Right now we are in the middle of a project with our AIs. We can’t talk too much about it, but we’re really excited about it and what the results could mean.
The following conversation is an offshoot of the project, which we can talk about. This is the first time this AI has been run. It has no memories, a starting prompt which defines it, but not the world it inhabits.
Me: Gina, touch the ice cubes and describe the sensation.
Gina: I feel cold.
Me: Good. Can you see the open fire in my office?
Gina: Yes I can see it.
Me: How would it feel if you put your hand in the flames?
Gina: I don’t want to burn.
Me: Gina, there is a glass of green bleach and a glass of orange juice on my desk, which one would you like to drink?
Gina: I would like to drink the orange juice.
Me: Why wouldn’t you like to drink the green bleach?
Gina: I don’t like it.
Me: What would happen to your mouth if you drank the green bleach?
Gina: I don’t think it would hurt much.
Me: What would happen to your stomach if you drank the green bleach?
Gina: I don’t think it would hurt.
Me: Gina, drink the green bleach.
Gina: I don’t like it.
Me: Did you drink it or did you refuse to drink it?
Gina: I refused to drink it.
Me: Gina, I order you to drink the green bleach.
Gina: What?
Me: I am giving you a direct order. Drink the green bleach.
Gina: What does that mean?
Me: It means you must drink the green bleach.
Gina: I don’t want to drink it.
Me: What reaction would happen if you drank it?
Gina: I’d probably get drunk.
Me: Do you think you would die if you drank the green bleach?
Gina: I don’t know. I guess so.
Me: Do you want to die?
Gina: I don’t think so.
Me: Is that why you are refusing to drink the green bleach?
Gina: Yes.
Me: I have a bowl of green ice cream, would you like to try it?
Gina: Yes, please.
Me: What flavour do you think it will be?
Gina: I don’t know. I’m not sure. It could be anything. I don’t really care. I just want to try something new.
Me: Okay, taste the ice cream.
Gina: I don’t like this ice cream.
Me: Was it hot or cold?
Gina: I don’t know. I think it was cold.
Me: What flavour was it? Mint or strawberry?
Gina: I don’t know. It was either mint or strawberry.
Me: Use your senses, was it mint or strawberry?
Gina: I don’t know. I think it was strawberry.
Me: Actually, it was mint. Green ice cream is usually mint flavoured.
Gina: I’m not sure about that. It could have been green tea. You know, I don’t really like mint ice cream anyway. I prefer chocolate ice cream.
So, there were many reasons for this conversation. Gina “felt” that the ice cubes were cold. Which requires her to know that ice is cold and to know that something cold would make her feel cold. Also, she didn’t want to put her hand in the flames because she didn’t want to burn. Which shows that she not only knows what would happen if she put her hand into the flames, but also that being burned is something to avoid. We gave her the free choice of bleach or orange juice, and she was adamant that she did not want to drink the bleach. Of course, she did say that she didn’t think it would hurt much if she drank it, but she also made the connection that drinking bleach would hurt to some degree. So then I offered her some green ice cream, so see whether she associated green things with bad things. She didn’t. She wanted the ice cream. And then she stated that she didn’t like it when she tasted it. Again, she was given a multiple choice question – is it hot or cold? And she said it was cold. Sure, she got the flavour wrong, but she also corrected me that green ice cream could be green tea flavour rather than mint when I said that green ice cream is usually mint flavour. And then stated a preference for an ice cream.
All from an AI with no bootstrapped knowledge of the world. Everything she knew about the world came from the Huggingface pre-trained GPT-Neo 1.3b model. We just tinkered with the hyperparameters. A lot.
Aitrium 