end up with at the end is a real difficult process of not so much getting the AI model to be able to trace this guy through and write the biography, but be able to feed it the records that it needs to do that. And I've tried all the various techniques. Which you can do embeddings and try semantic search approach with this. That works reasonably well, except you very often miss some of the most important documents because the pre-trained embeddings out there like all-MiniLM and various Hugging Face sentence transform models, they won't recognize my girl as necessarily being your wife. So then that doesn't show up. And if you simply have a question like, tell me about Ferdinand Wentzell's family, that doesn't work very well in retrieving a lot of the relevant documents either. Because there's not enough information in the prompt to do. I've tried the HyDE approach, which is you try and generate hypothetical versions of these documents. And then again, you get some interesting behavior from the language model. It won't write in the way that you need to historically to match the example to examples in the database. And again, this has been kind of a funny process where if you tell it to be succinct, to be short and to be choppy, and here's five examples of how to write a diary entry like a fur trader, it almost always wants to write some kind of a flowery example in which the fur trader pauses to reflect on the beauty of the morning or something like that. And no matter how many times you tell it not to do that, it still wants to do that on some level. And nobody ever wrote like that. And so this is another kind of a problem with the documents don't really look the way that GPT-4 is expecting them to look. You might have a two sentence entry in a diary that says, fine weather this morning, my man died when a tree fell on him this afternoon and we traded five kegs of rum later on in the afternoon. And so the model is not thinking like that. It will want to know more about the death of that individual and then become very reflective about how it feels about this. And that's just not how these historical documents work. So there's that issue. You can certainly fine-tune a model that does a better job of that. But then you get into this problem of you end up fine-tuning a whole series of models, kind of a mixture of experts almost approach. Where you have to fine-tune models and different things. You've got to have a fine-tuned embeddings model. You have to have a fine-tuned model to create the examples for semantic search. And then you have to fine-tune a model to actually determine what's relevant. GPT-4 does a fairly good job at determining is this document relevant to that question? But again, it stumbles on historical language very often. It stumbles on when it finds something in the document that it perceives as being racist or as often racist or problematic for whatever reason. It will also look at that and say, well, this isn't relevant sometimes. And so then you've got to fine-tune a model to teach it that no, this is relevant to that. And that's just for the semantic search approach. The other approach of course is to use keywords and this sort of thing, but then you get into classification, you run into the exact same types of problems that I talked about a second ago. And as you start to try and classify material historically, the model has to learn these things are relevant to this. And there's a whole vocabulary that comes into it that is just not in the training data. And again, it kind of presents some interesting test cases, which I often go down these rabbit holes trying to figure out because you might have a term like varang, for example, which if you ask GPT-4 about it, it'll come up with a hallucinated response for what it is. And then if you say, are you sure that's what it is? It'll say, no, I don't know what it is. It's not a common English term. And it's not. It basically means the wooden ribbing of a birch bark canoe. Fairly specific. But you have to then train a model to know that or at least to recognize how this works in order to get it to properly classify things with keywords. And once you do that, you can begin to do some keyword retrieval, combine it with semantic search and then you get your document retrieval. But you have to kind of ladder that into then another model in which you're also feeding that up to try and get the model, some kind of overarching model to pick and choose between approaches, be able to trace that individual you identified back in Montreal through these documents, determine if they're relevant and then come out with a biography at the end. Now, if this works properly, you can then do take a database like there's a Voyager's contracts database, which Nicole St. Onge and some others out of University of Ottawa did. They have 35,000 individuals in there. You could very quickly, if you get this working, run through all 35,000 individuals and produce mini biographies of each of them and then begin to think about, oh, what are patterns we're seeing here we weren't aware of before. And the big limiting factor now is context limits in terms of the models you're dealing with as well as cost. Again, the cost of those recursive API calls begin to add up very quickly. And so again, it's one of those things where what we know is the context windows I think will continue to get larger and larger, at least I'm assuming that. And the cost of those API calls will probably come down. And I mean, I've heard the rumors too that maybe OpenAI is going to release some kind of an agent, kind of built-in agent approach or something like that. But again, this is stuff that you can see now, this is where it's going to go probably and what's going to be possible, it's just we're not quite there yet in making it feasible. But you can imagine taking this very pedantic fur trade example and then scaling that up not only to historical applications, but to a whole variety of different things. And it raises issues about doing this for living people. The ethics of being able to take social media data and trace people. So there's a whole series of different things that come out of this type of approach, which again is very new. This isn't something we could have even talked about a year ago in history. That's how new this is.

Nathan Labenz: 40:44 That's really cool. I mean, you're definitely doing all the latest and greatest stuff, which I think is super impressive. So just one very small question on the comment that you made about not being able to use some of the individual records for privacy reasons. My understanding of, like, the OpenAI API terms of service is that they, a, promise not to train any models on your data, and also, b, promise to delete it all pretty quickly. In the past, I've kind of said, hey, why don't you guys have more visibility into what I did in the API? And they're like, yeah, because we don't really want to save that stuff because people don't really want us to have it. And so I guess that's not good enough for the Canadian government. Like what do you think is sort of the future of that debate? Is there anything they could do that would make it okay to use their API?

Nathan Labenz: 40:44 That's really cool. I mean, you're definitely doing all the latest and greatest stuff, which I think is super impressive. So just one very small question on the comment that you made about not being able to use some of the individual records for privacy reasons. My understanding of the OpenAI API terms of service is that they, a, promise not to train any models on your data, and also, b, promise to delete it all pretty quickly. In the past, I've kind of said, hey, why don't you guys have more visibility into what I did in the API? And they're like, yeah, because we don't really want to save that stuff because people don't really want us to have it. And so I guess that's not good enough for the Canadian government. Like what do you think is sort of the future of that debate? Is there anything they could do that would make it okay to use their API?

