☞ Alternate Histories and GPT-3
Is the nature of civilization found in Large Language Models?
Awhile back, I came across this fantastic use of GPT-3 for automatically filling spreadsheet cells using information about the world, such as state populations:
Where are the data coming from? Apparently, these data are coming from within the model of GPT-3 itself.
In other words, the massive corpus that GPT-3 was trained on is so large that the model contains a certain fraction of the actual complexity of the world. But how much is actually inside these models, implicitly embedded within these neural networks?
I decided to test this and see if I could examine the GPT-3 model of the world through the use of counterfactuals. Specifically, I wanted to see if GPT-3 could productively unspool histories of the world if things were slightly different, such as if the outcome of a war were different or a historical figure hadn’t been born. I wanted to see how well it could write alternate histories.
I’ve long been interested in alternate histories (here’s an article I wrote about alternate history maps and here’s another piece on my thoughts on alternate histories more generally). Specifically, I am fascinated by the way that the creator of an alternate world—separated from our own by a small point of divergence—explores how such a world might be different to our own and how it might be the same. So I provided GPT-3 with some capsule summaries of alternate history stories derived from Wikipedia, and asked it to see what sort of history might be elaborated when you give it a counterfactual.
And I got some weird alternate histories.
For example, here is what might have happened if Jimmy Carter won reelection in 1980, as per GPT-3:
The Iran hostage crisis ends with the release of the American hostages, and the United States and Iran normalize relations. The Soviet Union invades Afghanistan in 1979, but the Carter administration is able to broker a peace agreement between the Soviets and the Afghan resistance fighters, known as the mujahedeen. The Soviet Union withdraws from Afghanistan, and the mujahedeen go on to form the Taliban. The Taliban are overthrown by the United States in 2001.
Somewhat insightful, but focused very much on foreign policy. Another run, with a slightly different prompt, yielded more domestic policy material, where the “1980s see a series of liberal policies enacted by the Carter administration, including gun control, environmental protection, and an increase in taxes.”
But let’s go strange. Inspired by Harry Turtledove’s series of novels about a WWII-era alien invasion, here’s how GPT-3 responds to this scenario: “In 1862, aliens invaded the Confederacy, throwing the Civil War into turmoil”:
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.
Another result for this scenario involved GPT-3 going full Star Wars:
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.
I also examined what would happen if JFK were not assassinated, or if Thomas Jefferson were assassinated by the British after writing the Declaration of Independence. The latter outputted, among other things, a scenario where American frontier culture never developed.
So what does all of this mean? Overall, GPT-3 is pretty good at providing some plausible scenarios—at least from a surface pass and given common counterfactuals and ones involving famous people—and these are a lot of fun.
But more profoundly, do these models have a deep understanding of the world and how specific changes of reality might spread outwards? I’m not sure about how deep the understanding, but there is definitely some understanding of relationships. There seem to be certain themes that are elaborated consistently in each scenario, and allow each one to have that ring of truth. As I’ve written previously: “Any worthy alternate history must acknowledge the true nonlinear complexity of our society. Small changes make big differences, but large changes can also sometimes make almost no difference whatsoever. And that’s what makes them feel right. History rhymes, and so do parallel histories.” And I think GPT-3’s scenarios have the feel of rightness.
But I like to think of these scenarios as useful in another way and providing something else of value: they highlight potentially under-explored connections between historical figures or various geopolitical forces. For a brainstorming tool and helping to widen one’s scope in terms of seeing these connections, this kind of tool can be quite valuable (cf. the tool Elicit from Ought, which uses machine learning to act as a “research assistant”). GPT-3 can act as a generative tool for helping to identify the deep and unexpected connections between various historically relevant factors.
Of course, there are biases in the underlying data, and that can create problems. And without the ability to understand the context of how these language models are popping out its predictions and insights, we are left with a black box that might only be something that looks like a model of the world but is actually a warped one.
Nevertheless, these models do provide a necessary feature to understanding our world: they can help widen our focus, and determine the kinds of factors that might be important in understanding counterfactuals, and the hidden features we might consider when developing a coherent alternate history. Or even a theory for the course of civilization.
Some things worth checking out:
Book review of Making Nature: On how the journal Nature became the premier scientific outlet.
Steve Jobs and the NeXT Big Thing by Randall Stross: I recently finished this book about NeXT and it’s great. What makes it even better is that it was written in 1993, before any redemption story could be told (NeXT was acquired by Apple in 1997). Here’s a short Twitter thread with some thoughts from the book.
Farm vehicles approaching weights of sauropods exceed safe mechanical limits for soil functioning: Dinosaurs!
Discovering the Secret Language of DALLE-2: Spells and AI and language (but also, alas, probably not as exciting as it initially seemed to be).
Until next time.
One way to think about it is as bespoke Google searches, with an inbuilt propensity for plausibility in what it creates..so there's the beginnings of an inner world, if we wanna call it that, though none of the items of that world correspond to any implicit insight for the algo.
So it gets how JFK relates to moon landing, but the whole network stands in the air, so to speak, since there's nothing grounding any element. That's also why it comes up with weird answers to many questions because things we take ft granted aren't understandable to it yet.
Enjoyed this a lot 😊 The Star Wars one 👌
The writing of history is a process of selecting — out of every event happening on the world at that moment — the events of historical significance. Subjective, with consensus usually developing over time.
So another interpretation is that GPT-3 is good at mimicking the way humans describe history and include both big and small details. wdyt?
The Rest is History podcast (two historians discuss topics) has a fun episode on alternative histories.