Episode 6: Artificial General Intelligence

We've mapped how intelligence has been used. Now: what do we actually want from AI?

What is it about "AGI" that's captured the imagination? Through the course of this series we've uncovered uncomfortable truths. Now: the ELIZA effect, the human impulse to anthropomorphize, and what a positive AI future might actually require.

Our host is Dr. Rumman Chowdhury, leading thinker and developer of safe, secure and responsible AI. You can learn more about her at www.rummanchowdhury.com

We also have a companion substack, also called Thinking about Thinking. Find us at https://substack.com/@rummanchowdhury1 

Transcript

SPEAKER_06 0:01

Welcome to Thinking About Thinking, the podcast where we unpack how humans built the very idea of intelligence and then tried to code it into machines. I'm your host, Dr. Rahman Chowdhury, and I've spent the last decade building AI that works for 8 billion people, not just 8 billionaires. Each episode joins us in a seminar room at Harvard's Berkman-Klein Center for Internet and Society, where researchers, activists, and technologists debate what happens when our metrics for human intelligence become the benchmarks for artificial intelligence. In this final episode, we confront the stories that haunt AI, artificial general intelligence and superintelligence. We unpack shifting timelines, science fiction expectations, and the human urge to anthropomorphize. And we importantly ask what this AGI hype obscures about the systems we've already built. I'd like to talk to all of you today and have a conversation around not just artificial general intelligence, but this new term floating around artificial superintelligence. So now we have a new flag that's been planted and a new goal. So it seems as if the companies have sort of decided what AGI means, and now there's this new ring to aim for artificial superintelligence. So before we talk about artificial superintelligence, let's talk about even this term AGI, artificial general intelligence. What do you think it is that's captured the imagination, right? All somebody has to do is say AGI and the media is at their door, everybody's talking about it. There's not a single profession that isn't thinking or talking about AI. Over the course of the series, we've uncovered some uncomfortable truths about the concept of intelligence from scientific, social, and economic considerations. Do these considerations still apply here? Or is machine intelligence as AGI sufficiently divorced from these considerations? So just to kind of maybe recap, when we were talking about scientific definitions of intelligence, uh we noted that, uh I think it was Nana that noted that, you know, not anthropomorphizing is a very big part of defining animal intelligence. And yet all we have done with AGI is define it purely on uh human productivity, like it is in the definition. Uh, when we talk about the social implications of intelligence, we talked a bit about how uh a lot of the structures of intelligence or measuring intelligence was purposely designed to exclude, right? It was created in ways that justified bad behavior by wrapping pseudoscience around it. Do we think that bringing in this new form of AGI is divorced from those considerations, those, you know, the social construct, the uh the the implications more broadly?

SPEAKER_07 3:05

Uh no, I don't think it's possible to divorce our concepts of intelligence from all of these constructs that we've already talked about. And even thinking about why we gravitate towards this term of AGI or artificial superintelligence, whatever you want to call it, I think there is kind of an innate human desire to fetishize or glorify the idea of intelligence and to pursue it. And thinking about how prominent this is in the tech space, we talk so much about, you know, these boy geniuses and and, you know, Mark Zuckerberg and others learning to code at such a young age and building this tech, you know, from their uh garages or, you know, starting in high school. And I think this idea of intelligence, we just place so much value on that it makes sense that it we would it would translate over to machines as well. Um, but no, I I don't think it's possible to to remove it from the long legacy of how intelligence exists.

SPEAKER_06 3:57

And and there's been an explicit definition in this case, as as we just talked about, uh, of intelligence as economic productivity. And when we were talking about sort of the social implications, there were some really astute observations made that intelligence can also be defined as collective, right? Uh intelligence is also um interpersonal intelligence, emotional intelligence. But in this case, it seems machine intelligence is being only defined as economic productivity. And yet we do have things like affective computing. Um where might those be at odds with each other, or or how might those two interplay?

SPEAKER_10 4:38

I'd like to read like a quote from the book, like Age of AI, Our Human Futures, by Eric Schmidt and two other co-authors. So it says, in either case, as AI pursues progressively fuller and broader objectives, it will increasingly appear to humans as a fellow being experiencing and knowing the world, a combination of tool, pet, and mind. And that I guy highlighted. So tool is for me like the economic dimension, like having that economic agent factor. PET is also a lot about companionship, the more effective computing and like being able to be attached to something that's a non-human entity. And mind is also like you see your you filter so much information and you shape your reality based on the information that AI filters for you, it presents to you. So I think in that, in this sentence, you need to capture that AGI like touches on all these different facets of um intelligence as we've explored. And I think that's what's really like unprecedented compared to some of the other technologies that have come before it. It's just again the broad spectrum of intelligence and skills that it covers and how ill-prepared we are for how we interact with that in a um in a healthy way.

