(claps hands) - My name is Ellie Pavlick, and this is my take on artificial intelligence.

So AI is a system that's not human, but otherwise has human-like intelligence.

(soft music) I'm a professor at Brown University, and I study artificial intelligence, particularly language.

Starting around the 1950s, there was work on AI that is what we would call rule-based.

- [Announcer 1] Rules for decision making stored in the machine's memory.

- This would be things like playing chess.

We write down kind of rules that a computer could follow to be very good at chess.

- [Announcer 2] The sensory aid matches your skill.

- Then there was this shift to what we would call statistical or machine learning based approaches.

A really useful example to have in your head is something like a spam filter.

- [Automated Voice] You've got mail.

- The designer of the system would come up with the features of an email that might be useful signals of spam.

One feature that says is the subject line in all capitals?

Is there an image attached?

And then the model would learn statistically, given that there's capitalized subject and an image in the email, there's an 80% chance of it being spam.

The era we're in now is a continuation of this statistical approach, but what we would call the neural network based approach.

And that way of doing things, they just throw the whole text of the email to the system, and I say these ones are spam, these ones are not, and then the system kind of decides what features are useful on its own.

And Chat GPT is one of these neural network based systems.

Chat GPT is what we would call a large language model.

They're big statistical systems that are just learning to generation language one word at a time.

The president said on Monday that, you know, da-da-da.

And after going through a very, very, very large amount of text on the internet, they get very good at this, and then they can generate really good sounding texts.

Until recently, the only time you could have a chat conversation was when there was a human on the other side, and now all of a sudden, you're talking to a thing that's not a human with language.

It's really hard to turn off part of your brain that feels like it's a human.

- Men are all alike.

- In what way?

- This isn't the first time this has happened.

There was a very famous system called ELIZA, in these early days of rule-based AI.

- [Announcer 3] Does it understand what it's doing in the sense that we do?

It's easy to leap to false conclusions, as Professor Weizenbaum discovered when he created ELIZA.

- It was an AI therapist, or like a computer therapist, and it would talk over text.

- [Patient] He says I'm depressed much the time.

- [ELIZA] I'm sorry to hear that you're depressed.

- People got very attached to this system, even knowing that it was this rule-based system, they felt like it understand them, and it cared about them.

- [Announcer 3] Weizenbaum's secretary fell under the spell of the machine.

(keys clacking) - After two or three interchanges with the machine, she turned to me, and she said, "Would you mind leaving the room, please?"

- We have trouble sometimes interacting with a system that's doing human-like things and not feeling like it must be human-like underneath.

Something people always wanna ask, is Chat GPT intelligent?

Does Chat GPT understand language?

I don't think anyone can honestly give you a scientifically valid answer.

So if you think about a rule-based system, there is no uncertainty, no ambiguity about how the system is working.

So if I write down a set of rules for how to play chess, I know that the system is gonna follow the rules in the order I told it, exactly when I told it.

But the way that the current AI systems work is not like that.

There's this output of this large statistical learning process, and so when they do something weird or bad, you might know that the problem exists and not know how to fix it.

(soft music) I always like to emphasize how little we understand these systems.

We don't actually know what is happening inside, so it is very highly debated how much we should be able to attribute intelligence to these systems.

(bright music) (keys clacking) (claps hands) My name is Ellie Pavlick, and this was my take on artificial intelligence.