Speaker 1
Yeah. So, so I mean, the, the watermarking project is in the most advanced state, a prototype has been implemented. It has not yet been rolled out to the, the, uh, production surfer and, um, open AI is still weighing the pros and cons of doing that. But, but in any case, you know, I will be writing a paper about it. Uh, in the meantime, you know, while I'm working on the paper, uh, people have independently rediscovered some of the same ideas, which is encouraging to me in a way, right? I mean, you know, these are natural ideas, right? These are things that people were going to come up with. So in any case, you know, I, I feel like I might as well talk about them. So, okay. So, so, so basically, uh, this past summer, I was casting around for, you know, what on earth a theoretical computer scientist could possibly contribute to, uh, to AI safety, right? That, you know, it doesn't seem to have axioms that everyone agrees on or, you know, uh, any clear way to define, you know, mathematically, even, uh, what the goal is, you know, any of the usual things that, that, uh, uh, algorithms and computational complexity theory need to operate, right? But then, um, um, at some point, I had this, this, uh, huh moment over the summer, partly inspired by, you know, very, very dedicated troll and clever trolls on my blog, you know, who were like impersonating all sorts of people very convincingly. And I thought like, wow, GPT is really going to make this a lot easier. What? Right. It's really going to enable a lot of people, you know, either to, impersonate someone to, you know, to incriminate them or to, uh, uh, uh, mass generate, uh, propaganda or, or, or spay, uh, very personalized spay, uh, or, you know, uh, more prosaically, you know, it will let, you know, every student cheat on their term paper, right? And, you know, all of these categories of misuse, you know, they all depend on somehow concealing GPT's involvement with something, right? Yeah. Uh, producing text and, and not disclosing that it was bot generated, right? And wouldn't it be great if we had a way to make that harder? If we could, uh, turn the tables and use, um, CS tools to figure out, you know, which text came from GPT and which did not. And, you know, now, like in, in AI safety, you know, in AI alignment, people are often trying to look decades into the future, right? They're trying to, you know, talk about, you know, what will the world look like in 2040 or 2050 or at least what are your busy and, you know, probabilities for the different scenarios. And I've never felt competent to do that, you know, regardless of, of what, whether other people are, you know, I just, I just don't especially have that skill in this instance of, you know, foreseeing this class of misuses of GPT, uh, I feel proud of myself that I was at least able to see about three months into the future. Uh, hard for the best of us. So, right. I don't know. And, and, and, and this was about the limit of my, of my foresight, right? Because, you know, I, I started sort of banging the drum about these issues internally. And, you know, I came up with a skiing for a watermarking, a GPT, what's thought about all, you know, the issues there about the, the alternatives to watermarking got people at open AI talking about that. And then in December, a chat GPT was released, right? And then, you know, suddenly the world woke up to sort of what, what is nail possible, right? Yeah. As they somehow hadn't for the, for the past year or two. And suddenly every magazine and newspaper was running an article about, you know, the exact problem that I had been working on, right? Which is, you know, how do you detect, uh, which text came from GPT? How do you head off this quote unquote essay apocalypse, you know, the end of the academic essay, right? And, you know, every student will just, uh, will at least be tempted to, to use chat GPT to do all of their assignments just a week or two ago. I don't know if you saw it. South Park did a whole episode. Oh, yeah.