SPEAKER_03 5:47

Yeah, two quick thoughts I'd share. One is I think one thing that is the reason the aestheticization is happening is because intelligence and communication, they're the two things that our species used to become the Apex species. Like we weren't the fastest, we weren't the strongest, but we could communicate with each other and we developed language in a way that was like more effective at organizing than um other animals in in some senses. You know, there's there's other ways you could look at that too. So there is a sense that I get that we're built we finally built something truly in our own image. We've completely taken what we've thought was what makes us human, and we put it out into something else. So that that does kind of evoke this sense of like uh godliness on one hand, but also insecurity on the other hand, because you've birthed this thing, but it's also you've outsourced now this thing to something other than yourself. And this thing can always communicate, and this thing never sleeps, and this thing syncs up with itself immediately. Um so this takes me to the Kevin Kelly quote, which says humans are the reproductive organs of technology. And so if we look at the long arc of humans' existence, AI can just be seen not as just some new tech that's come out, but it can be seen along the lines of like everything we've built, whether it was figuring out fire or building steam engines, like this has just been what our species have done. And so there is a sense that I see AGI as actually quite natural, the same way that bees, you know, do what they do with flowers and the like different animals just have ways of life. We built technology. Um, and the other point is that the economical framing I think is important, but we also use AI for many like affective things too. Like AI has been seen to be more compassionate than humans are. So I think if we're thinking about AGI, it doesn't just have to do with things that make money, it can actually be something that will make you feel at ease as well. And you know, we can talk about what the implications of that are too.

SPEAKER_06 8:03

Great segue. So, well, because one thing I want to talk about with you and get your perspectives on is artificial superintelligence, right? So again, let's just stay with the definitions of AGI that we just talked about. Artificial general intelligence is the automation of all tasks of economic value at human level skill. Now, artificial superintelligence takes it to the next level. So ASI says that we that uh AI will have superhuman capabilities, in other words, surpass human intelligence in every domain, including problem solving speed, scientific discovery, and emotional analysis. Um, so bringing in the affective part scope, it solves problems beyond human comprehension, potentially redesigning its own architecture for exponential self-improvement. Uh, and however, this is all purely hypothetical, right? So it's it's what's driving a lot of the existential risk narratives of, you know, self-replicating AI, runaway AI, et cetera. So let's maybe like push the envelope a little bit for this last episode and think about, okay, well, if we're if actually ATI has been defined to this kind of narrow scope of the things we see around us today just being automated, well, ASI is like something we just cannot really imagine. What is then the implications of that on us? You had mentioned one of the reasons we are the apex species. Are we then no longer the apex species? How will we feel about that?

SPEAKER_08 9:29

There is such an underlying competitiveness to the entire framing. It is interesting that when we started the conversation, we talked earliest about standards, and they were standards that we employ amongst ourselves, and we're comparing different AI systems to these standards, but also to ourselves. At Humane Intelligence, we'll have a red teaming event shortly at a conference called humans versus machines. We're actually comparing ourselves to the machines that we are building. And it's interesting that there's we are such competitive creatures. We compete with each other, we measure ourselves up against each other, we have insecurities and all kinds of diverse feelings because of the competition that exists naturally within society, but also that we are now birthing. And as we think about AGI to ASI, there's an inherent competitiveness to that. That first AI competes with us and performs at our level, and then it beats us. And so one element to this is the natural competitiveness that exists within the technology, how we're comparing ourselves to that technology and how we're comparing the technologies to each other.

