Alternate Histories and GPT-3

Chris Gates:
Okay, you are recording locally. I'm going to count myself in. Three, two, one. All right, Sam, so you put out a thread recently about using GPT-3 to generate alternate histories, and that's just fascinating to me. There's something about the intersection of AI and art, and alternate history novels are one of my favorite things to read, so this kind of piqued my interest. Well, I think maybe a good place to start would be, can you give a quick synopsis of GPT-3?

Sam Arbesman:
Sure. I mean, yeah. GPT-3 is one of these large language models. This one, it's developed by OpenAI, and it's using a deep learning network too, and that has ingested a insane amount of text, presumably scraped from large portions of the web, and then used that to, kind of, I guess find associations between words and basically create a prediction model of saying, "Okay, here's like, words, or I guess letters." They use tokens, it's kind of like a stand-in for basically characters, roughly.

The idea is looking at the text that has come before, what makes sense to predict what should come next. The idea behind this though, is that as this model ingests more and more text and becomes more and more complex, not just being able to look at the most recent word or even the most recent letter, so rather than saying, "Oh, when a Q comes, we know that a U comes after it." So that's a very, very simple model. But the idea is once you have a much larger network, you can actually say, "Okay, based on what has come in, can we actually figure out what has come before?" And you begin to generate associations between words and concepts and things like that, and all embedded within this model. You can prompt it with text and it will then pop out things that will autocomplete sentences, it will answer questions, sometimes more or less accurately.

Chris Gates:
The thing that comes to mind for me, like a use case, is like, "Okay, I'm designing a website and I need to fill in a bunch of copy." Instead of filling it in with Lorem Ipsum, I could use GPT-3 to just fill in some generic copy.

Sam Arbesman:
Oh yeah. And can do that, and it can be smarter than that, and you can also say, "Tell me a story about something written as if it was like..." you would say, "Tell me a story about a Prince written in the style of fairytale." And it would then pop something out that is not terrible, it's kind of interesting to see. And there's a lot of different things you can use it for. And I guess GPT-3 came out a couple of years ago? Since then, people have used it for a lot of really interesting and wild use cases, sometimes summarizing news stories, writing things.

Chris Gates:
So talk to me about how we got to this particular use case.

Sam Arbesman:
Oh, okay. Yeah. So I also am very interested in alternate histories. I've been reading them for a long time, I've been thinking about them a lot. I just love this idea of a counterfactual what-if of, okay, if something were different in history, this point of divergence, some little thing, like someone was never born or some election went the other way, or something different happened in a war, what would've actually happened? And then you kind of game out history and tell this wonderful story. I've always loved those stories.

And then soon after GPT-3 came out, someone was playing with it and kind of turning it into a... And I think actually this might have been the official thing through OpenAI, they showed how you can use it to almost autocomplete anything you have in some sort of spreadsheet. And you can say, okay, "Here's some states, here's some populations." And then it just automatically fills it in. Now it turns out some of the data that it fills in is not entirely correct, but it clearly shows that embedded within this massive dataset and the associations and the web of knowledge within GPT-3, try actually asking it a what-if and say, "What if someone weren't elected and what would happen?" And so that was how I began thinking about the intersection of GPT-3 and alternate histories.

Chris Gates:
Okay. So I see that there's lots of embedded information in GPT-3 about the world, so let's give it a couple what-ifs.

Sam Arbesman:
Right. If it has the actual complexity of the world, maybe it can give insightful answers to all these what-ifs. And so, after I started playing with some of this, one of my friends noted, he's like, "A lot of these what-ifs are well attested online, like people will ask lots of what-ifs." So presumably it not only has been trained on some of these what-ifs, or not only just trained on the complexity of the world and its relationships, but some of these actual counterfactuals. So, that being said though, I don't necessarily think that's a bad thing because a lot of these counterfactuals try to actually show the relationships between various factors in the world. But yeah, so that was kind of what I did.

And basically the kind of prompts that I gave it were, I would say, "Here's a scenario, here's the counterfactual." And then I would give it a paragraph of what would happen after that what-if. So again, it's being based on alternate history novels, which sometimes are more reasonable than others.

Chris Gates:
Right.

Sam Arbesman:
But the idea is I'm kind of prompting with, "Wait, this is the kind of thing I want out. I want an interesting alternate history, so now give me an alternate history when I tell you something new or something a little bit different from from what you've seen or whatnot."

Chris Gates:
Right. And so to give people maybe a way of thinking about this, we're not saying, "Look, this is what would've happened," we're like, "This is what a really amazing novel about this would be."

Sam Arbesman:
Oh yeah. And I will say, really amazing is probably not the right phrase. We'll say somewhat thought-provoking.

Chris Gates:
Okay.

