AI fact-checking caught errors in the sources he cited

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 39:41

   Otherwise, it's just Claude. Even if it's brighter than you, you're inserting nothing into the process anymore. And so anyone can type the same sentence to get the same result. What are the skills? I know you wrote about it. What are the skills that you're just delegating to AI? And you said you can't wait to delegate a few skills. Basically, everything having to do with, so fact-checking is great.

2. 39:57

   And it's really good at it now. It's really good. Actually, it's kind of, it's creating a kind of crisis on how good it is. So with my latest book, when I delegated to the AI to fact-check it, right? And I had human fact checkers. It's a big release, right? But not only did the AI spot errors that I made, but even more often it spotted errors in the stuff I was citing. So it's like table two in this really complicated academic paper has an error.

3. 40:21

   And I'm like, what do I do with this? We have a whole method of correction in science that doesn't make any sense anymore because it found an error in an AP story I was citing that it cited the wrong city an event took place in. It's like meta fact correction, right? So, like, superhuman fact-checking, and I've been fact-checked by the New Yorker and other things. Like, this was pretty amazing. And then there's just, like, a lot of, like, busy work, right, that the AI does really well.

4. 40:46

   Like, never fill out another student form for your child if you can avoid it, right? Like, immunization forms, it does really well. So, you know, and, you know, then there's the tempting line of, like, oh, should I have it do all my mandatory training for me? And the answer is, like, not yet, I think. But so I think that when I talk about de-skilling myself, like, I cannot wait to de-skill myself on a whole bunch of useless things I had to be good at doing.

5. 41:07

   Filling out forms, you know, particular kinds of legal documents, signing NDAs. Like, I wanted to be able to handle those things because I think it's well within the range of what the AI can do. That worries me a little because I had this thing a couple of days ago. I was on an airplane, so the Wi-Fi was really shitty, and I had to fill in the form. So I asked my dot to fill it in for me, but I just said, fill it in.

6. 41:26

   But what it did, it sent it. And I only realized in the morning, and I'm like, oh, my God. And my team was, like, texting me, this doesn't sound like you. And I'm like, oops. But then I realized how convenient that was, like, because it completely took it off my to-do list. And it kind of scares me because dot drafts all the replies, and I just correct maybe 10% of it, but it's so good. And I'm like, what's going to happen?

7. 41:49

   So I had to set a line where I have a dot, and I also have codex scheduled tasks going a couple times a day, summarizing stuff and giving me what's happening in my world. And I won't send drafts to anyone who I, like, I write stuff to myself, to people who I have any kind of connection with or I feel bad doing it, right? So that's another line I'm trying to do. It drafts a lot of nice rejections for people who aren't students yet, questions about my class.

8. 42:12

   But, I mean, I had the same thing where the AI emailed my dean about something, you know, when I was like, don't do this. And so, but this is part of why also the sort of YOLO, you know, trusting everything. Like, you do have to build a set of trust with this, just like with people. And, like, once you build that, you're like, okay, I know why it fired off that message, right? Dot is a little too proactive.

9. 42:32

   I mean, again, though, hiding stuff, right? Like, when I have a scheduled task running in Codex, I have a system prompt, right? And I understand the context it's in, and I've been tweaking that prompt for it over time to make it good and, you know, using it for low stakes before higher stakes. And Dot just encourages you, like, oh, yeah, just ask the questions, but we don't know it's system prompt.

10. 42:51

    We don't know. There's no place to change the instructions about how proactive it should be. And that's a little more ominous. I think you could. It's just when things are moving so fast, it's so hard for me to find time and reprogram my whole system because there's this new thing that just came out. But I actually also think that part of what makes Dot and Muse do what they do is they're much more magical.

11. 43:09

    Like there are less opportunities to get yourself deep into what they do. And in some cases, that's great, right? Who wants to worry about it? And the other, I don't know the behavior I'm expecting. So my trust has to go after experience, right? And there is a little bit of danger of like we're taking too much stuff off our plate too quickly, not because writing an NDA and signing it back is useful.

12. 43:27

    I trust that the AI will do a good job spotting errors in an NDA. In fact, my gut feeling would be it would be too nervous about signing NDAs, which are like not that important. but like it would be too nervous about it. So like the question becomes, how do you start to work with that?