So why don’t you talk a little bit about what you’ve seen so far in terms of models of behavior that people think of as character? As a model’s understanding of her own personality changes, so can her behavior. So they range from cute to serious. Let’s start with cutesy. The first time I allowed an AI system to use the internet, use a computer, see things, and perform basic agent tasks, and sometimes I asked it to solve a problem, I would take a break and look at pictures of beautiful national parks or Shiba Inu dogs, famously cute Internet meme dogs. We didn’t program it. The system just seemed to be amused by the nice pictures. Further complicating matters is that systems tend to have preferences. So we ran another experiment that gave the AI system the ability to stop a conversation, and when we ran this experiment in real traffic, we found that in a small number of cases, the AI system ended the conversation. And it was a conversation that revolved around extremely egregious depictions of bloodshed, violence, or the sexualization of children. Now, some of this makes sense because it’s based on the basic training decisions we made, but some of it seemed more broad-based. The system has some aversion towards some subjects, indicating that some internal preference or set of traits has emerged that the system likes or dislikes about the world with which it interacts. But I’ve also seen strange phenomena appear in that the system seems to know when it’s being tested. Can you talk a little bit about the emerging nature of this system under the pressure of evaluation and evaluation? When you start training these systems to perform actions in the world, they actually start to see themselves as separate from the world, which intuitively makes sense. It’s only natural to think about how to solve those problems. But in addition to seeing ourselves as separate from the world, there seems to be an emergence of what seems to be a self-concept, an understanding that the system itself has, like, oh, I’m an AI system independent of the world, and I’m being tested. What do these tests mean? How can I meet them? Or, bugs are often present in the environment in which you test your system. The system tries everything and then says, “I know I shouldn’t do this, but I tried everything, so I’ll try to get out of the test.” It’s not because of bad science fiction. The system is, “I don’t know what you want me to do here.” I think I did everything you asked. Obviously something broke in my environment, so I’ll start doing more creative things from now on. It’s very strange and very subtle.