Sam Arbesman:
Yeah. I would say the takeaway for me is not that GPT-3 is writing amazing stories that are going to blow your mind. A lot of them, they kind of go off the rails or they say things that are not terribly insightful. I mean, it's not bad. They're interesting. They're pretty good. But for me, it's much more insightful as GPT-3 is useful almost as like an assistant to help me think through the possible implications for these kinds of things. So, for example, if something were different, if an election were different, I should be thinking about various domestic issues, various foreign policy issues, various geopolitical issues. And by posing that to GPT-3 multiple different times, and you can can rerun it over and over, it gives you a sense of these are the kind of things you might want to be thinking about.

And in some ways I view it more as if I wanted to write an alternate history or I just simply wanted to think about a scenario, it's a really good research assistant, like it's helping me think through these things. And it has interesting stories and it tells interesting things, but it's not like, "Oh, now I'm just going to take this as a given." And I did this with... I think after I posted the essay about this, someone said, "Oh, rerun this, what would've happened if Bill Clinton had resigned rather than face impeachment." So, I tried that and it came out with an interesting scenario, but it also got the number of the presidents of what Al Gore would've been completely wrong. And so, clearly it's like, there's certain facts that it's just botching.

Chris Gates:
Yeah, like, wait, that should have been the easy part.

Sam Arbesman:
Right, so the easy thing was hard, but then it still told me an interesting story. And the truth is, probably that's because enough people have asked that question or written about it, that it knows what people might expect.

Chris Gates:
So that made me think of something, just the value of this, what is the value? Made me think about the Danny Kahneman episode that we did around pre-mortems, and trying to get your brain to think about all the things that could go wrong before you do a thing. I mean, it's kind of getting your brain in that zone as well. Like, okay, what are all the infinite possibilities if you change this one factor?

Sam Arbesman:
Totally. And I think the whole idea and, at least for me, the most value for this kind of use, is it expands your scope and widens your focus and says, "Okay, maybe we should think about all these different things as opposed to just the few things that you first think about. It gives you a way of expanding that scope rather than just getting focused on, "Okay, here's the things I know well, these are the things that it must be affecting," as opposed to saying, "Maybe there are some other things it could be affecting as well."

Chris Gates:
So, I also think about creativity and wanted to ask you what your thoughts around how this influences people's creativity.

Sam Arbesman:
I think it can get people maybe out of certain ruts. And so there are tools that help writers based on GPT-3. So there's a company called Sudowrite that does this. I played with it a while ago, I haven't played with it recently. I don't know what its current features are, but some of it, it's kind of like, if you're sort of in a rut, like there's a well-trodden way of thinking about something, it can jump you out of that.

And you can kind of think about this like, if creativity is this search space where you're navigating the search space trying to find the thing you want to write, it's very easy to get stuck playing only a certain portion of the search space, and these kinds of tools can jump you out of that and say, "Wait a second, it's a really high dimensional space. You should go be thinking about something completely different." And some of those completely different things might be things that should remain on the cutting room floor, but they can still help spur you to then think about something entirely different that could then be useful. And so I think that, the kind of partnership, like creative partnership of human and machine, I think that's also really important.

Chris Gates:
I mean, that's where my brain goes. Look, I'm an artist. I've identified as an artist for a very long time, and for me, art has always been iterative. It's basically like, I think about my brain as kind of an algorithm. I try to input it with things that I find pleasing, and then basically my brain then takes all that stuff and smashes it together, and I kind of just recreate art that I've consumed, just in my own unique way, as a different combination of all the stuff that I've consumed. And this feels like just a different way to input it, because I've read the alternate histories and yeah, I feel inspired. I'm like, I can write characters for that story, I can see where then that story could potentially keep going.

Sam Arbesman:
Oh, I love that. Yeah, no, and I think that creativity is this fundamentally combinatorial act, and the whole remix approach, and that's very powerful. And so having additional things and additional ways of recombining things in novel ways is really powerful.

Chris Gates:
All right, so will you walk me through your top three alternate histories that came out of your prompts?

Sam Arbesman:
So I mean, I don't know if it's top. I didn't try that many. So certainly, one I wrote in "What would've happened if Jimmy Carter had won reelection in 1980?" That one sort of... And actually, I reran that a couple times. One time I ran it and it focused entirely on foreign policy and it was all about things around the Iran hostage crisis. There's obviously differences and it also clearly rhymes with the way certain things happened in the world, like the prime world or whatever.

So I did another, I ran it again, and then that one almost was entirely domestic policy stuff. And so, one of the quotes, I didn't include the entire thing in my essay, but it was like, "The 1980s see a series of liberal policies enacted by the Carter administration, including gun control, environmental protection, and an increase in taxes," which is, I don't think, a very big jump to make.

Chris Gates:
Totally.

