‘Artificial General Intelligence’ Is Apparently Coming. What Is It?

For years, AI companies have said that AGI is coming soon. But what does the term mean, and what is the science behind it?

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  WNYC is supported by Bear. Science is about asking questions, testing, trying again. It's not easy, but it's relentless. It evolves, and that's why science delivers. Learn more at sciencedelivers.com.

• 0:14Ira Flatow?

  Science Friday is supported by MathWorks, creator of MATLAB and Simulink software for technical computing and model-based design. MathWorks, accelerating the pace of discovery in engineering and science. Learn more at mathworks.com.

• 0:31Speaker 02

  Support for Science Friday comes from USA for UNHCR, the UN Refugee Agency. Here's a problem. How do you keep a house above freezing when the windows are gone and the power is out for 16 hours a day? In Ukraine this winter, the answer fits in a box. Reflective insulation, plastic sheeting to cover broken windows, draft blockers for the doors. Simple materials installed the same day a team arrives. Your gift delivers them, and until December 31st, it's matched. Donate at unrefugees.org/friday.

• 1:11Ira Flatow?

  This is Science Friday. I'm Ira Flatow.

• 1:14Flora Lichtman?

  And I'm Flora Lixman. Today in the podcast, tech companies are in a race to create AGI, artificial general intelligence. But what does that actually mean?

• 1:24Speaker 04

  I'm not fond of the term because it is so ill-defined.

• 1:32Ira Flatow?

  If you're not familiar with AGI, that's what we're going to do. We're going to try to get you in the loop. Let's start. OpenAI is an organization founded in 2015. It makes the chatbot ChatGPT. It has also said that its founding goal was to build AGI. That is a system that, according to the founders, quote, "benefits all of humanity and gives everyone incredible new capabilities." Others in the field, like DeepMind co-founder Demis Hassabis, have called AGI, quote, "a system that should be able to do pretty much any cognitive task that humans can do." On the other hand, OpenAI CEO Sam Altman said AGI would

• 2:15Ira Flatow?

  "matter much less than people think." So which is it? Will AGI elevate all of humanity? Or will it end up not mattering that much? And just what is artificial general intelligence supposed to mean anyhow? Lot to unpack here. Who better to answer these questions than someone who researches the intersection of machine and human intelligence? That would be Dr. Melanie Mitchell, Professor of Complexity at the Santa Fe Institute, where she researches cognition in artificial intelligence and machine systems. Welcome back to Science Friday. Been a while.

• 2:52Speaker 04

  Yeah, thanks for having me.

• 2:53Ira Flatow?

  Nice to have you back. All right. Now that we've been rattling off all these definitions of AGI, what is your definition of AGI? And if you think it is really definable?

• 3:05Speaker 04

  Well, honestly, I'm not one to give a definition of AGI because I'm not fond of the term, I have to say, because it is so ill-defined. And it makes the assumption that humans have something like general intelligence, which I think is not really true. Humans have very specific kinds of intelligence that are good for the kinds of environments that we find ourselves in, but not for everything. But AGI has been defined in so many different ways, it's almost lost any rigorous meaning.

• 3:41Ira Flatow?

  Really? So how is AGI then different than, let's say, ChatGPT or Perplexity? Or any of those services you use and ask it for a prompt or a question?

• 3:52Speaker 04

  Well, it depends on how you define it, of course. Originally, AGI was defined as some system that goes beyond doing a narrow kind of task or capability. So if you think back to, say, Deep Blue, which played chess, or AlphaGo, which played Go, they were superhuman in their abilities to play those games, but they couldn't do anything else. So they were what people call the kind of narrow AI. General AI was supposed to be machines that are more like humans that have the range of human intelligence and that are more, you might say, like the AI that we see in movies.

• 4:35Speaker 04

  That can do all kinds of things. So ChatGPT is more general than, say, AlphaGo or Deep Blue, but it certainly can't do the range of things that humans can do. So I think that when OpenAI says its goal is to produce AGI, what they mean in some sense is to get machines that can do really all of the kinds of things humans can do, but then they put a caveat on it. You mentioned at the beginning when you said cognitive tasks. So they want to separate the idea of being able to do all the so-called cognitive things that we do from the physical things that we do.

• 5:19Speaker 04

  You know, ChatGPT is not going to go fix your plumbing or re-roof your house.

• 5:25Ira Flatow?

  Right.

• 5:26Speaker 04

  So they don't count those interactions with the physical world in their definition of AGI.

• 5:33Ira Flatow?

  And lately, some people like Sam Altman have started throwing around the word "superintelligence" more often in these conversations. Not to open another can of worms here, but how is that different from artificial general intelligence?

• 5:48Speaker 04

  Good question. So I think, you know, we have this notion of human-level AI, which is sort of the equivalent to AGI, aside from all those physical things that I mentioned.

• 6:00Ira Flatow?

  Right.

• 6:01Speaker 04

  But superintelligence is AI that's better than humans across the board. You know, we already have AI systems that are much better than humans at many different tasks. We've had that for a long time, including playing chess or perhaps other, you know, navigating a city very quickly with maps and so on. But superintelligence in this context, the kind that Sam Altman is looking for, is AI systems that are better than humans at everything.

