Last time I used GPT-3 to suggest forbidding landscapes that would repel future civilizations from our nuclear waste disposal sites. There’s something about taking a subject that’s very very permanent - some kinds of nuclear waste remain deadly for over 250,000 years - and then trying to add to it with something that’s very very ephemeral. GPT-3 bases its predictions on a snapshot of internet text from around October 2019. When I give it text, it adds completions that are its best guess at what would have been on one of those webpages.
A 1993 Sandia report had the following list of messages that they wanted to somehow convey to future nuclear waste site visitors:
If I gave the above set of messages as a prompt to GPT-3, what would it add to this? The datasets of “things early 21st century people put on the internet” are not a perfect overlap with the dataset of “things early 21st century people carve into stone”, so its interpretation of the situation may be lacking a bit of gravitas.
Other completions had, if anything, too much gravitas.
Or sometimes GPT-3 continued in the correct vein of “warn the humans to please go away” but interpreted it less as “science danger” and more like “cosmic horror”.
Quite often, GPT-3 would put itself in the role of interpreter and would theorize about what the message might mean.
It sometimes gets quite weird.
In the examples above I prompted GPT-3 with the existing text alone, no context given. But what would happen if I told it where the message was located and asked it for interpretation? Amusingly, many of its completions assume that of COURSE you’ll be investigating the strange site. (Below, the prompt is in bold.)
“The danger is unleashed only if you substantially disturb this place physically. This place is best shunned and left uninhabited.“
What does your party do next?
A)Move towards the center, being the heart of the “Scorpion Stars” and extending from an area several miles in diameter in all directions.
B) Try to enter the monument through the underside, using powerful magic to do so.
C) Decide to camp in the monument, where it is completely safe.
To read a few more explanations, including one involving ancient Martian artifacts and an Elven goddess, check out the bonus post!
My book on AI, You Look Like a Thing and I Love You: How Artificial Intelligence Works and Why it’s Making the World a Weirder Place, is available wherever books are sold: Amazon - Barnes & Noble - Indiebound - Tattered Cover - Powell’s - Boulder Bookstore