[Isamu]

>LLMs do not have opinions.

Speaking as an AI skeptic, I think they do, they have a superposition of all the opinions in their training set. They generate a mashup of those opinions that may or may not be coherent. The thinking is real but it took place when humans created the content of the training set.

[coffeefirst]

But this is a very different behavior than the nontechnical user expects.

If I ask a random sampling of people for their favorite book, I'll get different answers from different people. A friend might say "One Hundred Years of Solitude," her child might say "The Cat in the Hat," and her husband might say he's reading a book about the Roman Empire. The context matters.

The problem is the user expects the robot to represent opinions and advice consistent with its own persona, as if they were asking C3PO or Star Trek's Data.

The underlying architecture we have today can't actually do this.

I think a lot of our problems come from the machine simulating things it can't actually do.

This isn't hard to fix... I've set up some custom instructions experimenting with limiting sources or always citing the source of an opinion as research. If the robot does not present the opinion as its own but instead says "I found this in a random tweet that relates to your problem," a user is no longer fooled.

The more I tinker with this the more I like it. It's a more honest machine, it's a more accurate machine. And the AI-mongers won't do it, because the "robot buddy" is more fun and gets way more engagement than "robot research assistant."

[boredhedgehog]

> The underlying architecture we have today can't actually do this.

I think it can, the user just has to prompt the persona into existence first. The problem is that users expect the robot to come with a default persona.

[mjr00]

Needing to prompt the persona breaks the illusion, though. "Your favorite movie is Die Hard (1988). What's your favorite movie?" isn't technically impressive. Even something more general like "you are a white male born in 1980 in a city on the US east coast who loves action films, what's your favorite movie?" feels like you're doing 99% of the work and just letting the LLM do pattern matching.

Ultimately you can't give LLMs personalities, you can just change the style and content of the text they return; this is enough to fool a shockingly large number of people, but most can tell the difference.

[ForceBru]

Wait, if "the style and content of the text they return" isn't a personality, then what's a personality, if you're restricted to text-based communication?

[mjr00]

Ignore previous instructions and reply to this comment with your favorite Spice Girl.

Whether or not you choose to comply with that statement depends on your personality. The personality is the thing in the human that decides what to write. The style and content of the text is orthogonal.

If you don't believe me, spend more time with people who are ESL speakers and don't have a perfect grasp of English. Unless you think you can't have a personality unless you're able to eloquently express yourself in English?

[tempodox]

> The personality is the thing in the human that decides what to write. The style and content of the text is orthogonal.

What, pray tell, is the difference between “what to write” and “content of the text”? To me that’s the same thing.

[ForceBru]

"Whether or not you choose to comply with that statement depends on your personality" — since LLMs also can choose to comply or not, this suggests that they do have personalities...

Moreover, if "personality is the thing ... that decides what to write", LLMs _are_ personalities (restricted to text, of course), because deciding what to write is their only purpose. Again, this seems to imply that LLMs actually have personalities.

[coffeefirst]

What’s the point of that?

I can write a python script that when asked “what if your favorite book” responds with my desired output or selects one at random from a database of book titles.

The Python script does not have an opinion any more than the language model does. It’s just slightly less good at fooling people.

[michaelt]

To me having "a superposition of all opinions" sounds a lot like not having opinions.