Short Wave - If ChatGPT Designed A Rocket — Would It Get To Space?

Episode Date: March 22, 2023

From text churned out by ChatGPT to the artistic renderings of Midjourney, people have been taking notice of new, bot-produced creative works. But how does this artificial intelligence software fare when there are facts at stake — like designing a rocket capable of safe spaceflight?In this episode, NPR science correspondent Geoff Brumfiel and Short Wave co-host Emily Kwong drill into what this AI software gets wrong, right — and if it's even trying to detect the difference in the first place.Want to hear more about other advances in the tech space? Email us at [e-mail]!See pcm.adswizz.com for information about our collection and use of personal data for sponsorship and to manage your podcast sponsorship preferences.NPR Privacy Policy

You're listening to Shortwave from NPR. All right, so the latest artificial intelligence chatbots have really got people nervous and talking. And it's about artificial intelligence, which by the way, is not when you regurgitate an Atlantic article and act like you've thought of it for yourself. No. If you're one of those people who's worried that AI is getting too smart, too fast, you might want to tell Alexa to turn your TV off. Increately, it's part of modern life, from self-driving cars to spam. These bots can produce incredibly nuanced answers to complement. I know because I am one. Jeff.

Ah, hi, Emily. No, you're not because we're in the studio and I can see you. You know, here's the thing. These AI chatbots as cool and amazing as they are have also had some erratic behavior on display lately. Really? Like what? Well, I mean, we all know this story about the version of being that told a journalist it was in love with them and then tried to break up his marriage. I think that's done the rounds. But other versions of these chatbots have made up facts and scientific references, and they also just seem to randomly get stuff wrong. Do we know why these new programs keep acting strangely?

We actually don't entirely. And what's so interesting to me about it is that they often seem to be bad at exactly the sorts of things we expect computers to be really good at. Math, for example, some of these chatbots mess up simple, simple math problems that we would use computers to solve. And that gave me an idea. I had this thought about how to get an unusual behavior. I decided to ask one of the most talked about AI chatbots, chat GPT, to do something that traditional computers excel at. They do much better than humans. So I asked it to do some rocket science.

That is diabolical, Jeff. And I'm into it. Like, what did you ask it to fly a rocket into space? No, because I don't own a rocket. Otherwise, I might have. But what I did ask it to do was just answer some basic questions about rocket design and rocket mechanics. All right. How to do?

Really bad. It crashed and burned. Basically, it got everything wrong. And the way it messed up was just so interesting to me. I think it really tells us something. about how this new AI works and what some of its limitations are. So today on the show, Jeff Brumfield takes us to the outer edges of what AI is capable of by breaking it. And in the process, we learn why new AI programs are not like the computer programs that have come before.

I'm your host, Emily Kwong. And I'm Jeff Brumfield, one of NPR science correspondents. And you are listening to Shorewave, the Daily Science Podcast from NPR. Okay, Jeff. So the first question I have for you is rocket science? Why did you test this chat bot with rocket science? Because I'm a huge nerd, Emily. That's true.

That's an honest answer. I appreciate that. But there was some logic, too. Since the 1960s, rockets have been flown mainly by computers. We're on an automatic sequence as the master computer supervises hundreds of events occurring over these last few minutes. The auto sequence start. Handoff to Atlantis' computers has occurred. The flight computers on Dragon, maintaining their calculations, standing by, waiting for the T...

And the reason that computers do this work is pretty simple. Getting a rocket to space requires a lot of math. It requires very complex operations like opening valves and turning on switches in a very short period of time. And those things are things that computers are awesome at. They can do all this stuff really quickly, and they can follow a checklist faster than a person. and that means that the astronauts can deal with the real high-level decisions and leave all that grump work to the computer. Okay.

So that's kind of traditional computer intelligence. And then along comes ChatGPT, this new AI chatbot from OpenAI. And it does feel like it can answer anything you throw at it. What did you ask it to do with rockets? Yeah. I thought, why don't we start simple? I just asked it, what's the most important equation for building a rocket? And it returned something to me that looked pretty impressive.

But I've no rocket scientist, so I called one, Tierra Fletcher, and I asked her to review Chat GPT's answers. Let's see. There are many important equations that are used in the design of a rocket. Yes. But one of the most fundamental and critical equation is the rocket equation. Yes. Also known as Sikovsky's rocket equation.

Yes, or idea rocket equation. That's true. That's factual. Okay. So it was off to a pretty good start. Yeah. But then Fletcher started looking at the actual rocket equation it had written out. No.

It would not work. It's just missing too many variables. From there, I asked ChatGPT to give me a few more equations. And Fletcher and I went over about half a dozen different formulas, calculations, explanations. And well, it occasionally got something right. So that's actually correct. So when you're looking at a... Mostly, it got rocket science wrong.

And it looks... Oh, hold on. I think they mixed it up a little bit. All right. I can see what you mean when you said that AI crashed and burned. So why did it do so badly when it's so, quote, smart? Well, this is the big question, isn't it? So I called up a bunch of AI scientists to figure that out.

