Post

I agree and it always makes me think of how people interact with the ship’s computer in Star Trek. It never uses first person pronouns. Instead of “I can’t do that,” it says “unable to comply.” Instead of “I don’t know,” it says simply “unknown.”

The ship’s computer in Star Trek is a much better comparison to LLMs than Data from TNG or the Doctor from Voyager. The makers of LLMs desperately want us to think of those characters because that’s more exciting, but their technology is closer to the non-sentient computer that’s just an interface.

In fact, seen in this light, it’s clear that the impersonal and non chatty tone of the ship computer is a better design for the task it was set up to do. How insufferable would it be if it was constantly responding to users with “That’s a great question. I can help you with that.”

The tone of LLMs only makes sense when you consider that they are intentionally set up to fool you into thinking there is a mind there. (And i guess also when you consider that they are designed to spend as many tokens as possible.)

they are designed to spend as few tokens as possible, though? frontier labs are benchmarking their models on how cheaply they can complete tasks

"They're designed to spend as few tokens as possible" like social media is designed for you to use it as rarely and shortly as possible. (Like maybe they're being optimised to provide responses to prompts on the shorter side of "tokens", but overall, maxing engagement is what they're going for).

those are two entirely separate things

Yes, but the two things mentioned here - getting people to think its a person-like entity responding & maximising engagement, ie "token spend" - are entirely linked. (addmittedly OP may have meant the latter a little differently - the overall point stands though). bsky.app/profile/yanb...

I guess I should have been more precise. I meant that LLMs are designed to output a lot of text. They could be adjusted to be less verbose and the fact that they are not indicates that those who design the systems want them to output a lot of words. Everyone is free to speculate as to why that is.

i guess i am trying to understand the incentives because more tokens = more cost for the labs. do you have a source for this?

Have you ever spoken to a chatbot? You ask a simple question and it says: "Great question! I can definitely help you with that." Those words are 100% superfluous. It could just skip ahead to the answer, but it doesn't. This is self-evident.