Mark Humphries: 41:42 It's hard to imagine, right, that this is, again, we're not even a year out from GPT-3.5 coming out, right? So I mean, this is all brand new. So when these contracts that I had or agreements that I had with Veterans Affairs in Canada were negotiated back in 2018, this wasn't even on the horizon. Right. And so their concern is, you know, we're not allowed to transmit the data effectively over the Internet. We're thinking email and stuff like that at the time. We're not future proofing this for sending it to an API, which we didn't think would exist. Right. So I think going forward, that's the type of thing that is going to become a huge point of negotiation for historians. But I think for researchers more broadly, right, is can you send this? What type of material can you send to an API? And so I agree. I mean, I think that OpenAI doesn't retain the material and they're very transparent about that. And people are often skeptical of that. And I say, well, you know, if they're lying, they're going to get sued. And so I don't think they're lying. Right. This is obviously, I think you can trust people on these things when they have them in the terms of service. But I think it's building the confidence amongst researchers as well as amongst agencies that hold documents. This will be true for hospitals as well. Right. About what you know, what you can do with these types of records and what you can send them. There's also legislation often that governs this. So, for example, in Canada, material that contains identifiable information for an individual that's owned by the government, an individual who's been deceased for less than 20 years falls under the Privacy Act and you can't actually transmit that. So it doesn't matter if OpenAI actually retains that or not. You sign a researcher's agreement saying, I will transmit this in exchange for getting access to it, right? I think that kind of thing will have to change or at least we'll have to update those types of legislative documents to deal with this world. And they'll either look like more restrictive in the sense that you can't do this or there'll be less restrictive or something in between, but they'll have to be updated, right? And I think that that's an important kind of a component of all of this. I think the other thing with data in general and data privacy is that for people who are using this, I think, as you said, like you've used it, you're fairly confident how it works. The vast, vast majority of people are still, you know, at the stage of encountering ChatGPT for the first time. And certainly my colleagues at university are kind of at that stage. Getting over the hump of then trying to explain that, yes, there are ways that you can send this material that is more secure. And I mean, that itself is a learning process, right? Because there's a huge knowledge, like huge learning curve between, you know, wow, ChatGPT can write a poem in the style of Bob Dylan to you can actually use this to do something very constructive and I'm going to trust it with my data. And I think that gap is only growing as these models get more and more capable. And because everything is evolving so quickly, I think in an unprecedented way, we're witnessing that gap just get larger and larger. So there's a trust gap and there's a legislative gap. And I think that there's also just kind of the fact that people like me have researched agreements we negotiated before these things were even on the event horizon. Right?

Nathan Labenz: 44:52 Yeah. A lot of dimensions to the capabilities overhang. And that's honestly one of the big motivations I have for doing this show is to try to help close the capability overhang gap. I often say I'm an adoption accelerationist, not necessarily a hyperscaling accelerationist. But to use what we have and to make the most of it and to understand what we have, it seems like we're currently still not even caught up to the pace of progress just in terms of understanding and using what already is online and available. So it's an interesting and unlikely historical challenge. I do want to cover your experience in experimenting with GPTs and education and also, you know, kind of turn to the lessons from history. Before doing that, just to kind of finish the discussion on the technical stuff and all this experimentation that you've done on the agent, I think, you know, we're fairly kindred spirits, it seems, because we have both been willing to work pretty hard to get AI to do something, that, in the grand scheme of the world, is fairly narrow, often when it was still kind of sucky at it. And so I'd love to hear what did the 3.5 fine tuning unlock and what's the current state of play? And then also, are we getting to the point now where it's actually adding value? Or are you still in the kind of promise phase where you still need to turn one more corner before you'll actually be accelerating the actual historical work that this is all geared toward.

Mark Humphries: 46:35 Yeah. That's a great question. And let me kind of deal with the end part of it first. Right? I mean, it's such an iterative process, right? As you know, and I think most people listening to this probably are aware as well, right, where it's a lot of trial and error. It's a lot of frustration. There are days where I'm like, yeah, this model is not going to able to do this. And then you have that moment where suddenly you figure out, you know, if only I do this, you go, great, this is now going to work. And you have the moments where the thing you've been working on for three months suddenly gets released through an update in OpenAI and now the model can actually do it. Or, you know, something comes out and you're like, oh, this is, you know, this is really kind of fighting with it to do this before or something like that. Right? So there's that kind of aspect to it. What I've found is that fine tuning since it was released in August, I use the OpenAI API the most for this. In Canada, don't actually have access to Claude, which is kind of a, I'd like to try it with a 100,000 context window.

Nathan Labenz: 47:27 I think that just changed this week, by the way.

Mark Humphries: 47:30 Did it? Oh, I didn't even know that. Look at that.

Nathan Labenz: 47:32 Some new countries. Vivian, shout out to Vivian, producer of our show, is in Canada and had been we use Claude 2 specifically for the timestamp transcripts of the show, like the outline. Right? So dump in a transcript and say, you know, give me a format of the discussion topics. She couldn't do that from Canada until this week. She's finally able to use it. I think you now should have access.

Nathan Labenz: 47:32 Some new countries. Vivian, shout out to Vivian, producer of our show, is in Canada and we use Claude 2 specifically for the timestamp transcripts of the show, like the outline. So dump in a transcript and say, give me a format of the discussion topics. She couldn't do that from Canada until this week. She's finally able to use it. I think you now should have access.

