Scott Alexander in Astral Codex Ten:
Janus (pseudonym by request) works at AI alignment startup Conjecture. Their hobby, which is suspiciously similar to their work, is getting GPT-3 to do interesting things.
For example, with the right prompts, you can get stories where the characters become gradually more aware that they are characters being written by some sort of fiction engine, speculate on what’s going on, and sometimes even make pretty good guesses about the nature of GPT-3 itself.
Janus says this happens most often when GPT makes a mistake – for example, writing a story set in the Victorian era, then having a character take out her cell phone. Then when it tries to predict the next part – when it’s looking at the text as if a human wrote it, and trying to determine why a human would have written a story about the Victorian era where characters have cell phones – it guesses that maybe it’s some kind of odd sci-fi/fantasy dream sequence or simulation or something. So the characters start talking about the inconsistencies in their world and whether it might be a dream or a simulation. Each step of this process is predictable and non-spooky, but the end result is pretty weird.
Can the characters work out that they are in GPT-3, specifically?