4 Comments
User's avatar
Rudy's avatar

Ted Chiang's article is incredible. Genuinely one of the best things I've read in a while. He is so spot on with the concept that you are technically co-authoring a document with an LLM, and any user/agent conversation in an instruct-tuned way could very easily be formatted as a conversation between Julius Caesar and Genghis Khan.

I wonder why if these agents are able to assume the form of multiple different people within the same context window, like a User/Julius Caesar/Genghis Khan dynamic where you prompt the agent and it responds with a dialogue of Julius Caesar and Genghis Khan responding to your prompt

Adam Keys's avatar

> The New Yorker archive and read some of the Greatest Hits of the last 100 years

I want to go deeper! Is there a link to this? Or do you just flip through your favorite writers?

Thorsten Ball's avatar

https://www.newyorker.com/100

Scroll down there to "Dive In".

Aaron Tagliaboschi's avatar

> Conscious or not, does the distinction matter on a practical level?

Something I've been thinking about a lot is that how people are trying to treat LLMs and AI as independent entities as a way to avoid culpability or accountability when they are used irresponsibly and something goes wrong, much in the way that so very high ups in a corporation will avoid accountability by shifting the blame and punishment onto workers, and this question sort of plays into that in a really sneaky way.

In a sense, I feel like there are people who are trying to make this argument because they want to be able to exploit it like they exploit workers.