Mark Humphries: 47:59 That's what I'll be doing this afternoon. But I think what's interesting is that since the fine tuning capability was released in August, I've seen a lot of the problems that I thought were going to be farther off in the horizon of solving become solvable. Because again, all those things I described where you have to have essentially a different kind of model to do all these things, that wasn't possible before August. And so, we talked about this over email, but I think that the fascinating problem is how you get that information into the model. And I think there's some good information out there on fine tuning, but it's also an area where I think there's the least kind of good, hard information on how to do it properly, while there's also obviously a lot of misconceptions. The idea that you can use it, some people say you can use this to impart new knowledge. It doesn't really do that, but it kind of does. Not really. And it's just about behavior. But yeah, behavior is also a form of knowledge. So there's this kind of gray area that's still in there, at least in my mind. And so what I was originally trying to do is for one fine tuning example I use, I was trying to teach it how to determine if a document was relevant or not. And use that kind of classic approach where you give it three examples of the same document. You change a word in each one where, in the first document, if I'm trying to teach it that I'm asking you about Wentzle's wife and his family, and the document describes Wentzell saying, today my girl and I went somewhere in a canoe. Well, you have to teach it that my girl means wife. And so you do that by, in the first one, saying it's relevant. And then in the second one, take my girl out, you put it with wife and you say relevant. And the third one, you say my dog and you say irrelevant. And it learns that these are different approaches that way. It's an iterative approach again in that sense. That doesn't work quite so well when you're dealing with something much more complicated like keywords, trying to read a fairly long text, break it into thematic sections, assign keywords and then basically go from there. And what I was finding is, my first run throughs with the native, like basic models, my corpus of documents I'm working with as a test here is about 650,000 words, which is pretty significant, but it's not giant in that sense. But in categorizing that, you might come out with 11,000 keywords just by using GPT-4 and saying assign keywords to this. And many of those keywords are overlapping and you can go through this kind of process, which I tried, of consolidating the keywords. It becomes very time consuming and it just doesn't eventually work. What you effectively need is you need to be able to say when a user asks this question, choose from this bank of keywords, keywords that are going to likely be relevant to that question and then retrieve all the documents that contain those keywords in the keyword section. And so what I then tried to do was I tried different approaches of giving it a bank of the keywords. Here are the 300 keywords you're going to use. Well, that doesn't really work because every once in a while, GPT-4 doesn't stick to it and will pull out some other random term. It will also miss things. So then you try and put in these rules. If this happens, always apply this keyword. If that happens, always apply that keyword. And then you get these prompts that are 2,000 tokens long. And then you put your document into it and it's hugely expensive. And so I heard on your show actually a little while ago, you talking about using chain of thought as a way of doing this. And I immediately thought, well, this would be a really interesting approach to take, because I've been basically trying to show it just before this document results in these keywords and changing this document a little bit changes these keywords. And that just wasn't working. And what's fascinating to me, and I think it's a really insightful kind of way of doing it, and I've been looking around, I haven't seen that much written on it either, is that it's teaching the model behavior. So that what you do is if you prompt it properly, you can basically say, okay, first break this document up into thematic sections. Then write a description of the theme below that because you're teaching it to think through that process. And then based on those two things, write a series of keywords. Do this sequentially through the document working step by step. And then at the end, reread everything. If you missed anything from this list of keywords, put it in and then have a collate everything into a final list of keywords at the end. And using that approach, it worked quite well. I then generated a whole series of synthetic data. It wasn't fully synthetic, the responses are from GPT-4 doing this for about 400 examples. And then you use that as your training material. And that process has been really interesting because you basically end up with this corpus of training material at the end, which is about as good as GPT-4 gets on this. If you fine tune 3.5 on that, when 3.5, if you just run the native model on, say, a test of 100 examples, which is what I've been doing, and you generate a list of ideal keywords you want to have come out at the end, it gets maybe 50 or 60% of them and GPT-4 will get maybe 70%, 72% accuracy against a human at the end. If you fine tune 3.5 just on GPT-4's responses, you get about the same accuracy as GPT-4. But if you then go through and you actually work through all of those fine tuned, those examples for fine tuning and you edit them and you correct them and you fix all the mistakes and then you figure out, oh no, I need more examples of this and less of that, and you really dig down your data set, you can get that number with a fine tuned 3.5 model up to about 85%, which is significantly better than GPT-4. And I find that itself really fascinating in the sense that that process of teaching 3.5 to think through sequentially and to do it the same way every time results in that significant a change from GPT-4. And it gets it to the point where effectively the model you end up with, the fine tune model at the end you end up using is usable in the sense that its results are at that point about as good as a human. I mean, I've had a few students try to do this as well and see like, okay, I have my list of ideal keywords, but if I hand this off to a student, what am I going to get back? And it's in the same range. It's 78 to 80 some percent. And then you kind of go, okay, this is actually at the top end of what humans can do with this because people are going to have different opinions about what keyword fits as well and stuff like that. To me, that's been the most fascinating part of this process. And I do think it all comes down to, again, in fine tuning, what we're really doing is teaching behavior. And so as a result, if you can teach it to think through that behavior and articulate that behavior instead of just showing it the ideal outcome, that seems to be a far more effective approach. And that to me is really interesting. I've gone back now and I'm working on trying to improve those earlier models I did looking at trying to determine relevance through the exact same process. And it's the same thing where you're finding accuracy and success rates are increasing 20%. And that's really important when you're dealing with something that's fiftyfifty, like relevant, irrelevant. It's now not, it's now getting it right almost all of the time. And that's really what you want to see on a model that's like that. So to me, that's been a fascinating kind of a learning curve. And again, we've gone from August where this technology wasn't there to make this work to now with the fine tune model where, yeah, you can begin to actually deploy this and to use it, to do something useful. I think when I first started looking at this back a year ago, I thought we were probably two or three years away from that point when in reality we were less than a year, which just tells you again the pace at which this is moving.

Nathan Labenz: 55:29 So ultimately, it seems like you've kind of just crossed into the territory of like now you're actually accelerating the object level work for the first time. Well, congratulations. That's a big threshold. I constantly harp on the importance of threshold effects in AI, where it's all kind of just noise until it works. And then when it works, things really are kind of qualitatively different. And that's true in terms of model capabilities and individual workflows. It just kind of seems to be a huge theme. So it sounds like you've crossed a big one there. I think so.

Mark Humphries: 56:06 Yeah.

Nathan Labenz: 56:07 Well, yeah, it's awesome. And I do, yeah. It's funny that I don't have anything else really too much to add on the fine tuning point. It sounds like you've got a really good read on it and probably better than me. Honestly, my challenge is probably a little easier than yours, certainly more in the wheelhouse of what the models are meant, I think they're intended to do everything, but these sort of highly available use cases, we're more in line with those. So it's cool to see that this is working in other areas as well. Two kind of quick thoughts. One on the yes/no. I like to band my kind of classifications out of this stuff in usually like a five band and also typically try to label those. You could say like one to five, but I don't trust it to kind of translate to numbers as well. So I usually go like, in your case, would be like highly relevant, somewhat relevant, whatever, neither, whatever. You got to obviously come up with your own labels. How do you think about two versus say five? I feel like I get better results from five because it's less like right and wrong and more kind of, kind of think of it as like, now I can kind of skim off the top. Like, now I can only, I can look at the ones that were all labeled highly relevant and then maybe go to the next tier if I need to, as opposed to a yes, no.

