Answered raw tape
D 5 · C 5 · P 5 · Cm 4 4.85
Q a recent example of this is ChatGPT, which is throwing out a lot of correct answers, but also a lot of incorrect answers, and humans are just kind of taking it at face value. So maybe we don't even need human-like bots to trick humans. It seems like less evolved bots are already doing that. So what do you think about that? Have we already kind of crossed that chasm?
A Oh, absolutely. Um, and I think the question here is, like, at what point do you cross the chasm in whatever domain to deceive humans? So, for example, it's like, you used to read the words, like, don't believe everything you read on the internet. Now, I think this applies to AI. Like, don't believe everything you read from AI. Um, the Turing test, as you say, is like a flawed metric, and the reason for this is in part because It is too low of a bar. It is a binary bar, so it's either intelligent or not intelligent, and it tests for deception, not real understanding. What these AIs do is they Try to understand the underlying rules that generate the dataset that they're trained on, um, and then generate an approximation of what they believe is the next token. But it doesn't mean that it's right. Like in the case of chat GPT, it communicates with a lot of confidence. A great example of this might be say, asking chat GPT to multiply a 153 times, 257. Like it might get an answer that's close to correct and it communicates it like it is correct, but it's not correct. And this is a deterministic domain where you know what the right answer is. There are a lot of domains where there isn't actually a right answer. The right answer is much more amorphous. Say like you ask about China's, what should China's AI strategy be? Like this is a very complex question and it is dictated by say lik…
AI assessment note: “Oh, absolutely.”
Answered raw tape
D 5 · C 5 · P 5 · Cm 4 4.85
Q a recent example of this is ChatGPT, which is throwing out a lot of correct answers, but also a lot of incorrect answers, and humans are just kind of taking it at face value. So maybe we don't even need human-like bots to trick humans. It seems like less evolved bots are already doing that. So what do you think about that? Have we already kind of crossed that chasm?
A Oh, absolutely. Um, and I think the question here is, like, at what point do you cross the chasm in whatever domain to deceive humans? So, for example, it's like, you used to read the words, like, don't believe everything you read on the internet. Now, I think this applies to AI. Like, don't believe everything you read from AI. Um, the Turing test, as you say, is like a flawed metric, and the reason for this is in part because It is too low of a bar. It is a binary bar, so it's either intelligent or not intelligent, and it tests for deception, not real understanding. What these AIs do is they Try to understand the underlying rules that generate the dataset that they're trained on, um, and then generate an approximation of what they believe is the next token. But it doesn't mean that it's right. Like in the case of chat GPT, it communicates with a lot of confidence. A great example of this might be say, asking chat GPT to multiply a 153 times, 257. Like it might get an answer that's close to correct and it communicates it like it is correct, but it's not correct. And this is a deterministic domain where you know what the right answer is. There are a lot of domains where there isn't actually a right answer. The right answer is much more amorphous. Say like you ask about China's, what should China's AI strategy be? Like this is a very complex question and it is dictated by say lik…
AI assessment note: “Oh, absolutely. Um, and I think the question here is, like”
