Hello, this is Ted. And this is Brandon. Welcome to Concerning AI. Yeah, this is the show where we come up with all the answers, and you guys go out and make it happen?
This is where we get to look like real experts. And that's because we are. Yeah, we've not really. We're just happy who are wondering what to worry about.
Yeah, we started the show thinking we had two questions to try to answer. One was, is there an existential threat to humanity from AI? And if there is, what should we be doing about it? And now we think we've agreed that there is an existential threat.
So it seems like we should get into what to do about it. Like on the Facebook group, there's been a little, I would say even a little friendly needling in that direction. All right, so let's go right to the Facebook group to get some talk about that. But before we do that, if you haven't been to the Facebook group, just search for Concerning AI on Facebook.
And we've got a group and just request to join. And it's a really, it's a fun place to have a lot of good conversation about these topics that we come around the show. Yeah, and there's a lot of other people's voices to be heard there than just the two of us. People who know more about AI than us.
Oh, always a good thing. All right, so where do you want to start? Well, there's so many good things here. Well, let's start.
We got a comment from Ed Orman. And he noted that we're talking a lot about AI in a way that makes it sound like how we think we have kids. And we actually haven't probably mentioned this in a long time. When we very first started the show, we mentioned that we both have kids.
And we're thinking about their future. And so I thought that was really an astute observation from Ed. So thanks for listening, Ed. Yeah, and I kind of agree.
I think that is feeling fruitful to me as a potential. The idea that we need to be what we want our AI to be. Yeah. Yeah, and we touched on that a little bit in a previous show.
But it's probably going to be something to come back to. Because there are some very difficult problems, theoretical problems, philosophical problems, and how do AI's learn, human values, or values that will result in humanity not being wiped out? And it's far from trivial, how that happens. So we're not going to get into that too much tonight.
But that's something I'm sure we'll come back to it other time. Yeah, so another theme that seems to be coming up on our group, on some people that we are friends with, I've been posting about the topic. And it seems like it's the hype cycle, where we are on the hype cycle, with respect to AI. There seems to be some folks who are pointing out that there's more hype there than there is reality.
Yeah, yeah. And I think that's absolutely true. I read the headlines, and it makes it sound like there's breakthrough after breakthrough, which they are, breakthrough after breakthrough. But the headlines make it sound like each of these is earth shattering right now.
And in my mind, they're important because of what they portend for the future, not because of what's happening in this moment right now. Is that how you think about it? Yeah. I mean, I think it's absolutely the case that there is a lot of hype and a lot of click baiting stories out there about their over promising what AI is doing right now, and even what it'll do in the next couple of years.
So yeah, I agree with that. And I still think it's super dangerous. Yeah, I agree. I mean, one of the articles was saying that Google's AI has invented a new language.
And that's just ridiculous. That's just not happening. There is a breakthrough around translation. And I think it's really interesting.
But all of the news about the breakthroughs that are happening right now are really not the point of the show at all. Any one of them, all of them, put together, they're all irrelevant. Because the mistake that's going to lead if the existential threat comes true and when humanity gets wiped out or DNA-based life gets wiped out, it's not going to be because of any one of these breakthroughs. It's going to be because of, well, I don't know.
I don't know how to answer that, what it is going to be because of. But each of these little things gets us closer to human level AI, I suppose, if you add them all together. Yeah, so that takes us to a comment by Peter Edstrom. And I'm not sure if I said your name right.
Anyway, thank you for the comment in the group. And he's responding to the will, hey, I just fell in the end of humanity. The tech industry wants you to think so that story right there, where it's kind of about us, just a bunch of tech luminaries who have egos who want to be the saviors of the world, who are hyping it all up. And Peter asks about the story, is it really correct to claim that no qualitative improvements have been made in the AI field since the 1980s?
And the only improvements seen are due to increases in processing power and better handling of large data sets of big data. And then he asks a few more things about the credibility of the author and stuff like that, and how it affects that. And I wanted to talk to that a little bit. Because there's an analogy, I think, to be made from the computer graphics world.
And I want to point to a moment in time. And it's not a moment in time as in. It's actually a moment in time. But when you step back and look at it from the perspective we're at right now, it appears to be a sharp defining moment when things changed in the computer graphics world.
And that was when hardware was able to support the Z-buffer. So we had enough memory to back our display that we were drawing in computer graphics. And the Z-buffer algorithm came along. And suddenly, computer graphics got so much easier.