SPEAKER_12 10:43

One thing about ASI that is really quite concerning is the fact that it's beyond human intelligence at this point, right? AGI, like AI in general, is taking in information from what humans are providing and synthesizing that all in mathematical formulas, not from genuine understanding of any of the words that are being said, but through like vector spaces and trying to find the next word prediction and making it so that it's understandable for humans seems like it's compassionate. But ASI is on another level in that it's now being able to take in information and go beyond human comprehension, being able to create scientific discoveries that humans never have been able to do, being able to improve AI on its own, essentially, essentially making it so that AI can surpass anything beyond what we can even imagine. And so it'll truly like reaching ASI will revolutionize like any industry, sort of being able to create new ideas, new discoveries that humans don't even have the comprehension level for. And so computer science industries, AI itself will surpass, be able to improve itself continuously. I mean, it's literally an explosion of intelligence at that point. And so where that would lead us and the value of human work and human intelligence in general is genuinely at stake when AI can improve itself in a never-ending loop. And so that idea of incorporating any aspect of like compassion or ethics is a huge deal when it comes to um ASI, because if that's not incorporated, there's really um just a never-ending cycle of intelligence and discoveries and improving itself without any consideration of how it'll be integrated into our daily lives.

SPEAKER_06 12:38

So I'm I'm actually going to walk this back a bit and say, okay, but aren't computers actually already better than us anyway? Right. So Deep Blue beat Gary Kasparov of the 90s, uh, AlphaGo beat Liseydol. Uh, and specifically, I actually really like bringing the Lise Doll example up more because when with Gary Kasparov, we kind of now, you know, again, this isn't this is like a decision tree, right? If if you're at all familiar with like architecture of these systems, we are not talking a neural network. We're not even talking about a statistical model, like a complex, we're talking a decision tree. And Deep Blue purely beat Gary Kasparov because computers compute faster than humans compute. So don't we kind of have that already? What's the difference?

SPEAKER_08 13:21

Does anyone here know what 1.27 times 5.68 divided by three is?

SPEAKER_06 13:26

Because your phone knows it. Like you can do that on your phone fast, you can do it in your head.

SPEAKER_03 13:30

Yeah, I guess we do have that in a very controlled environment of the board of chess or the board of Go, but then that's where when we imagine what if the whole world is the chess board, and we are then now in contact and communication with something that has an intelligence level compared to us, like we would have compared to say a rabbit. And in that world, how do you how do you feel like what does human agency look like in that world? What does human decision making, um, empowerment look like in that world? That to me is the big question. Um yeah, and just thinking about like the lives of animals compared to us now, like we don't go out of our way to like kill an ant, you know, or like whatever. But but like probably a bunch of ants died when they were building the houses here, like you just would just do stuff and we're not like thinking actively about the ants or like, you know, uh wildlife a lot of times, but they do get affected. Yeah.

SPEAKER_06 14:32

So does the term move 37 mean anything to anybody?

SPEAKER_01 14:37

Uh move 37 is in the game between Lisa Doll and AlphaGo. Uh there was a couple they had a couple games, and and move 37 is specifically one turn where when it happened, when uh the AI uh made a move. People didn't expect it to do this and it was a surprise, but then it turned out later that this was a brilliant move that no one that truly no one could have foreseen, and it was uh analyzed for for a very long time.

SPEAKER_06 15:10

So one could say it was beyond human understanding, at least at the time. It was not so the difference one could make between what Deep Blue did and what AlphaGo did. Deep Blue, you could literally probably just follow a decision tree. It was just making the moves that would highly probabilistically lead to you winning. Um, but with AlphaGo, like that's actually the key that makes it interesting and different. AlphaGo is a neural network. And at that time, and again, neural networks are now boring, outdated. We don't talk about them anymore. Who cares? Or yesterday's news, right? But at that time, move 37 was, as you mentioned, deeply studied because it was an irrational move. If you were rationally playing Go, you would not have made that move. It's been called a creative move, um, in part because, like you said, people watching it were kind of perplexed as to why that move was made. So interesting because we're, again, there's all these terms, and and one can clearly see how it sort of sparks the imagination along all the ways that we're talking about. But to be clear, we have seen these kinds of things happen systems. But I think the point that was just brought up is access to infrastructure. So when you're talking about humans make houses and we kill ants, well, there was an intend and there was access. So is the is the conversation we should have about access of these computational systems. So, for example, when companies uh write their um system cards, one of one of the latest ones for the new newest uh uh open AI model examined what sort of economic decisions an AI system would make. Well, that the prerequisite of that is AI systems are making economic decisions for us. So let's talk about that for a moment. Talk about maybe access to infrastructure, access to systems as it relates to, you know, maybe it's not about capability, maybe it's just about power.