• 6:36Ira Flatow?

  Is this like Ray Kurzweil's singularity moment? You know, when AI surpasses human intelligence and we become the robots?

• 6:46Speaker 04

  Yeah, in some sense, that's that's right. So Kurzweil's singularity is the the moment that AI systems become smarter than humans. And in that, his view, they're going to be able then to improve themselves. And so you get this kind of feedback loop, positive feedback loop, where they're getting smarter and smarter and smarter. And then we have this singularity that where machine intelligence sort of becomes, in some sense, incomprehensible to humans. That's Kurzweil's vision.

• 7:20Ira Flatow?

  Yeah, so that's not the super intelligence Sam Altman is talking about, or is it?

• 7:25Speaker 04

  I think it is, you know, and Sam Altman has has this idea that once we get superintelligence, we'll get super, superintelligence, and then super, super, superintelligence. And then we'll, you know, have machines that have cured cancer and figured out how to colonize Mars and all of these things.

• 7:45Ira Flatow?

  Right. Let's let's start talking about this word, intelligence. And this is a big question. How do you study intelligence? What do people who do this work think about the concept of intelligence itself?

• 8:00Speaker 04

  Right. So intelligence is a kind of an umbrella term for a lot of different kinds of capabilities. And, you know, we think of our abilities to reason, to figure out what's causing what in the world, to figure out how to interact with other people and understand them better. You know, I don't think intelligence is any one thing. It's a whole host of capabilities. And some people have strengths in some areas of intelligence, and others have other kinds of strengths.

• 8:34Ira Flatow?

  Right.

• 8:35Speaker 04

  Some animals are more intelligent in some areas than even humans, but humans have this capability for reasoning and for reasoning about our own reasoning and kind of being able to understand the world i- in a deeper way than perhaps any other species.

• 8:57Ira Flatow?

  Does that include self-awareness?

• 9:00Speaker 04

  I think it does. I think self-awareness is a key part of intelligence because it helps us understand and reason about our own thinking. And I think self-awareness is something that current machines lack.

• 9:15Ira Flatow?

  Right.

• 9:16Speaker 04

  Chat GPT doesn't have self-awareness. It doesn't have the concept of itself as an entity, I believe. And that's part of what's keeping it from being more intelligent. You know, it doesn't have any sense of whether what it's saying is true or false, or whether it has more confidence versus less confidence in some of its statements. And from that, we see that, you know, it can produce untrue things that are it's just as confident about as the true things that it generates.

• 9:52Ira Flatow?

  Well, you know, people who claim to be intelligent also can produce untrue things.

• 9:58Speaker 04

  Absolutely.

• 9:59Ira Flatow?

  They know to be untrue.

• 10:00Speaker 04

  Right. And that's when you say they know to be untrue, that's a kind of self-awareness. That's an intention. Whereas these systems like Chat GPT don't have these kinds of intentions. You know, they don't have the intention to be deceptive or to be truthful. They're really just generating text according to some probabilities that they've calculated about what's the sort of most likely kinds of things that they should be saying.

• 10:35Ira Flatow?

  After the break, did you know?

• 10:37Speaker 04

  There's a probably a pretty good correlation between the people who are drawn to study AI and the people who like Star Trek.

• 10:49Speaker 00

  WNYC is supported by Bear. Every day in America, science delivers, not just in the headlines, but in the moments that matter most. Science is a journey, one of curiosity, persistence, and progress. It's about asking questions, testing, learning, and improving again and again. Science delivers harvests that feed families, tools that help farmers grow more with less, and new ways to protect land for generations to come. Let's keep asking, keep testing, keep discovering, so science can keep delivering. Learn more at science delivers dot com.

• 11:24Speaker 05

  I'm Kevin Roos. And together, we've been tech journalists for decades. Now we're your hosts of Machine Gods, a new podcast from NPR that makes sense of what AI means for your work, your life, and all of our futures. Twice a week, we'll be exploring the biggest moves in tech and give you unmatched access to the real titans of industry. Listen to Machine Gods from NPR on the NPR app or wherever you get your podcasts.

• 11:54Ira Flatow?

  Let's talk about the history a bit behind this term AGI, but also AI itself. I know you've been studying this for a while. How far does it go back?

• 12:04Speaker 04

  So the term AI goes back to the 1950s, when a group of people had a meeting at Dartmouth College about this new field, and they had a kind of arguments about what to call it. And one of the founders of the field, John McCarthy, suggested artificial intelligence as a way to distinguish the field from other kinds of fields that were studying intelligence at the time. He later regretted calling it artificial intelligence, because, you know, why are we calling it artificial? We should be seeking actual, real intelligence.

• 12:43Ira Flatow?

  Yeah, makes sense.

• 12:44Speaker 04

  Yeah. But other people proposed other terms for this field. One of them was Herbert Simon, who suggested complex information processing, which avoided the sort of anthropomorphism of intelligence, the notion of intelligence. So you can imagine, maybe the way that we think about these systems might have been a little different if we didn't call them artificial intelligence.

