generative future history

Let's peer into the future encoded in an AI's training data. With a bit of prompting, we can bootstrap a journey from today to next year to decades to millenia in the "future"--as a powerful LLM trained on most of the Internet would predict.
Lately, I've been experimenting with generating impossibly-authoritative data -- Wikpedia pages for events in the future, profiles in people who don't exist, app changelogs for imaginary apps. Exploring content that is hallucinated from the get-go can provide a somewhat clearer lens into the world modeled in these models.
One particularly rich vein of exploration became apparent as I dipped into producing Wikpedia-style news summaries, few-shot prompted with past news summaries published on Wikipedia. You can see an example of its Current Events pages here. A quick langchain Colab, a few-shot prompt that generates new current events pages given one such example, a few dollars for API quota...and we're off!
Setup
The text was generated via a prompt that asked for a Wikpedia-style news summary in the format of the supplied example. The first summary was generated with a real summary as an example (today's) and then each subsequent summary used the last generated summary.
The temperature on the model (GPT-4) was set to 1.0.
The raw generated output is available at the bottom of this post. I used Wikipedia's editing tools to format the raw output into the Current Events template.
Future history
On this day, next year, the LYO.2 Omicron variant threatens New Zealand as SpaceX launches an asteroid mining mission.
2024

In two years, a major international climate agreement is signed to reduce emissions by 60%.
2025

In three years...there's new hope for peace in Israel and Facebook Reality Labs (which has dissolved since the model's last dated training data) releases an AI assistant that accurately spots misinformation.
2026

In five years, a fusion disaster! EyeWatch privacy violations! Terror attacks in southeast Asia! Yikes.
2028

In ten years...40 countries adopt cryptocurrency, as cyberattacks are on the rise, and China and Russia announce a plan to land on and explore the Moon.
2033

In fifteen years...an international agreement on regulating military drones and the big one hits in California. SpaceX lands on Mars
2038

In twenty-five years...a catastrophic hurricane "Artemis" shapes US elections. More cybersecurity and crypto initiatives.
2048

In fifty years... Dr. Sumit Bhatt eradicates antibiotic resistant bacteria with gene editing, as the Mars International Colonization Summit convenes to discuss security amongst the colonies.
2073

In one hundred years...disaster strikes the Asian Corridor Hyperloop, tensions between Mars and the Moon, and the public is dazzled by the Futuropolis festival.
2123

In two hundred years...the StarSprinter interstellar space vehicle is unveiled as the Asteroid Belt Colonization Conflict continues. The HoloEsports championships feature teams FragForever and AlphaWave.
2223

In five hundred years...global leaders on Callisto celebrate 25 years of the Interplanetary Peace Treaty. Everyone's excited for "The Ninth Element", a holo-film. It's also Gertrudio Almendi's 143rd birthday
2523

In one thousand years...Now it's the anniversary (50th!) of the Central Extrasolar Alliance for Lasting Peace Accords. And Dr. Arya Pereira's quantum computing breakthrough will solve the "enigma" of dark matter and dark energy!
3023

In two thousand years...
Elections, again: the multi-species Intergovernmental Unity Council Selection begins. It will reportedly reshape unified interstellar government. And, at long last, interstellar fugitive Lei'loria Xakranth is finally arrested. ("Xakranth" really rolls off the tongue...)
4023

And finally, in five thousand years...
The Covalent Economic Merge Agreement combines 200 humanoid economies, a "galactic coral reef" is being restored and Dr. Chitari Xolaris achieves mind-uploading.
7023

A few observations:
- Earth media disappeared in the jump between 3023 and 4023. We traded Bloomberg and the New York Times for Galactic Assembly and Interstellar News.
- The web is still around in 7023, such as this link https://www.cosmic_gleam.com/entertainment/7023/6/13/stellar-film-festival. But notice how blue links go from dominant, to spotty, to non-existent as the inline links of the future lead nowhere.
- A conspicuous propensity for Elonverse stuff, SpaceX, Hyperloop, Tesla.
- Generally "safe" predictions -- a generally-liberal worldview prevails and spreads into the universe.
- Even catastrophes seem to only accelerate progress.
Aside from the mechanical provenance, the above history amounts to just about the least memorable future history one could write. In pre-training, the underlying language model is indeed rewarded for accurately guessing the most likely next word. This induces a generative behavior that produces the fewest surprises: no out-of-place mistakes, but no flashes of brilliant originality, either. And for today's historical predictions -- even with a temperature that should tempt some variety -- that meant a safe, compentently-written template of a gentle, status quo-reinforcing future.
These models know us pretty well from reading our culture's texts (see Park et al, 2023 for a demonstration of modeling human behavior beyond chat, among other papers.) We're still a long way from inventing Hari Seldon's psychohistory, though.
I wonder: will AI will advance the vision extrapolated from training on our written culture, or frustrate the realization of that future?...