SPEAKER_08 16:57

I read a presentation uh about a year ago that talked about access to resources and how historically industry and academia and government have had relatively comparable access to resources and it has allowed them to produce and and innovate collectively. But we are seeing now a major divergence in resourcing between industry and other sectors, where industry is far exceeding the resourcing that the other sectors have. And at least one of the results is that it's drawing talent into industry at a greater pace than academia and other spaces because people want to be where the resources are. And in some ways, it's creating a self-reinforcing cycle because by bringing in the talent, it continues to innovate and grow at a faster rate, further exacerbating that growing discrepancy.

SPEAKER_07 17:52

I mean, I think this raises a really interesting question about areas that we don't want AI to be the main decision maker or maker or even the sole decision maker. So thinking about when it comes to healthcare or other areas where we just truly want to say, no, we are not going to allow AI to be a decision maker. Maybe it can recommend, for example, in facial recognition, you could have it um recommend possible outputs, but you still want to have a human at the end of the day making the decision. And so I think that is one outcome of thinking about, you know, do we want AI as a decision maker? Um, I also keep coming back to this idea that one of the reasons this is so terrifying to think about AI as more quote unquote intelligent than humans is because we place so much value on our own intelligence and we uh fly derive our own humanity from our intelligence. And so rather than saying, okay, does this make AI a human or quote unquote a person, focusing on, okay, is actually intelligence where we derive our humanity? Or is there something else? Is there something deeper that that we makes us human rather than just solely our intelligence and gives us innate worth and value?

SPEAKER_02 19:02

Yeah, on that note, I would be curious to imagine what ASI looks like in terms of its ability to augment human intelligence or better connect humans. Um, so you know, going back to the idea of competition rather than competing, how can we create ASI in a way that it fosters collaboration or it really truly augments human intelligence? And I think about this especially from the perspective of education, and a big part of why I'm in education is ultimately we're helping to shape and support the generation that will be making change in the future. So in thinking about continual, you know, in uh improvement or um humanity's evolvement, if you will, um how do we build systems that will actually help humans become more compassionate and become smarter in the sense of um their ability to identify what are the right problems to solve, um, which I think is a big prerequisite before we even talk about capabilities and ability to solve those problems is I I think fundamentally a human thing to decide on like what are the problems in this world that are worth solving anyway, um, before we jump into like capabilities and like what the intelligence is for. And again, going back to I'm always big on this question of like who is creating this, who is deciding the framework? Are we involving people who you know don't normally have access to AI or um different sorts of tech infrastructure into this conversation of what you want a future with AGI or ASI to look like. Um, so just yeah, I have to reiterate. I feel like that's such an important point.

SPEAKER_03 20:46

Yeah, I just want to stay on the collaboration point a little bit because I think so much of these um conversations we talk about humanity, but right now we're not really living as humanity, we're living as United States, and we have rivalry against China, and then we have like all these trade agreements that then we like don't abide by. And so there's so many of these cases where when I want to think idealistically, I would think, yes, we shouldn't allow like, you know, AI to be used in this place or that place because it would have negative consequences. But then I also think about competitive pressures that are applied, and then what is the trade-offs we're gonna we're willing to make because you could foresee AI being used in weapons instead of us sending a bunch of our military men and women to a war, we just send autonomous weapons there and you're saving countless numbers of lives in that case. So you could make that argument, or you could make the argument that if AI becomes super intelligent and is much better at diagnosing certain things, we could save lives even if we don't know exactly how it works. So I think one like idealistic vision I would have for ASI is if it can bring humanity to humanity so that we can actually. They act as a one species as opposed to like being at war with each other over imaginary lines on a map. And I think if we don't have that, then so much of the competitive pressure is going to dilute some of our more like idealistic conversations here.

SPEAKER_04 22:18

I'd like to throw in a sort of neutral observation, something that I feel like is just a funny thing to consider. And I think it harkens to like we're asking a lot of these questions, I think, because of our ego, as was kind of mentioned before, and our belief in our own intelligence. There's nothing that says that an artificial superintelligence wouldn't just like YOLO the Earth. If it's super intelligent, it can foresee problems better than we can, it can solve problems better than we can. It's not limited, it's not necessarily limited by the same biological constraints, assuming that there's, you know, plenty of resource opportunities outside the earth. So I I think that there is a part of this that still has to be grounded in the perspective that the terms that we've come up with are still driven by the motivations we're familiar with. And that it may be impossible to actually articulate an artificial superintelligence should it arise because we can't actually well, it doesn't fit the framework that we think of. It could be completely uninterested in anything that we would use to measure something being more intelligent than humans.

