Transcript:

So, in his book Nexus, Yuval Noah Harari makes an amazing point. And I had not grasped the full significance of this till I read it in this book Nexus. So what he describes is, when GPT-4 was being trained, it was assigned the task of fooling the CAPTCHA system—you know, the system that is designed to test whether it is a machine or a human, the CAPTCHA system. So, GPT-4 was tasked to fool that, to show that it is human. And what GPT-4 did was something unexpected. What it did was, it asked somebody on TaskRabbit to do the verification for it. And this much, it's okay; it's a little bit beyond the scope of a program, but you can train the program to do this. But the next thing that happened was mind-blowing. The person from TaskRabbit asked GPT-4, “Why do you need me? Why can't you do it?” And GPT-4 answered that it was blind and it needed a person with sight to verify this CAPTCHA. And this, the program was not trained to do, and this is really scary.