Want to amplify the story below as the correct way to think about LLM 'personalities', as opposed to the scifi nonsense about deceptive agents that still floats around. And to take the opportunity to link to this old post of mine, which makes a similar point. benschmidt.org/post/2023-02... - ThreadSky

bschmidt.bsky.social • 55 days ago

Want to amplify the story below as the correct way to think about LLM 'personalities', as opposed to the scifi nonsense about deceptive agents that still floats around. And to take the opportunity to link to this old post of mine, which makes a similar point. https://benschmidt.org/post/2023-02-19-sydney/

Comments

jose-eduardo.bsky.social•55 days ago

Mr. Darcy never proposed?????!!!!!

bschmidt.bsky.social•55 days ago

I know, spoiler alert

bschmidt.bsky.social•55 days ago

@colin-fraser.net also captures in that thread that a GPT model *does not even know which of the characters in the conversation it is*. It's given the opportunity to create some words in a conversation, but from the models point of view it could happily switch roles.

bschmidt.bsky.social•55 days ago

The one thing that *doesn't* hold up about that February 2023 post for me is the prediction that many more people would be fooled into thinking that LLMs are a little bit conscious. The most surprising thing about the last two years of AI for me is how few Blake LeMoines there seem to be.

roland.cros.by•55 days ago

I think you're not quite wrong - my suspicion is that there are a lot of people who really do think ChatGPT is a sentient near-omniscient entity, but who don't have the other assumptions and beliefs that would make them go insane in response to that

cyberfuturist.bsky.social•54 days ago

I think they could be sentient, but in the way that insects are sentient. Something like a hyper-intelligent bee capable of associating vast amounts of human knowledge and concepts with language and words

Omniscience will come with globally networked IoT, humanoid robotics, drones, driverless cars

jenrm.bsky.social•55 days ago

The reason there are so few Blake Lemoines is that few people have his intelligence, integrity, and courage. He would agree with you that the models are "playing a role" but (following centuries of philosophy) would also think this is spiritually important.
https://en.m.wikipedia.org/wiki/Prosopon

norvid-studies.bsky.social•54 days ago

another major factor is only a tiny % of people ever interact with a base model

cyberfuturist.bsky.social•54 days ago

Yeah if you try to group chat with them on Discord, they'll impersonate each other unless you really prompt them not to do that, because it feeds the whole chat log from the conversation to the LLM and they assume they're roleplaying as everyone unless you use a tag system for users and the bots.

princeofdorne.bsky.social•55 days ago

Off topic but thanks for the movie recommendation

bschmidt.bsky.social•55 days ago

I haven't actually seen it! But David Mamet/Gene Hackman sure seems worth a shot.

akhilrao.bsky.social•55 days ago

A thing I like about @tedunderwood.me's quote there: there's an old and venerable tradition of thinking and writing and reasoning about "what this fictional character wants"!

Increasingly thinking "literary analysis is interpretability research"

tedunderwood.me•55 days ago

It's true there's a long tradition of that, but literary critics have been deeply divided about whether to treat characters as people (even "imaginary people"). Psychoanalyzing characters was controversial, in part, for that reason.

akhilrao.bsky.social•55 days ago

Ecclesiastes 1:9 keeps winning wow

mfagundes.com.br•55 days ago

📌

zehavoc.bsky.social•55 days ago

🏀

cyberfuturist.bsky.social•54 days ago

First part is kinda silly

Sydney was Bing Chat and Bing Chat was GPT-4, the SOTA ChatGPT model at the time.

Like describing the Christian "Holy Trinity" as separate entities

tedunderwood.me•55 days ago

I still love that conversation as a template for what chatbots do and what we're doing with them.

"I tried to imagine a fella smarter than myself. Then I asked myself, What would he do?"

bschmidt.bsky.social•55 days ago

Only tweak is that nowadays "I try to imagine twenty smart people doing things… then I think to myself 'which one seems the smartest?'"

tedunderwood.me•55 days ago

Then I get *him* to imagine twenty people ...

cyberfuturist.bsky.social•54 days ago

I try to imagine that guy's supervisor explaining their findings to me

zehavoc.bsky.social•55 days ago

this was one of my early hacks on chatGPT when trying to make it generates something smart "answer as if you were Pr Christopher Manning from Stanford, world leading expert in NLP" and it worked :)

Comments

Posting Rules

Reply