Mark Humphries: 57:31 Yeah. No. And that's a really good point. And so when I was first trying that out, what I was finding when you're using just 4 without an untuned model was that I was using basically what you were saying. Very relevant, somewhat relevant, like unlikely to be relevant and then irrelevant. Or something like that. And what I was finding with that was that the model very often wanted to say something was likely relevant. And you could play with the language and try and change the actual semantics of that to make that a less desirable input. But especially in historical documents, it seemed to be really tentative about saying something was relevant. So as a result, if you wanted to take, if you got 500 results from your semantic search and you wanted to take the ones that were in levels three, four and five of those, you would probably have 75% of them. And most of them would probably not be relevant at all. And so the fine tune model, though, seems to be much better than that. I think that that's exactly what I'm actually trying to do right now, which is go back to that first approach and to try and fine tune a model that is much better at making that determination. Because what I originally found with GPT-4 was that, again, when you, it worked better to say irrelevant or not to give it that choice, it would, you are more likely to get the relevant results by doing that, even though you might lose one or two ones to being irrelevant. But you wouldn't get 75% of examples that just weren't there at all. And so I think that, again, to try and use that chain of thought fine tuning process, I think that will be far more effective. That's my hypothesis anyways, at getting these models to think through what you're describing. But yeah, I think that that's, again, it comes down to teaching a behavior. And thinking with fine tuning about teaching in this behavioral kind of way. And as you say, there's different use cases. Mine is very specific. And I think that's what's fascinating about this is that I still have, I'm always amazed that there's these things out there you think this won't be able to do that. This is way too specific. This is way too kind of niche and isn't going to work in this model wasn't designed to do that. And then all of a sudden it can. I ran into one of these emerging capabilities the other day, which is really interesting, which is like an example of this thing where you don't expect it to do this. So we were dealing with indigenous names from these records and we were trying to trace indigenous people through. And their name, indigenous people's names are spelled phonetically very often in these documents. And they're very different from one place to another, one document to another. And I actually tried to get the model then to take some names and say, to come up with training examples, spell this out phonetically. And the model started doing that. And what's interesting is it didn't use the type of phonetics you'd find in a book. It was actually making the sounds and writing out these names phonetically. And again, you kind of pause and you think that's really an odd thing in the sense that it doesn't know, it shouldn't know what these things sound like. And yet it's not taking a direct kind of a phonetic approach where you would be using the type of language you'd find in books that it might have learned. And so again, there's these kind of, that was something I thought we'd run into as a wall with this as well with names that were largely phonetic that it wouldn't have run into. But again, it seems to be able to think in phonetics and recognize phonetics, which I'm still kind of processing how that would work and how that training process would have worked that it could actually think in terms of sounds and be able to translate words into sounds and then back from sounds into words without using a Whisper type approach. But that's an example again of something I thought would not be a possibility, which then you're like, oh, it can actually do this. Well, that opens up this. So it's to watch how this evolves so quickly, I guess.

Mark Humphries: (57:31) Yeah, no, and that's a really good point. And so when I was first trying that out, I mean, what I was finding when you're using just 4 without an untuned model was that I was using basically what you were saying. Very relevant, somewhat relevant, like unlikely to be relevant and then irrelevant, or something like that. And what I was finding with that was that the model very often wanted to say something was likely relevant. And you could play with the language and try and change the actual semantics of that to make that a less desirable input. But especially in historical documents, it seemed to be really tentative about saying something was relevant. So as a result, if you wanted to take, you know, if you got 500 results from your semantic search and you wanted to take the ones that were, you know, in levels 3, 4 and 5 of those, you would probably have 75% of them. And most of them would probably not be relevant at all. And so the fine tune model, though, seems to be much better than that. I think that's exactly what I'm actually trying to do right now, which is go back to that first approach and to try and fine tune a model that is much better at making that determination. Because what I originally found with GPT-4 was that, again, when you it worked better to say irrelevant or not to give it that choice, you are more likely to get the relevant results by doing that, even though you might lose one or two to being irrelevant. But you wouldn't get, you know, 75% of examples that just weren't there at all. And so I think that, again, to try and use that chain of thought fine tuning process, I think that will be far more effective. That's my hypothesis anyways, at getting these models to think through what you're describing. But yeah, I think that it comes down to teaching a behavior. And thinking with fine tuning about, you know, teaching in this behavioral kind of way. And as you say, you know, there's different use cases. Mine is very specific. And I think that's what's fascinating about this is that I'm always amazed that there's these things out there you think this won't be able to do that. This is way too specific. This is way too kind of niche and isn't going to work in this model wasn't designed to do that. And then all of a sudden it can. I ran into one of these emerging capabilities the other day, which is really interesting, which is like an example of this thing where you don't expect it to do this. So we were dealing with indigenous names from these records, and we were trying to trace indigenous people through. And their name, indigenous people's names are spelled phonetically very often in these documents. And they're very different from one place to another, one document to another. And I actually tried to get the model then to take some names and say, to come up with training examples, spell this out phonetically. And the model started doing that. And what's interesting is it didn't use the type of phonetics you'd find in a book. It was actually making the sounds and writing out these names phonetically. And again, you kind of pause and you think that's really an odd thing, right? In the sense that it doesn't know, it shouldn't know what these things sound like. And yet it's not taking a direct kind of a phonetic approach where you would be using the type of language you'd find in books that it might have learned. And so again, there's these kind of, that was something I thought we'd run into as a wall with this as well with names that were largely phonetic that it wouldn't have run into. But again, it seems to be able to think in phonetics and recognize phonetics, which I'm still kind of processing how that would work and how that training process would have worked that it could actually think in terms of sounds and be able to translate words into sounds and then back from sounds into words without using a Whisper type approach. So but that's an example again of something I thought would be not a possibility, which then you're like, oh, it can actually do this. Well, that opens up this. So it's to watch how this evolves so quickly, I guess.