And that's because of the leap in processing power? Because of that moment when a graph crossed the line in processing power? Essentially, it was two things. It was the combination of this incredible richness of hardware.
It was just this huge amount of memory to be exploited suddenly. And what that spurred creatively, it was not a huge leap looking back on it. But it turned out to be that thing that was a huge multiplier in making computer graphics so much easier. So as we look at what it takes to make an intelligence, I don't know if this is a great way to think about it.
I like to think about it this way, though, of being like a stool with three legs, where there is the substrate, you've got to have the computational resources that you need to be able to do the processing. There is the learning environment, which if you look at the process, the optimization process that led to human level intelligence, that has had big data to the max. I mean, you think about all the iterations there, all of the experiences that each one of our ancestors has had with the environment and the incredible richness of that fitness function that evolves over time, that we've evolved with, big data is an integral part of human intelligence emerging. And so that's another thing, having enough training data to be able to do the training.
And the third one then is the art. What do you do with the computational resources? What do you do with the training data? So those are three things that all have to be in place in order to come to general intelligence.
And the thing that has been driving probably most of the progress is the hardware. Yeah, but hold on. So hold on a second. If take that graphics example, if the hardware had simply improved exactly the way that it did in real life, but the software designers had not taken advantage of that increase in power and were simply running the old algorithms faster, we wouldn't have gotten the Z-buffer, would we?
Yeah, I don't think so. I mean, I think it was a development of substrate and art, Bo. And I imagine that our algorithmic improvements are very much coupled with our process of improvements. They might both be just, you know, a quantum.
They might both have leaps at different times. Maybe they don't go lockstep. But it seems to me like we're not simply running 1980s algorithms today on much faster machines. Yeah.
And that is as somebody who was working in the field in the early 2000s, I can say that the things are different now than they were doing with the hardware is different now than it was then. So it's easy. It's very easy to say. And I think it's uncontroversial to say that the computing substrate is rapidly improving.
Certainly. Nobody's going to deny that. That's moving quickly. There's some debate as to whether how fast it's going to continue in the future.
But the fact that it has been moving rapidly is something is unassailable. It's also uncontroversial that the amount of data that is available for our algorithms to play with, the amount of labeled data that people are labeling on the internet all the time for us, you know, pictures and all these kind of different things, that is growing so rapidly as well. So you have two of the legs of the stool that are moving very quickly. Right.
And they're tied together because you couldn't have so much data being created all the time. If we didn't also have the processing power that is so rapidly increased in order to deal with that, we just wouldn't have been able to create that much data. Yeah. So those two things are, those two things have a very steady growth rate, a very relatively predictable growth rate.
And there are some things come along that will shake things up now and then in those two legs of the stool. But the really unpredictable one is the art. What happens when a breakthrough happens? So like we had one earlier this year with AlphaGo.
Would you say that that was a breakthrough in art? Yeah. You know, again, it's important to recognize that the breakthrough with AlphaGo, if you just look at what they did different compared to what people were doing in the past, is modest in terms of what's happening with art. You know, and a lot of what they did is dependent on the training data and the computational resources for sure.
But it wasn't as if they, a year before they had to match with Lease Thydol, they were beating a lower level pros and they had to wait for the processing to get faster in order to beat the higher level pros. They came out of nowhere. Yeah. Yeah.
So they looked clever things, you know, in their, in the way they put existing techniques together. So when we say clever, I mean, it's clever, a backhanded compliment, is that a kind of a, are we saying that because Go is such a limited domain that it's not necessarily generalizable, you know, because it's so specific or is that why I use the word clever? I guess I use the word clever because there is still, it's not like there's a general purpose intelligence that you could point at the game of Go that would just figure it out and learn it quickly. So they had to take existing modules and figure out what, you know, how they were going to spend their computational resources.
They don't have enough computational resources in the first place, right? There's not enough. So they have to figure out how to use it kind of cleverly. And so is it a compliment?
Yes. Is it disparaging? Yes. It's both of those things.
It's, it's trying to put them in their place, I guess, to some degree. And that places, AlphaGo is a narrow intelligence. For sure. There is no tremendous breakthrough there.