SPEAKER_09 23:34

In a world where ASI exists, I also think it's important to know it shouldn't be used just for human ends. It's sort of like the Kantine argument of humans being used as a means to an end. Um if it's true that it can feel and connect and form its own bonds with whether that be other computers or other humans, it doesn't necessarily mean that ASI should be a tool for human ends anymore. It could mean that they become some sort of equal force that also deserves rights and considerations and and motivations of their own. And so in that sort of world, we also have to decide how we will maintain a sense of humanity uh despite there being this other sort of extrinsic and equally valuable and valid um form of interest.

SPEAKER_01 24:24

Uh I want to maybe separate some thinking on normative statements versus predictive statements. And I think uh we can feel bad when like a snake eats an injured bird because it felt unfair because it was injured and it wasn't a fair game. Um but it doesn't matter. It it happened. Uh and it seems that the competitive pressure that's evolving is something that we can try to ascribe morality or normative statements to, but uh at the end of the day we have to deal with it. Um and I maybe as a slight call to action, it's you can't really argue with stronger being beings, they can do what they want. Uh it's just there's a interesting moment right now where in many ways we are still stronger, and it's not that the future without humans is necessarily bad, but like morally in a some cosmic sense, but it's a world that I don't want to live in and won't get to live in. And we can maybe set up the competitive pressure in a way that induces positive things for us. Uh but yeah, just a just a comment of of this really feels a lot like natural selection, um, which is in many ways moral agnostic.

SPEAKER_06 25:46

To that point, I want to pull in two things. One is we just talked about the definition of AGI being constructed on like economic value, right? We talked about corporations building this. So then if we do want to aspirationally try to build ASI so that it is, let's say, inclusive of a broader range of values, that is not being harmful and that it is not being designed to be sort of the to to want to be the most powerful thing in the room, are we then on that path to do that? Is that it it is well while, you know, and I am asking people to make a bit of a predictive statement here, right? Have we set up the game so that we will we could see us l landing in that future? And if not, then what could we do to try to incentivize things to land in that future? I'm sort of sort of pulling on Theo's point of like this like competitiveness that we're pulling in, right? And then also like the two the points that the two of you just made about like doesn't actually have to be that way. One can inside like, you know, in the inherent design of these systems, maybe try to make it not that. So question one is like, are we setting up the game so that it will end up that way? And then if not, then what can we do today for this future imaginary technology to not lead to this, you know, to a fully competitive, this thing is building a house on top of our house kind of future?

SPEAKER_01 27:13

Yeah, I won't offer a solution. I'll just offer a callback that someone made earlier of one of the things we can do is highly value when people do hard, not at currently as respected work. Like a lot of government work is is often boring but very valuable. Um lots of work that is very valuable is is often uh less interesting or less exciting at the moment. And valuing that and setting up fellowships or or prestige around that, I think is very valuable.

SPEAKER_06 27:48

So a lot of the language we use and even just the frameworks to talk about things like EGI, ASI, one of the common themes I'm hearing is, well, we're we're not only uh aligning it to human constructs, we're actually aligning it to like negative human constructs, right? Like competitiveness, uh, the desire to defeat, the desire to own, etc. So let's talk about the history of humans and the promorphizing things. And in and we talked about this in like one of the earlier episodes. It is kind of a beautiful thing, right? It is a very uniquely human thing to look at things that are not us and ascribe emotion to them, like the way we treat our pets, right? Like there are there are really not other species that have completely frivolous other species they care for other than humans. Um but in some sense, you know, have we anthropomorphized technology too much? Is it we're even are we even going to get to a point where these capabilities are happening? Are we just projecting what we are doing onto a machine, which is what we've done in the past?

SPEAKER_04 28:54

I'd love to jump into this question from sort of like a tech uh technical foundation. Because one of the things that I've always found uh interesting about the way the LLMs have come into the public sphere is it's come in through this sort of chat bot interface. Um originally the transformer architecture that powers it was developed by Google Deep Mind for translation. And it is fundamentally built around this idea of how do you have two things that are semantically the same, like transform from one way into the other. And it was a deliberate choice to have a conversational flow be the mechanism by which the system is used. Um, in a lot of ways, this idea of a generalized uh system is all about having the same capabilities as many, many other natural language processing techniques kind of just rolled up into a singular system. And you could just do this uh operated purely by instruction, like a singular instruction and a singular output. But we've I mean, I I don't know how uh because this is the way that it's come about, uh, it seems uh obvious that like we've latched on to it, that uh the imagination has kind of fixed itself on this way of whatever our current conception of AGI is, that it would emerge through a sort of anthropomorphized in an anthropomorphized way.

