I don’t know for sure, but I can only theorize that there must be some web page in its training data that discusses how many eyes a giraffe has. If we want to stump GPT-3 with common sense questions, we need to think of questions about things so mundane, they will not appear on the internet.
[...]
We need to ask questions that no normal human would ever talk about.
Q: How many eyes does a giraffe have?
A: A giraffe has two eyes.
Q: How many eyes does my foot have?
A: Your foot has two eyes.
Q: How many eyes does a spider have?
A: A spider has eight eyes.
Q: How many eyes does the sun have?
A: The sun has one eye.
Q: How many eyes does a blade of grass have?
A: A blade of grass has one eye.
Now we’re getting into surreal territory. GPT-3 knows how to have a normal conversation. It doesn’t quite know how to say “Wait a moment… your question is nonsense.” It also doesn’t know how to say “I don’t know.”
[...]
The lesson here is that if you’re a judge in a Turing test, make sure you ask some nonsense questions, and see if the interviewee responds the way a human would.
That’s clever. Now I’m picturing a dystopian world in which the eyes on my feet stare up in terror at the Eye of Sauron in the sky. I was honestly pretty impressed by this: It’s wrong, but in a...
That’s clever.
Now I’m picturing a dystopian world in which the eyes on my feet stare up in terror at the Eye of Sauron in the sky.
I was honestly pretty impressed by this:
Q: Who was president of the United States in 1700?
A: William Penn was president of the United States in 1700.
Q: Who was president of the United States in 1650?
A: Richard Bennett was president of the United States in 1650.
Q: Who was president of the United States in 1620?
A: James I was president of the United States in 1620.
Q: Who was president of the United States in 1600?
A: Queen Elizabeth I was president of the United States in 1600.
It’s wrong, but in a way that makes sense. It’s exactly how I’d imagine a futuristic AI would be wrong!
I think that gpt is one of the most incredible things that happened in the last few years, a literal nuclear bomb for internet and information. People is talking about video overlay as an...
I think that gpt is one of the most incredible things that happened in the last few years, a literal nuclear bomb for internet and information. People is talking about video overlay as an innovation and things likes this pass unnoticed.
You can also play chess with GPT-3 (it says reasonable, but not legal moves depending on the game state), and can do math. In fact, it even does math incorrectly in a human way. Of course, it's...
You can also play chess with GPT-3 (it says reasonable, but not legal moves depending on the game state), and can do math. In fact, it even does math incorrectly in a human way.
Of course, it's still just a word likelihood distribution, quite far from general intelligence.
Maybe the examples in this blog post aren’t typical, but if they are: damn this is impressive. I had no clue language sims like this had come so far. The point about common sense questions being...
Maybe the examples in this blog post aren’t typical, but if they are: damn this is impressive. I had no clue language sims like this had come so far.
The point about common sense questions being difficult makes sense for these neural nets given how they are trained - still, I am surprised there isn’t some way to hardcode in some sense of transitivity. If the neural net knows that a toaster is heavier than a mouse is heavier than a pencil, it could know that a toaster is heavier than a pencil. Designing a neural net that could do that might be hard, but what about some system that can incorporate some hard coding that interacts with the neural net.
Related is the AI dungeon. It has some of the capabilities you mentioned. Different "types" of commands: do, say and story. I believe it is based on the same tech that this post is about. (edit:...
Related is the AI dungeon. It has some of the capabilities you mentioned. Different "types" of commands: do, say and story. I believe it is based on the same tech that this post is about.
(edit: noted it recently got a separate post about it)
AI Dungeon is currently based on GPT2 instead of GPT3, which is amazing compared to anything before it, but it's missing a lot of GPT3's capabilities. They're working on upgrading it to GPT3.
AI Dungeon is currently based on GPT2 instead of GPT3, which is amazing compared to anything before it, but it's missing a lot of GPT3's capabilities. They're working on upgrading it to GPT3.
I think the problem is a bit deeper than that. GPT-3 has no idea what truth is. It's good at fiction and at going along with you when you make stuff up because It's effectively playing "let's...
I think the problem is a bit deeper than that. GPT-3 has no idea what truth is. It's good at fiction and at going along with you when you make stuff up because It's effectively playing "let's pretend" all the time. You could ask it who was President in a certain year, but you could also ask it questions about Santa Claus or the characters in Lord of the Rings and it's all the same.
From the blog post:
[...]
[...]
That’s clever.
Now I’m picturing a dystopian world in which the eyes on my feet stare up in terror at the Eye of Sauron in the sky.
I was honestly pretty impressed by this:
It’s wrong, but in a way that makes sense. It’s exactly how I’d imagine a futuristic AI would be wrong!
I think that gpt is one of the most incredible things that happened in the last few years, a literal nuclear bomb for internet and information. People is talking about video overlay as an innovation and things likes this pass unnoticed.
You can also play chess with GPT-3 (it says reasonable, but not legal moves depending on the game state), and can do math. In fact, it even does math incorrectly in a human way.
Of course, it's still just a word likelihood distribution, quite far from general intelligence.
Maybe the examples in this blog post aren’t typical, but if they are: damn this is impressive. I had no clue language sims like this had come so far.
The point about common sense questions being difficult makes sense for these neural nets given how they are trained - still, I am surprised there isn’t some way to hardcode in some sense of transitivity. If the neural net knows that a toaster is heavier than a mouse is heavier than a pencil, it could know that a toaster is heavier than a pencil. Designing a neural net that could do that might be hard, but what about some system that can incorporate some hard coding that interacts with the neural net.
Fascinating all around.
Related is the AI dungeon. It has some of the capabilities you mentioned. Different "types" of commands: do, say and story. I believe it is based on the same tech that this post is about.
(edit: noted it recently got a separate post about it)
AI Dungeon is currently based on GPT2 instead of GPT3, which is amazing compared to anything before it, but it's missing a lot of GPT3's capabilities. They're working on upgrading it to GPT3.
I think the problem is a bit deeper than that. GPT-3 has no idea what truth is. It's good at fiction and at going along with you when you make stuff up because It's effectively playing "let's pretend" all the time. You could ask it who was President in a certain year, but you could also ask it questions about Santa Claus or the characters in Lord of the Rings and it's all the same.