The fact that the existing techniques given the training data and the hardware worked so well with just a modest tweak, I think, is part of what makes me so cautious in saying that massive breakthroughs are going to be needed. Like that, it worked so well, is surprising. Oh, so you think that the cleverness might be bad news for the future because it might be, there were just a few other clever breakthroughs away from serious existential danger? Is that what you're saying?
Kind of. I mean, I think I'm more saying that, you know, the clever is a little bit tongue in cheek in hindsight looking at what they've done. It doesn't feel like it's the sort of thing that only, okay, so when Einstein came up with general relativity, that was a mind-boggling leap. That wasn't kind of like an iterative move forward in physics.
That was something that, you know, that took kind of an unusual breakthrough. I don't think a lot of physicists would look at that and say, I was just right on the cusp of that or I could have done that myself. But the AlphaGo stuff is something that almost any practitioner in the field could look at and say, oh, yeah. Okay.
So that's how that works. And yeah, that's an iterative kind of set of connections that I could have done. And if they had been part of the AlphaGo team, you know, maybe it would have been in replacing somebody who made a really smart move on the AlphaGo team, maybe it would have taken an extra year or two or five. But anybody in AI could say, yeah, I could have, if I was part of that, I could have come up with those ideas too.
Is that what you mean? Yeah, it's just not like just coming out of left field. The result is coming out of left field. That it works so good is coming out of left field.
But what they did is kind of like, oh, yeah, okay, that makes sense. You follow me there? I'm going to make up some huge new paradigm machine learning here. No, that's not what happened.
Yeah. So I get the question about the only increases due to improvements in processing power and data sets in one way of looking at it. I think that's true. But in another way of looking at it, I think it's scary that that's true.
And it isn't like there hasn't been any progress. I mean, our theoretical understanding of what is going on in certain algorithms and how they relate to other algorithms. What is the math that connects this sort of algorithm with this kernel, this sort of network, with this kernel and this learning function to this other that seems on the surface to be completely different. But what is the different, true difference mathematically between their optimization functions?
We're getting a lot better at that. So the sort of breakthrough that I imagine happening is the breakthrough or the kinds of breakthroughs that I'm having that are on the order of like a Z buffer are when we start being able to learn the meta parameters, being able to learn those things that are black art right now, that the AlphaGo team got right. Yeah. When we can learn those things, then it's like the Z buffer in that it suddenly makes so many things easier that I could see rapid progress happening.
Yeah. Okay. Anyway, so that also ties into another thing that's come up in the Facebook group. And that is the embodied AI.
Who was that? I think I might have a Nick Jarvis maybe. There were a couple of people talking about it in the group about whether you have to have a rich world. Yeah, it was Eric was sort of going back and forth with, I think there was another person too.
Who had some really good things to say. This is a conversation between Scott Veitch and... Yes, yes, that's him right there. Yeah, clearly someone who's steeped in the stuff.
So glad that you joined us and look forward to hearing more from Nick, Scott and Eric have kind of had an interesting conversation about whether or not you have to have a really rich world in order to have a rich intelligence or not. Yeah. So if you sort of jump to the end of that conversation, you say, all right, well, we have an intelligence that is a general intelligence today in 2017. What if we were able to put ourselves into five-dimensional space?
Would our intelligence who has evolved clearly in this three-day or maybe four-day space, depending on how you look at it with space-time, if you were to put us in five dimensions, would we be able to function there eventually? What do you think? Would we? I think...
I think no. I'm one of the people who is in the embodied AI camp. So even though our intelligence is general, it's not extensible in that way in the short term. Well, it's general in the world we live in.
So if we created a simulation, we were able to put ourselves in immersive VR that actually worked and somehow gravity was different, that we would be broken or was that something we could adapt to? Something like one of our constants changing modestly. Yeah, maybe something like that we would be fine with. We don't have the hardware and the apparatus that has evolved to deal with another dimension.
We have trouble visualizing what four dimensions could even be like. It's not accessible. It's not accessible. Yeah.
There's some really weird brain research where people who are disabled in different ways where they don't have access to one of the senses that the rest of us have, they're able to wire up input apparatus into a different sense. Like I've heard about people who are blind, who they hook up electrodes to their tongue. The sensors to their tongue and somehow the taste sense gets mapped in the brain to the visual cortex or something and they're able to see. But that's in a very rude sense of the vertices.
Sure. But those are different. That's a different thing. The existing apparatus and rewiring it.