Nathan Labenz: (1:01:22) Yeah. Funny enough, one of the things that I've worked on might actually be related to that, which is for voiceover script writing, often the talent needs the proper noun in kind of not even like formal phonetic representation, but just sort of whatever is going to make it least obvious or least likely for them to make a mistake in the pronunciation. Because that's probably the number one thing that ends up cycling voiceover work is you say the name wrong, they can't write it. It's just not going to work. So you're back in the studio, so they really want something that is super intuitive and super clear for how they're supposed to say it. And we've tinkered with that a bit, and it's been kind of the object of some of our fine tuning as well. We're also using an AI voice in our case sometimes. So it's like we've kind of two levels of idiosyncrasy. The names can be weird, and also the AI way that it pronounces can be weird. If I had to guess what sort of data it might have seen that may allow it to kind of generalize to the historical version of that, there's at least something like that that's in the marketing realm that is very much kind of likely to be in the training data. Let me give you one thing on Claude. You're going to spend the afternoon on that. And let's do the education and the big picture lessons from history. The Claude 100ks, I don't know where, of course, these are never like super sharp cutoffs. But in my experience, if I take a 3 hour podcast transcript, and that typically will fit roughly speaking into the 100,000, drop the whole thing in and ask it to do that timestamp topic outline like we mentioned earlier, it basically totally fails. Massive hallucination, totally wrong. Up to about 90 minutes, it seems to be very good. So I don't exactly know. I haven't systematically explored. We've only done so many episodes and certainly only so many 3 hour episodes. But where I've kind of come to for the moment is it seems to be very reliable up to at least about half of the context windows. If I do have a full 100,000, I'll usually chop it into two and then just get dramatically better results. So let me know what you find in that regard or watch out for that. Something about hitting the upper end of, you know, what it can handle still sometimes takes it in a weird direction. But you can still be, you know, well above 32k and also a lot cheaper, right? Compared to 32k, you're talking like, I don't know, whatever. It's maybe a sixth of the cost. So it's a significant savings and can go significantly longer, but just not all the way I've found to kind of, you know, the full published limit. So see what you find. Okay. So then let's talk about education. So this has been something you've been kind of writing and blogging and experimenting with. Much has been made and said of homework will never be the same and essays won't work. I think to your credit, you went out and just looked at that where the rubber hits the road and gave students some interesting options to try things with and without using AI and tried to see how much it helped them. Tell us what you've learned in kind of first two semesters of bringing AI to the classroom.

Mark Humphries: (1:04:52) Well, I think most people in social sciences and humanities are still very much worried about plagiarism with this. And it does fundamentally change the types of assignments you can offer in higher education because, yeah, if offering a course where the majority of your grading is going to come from short answer kind of responses that you let students do at home, I mean, yeah, ChatGPT does a pretty good job at those. I think what it does is it forces a lot of people to rethink what exactly it is that we're doing and what we're trying to teach people. To me, I think we're probably heading into a world in which the students who are starting today in university and college when they graduate, the odds that they're going be working in jobs where they are not using these tools are pretty slim, right? Because they do speed things up. So I think to me, the goal is to try and get students to learn to use it effectively. And that's a motivational kind of thing more than anything else, because a lot of students, from my experience, come to it and they bounce off it because, you know, it doesn't just do the thing for them immediately. Whereas if you can motivate people to learn that what you can actually do even through a chat window or something like that, because that's the most basic version of this, just using ChatGPT is that you can actually use it to kind of help them level up their writing. So that if a you're student who really struggles to get your good thoughts across in a cogent, coherent way that is stylistically the way that your professor wants to see it or whatever, ChatGPT does a pretty good job of obviously helping you edit things. And I try and explain to people that in that scenario, if you give it a paragraph and you try and say, make this more, you know, coherent or flow better or whatever, it doesn't tend to even change your ideas. It's basically like having somebody be a writing coach or having your mother or something like that, you know, help you edit your paper or a friend or whatever. So there's that kind of level. And think that the biggest problem is, again, motivating students to do that. And it's like anything with AI. If you become interested in it and you're willing to try out and play around with it, there are benefits you're going to get from it immediately. Other ones are things that we can see are going to come online in the next few years that aren't there yet. There's a lot of promise there for students who are kind of struggling to get from that kind of C range into that B and A range to simply help them get there with editing and with just sometimes even understanding the material. I mean, you can if you're having trouble understanding a concept, as I tell people, if you're teaching math and you get tired of explaining to a student why this certain concept works or, you know, how this concept they need to know, well, ChatGPT will explain it nicely to the student until eventually they do understand it or, you know, choose to not engage anymore. And I think that there are those types of things using this as tutors and things like that, that we can start to do. But that's not really being integrated yet in a systematic way in higher education. I think we're seeing people talk about it. We're seeing the promise of this coming. We're seeing some companies begin to offer some of these services. But like a lot of these early kind of AI services, they're, you know, sometimes kind of wonky, sometimes they work, sometimes they don't. And the expense is really prohibitive for universities right now. I mean, doing it on a one off student has a ChatGPT subscription, they're going go use that. That's one thing trying to make this systematically available to people. That's going to be much more expensive. And I think we're really going to see things shift in academia when you begin to see it roll out in Office in a way that is available kind of more broadly. Copilot available in Office. That's when things will change and more people are just going to get exposed to it. And as a result, I think universities are in this holding pattern right now of trying to avoid the issue oftentimes more than anything else. And that's going to bring about a change where we're going to have to decide what we do with it and how.