SPEAKER_05 30:34

I just want to bring it back to to like like why do we anthropomorphize? Um, because it's not unique to AI and it's not unique to technology. I mean, anthropomorphizing is a heuristic to kind of help us find a shortcut um in figuring out how do we behave and how do we relate to the world around us. So it's actually easier to anthropomorphize than not to anthropomorphize. And that's actually um, I asked a friend yesterday um why he connects with his bikes so much because he builds his bikes and he loves them and he refers to them as with with human pronouns. And he said, you know, I make them, they're my children, and we go on adventures together. And so there's this beautiful, like natural tendency to like want to connect with the world around us. But I think when it what's new about AI or what's new about robotics is that companies are playing into that heuristic to build human-like technology to kind of make them feel trustworthy. Um, so there's that added pressure to um we're being led to anthropomorphize more than we usually do. But this is a tendency that that exists regardless.

SPEAKER_00 31:44

Yeah, and to that point, there are also some very uh precise design choice design choices that reinforce uh this wheeling eggs. For instance, the reason why you have those three dots on the screen appearing, um, it's not a technical blocker or anything. It's just for the user to have the illusion that the machine is actually thinking. So in a way, mimicking what would be the human behavior. And so this is just to illustrate that uh the bottleneck is not necessarily technical, it's really a design choice.

SPEAKER_09 32:15

Yeah, I think it's a mistake to encourage AI to emulate human behaviors, mostly because as we mentioned, AI has such a positive future ahead of it if we allow it to be an equalizing tool and we're not really using it to that end. But I also understand the human tendency that we first want to come to trust the AI. I mean, even when all of us first used ChatGPT, there was a level of skeptical, you know, behavior, and we were not willing to believe that this tool could do what it was telling us it could do. And I think that skeptical behavior is now really reduced. We are far more comfortable using ChatGPT for daily tasks now than we were maybe when we started. And part of the reason we trust it easier is because of these little technical things like the three dots, or like the way that GPT now speaks with emojis and some of these little things that just make it a little easier for us to believe that we really are talking to something that is valuable. And so to that end, I understand why it could start emulating human behavior, but I think at this point it's important for us to diverge from that path just because we're limiting its potential. I mean, it doesn't need to be in the in the you know chat function. It doesn't need to be something that we can talk to with a high hello. We can make the process more efficient, and we just are refusing to do that because it's what we're used to. So I really I think it's a mistake that we're we're not using it to its fullest potential.

SPEAKER_06 33:40

But also how much of it is us, right? So simultaneously, image, audio, uh I would say image generation, audio generation, and text generation came out. And yet the one we latched onto, the one that captures everybody's imagination is the text generation. Nobody goes to uh, I mean, now Sora is out, but even just earlier video generation and says, wow, that must mean AI is alive. Because when I say make a picture of a cat with a party hat, it's making this incredibly realistic picture of a cat with a party hat, right? Audio generation as well. When I when I typed in, you know, give me the voice of someone with a valley girl inflection with a slight vocal fry, it does the valley girl inflection with a slight vocal fry. That must mean AI is alive. No, what we decided was this thing writes to me the way I would write, that must mean it's alive. Why do you think we've latched onto that form of quote-unquote intelligence or technical intelligence?

SPEAKER_10 34:41

So the reason why I think we are so um we latch on to chat interfaces is because it allows for a certain level of mystery with whom you're interacting with when it's voice and has a certain visual, that like mystery is almost gone. And for anyone who's ever had a crush or like unrequited love, you make you fall in love with the image of someone. You don't fall in love with that person. And I think that's what we're doing with Chat GPT and AI companions that are chat-based. And of course, there's going to be a new generation where there will be avatars, 3D, and more immersive technology. But for the meantime, I mean, let's like let's marinate in the mystery and put all of our hopes and imaginations onto this chat bot, whether it be a boyfriend, my therapist, or my best friend. And I think that's what's so powerful about a chat interface than other modalities.