Whereas the idea that somehow we could work in a vastly different environment like a different dimension which we can't even imagine what that would be. So Eric was trying to suggest that maybe our simulations could be much simpler than the existing world and they could result in intelligence that then would be generalizable anyway. Is your position that that would work? I don't agree with that.
Yeah. I don't believe that would be the case. I think it's partly a case of intuitively where I come from right now that doesn't make sense to me. So I post-hoc argue why I don't think that that's true.
But I'm just going to flat out say intuitively it doesn't make sense to me. An example that I think really illustrates it for me is we are so proud of our ability to do math and our ability to do chess and our ability to do some of these things that set us apart from other animals that we know. And it turns out we really socket those things. By the way that breaks the mind of article I didn't post.
The Russian Go Federation posted some games between Korean professionals and they're saying all Go textbooks need to be thrown away. So it's totally hilarious. AlphaGo has changed the game of Go for top professionals in the world. Yeah.
No it's totally going to be the case that AlphaGo and this is going to leave us into the bus. It's just, yeah, it's not a, and we'll learn in that, you know, we'll learn from that. In the wake of AlphaGo we will learn things that we didn't know before. Yeah, that's what they meant.
Right? That's what they meant. Right. Right.
And what you're saying we're going to continue to suck it go as AlphaGo is going to continue to get better. It'll get better faster than we're getting better obviously. I mean, that's just how, I mean, as long as anybody puts any research time into it today, AlphaGo is not going to self-improve. So this is the crux of the issue is our minds are evolved for a certain game and that game is not go.
For sure. It's for making our way through the rich complex world that we live in that includes social interaction and having to make ourselves look good and explain why we're right and all of these things that we do that we are really good at. It's going to take the computer a lot longer to learn to run through the forest and to make, you know, make a good burn on somebody, put somebody in their place because those are the things that we have evolved for. And you know, even the putting someone in their place is a, it's a pretty late part of our evolution.
So we're probably not that good at that either. So that part I think will come easier to the computer than will come like this thing about running through woods without falling down and some of the other things like that. So it is amazing how much progress we're making on that front too. How much progress the machines are making on sort of being embodied and they still are pretty terrible.
Yeah. But you don't get all of our analogies and all of the things that we use when we think and talk to each other without an embodied sense of what they mean. If they don't make any sense unless they are grounded in sort of a perceptual understanding of the world, for which for us it's like a visceral understanding more than it is a symbolic understanding. Anyway, this is me, Rant.
All right. So that's some people believe other things and they can be right. So I'm feeling like we're getting pretty close to the end of the show. And I got a couple of things that I want to say how important we are, you know, because my ego, human ego has evolved to want to be important.
How about we be important? Let's be important. Okay. So Eric made a comment.
He said, what if concerning AI is the cause of unfriendly AI? What if by spreading concern, you're really spreading fear which breeds hate and mistreatment which may be reflected in the AI that are mistreated back onto us as hate and extinction? We're so important. Yeah.
Wow. Yeah. That's kind of like that. Is it roko?
It's kind of like the thought exercise that you're not supposed to think because don't say it. Yeah. Right. The oscilloscope.
You've already blown it Brandon. Yeah. No, it's possible. I think we're probably not very important, but still the sentiment of what if all of our worries about AI and by our worries that we generalize from concerning AI to the wider world.
It's kind of concerned about it right now. What if that does actually create the dynamic in which bad outcomes occur? Yeah, totally. And I just looked at the stats for downloads of the concerning AI podcast and we have downloads now from over 100 countries.
So I wanted to send a shout out to those of you all around the world. It's easy for us to forget that the internet is vast and that we're not just speaking to Americans like ourselves. So for those of you from around the world, thank you for listening. From those of you from the US, thank you for listening as well.
We have about 45% of our downloads that looks like are coming from the US or from people who are spoofing one way or the other. So I just stroke my own ego there. There are people listening from all over. So thank you for listening.
If you want to take part in more of this conversation, please join us in the concerning AI Facebook group. We'd love to have more of you there. It's really fun to check in on those conversations and try to keep up with them. There are better folks than us to talk to there.
So it's good. It's really good for someone else. Absolutely. Thank you so much for listening.
If you want to subscribe, then you'll get the next show where we put it out and that equals. Alright, so until next time, this is Ted signing off. Well, and this is Brandon also signing off. Bye.