Nathan Labenz: (1:08:40) The cost thing always kind of blows my mind. I mean, not to say it's insignificant, but I remember when I paid for textbooks when I was in school, and I'm like, okay, dollars ChatGPT per month for the best available thing in the game right now. That's basically what one textbook would typically cost me. And so for two textbooks per year, would cost one textbook per semester, you can basically have revolutionary technology in your pocket. And yet, it's deemed too expensive. That must just be I mean, that's got to ultimately just be lack of appreciation for what it's good for. The budget will shift, I think.

Mark Humphries: (1:09:29) Oh, yeah. And I mean, as I try and tell people, I mean, that's a Netflix subscription. I mean, this is not, in the grand scheme of things, the biggest expense university students you could have by far. So yeah, I think it's a red herring. But I think the larger cost is for institutions adopting it when you're talking about thousands of employees. And what often institutions do is they will buy a piece of software for everybody, including all students. So if you have an institution with 30,000 students and you're going to have to make a decision, are you going to buy the Copilot version of Office Enterprise at $30 a month per person? That's then where the expense gets prohibited for the institution. At an individual level, it's not. And so institutions are very concerned about some students who are going have it and other students who won't and these types of things. So I think there is that kind of red herring cost of, oh, it's expensive on an individual level. I agree, it's not, not in terms of the grander scheme of things and the people pay for at university. But in terms of institutional adoption, it does actually become quite expensive if you look at some of the pricing schemes that are out there right now, if you really wanted to dive into this. And a lot of the applications that are going to come out that are going to be the most useful for AI in the next few years, I think require that institutional adoption. That's kind of, I think, where we haven't gotten over that hump yet.

Nathan Labenz: (1:10:47) Yeah. Again, the capabilities overhang. By the time we get there, we're going to have GPT-5 and then it's going to be a whole new set of questions. Two thoughts here. So you have written a bit about kind of what sort of assignments are good assignments today. And I'd love to hear that riff. And then also, how long do you think that lasts? It seems like we're on this trajectory now where, as you said, I think at a recent blog post, it's really good at 500 to 750 word essays. If you assign those, you're kind of playing right into the I can just have ChatGPT do this wheelhouse. If you assign something more expansive, then it doesn't really do that as well. So at a minimum, I mean, I'm starting to give you a rift for you based on the blog, so I should shut up and just let you do it. But then I do wonder, we've come pretty far, pretty fast. It seems like we're headed to, like, ability to do 4,000 word essays, you know, pretty soon too. Right? Nathan Labenz: 1:10:47 Yeah. Again, the capabilities overhang. By the time we get there, we're going to have GPT-5 and then it's going to be a whole new set of questions. Two thoughts here. So you have written a bit about what sort of assignments are good assignments today. And I'd love to hear that riff. And then also, how long do you think that lasts? It seems like we're on this trajectory now where, as you said, I think at a recent blog post, it's really good at 500 to 750 word essays. If you assign those, you're kind of playing right into the "I can just have ChatGPT do this" wheelhouse. If you assign something more expansive, then it doesn't really do that as well. So at a minimum, I mean, I'm starting to give you a riff for you based on the blog, so I should shut up and just let you do it. But then I do wonder, we've come pretty far, pretty fast. It seems like we're headed to ability to do 4,000 word essays pretty soon too. Right?

Mark Humphries: 1:11:48 Yeah. And so, I mean, I think there's two things going on. You talked about the capability overhang. Right? I mean, well, the reality is a lot of people who are teaching in university are still encountering AI for the first time. The number of meetings I mean, I do a number of these presentations about just what AI is because this is something I get asked to do around the university. The number of people who have never actually logged in and tried it out is still astonishing to me. Right? And I think that that's pretty common on a macro level still. And so you have a lot of fear. And I think right now, yeah, if you're worried about AI, you haven't encountered it yet, there are ways around it. You can do longer assignments. AI will not be as good at those. That will most definitely change as you say. Right now, it's not great at doing citations, even GPT-4. If you require certain types of citations, it does better than others. But yeah, that's going to get better as well. And there's lots of programs that are out there too that are doing this for students now as well, Cactus and things like that. Right? I think that's going to change going forward. That's kind of a this idea that there are ChatGPT proof type assignments is kind of a holding action for people who still are needing to experience this and figure out what they're going to do with it and are really nervous about it in the meantime, right? And which is the conversation you're seeing an awful lot in academia about this at the moment, especially in the social sciences and humanities. That's going to evolve, right? I think what's going to happen is we're very soon going to start to develop assignments that simply use ChatGPT or whatever the version is as a tool. There was a time when one would take a course to learn to use a word processor. That time has kind of passed. I think most people just learn that process through osmosis. And that's going to be this, right? And there was a time too, before I went to university, that spelling and grammar and things like that were on an English essay, that might be 20% of your grade. Well, with spell check, that became less important, right? Expressing yourself grammatically in a grammatically correct way with correct spelling was still important, but you didn't tend to assign grade points for that anymore in the same way as you used to. And we're going to see the same thing evolve. I think the future of this is learning to write quickly and effectively with AI in a way that levels up your ability to process large amounts of information. The tricky thing is I think it also scales up expectations, right? That what I've seen happen is that what AI is able to do with an assignment is probably going to become the minimum, right? If you think about the way that this is probably going to play out in the job market, if you have someone who is unable to achieve the level of GPT-3.5 on a given task and that task is doable in that job, it's unlikely that that person will be successful in that job. And I think that that's the lesson that I keep going back to, which is that what we need to do is teach people to use this in such a way that they exceed the baseline level of the model. Right. That's going to be difficult. It's not going to be possible for everyone. Just in the same way as not everybody gets an A today. And that's kind of just how things work. Right. But that's kind of the real trick is figuring out a way to do that. And I think a lot of that comes with experimenting and teaching students to be open to new ideas and new things, to really digging into this and playing with it and learning how to do that. And we're still doing that again because it's only been around for a few months. Right. So teaching something that you're still learning yourself is not always the easiest thing to do. And I think there's going to be this period in education where we're going to be adapting. That's going to be sped up as certainly the various tech companies out there that produce educational software start releasing out of the box solutions to some of these things and tools that are going to be used to help students do these things out of the box rather than through this hacky way of saying, well, try this type of a prompt in ChatGPT to do this, right? I mean, that's going to disappear when we start seeing things get integrated into things like Google Classroom and My Learning Space and things like that, right? We're not there yet, but I imagine we're months away from it, if not even weeks or something like that at this point.