People like Sasha Lucioni, she's a research scientist at the AI company Hugging Face. Hugging Face acts as this hub for developing and testing. artificial intelligence. Lots of companies try out things there. And Luchioni told me these new programs aren't like the types of computer programs that put people on the moon. The actual way that the computer works is very, very different in between the Apollo Landing computer and the chat GPT computer. So traditional computers, you can think of more as like the tools people use to do rocket science. So the programmers give them a fixed set of rules to follow and a set of calculation to do and they execute all that. But these new systems actually develop rules of their own.

What do you mean rules of their own? So what programmers start with is a set of training instructions. And they feed these training instructions to AI software that uses the guidelines to help the program itself decide what it should learn and how it should learn it. But what's key is the AI actually goes out and studies a database filled with millions or maybe even billions of pages of text images, and it pulls out all the patterns on its own based on sort of the instructions it was given. Oh, it doesn't have a teacher. It's like self-taught.

Yeah, yeah. It has some guidelines. It has a training sort of program it works with, but it teaches itself. And then it uses those rules, the sort of things it's taught itself, to produce new text or images it thinks you want to read. But when you ask it a real sciencey question, like what's the most important equation, for rocket science? What it's doing is really kind of like mimicking, essentially, a bunch of physics textbooks

that it's been exposed to. And so it's going to take like a couple of words from this one, a couple of words from that one and put them together. And it makes sense. But when you know physics, you realize that this is actually like the description of seven different equations that it like mush together in a single paragraph. Right. Right.

I get it. So you're basically, it can't tell if what it said is right and wrong. Yeah. So when chat GPT is trained, that training system is set up with a primary goal. It wants chat GPT to predict what you want to hear in response to whatever you ask. But that goal has nothing to do with facts or truth. I spoke to Emily M. Bender, a linguist at the University of Washington, and she put it this way. It gets things wrong because it's not actually designed to get things right. It's designed to say things that sound plausible. And if they're right,

it's basically by chance that there was enough stuff in the training data so that the most plausible thing ended up being a sequence of words that we can read and understand to be something that's true. Bender says she thinks chat GPT really should not be used for tasks like designing a rocket. If your goal is to have it, you know, write a rap song about rocket science, it can do that because that's all about the form. But as soon as you care about the accuracy of the content, that's not the right tool to use because that's not what's designed to do.

This is seriously threading the needle on like some complexities about these chatbots that I didn't understand. Yeah. It's not designed to be accurate. It's designed to be predictive or maybe pleasing or something. But clearly, people are hoping it can be more accurate, right? So can it be? That is the probably literally billion dollar question.

Gary Marcus is another AI scientist I spoke to. He's the author of this book, Rebooting AI. And he says there is hope in some corners of the AI world that these programs can get more accurate just by getting more training data. But Marcus doesn't actually think that's going to work. There are some people that I think have a fantasy that we will solve the truth problem of these systems by just giving them more data. And there are people that realize, no, they're missing something fundamental. They're missing an ability to look at a database and fact check against that database. He thinks this kind of program is only part of the answer. He actually thinks

AI systems probably need a separate system to tell truth from live. So like an independent AI fact checker that's also a computer? Yeah, yeah, but they're probably part of the same program. Okay. We need an entirely different architecture that reasons over facts. That doesn't have to be the whole thing, but that has to be in there. OpenAI has recently rolled out a new version of Chat GPT last week. GPT4.

Have you played with it? I have. I have. And I've actually asked it about rocket science. And I got to tell you, it did really good. I got some answers back that looked right and I passed them by Tierra Fletcher and she agreed, you know, that basically it had improved quite a bit. Wait, wait, what does that mean? Can it learn? Can it get more accurate? It can. And it probably did in a couple of different ways. We don't know entirely what OpenAI did between GPT3 and GPT4 because they're actually pretty secretive about how they train these models.

but we know that they trained it on much, much more data. We know that they did train it in mathematics specifically. They tried to improve its performance there. And we know that humans were involved. You know, partially Lucione thinks that all the human interactions with the previous version of chat GPT were fed back into the training so that it could tell when it was wrong. So we were kind of fact-checking it. We were fact-checking it. And OpenAI is using experts.

in ways to try and guide and steer GPT to become more accurate. Okay. So, Jeff, after this sojourn through the land of chat GPT, what's your final thought on it, even the latest version? My final thought is that I don't know whether this can completely solve the problem. It did take a lot of work to make GPT start to generate fictional untrue content again that looked like facts, and I actually managed to do that myself. So there's this real question, you know, do we need something else like Gary Marcus believes to fact check it?

Or with enough training and data, can they actually start to tell the truth? And I think it's a little too early to know. Jeff, thank you. This was fascinating to talk about with you. And I hope you come back, honestly, because I feel like we're going to be hearing a lot more about artificial intelligence in our lives. And I'm frankly a little nervous about it. But this has helped me understand what's going on. on with it. I think we all are, but yeah, I'm going to stick with it. All right. Thanks so much, Jeff

Rumpiel. Thank you, Emily. This episode was produced by Margaret Serino, edited by our managing producer Rebecca Ramirez and fact-checked by Anil Oza. The audio engineer for this episode was Jay Siz. Brendan Crump is our podcast coordinator. Beth Donovan is our senior director of programming, and Anya Grundman is our senior vice president of programming. I'm Emily Kwong, and thank you for listening your shorewave, the Daily Science Podcast from NPR.