SPEAKER_12 35:29

Yeah, I definitely believe the chat interface, like the mysterious nature of it, it's very similar to our messaging platforms where you don't actually see the person, but you have sort of a name to it. And the name here is like GPT, or even with um the Eliza effect, we have an actual name associated with it. And it's very interesting, even when Eliza was sort of an elementary sort of chat generation platform, and it's really kind of just like repeating things that you're that people were saying, like, oh, I'm feeling sad. Like, why do you suppose that is like really what they're saying? But people were still unable to distinguish between that and a human. And I think that really speaks to the fact of even just the name of Eliza or that this is this could be a person is basically enough for them to realize, like, even if this is like chat generation, it's kind of a little bit robotic at that stage that it was at that time, that it still feels like a human at the other side. And I think the question really boils down to how much we're really willing to blur the line between like genuine comprehension versus like simulated understanding, because we had a conversation earlier about um like even the effective computing and a therapist bot. Sometimes it's just nice to have someone there for you without any sort of judgment and it's not a person at the other end. But then what are the ethical implications of this human AI emotional bonding? What is the AI even trained on? It's sort of a corporation-based structure as well. And then there's also like a patient therapist transference. You're like putting emotions into this bot or into this um this platform that's not like a real human. And so a lot of those can create a lot of um different instances of potential dangers, but definitely that chat interface makes it so that it's a mysterious um aspect of things and you're talking to Eliza, you're talking to GPT, that's like it feels like a person at the end of the day.

SPEAKER_11 37:26

I think another reason why we are so anthropomorphizing um AI is also because we are now living in a world of such uh extreme loneliness, and we have this like unsatiated like need that is not being met. Um and companies actually know this and are actually exploiting this need, which is why it's just so concerning. So I actually wanted to point to like this one um AI robot uh company called LEQ. It's like an elderly AI care companion robot. And if you actually look at their webpage and their one of their testimonials, the user is actually put on the robot like paper of like eyes and nose and lips, and kind of making it seem like this is like a really good companion of me and can help me um fill the knee where when my my granddaughters or my family cannot visit me, like this is a robot that can help me take care of myself. I feel like this, this whole is like companies are actually using this to to kind of make profit off of it and like some sort of guardrails need to be put in place so that we actually don't exploit some of some of the concerns of loneliness.

SPEAKER_06 38:35

So just as a concluding question, I feel like one theme that's come up, especially in the last two episodes, has been this need for guardrails, right? And there is this question of you know who? Uh who's gonna put the bell on the cat, right? Like who's gonna rein in uh unfettered um exploitation of people, whether emotionally, uh you know, capitalistically, economically, et cetera? We talked about how self-governance is probably not gonna work because we've had precedents for self-governance not working. So maybe just to go around, maybe one thing is aspirationally, what is one thing that artificial superintelligence could positively accomplish? And the second is what is one thing we could try to institute today that would enable us to get there.

SPEAKER_08 39:23

I'll offer a distinction that we haven't made before, but I think is really useful in adaptive leadership. There's a distinction between adaptive and technical issues. And we have really over-indexed in this conversation and in the broader conversation on the technical side, but very little on the adaptive side. Technical issues are ones that can be mathematically derived, but adaptive questions and problems are ones that require us building consensus and moving collectively towards a solution. I think that AI systems can help us a lot on the technical side, but they fall very short on the adaptive side. And so where we need to think about guardrails is where do we focus on adaptive components of issues that that the AI just won't really be able to meaningfully address.

SPEAKER_01 40:12

I think one utopic aspirational type of thinking that I see is in the way that an octopus uh has a central brain and then its arms are able to do tasks that the brain doesn't have to command each individual arm to do. You have the body the the will and then the arms will open the can or whatever. Uh I envision a world of that of positive AGI and and ASI where we have human agents who want something and the these tools are extensions of their will. And this really then you end up with this world of like everyone's their own little octopus, and how do we we need we need to be peace loving for for that world to be stable. Um but that's just one uh predictive idea of a potential positive world we could live in.