• 13:09Ira Flatow?

  Right. And you've written how Star Trek has had a kind of outsized influence on the direction of the field as a whole.

• 13:18Speaker 04

  Yeah, exactly.

• 13:20Speaker 04

  As I said in the book I wrote on AI, that there's a probably a pretty good correlation between the people who are drawn to study AI and the people who like Star Trek. And one of the elements in the early Star Trek episodes was a computer, it was just called computer, that would answer any question. You know, you could ask it anything and it would give you a very cogent, concise, correct answer. It knew everything. And this computer was really what a lot of people in AI said that they felt it was sort of their north star for building AI systems. They wanted an AI system that was like the computer in Star Trek.

• 14:03Ira Flatow?

  But we're close to that. I mean, on a superficial level, aren't we? You can ask ChatGPT and speak to it and get an answer.

• 14:11Speaker 04

  Yeah, so we are closer than we've ever been, for sure. But ChatGPT and these other generative AI systems lack something that that computer had, which is trustworthiness. You know, often you could trust anything the Star Trek computer told you. But with ChatGPT, while most of the things that it tells you are correct, it does have a tendency to do what people have called hallucinate, which is to generate very confident sounding answers that actually are untrue.

• 14:49Speaker 04

  So this is, I think, the next frontier, if you will, is the final frontier is trustworthiness with these kinds of machines.

• 14:59Ira Flatow?

  Do we need another breakthrough somewhere down the line to get this to be more of self-aware? I guess it's the intelligence that we can't define, you can't really define AGI right now, so that, you know, it's the old phrase, I don't know exactly what it is, but I'll know it when I see it.

• 15:19Speaker 04

  Yeah, I think we need a couple breakthroughs. One, it would be in how to make these systems more trustworthy, more self-aware, more have a better notion of what they're talking about and whether it's true or false. We also need to better understand what we mean by intelligence. As you know, you've said, we know it when we see it. But one of the problems is that we've thought that for maybe millennia. And it turns out we've often been wrong. So just as an example, people used to think before the age of Deep Blue, the chess-playing computer, that in order to play chess at a grandmaster level

• 16:01Speaker 04

  or a superhuman level, you'd have to have superhuman general intelligence. But it turned out that you could get a computer to play chess at this superhuman level without anything like we call general intelligence. The same thing has been said of things like speech recognition and conversation, like the kind that we have now with ChatGPT, that to get those kinds of abilities, you'd need something like general human intelligence, human-level intelligence. But it's turned out that we can accomplish these kinds of capabilities without having this idea of general intelligence,

• 16:44Speaker 04

  the kind that the pioneers of AI really were looking for. So it's really taught us a lot about how hard it is to define what we mean by intelligence and to know when we have a system that's close to having that.

• 17:00Ira Flatow?

  Well, with that caveat in mind, especially about how wrong we always are in predicting the future, I want you to predict the future for me. Here it is, beginning of 2025. What should we be paying attention to this year? What kind of developments are you expecting to hear about in AGI or from any of these companies?

• 17:24Speaker 04

  Well, I think, you know, one thing is that, you know, this word AGI has gotten so much cachet.

• 17:31Speaker 04

  That people will be trying to sort of redefine it into existence. To say, well, what we have at the end of 2025 is clearly AGI. Because and then have some definition that captures what we have. So I, I do think that that is likely to happen. But I also think that people are going to realize that these systems are actually lacking a lot of the very important aspects of what makes human intelligence more sort of trustworthy. When it is and more general and that those things people are going to start focusing much more effort on in their development of AI systems.

• 18:16Ira Flatow?

  Yeah, because people in general are afraid. They're fearful of computers from what they see in science fiction and what they're watching in their real life. Now you have to become more trustworthy.

• 18:28Speaker 04

  Absolutely. You know, there's the fear that the machines will get too intelligent and will, you know, take over. But there's also the opposite fear that the machines won't be intelligent enough to do the things we give them to do. But we'll trust them too much and they will fail in ways that we didn't expect. So I think both of those fears might be worth considering.

• 18:55Ira Flatow?

  Yeah, I can see that in both in good and bad, like in medicine and even in computer warfare, you trust your machines, right?

• 19:03Speaker 04

  Yeah, we don't want to trust them too much when they haven't in some sense earned our trust.

• 19:08Ira Flatow?

  Well, you have earned our trust, Dr. Mitchell. I want to thank you for taking time to be with us today.

• 19:13Speaker 04

  Well, thanks very much. This is a very important topic and I'm thrilled to be able to talk about it with your audience.

• 19:20Ira Flatow?

  Well, you're welcome. We're happy to have you. Dr. Melanie Mitchell, Professor of Complexity at the Santa Fe Institute in Santa Fe, New Mexico.

• 19:31Ira Flatow?

  That's about all the time we have for now. A lot of people helped make this show happen. Jason Rosenberg. Melissa Mares. Beth Ramey. Sandy Roberts. I'm Ira Flatow. Thanks for listening.