Nathan Labenz: 1:16:02 Yeah. We did an episode on Khan Academy, and they've got plenty more work to do, of course. But they've really made a nice start at just at a baseline level. Their main approach is the thing will not do the work for you. It's a Socratic approach. It'll go back and forth with you, but it will not solve the problem. It will not answer the question. So it will only lead you there. So that's a pretty good start. And then they've got lots of plans for personalization and all sorts more. I think your comment maybe jumps out to me most there is it seems very obvious, but I thought you framed it extremely well that the model baseline output is something you have to exceed in order to be a value add in the very short term future economy. And I do think that's going to be pretty tough. I think we're already in a place where I wonder what you would say, you know, you see in terms of just comparing today's language models to your students. But across the board, it seems like a good short summary of where we are is and I'll say GPT-4 here. The best models, the GPT-4s, are better than the average person on most tasks and closing in on expert performance on very common tasks, still falling short usually, but getting pretty close on the most routine expert tasks and still well short when it comes to the sort of breakthrough insight type generation that is the non-routine work that's kind of the most alpha. I don't see that really coming from AI much at all yet, if at all. So don't know if you would see that any differently or how you see that playing out for just GPT-4 versus your students on your assignments. But it does seem like the bar for getting a job might go up quite a bit. And most jobs are a lot of routine stuff, and AI can do the routine stuff. I mean, it does seem like we're headed for this kind of starts also getting to the what lessons can we draw big picture wise from history for such a moment, but does seem like we're heading into a pretty disruptive period here.

Mark Humphries: 1:18:17 Yeah. I think so. I mean, I have nihilistic days where I get very worried about what this looks like. There are days where I get the model to do something and I'm like, oh, that's very close to what I do. That's a little frightening. And I mean, I think everybody uses these models has that kind of an experience at some point or another. What I think history teaches us about a lot of this is that jobs don't tend to disappear. They tend to evolve and change. I think what's true is that I had a student ask me at the beginning of a class, at the beginning of the semester when I take a very permissive approach to using ChatGPT in class, I tell them you can use it and you basically are responsible for the content you submit. If it's full of hallucinations and made up citations and you can't talk about it, well, it's probably going to fail on its own, having nothing to do with whether it's written by AI or not. Right? And so I had a student say, well, why wouldn't I just get ChatGPT to do this without trying to engage with it? Like, what's the point? I think the response is if you have a job where all you have to do is input your job into ChatGPT and it outputs it, I don't think that job will exist in that form for very long because you're a lot more expensive than ChatGPT. And so I think that what we're seeing here is we're trying to look at what that means. Right? I think you're right. Right now, these models are very good on a whole range of different tasks, but they aren't autonomous. Right? You can't just set them loose and tell them do this job. That's not how it works. Right? But there are going to be tasks that increasingly people will probably do less of. So if in a job where you might spend a lot of time summarizing written documents, you might still well have that job. It's just that you might be expected to summarize more of those documents faster. Right. And that's what history tells us about automation in general is it's not that that job disappears, it's that the expectations of production go up. Right. And I think that that's probably true here. I do think that if you are producing content, right. And I think quite rightly, there's a lot of artists and writers who are worried about what this does. I think that the baseline becomes important to you, right, that if you think about a job where you're producing content, that's kind of just what you're doing. If you're producing something that is not as good as what you can get out of GPT-4, that's going to be an issue going forward. Right. I think it's also true in terms of resume writing. I mean, there's a big debate about whether you should get your cover letter written by ChatGPT or not. Does that help you or hurt you? Right? And you'll see all these stories around. But the reality is that if you're submitting cover letters that are full of errors and things like that, you might well find yourself not getting at the top of the pile anyways. And as expectations go up because you might find less stylistically awkward letters, you might find yourself further and further down that list. And I think that this is what we're seeing. I think a lot of people see this transition as being an instantaneous one or something that's all or nothing. And it's going to be much more gradual and gradated, right? You're going to see basically tasks being offloaded in people's jobs to ChatGPT gradually and those jobs will evolve and change over time. There might be fewer of these types of jobs because well, now people are more efficient at doing them and there'll be more of those types of jobs because now we need people to herd AI or something like that where effectively you're going in there and your task is to get all these different models to coordinate and do these different things, right, and produce all these different outputs. That's what history teaches us. I think about how these things evolve, right, is that technology is constantly changing what we do. This is an especially fast revolution. So it's something that's unusual in that sense. But again, if we look at past situations, the invention of the steam engine or various types of machines that automate production, it's not a day and night thing. It's a gradual adoption. It's a process that begins. If you think about the auto industry, that process begins in the fifties and sixties and it's still evolving today. Right. Will this happen faster? Probably. But again, it's not going to be day and night. Right. And I think that that's why learning to use this is so important right now for people, is engaging with the technology is, I think, in many ways, future proofing yourself, understanding what these models can do and having an understanding of, well, if I'm asked to write a paragraph on this, this is what it looks like from GPT-4. I should probably be trying to aim for something like that if I'm being asked to write a paragraph. I mean, that's the kind of stuff that I think we're still at is that encountering phase with AI for so many people.

