Ezra Klein interview on Artificial Intelligence
Last week, Ezra Klein interviewed Kelsey Piper, a senior writer at Vox who has been especially active covering A.I.: “A.I. is About to Get Much Weirder. Here’s What to Watch For.” The expansion and improvement of A.I. models just since we talked about the subject earlier this semester has been remarkable and unsettling. Piper explains at one point, “I think that a lot of people anticipate we are not very far from systems that can do any job you can do remotely, anything you can do on your computer, so coding, like secretarial work, receptionist work, journalism, everything like that. I think that some of those same people would say we can invent better scientists and do science a lot faster.” (Though it’s not perfectly clear within the context of the interview, “not very far from” here seems to mean roughly 5 to 20 years away.) When we discussed ChatGPT earlier this semester, version 3 was available to the public and it scored in the 10th percentile on the Bar exam. Version 4 has since been released and it scores in the 80th percentile on the Bar exam. There is a ton of work being done on training A.I. to code A.I., with the expectation that once that breakthrough occurs A.I. development will increase exponentially faster.
The central point that I focus on in that conversation is this:
And I think something most of us believe about the internet at this point is it would have been better, all things considered, if surveillance advertising had not become the dominant model for all of the biggest players. I’ve said before on the show, I worry about A.I.s of this magnitude being hooked up to surveillance advertising. But there’s other things I also worry about them being hooked up to. And I think people are really unreflective at this point or underfocused on this question of regulating how you can make money within machine learning is actually very important for affecting the path of machine learning in the long run.
And I’m not sure that I know — in fact, I’m sure I don’t know — what is safe and what is not safe here. But a world where there billions of dollars to be made because the U.S. government had set up a series of prize competitions for things like the protein folding problem, which DeepMind was able to solve, you’d get A.I. systems that could do one thing, and a world where there’s billions of dollars to be made in trying to get me to buy things and manipulate my behavior, like you’ll get A.I.s that do another. I really worry about the most obvious profit models being a spur, then, to create A.I.s.
I mean, if you create A.I.s, and the main way they make money is manipulating human behavior, you’ve created the exact kind of A.I. I think you should fear the most. But that’s the most obvious business model for all of these companies. It’s what they already do, whereas that’s not what, say, Pfizer does. But Pfizer’s is the one creating the — this seems obvious to me, and I can’t seem to get anybody to care about it.
And yesterday, this article at New York Magazine wonders “Will ChatGPT Become Your Everything App?” after OpenAI announced plug-ins for their software, effectively creating the possibility that they can “reorient the entire internet, as well as users’ own devices, around chat.” The goal, according to Herrman, is to “manifest the Silicon Valley dream of an everything app: one piece of software through which you use all others, an operating system for everything, a fundamentally different way of interacting with our computers and phones, and a thing that replaces the home screen, search box, feed, or maybe all of the above.
