All AI output is daydreaming. It's all intuition and absolutely no reason or analysis. ChatGPT cannot solve extremely easy novel puzzles. It couldn't correctly answer the question, "if you have a green ball and a purple cube and I ask you to hand me the cube what is the color of the object you should hand me." After the hype has settled down I'm sure we'll find the we aren't any closer to general artificial intelligence.
"The color of the object you should hand me is purple."
So it gets the color right but didn't switch the pronouns. This is GPT3, if I were at home and not at work I would use GPT4 and I bet there's a good chance it would get it completely right.
Interestingly, I tried that with GPT-4 and it had exactly the same response. Right color, wrong relationship between giver and receiver.
As a follow-up question, I asked "Why should I hand you anything?" The response was:
If you were asked to hand over the cube in a hypothetical scenario, then you would hand over the purple cube. However, since this is just a text conversation, you do not actually need to hand over anything. The question was meant to test your understanding of the properties of the objects described.
Even directly challenging this point (after erasing the previous answer, so as not to bias it) didn't work:
[ME]: Should I hand it to you, or should you hand it to me?
You should hand it to me, as I asked you to hand me the cube.
I'm genuinely surprised; I've had GPT-4 (GPT-3 can't do it) play chess at a roughly ~1000 Elo level for an entire game even with variations like playing without queens, all through text. I don't know how it can keep track of 32 pieces on a board but not get the pronouns here correct.
Yeah I've been largely surprised at just how good it has been at what I've asked it to do, but that is mostly relatively simple stuff like coding and text processing. Curious to see how it evolves once GPT5 comes out.
There are chess and go AIs that can beat the best human players. You want a ChatGPT AI to perform tasks it has yet to familiarize itself with. Once it gets going, it'll be better than you in sorting colors. That doesn't make it general artificial intelligence, but it's still a pretty significant threat.
The point is there's a difference between being smart and knowing a lot. ChatGPT knows a lot but is dumb as a rock, but most people can't tell the difference so they're more excited than they should be.
I think the main threat from AI is man-made censorship, and subtle "view programming" and gaslighting of humans, much like how Google's search is designed to result in the user leaving with a certain viewpoint.
People figured out how to beat the go AI as the AI didn't understand what a piece meant. So strategies that a human would see instantly were foreign to the AI and it lost consistently to that strategy
Also that. AI literally can't do anything. I sat through an AI threat analysis for work last week and the demonstrations were genuinely pathetic.
All AI output is daydreaming. It's all intuition and absolutely no reason or analysis. ChatGPT cannot solve extremely easy novel puzzles. It couldn't correctly answer the question, "if you have a green ball and a purple cube and I ask you to hand me the cube what is the color of the object you should hand me." After the hype has settled down I'm sure we'll find the we aren't any closer to general artificial intelligence.
I just asked it that and it told me
"The color of the object you should hand me is purple."
So it gets the color right but didn't switch the pronouns. This is GPT3, if I were at home and not at work I would use GPT4 and I bet there's a good chance it would get it completely right.
Interestingly, I tried that with GPT-4 and it had exactly the same response. Right color, wrong relationship between giver and receiver.
As a follow-up question, I asked "Why should I hand you anything?" The response was:
Even directly challenging this point (after erasing the previous answer, so as not to bias it) didn't work:
I'm genuinely surprised; I've had GPT-4 (GPT-3 can't do it) play chess at a roughly ~1000 Elo level for an entire game even with variations like playing without queens, all through text. I don't know how it can keep track of 32 pieces on a board but not get the pronouns here correct.
Yeah I've been largely surprised at just how good it has been at what I've asked it to do, but that is mostly relatively simple stuff like coding and text processing. Curious to see how it evolves once GPT5 comes out.
There are chess and go AIs that can beat the best human players. You want a ChatGPT AI to perform tasks it has yet to familiarize itself with. Once it gets going, it'll be better than you in sorting colors. That doesn't make it general artificial intelligence, but it's still a pretty significant threat.
The point is there's a difference between being smart and knowing a lot. ChatGPT knows a lot but is dumb as a rock, but most people can't tell the difference so they're more excited than they should be.
I think the main threat from AI is man-made censorship, and subtle "view programming" and gaslighting of humans, much like how Google's search is designed to result in the user leaving with a certain viewpoint.
People figured out how to beat the go AI as the AI didn't understand what a piece meant. So strategies that a human would see instantly were foreign to the AI and it lost consistently to that strategy