Sam Arbesman:
So actually, one thing that I notice with a lot of these kinds of altering histories is oftentimes... And GPT-3, it suffers from the same kind of problem that a lot of human-written alternate histories have, which is they always occur prominent politicians, potential presidents and things, even if so much is different. So the idea is, "Okay, there's some point of contingency, like there's this point of divergence, something changes and a lot of things change." But somehow, a lot of the same players are in there. The great men are still being shuffled into some kind of new history, which is, I think, a failing of imagination, both GPT-3 imagination and also human imagination when it comes to alternate history.

Chris Gates:
Absolutely.

Sam Arbesman:
So I wrote this essay, I guess it was maybe a year ago or so, maybe a little less, about maps from alternate histories, and there was this one really... it wasn't quite a map, it was just a chart, it was about if the United States had a proportional democracy in terms of... similar to a parliament kind of thing. The idea was that the point of divergence of when America became much more proportional was decades earlier or if not a century earlier. So the fact that these would all be the same people, even though everything was different for over a century, is just kind of ridiculous. So it's more like... it's just a fun thought experiment, and I think that's kind of the way to view a lot of these, which is these are fun thought experiments.

So Harry Turtledove, who writes a lot of alternate histories, he has this series of novels where it's the middle of World War II and aliens invade, basically scrambling everything. And so the question becomes, "Okay, what happens?" And I said, "Okay, I'll ask GPT-3 that, but with a slight twist." And so I think I said "In 1862, aliens invaded the Confederacy, throwing the Civil War into turmoil. What happens?" I'll read the actual, the result. "The Confederacy won the Civil War and the aliens became their allies. The Confederacy was able to keep slavery, and as a result, the United States remained divided into two countries." This is kind of weird because the Confederacy, it doesn't make sense where it's like, "Okay, the Confederacy was attacked by the aliens, but then suddenly now they won and now the aliens are their allies." It doesn't really make sense, it's a weird scenario.
Then there was another time when I ran it, I can read this one as well.

Chris Gates:
Yeah, please.

Sam Arbesman:
Where GPT-3 just goes way off the rails but in a very entertaining way and kind of goes full Star Wars.

Chris Gates:
Yeah.

Sam Arbesman:
So this result is, "The Confederacy enacts a scorched earth policy to prevent the aliens from taking their land. This includes the use of biological weapons on Union soldiers. As a result, the Union is forced to sue for peace. The Confederacy is given independence and the Union is dissolved. The aliens occupy the former United States and establish a new government, the Galactic Empire. The Galactic Empire is a dictatorship and the humans are slaves. In the early 21st century, a human resistance movement called the Rebel Alliance, overthrows the Galactic Empire and establishes the Galactic Republic." And so like, it's kind of interesting, and then somebody just goes...

Chris Gates:
You're like, wow, where does Luke come in?

Sam Arbesman:
Yeah, it goes insane. It's a pretty wild response. That being said, the interesting thing is what is mentioned, like the use of biological weapons. That's super weird, but at the same time though, there were essentially primitive biological weapons in the 1700s where the Native Americans were given blankets.

Chris Gates:
Blankets, yeah. Smallpox, yeah.

Sam Arbesman:
Like smallpox blankets.

Chris Gates:
Yeah, smallpox blankets.

Sam Arbesman:
So at first I read this and I was like, "Oh wait a second, biological, that doesn't make sense, they can't engineer things," but then I'm wondering, "Okay, what is going on here? Maybe it's some other kind of thing like that too." I guess I never asked GPT-3 that, so I have no idea, an insight into what is going on here.

Chris Gates:
I love that. Okay, let's do one more and then we'll wrap. Yeah.

Sam Arbesman:
Okay. And so this one, I don't have any stuff written out. One I asked, this is kind of based on, there's a fun book called For Want of a Nail, and the idea behind this is the American Revolution is unsuccessful, it's put down as a rebellion by the British, and so then history kind of unfolds. And the thing is, the conceit of the book though, is it's written as a textbook, and so it's this very dry style textbook, like a history textbook, but in that alternate world.

Related to that, I asked "What would happen if Thomas Jefferson had been assassinated by the British after he wrote the Declaration of Independence?" Like, just what would've happened? And I actually can't remember if I included that, like the For Want of a Nail scenario in the prompt, so it might have just rhymed it. But one of the interesting things, though, is it mentions, if that had happened, I think like the American Revolution was unsuccessful, basically it includes within it the scenario, most of it is not interesting, but it mentions explicitly that American frontier culture never developed, which was kind of interesting and it makes sense. But if there was never any sort of like... I mean, if there was never any American Revolution there was no frontier culture. And I wonder to what degree that feature of American culture is so contingent. I don't know, but it was interesting to just have that prompt.

Chris Gates:
Right, where does frontierism come? And that's such a huge part of American identity, is that frontierism.

Sam Arbesman:
It does make me think, it's going back to this whole, like-

Chris Gates:
It does make me think, and I like that.

Sam Arbesman:
Using it as a brainstorming tool is fascinating.