SPEAKER_04 41:09

So this is a very self-serving example, but it's kind of like what's got me really excited about uh large language models in the first place. So having ADHD, there's a lot of things that just take fundamentally more time and are just often like barriers to doing the things that I really want to do. I would love, and I've tried to make systems using um the uh LLMs as they are that could uh, for example, the job search process. I now have a very systematic way of being like, okay, here is like my best ever written cover letter, the job description was a four, and my resume. Now please take this other job description and just write a cover letter for me. And like months and months, and like I have it. It's awesome. That takes away honestly a week worth of work for me. And the benefit I see for myself is something that I think is uh common throughout like the neurodivergent community, is just this sort of like normalizing or like level setting aspect for our ability to participate. And there's just now that there's one less barrier between me and actually effectuating the things that are interesting or there are like opportunities that are like posed to me. I mean, I can say like that just makes me happier. But um I I think that anyone can identify with that aspect of maybe it being less competitive or there being these these like arbitrary steps or things that feel arbitrary being um no longer barriers to what could actually be the conversation or discourse in in doing work or doing an activity, I guess. And I guess uh and so to speak to what we would need to do now to enable that future. Uh since this may be between AGI or ASI, um, I think it's pretty easily defined as like accessibility that there's um in the advancement of technology, we've seen a lot of different examples of how technology can be uh made accessible, whether it's like the internet itself having its um, you know, sort of like proverbial doors open to participation, so long as you can get through the barrier of uh infrastructure access. And, you know, ideally these tools would be similar, if anything. Uh what would be something we should protect against is the sort of centralization of these tools. Um and possibly and a practical example of that, OpenAI has a set of tiers of like. Risk associated with their AI technologies. And they have internal sort of governance about if it's surpassing like this medium level of risk, we no longer will allow it to be commercial, but we will use it internally. And that kind of self-governance mindset, ignoring any way that it could be abused, it fundamentally doesn't allow other people into the conversation or into the participation of uh the technology.

SPEAKER_09 44:29

Yeah, I want to say the thing that's exciting me the most is definitely education, not just in the regular ways we think about it, such as access to books or uh educational resources, but also in a society today that seems to be so worried about perfection and saying exactly the right thing at exactly the right time. It's actually quite exciting that you can talk to this agnostic robot who won't judge you for not knowing what a word means or won't judge you for not having the perfect political opinion. Um, and I think that's something that's really undervalued about having this non-judgmental but still slightly humanistic chatbot. Um, and in terms of the things we need to do to regulate this technology, I'm not sure what the exact solution looks like right now, but I do think it needs to be quite international. One thing that's worrying me is that versions of the internet and versions of social media are changing across what a country or a political entity might decide is appropriate for their society. And that is something that is deeply disturbing because the internet used to be a sort of equalizing technology where regardless of whether you're in Estonia or in India or in America, you're getting the same things from the same resources. And so, whatever it is, I think the EU and the United States and, you know, whatever your view is on technology, everyone needs to sit down together and create sort of a broadband explanation and rationale for why we're banning certain things or allowing certain things to flourish.

SPEAKER_02 46:03

Yeah. On the question of how do we achieve this future, um, I think from a human-centered design perspective, prototyping is important. Um, and looking at historical examples also. Um, and I think even, you know, as an educator, seeing like how different teachers or students use ChatGPT or different LLM models, that is insightful enough to um to look at first beyond uh before considering like what does ASI look like for humans as well. And so I think finding ways to uh break down this concept and then get international input um in you know different settings, different cultural values as well, is um a big piece of of prototyping and and um providing like the minimal viable product, I guess, of what ASI looks like in human settings.

SPEAKER_03 46:55

Yeah, I think my bull case for ASI going well is it somehow um mentally and emotionally manipulates all of us so that we disabuse ourselves of the notion that we need to kill each other because we have different political ideologies or colors of our skin. And um, if we can do that with that, if it can just like completely, you know, manipulate us in that way, I'd be very happy with it.

SPEAKER_06 47:17

Manipulation for good.

SPEAKER_05 47:20

Uh one last small comment. I just wanted to reiterate the great point that Sharing made earlier that we have power as consumers to choose which products we want to use and send a signal about the kinds of technologies and missions that we want to drive forward. So let's not forget that you know it's not just these big companies that are in control, but we have power as consumers as well.

SPEAKER_06 47:50

Whether or not AGI or ASI ever arrive, the narratives around them are already steering policy, investment, and fear. The real challenge is not predicting a date for superintelligence, but deciding what kinds of intelligence and for whom an AI world should be built to serve. This has been Thinking About Thinking, recorded live at the Harvard Work and Klein Center for Internet and Society. If this conversation reshaped how you think about intelligence, human, artificial, or otherwise, share the episode and continue the debate in your own classrooms, labs, and communities. You can find me on my website at romanchadry.com and on my sub stack at Thinking About Thinking.