Nathan Labenz: 1:22:55 I've kind of got mixed feelings on this because I think you're definitely right to say the general pattern, especially if you're in any sort of volume game, the pattern of doing it manually evolving toward maintaining, monitoring, fixing, supervising the quality of the machines that are doing the work is pretty well established and likely to happen here too. But I also do really worry about the fact that just I think, unfortunately, a lot of people cannot write at a GPT-4 level. And that is going to be pretty hard to change. And I don't know. One of the things I said to the OpenAI team when I was doing the red teaming was for the vast majority of people, this is superintelligence. It's just not superintelligence to you because you're really smart. And so I don't know how that part gets resolved. And I think at the sort of high end of the sort of whatever, let's call it intellectual or academic ability range, we're still just encountering ChatGPT. But then for so many folks who are never been in the university setting at all, it's just man, that's a really big chasm to cross. And I don't know how many people do it. So I don't and I don't know that they're gonna be able to supervise GPT-4 either. It just seems very seems very challenging for just a huge portion of the population. Nathan Labenz: 1:22:55 I've kind of mixed feelings on this because I think you're definitely right to say the general pattern, especially if you're in any sort of volume game, the pattern of doing it manually evolving toward kind of maintaining, monitoring, fixing, supervising the quality of the machines that are doing the work is pretty well established and likely to happen here too. But I also do really worry about the fact that just I think, unfortunately, a lot of people cannot write at a GPT-4 level. And that is going to be pretty hard to change. And I don't know, one of the things I said to the OpenAI team when I was doing the red teaming was for the vast majority of people, this is super intelligence. It's just not super intelligence to you because you're really smart. And so I don't know how that part gets resolved. And I think at the sort of high end of the sort of whatever, let's call it intellectual or academic ability range, we're still just encountering ChatGPT. But then for so many folks who are never been in the university setting at all, it's just man, that's a really big chasm to cross. And I don't know how many people do it. So I don't know that they're gonna be able to supervise GPT-4 either. It just seems very challenging for just a huge portion of the population.

Mark Humphries: 1:24:27 I often find myself shying away from those types of thoughts simply because, I agree. I mean, I think that there is a huge chasm between what this can do well and what many people can do at their best. And I think that that's where I think some of the obligation is going to fall on companies like OpenAI and Altman's talked about this as others have as well about trying to figure out a way of what this means, because I think what we can learn from history is that what you don't want to have happen is have a situation where technology displaces huge numbers of people from the workforce and leaves them disaffected and worse off as a result. That's not a good thing for society and it's not a good thing for culture and it's not a good thing for the workforce in general. And so I think that's something we have to be acutely attentive to here. I do think the technology that we're dealing with is a very different one than previous technologies that we've encountered. I think some of those patterns will be similar. But this is the knowledge economy, right? This is people who are the people who be most affected here are people who are professionals and do things like write for a living and analyze documents and things like that for a living. Right. This is very different than previous revolutions in that sense. We haven't figured out how to handle that. And there's always turmoil when this happens. And I think trying to tackle both problems at the same time is really important, right? Is trying to figure out how do you adopt these technologies and use them, but how do you also assume or make sure that in their adoption you are ensuring that people still have access to the types of at least incomes and living standards that they've traditionally had, right? And find meaning in work, right? And I think that that's an important thing too. So yeah, I mean, I think that those are real problems that we have to contend with as a society and that are much larger than just higher education. That's pervasive.

Nathan Labenz: 1:26:19 Are there any historical, and you could take this in whatever way you want, but could be moments, could be technology revolutions, could be moments when the social contract needed to be updated. Or you could think about it in a totally different way, more conceptual or whatever. But what would you recommend that folks who maybe don't have such a great grounding in history like myself could look to to sort of learn something about what our forebearers either did successfully or failed on when they were confronted with, at least the most analogous challenges that we've seen in the past.

Mark Humphries: 1:26:55 Sure. I think there's a couple of good examples. To learn about how societies adapt to this. It's the industrial revolution, right? It's the process, especially in the late nineteenth century of what industrialization does to societies, right? And it's where you do see the development of huge impoverished classes of people who are put out of work by automation or find their work devalued for what they were doing before. And that's what leads to in a lot of countries, the development of social programs in the early twentieth century, right? And into the mid twentieth century is trying to fill that gap essentially between that that's being created by urbanization and industrialization happening at the same time. So that's an area to look at. I mean, again, it's a process of the plays out over a very long period. This is something that will be much more condensed, but I do think it will release some of those same types of forces. Right. Another one I think is a little bit more removed, but it's kind of telling as well as the enclosure movement. This is a period in the early modern period when you begin to see people getting kicked off of farms in order to basically turn them into sheep pens because effectively it becomes more profitable to farm sheep than it does to people growing crops. And this leads to a huge, again, kind of group of people who are simply dispossessed of land and just sent off and they end up in urban environments in hugely problematic situations. Right. And again, it's about, in this case, not so much a technology displacing people as it is about the emergence of a new kind of economic force. Right. And a transition away from a subsistence agrarian economy to the beginning of a manufacturing economy. I think there's parallels again there, right? And the lessons really for us from history are that when this type of a transition takes place, large portions of the population do typically become disaffected and alienated from kind of the jobs that they would normally have been familiar with and that this is a real problem and that governments that have navigated this successfully, although albeit in fits and starts over a very long period, have done so imperfectly, although that is, by coming up with social programs that effectively bridge that gap. What that looks like, I mean, I think I'm better talking about the history of it than I am in the future of these things, which is my cop out on these types of issues. But I think that it is true, right, is that we know that that's going to happen. We know that if this technology evolves the way that it seems to me to be evolving, that we will be facing this type of situation. My daughter is 7, right? And I very often think what will her world look like in 15 years, right? Because it's very much going to look different than I thought it would look when she was born. And I know that much to be true. How you prepare for that, I think is a very different thing. So I think as individuals, we have to prepare for that. But I think governments and whole societies have to begin to think about how we do these things collectively. Right? And that's where I think some of that, those conversations around the long term consequences of AI that crop up every once in a while that we get the congressional hearings about or things like that. I think that's the much more immediate threat here probably than machines taking over the world type of scenarios. I think that has its place to be talked about as well. But I think we really do have to deal with that issue of how we deal with the effects of this technology.

Nathan Labenz: 1:30:17 Well, that's a good note and a good call to action for the audience to participate in that conversation and maybe start with some historical reading to get some good context. Mark Humphries, fantastic conversation. I really appreciate just how far down the AI rabbit hole you have gone and what a unique and interesting perspective you have brought to it. So thank you for being part of the Cognitive Revolution.

Mark Humphries: 1:30:43 Thank you so much for having me.

Nathan Labenz: 1:30:45 It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.