Welcome to Concerning AI, this is Ted. And this is Brandon. This is your first time joining us. You may not know, this is a show where we talk about two questions.
One is, is AI, is the development of super human AI potential existential threat to humanity? And we think the answer to that is... Yes, we think so. And so then the second question is what do we do about it?
Yeah, and that one is hard, but we've got a couple of nuggets that I think surprise the old Brandon, and maybe the old Ted, definitely, the ones that hadn't done the podcast yet. And the most surprising one to me is the idea that we should go fast. Yeah, because if we go fast and develop AI before... Superhuman AI before the hardware gets so far out ahead of us, then we may still have a chance to prevent the calamity.
Yeah, and that one, I don't know, it feels like a little bit like wishful thinking, but... I should think it's potential. Yeah, we're not going to really talk about that today, but we talked about previous shows, just want to catch you up to where we are. And then the other potential answer is actually the best example of it is Elon Musk's company, Neuralink.
He's not the only one who thought of this, of course he didn't make it up. But the idea that with fast brain computer interfaces, so the augmentation route might be a way... If we augment ourselves fast enough, we might delay the crossover of human and machine intelligence, and therefore be smart enough ourselves to solve the AI safety problem. Yeah, do a better job in time.
Right. Yeah. Yeah, and that one is... I'm enjoying.
Yeah, it's exciting. And there's also the locality principle that is poorly named, but I like and Ted is less sure about that there are maybe fundamental constraints in the universe, like the speed of light that end up creating a proliferation of minds, because stuff that's close together is so much more connected than stuff that's far apart, that you necessarily will have these sort of dense consciousnesses that you can't spread out, spread way out and still have operated at the same efficiency and speed. So there are not a lot of advantages to turning the whole universe into computernium, because it will just be making a bunch more copies or a bunch more other intelligent machines. Yeah, so it feels like we've exhausted kind of how to talk about that for now.
So why don't we leave those three concepts in the archives, right? And if you want to go back and listen to those, you're free to do that. We thought today we would talk about just speculating, how could we get from here to general intelligence, to superintelligence, artificial general intelligence? Yeah, yeah.
Like what is... If we had to predict from the vantage point of today, what will generate the first general intelligence, the first artificial general intelligence? Yeah, I was thinking we could just throw out some random speculations from whatever data points we like, and just see what builds on what? Like, for example, we could say, well, let's start talking about Google.
Let's start with what they've done so far. What might they do next? Would self-driving cars have anything to do with artificial general intelligence? Or what about deep mind?
We've seen some examples of advances in the field of artificial intelligence. We've got tools. We're making them better. What happens when we make them better the next step and the next step?
Do you have a different theory about how we should talk about this? Yeah, no. Well, that's good. Let's start with AlphaGo, because there's something there, and there's parts of a solution there.
So one of the things is the ability to do fast association. I perceive X that implies Y, and that might be that implies I perceive X and that implies that doing this thing would be a good thing. That could be an implication that Y is an implication that something is about to happen. So the ability to understand and make an instantaneous assessment of the state of the world.
And the world of Go is a constrained world, but it's too big to be fully comprehended. There's too many possibilities, even though it's a constrained world, for the purposes of the AI, it is as if it were constrained. Is that true? Is that an accurate statement?
Yeah, you can't know the true state of the world because that would require a truly exhaustive search to enumerate all of the possible pathways and evaluate based on enumerating every single possible pathway, which ones lead to the outcome you desire. That's not possible in Go. So we could generalize that to other problems which don't seem to be bounded. Problems like, I don't know, diagnosing diseases.
Or yes, diagnosing diseases is a good example of a somewhat stationary prediction problem. Given this moment, how accurately can I assess the state of the world? And how well can I predict, yeah, how can I predict hidden variables, which in this case might be what disease is causing the state of the world? Or lack of disease or whatever?
What's the health of the underlying person that is leading to the observations of the world that I see? And we're talking about AlphaGo now, but we know that Watson, the computer that won at Jeopardy, has been, there's a lot of experimentation going on. IBM is selling this as a product. And in the medical world, they're doing a lot of experimentation trying to figure out whether it can help predict disease.
Yeah, so I like to think of this as a function. Give me an input and I'll give you an output. And the input is going to be some state and the output is going to be some assessment of that state. Maybe fill in in some gaps, maybe suggesting how well we're doing or predicting, predicting some underlying state that's not visible, that's part of our model, it's this instantaneous kind of thing.
So that is deep learning and neural networks and all different kinds of classification algorithms and the regression algorithms as well. So predicting a set of labels versus predicting some real value functions, it doesn't really matter. It's the ability to predict and fill in the world instantaneously. So that's one part of AlphaGo.
Do you want to talk about the other part of AlphaGo? Or would I? I do. I want to get to the other part because then maybe we can then start to look at what's missing.
So that, I guess the last thing I want to say about that first part is I can imagine that just getting better and better to the point. And we're close to superhuman performance there on that whole class of tasks, mapping some input to some output. An example of that might be tagging photos. That's with the label of who they are or the search.
The algorithms are doing pretty well there. They're close to human level performance and I can imagine them doing better. And they do better in some cases than worse than others. So that part feels like it's not that far off.
But that just feels like one part. So another part is that AlphaGo used is reinforcement learning. And that's, I haven't done a lot of work myself with reinforcement learning, so I might get this wrong. But it's really about these long chains of states and action pairs and future rewards.
So how do you get from, it's a bit about planning and finding your way through. It has time built into it, I guess, in a way that is kind of absent from that other task. It's an instantaneous function. Give me X and I'll give you X as an input and I'll give you Y as an output.
And the better I can do that, the better I can do. But how do I get to a great vacation in Hawaii? Well, from here, it is not that there is, that's not a function that maps an input to an output. That requires a whole sequence of actions and responses to changing state where I'm trying to steer the world toward the outcome of sitting on the beach in Hawaii or snorkeling in Hanama Bay.
There's some outcome I'm trying to steer the world to. But it's a long-term dynamic process where the reward is way, way, way out in the future. Reinforcement learning helps to propagate future rewards back to the actions that led to them so that you can learn good actions given a particular state. And if you think about Go, an example of that is given that I'm in a particular board possession, how do I know what a good move is?
Well, I don't know. As soon as I make that move, I still don't know whether it was a good move or not. I'm not going to know whether it was a good move or not until another 100 moves down the game or 50 moves down the game or whatever. At which point I'll have one or lost.
And even then, I'm not sure that it was a good move because maybe there was a mistake made on the other side that they didn't find an optimal way through. So that framework for taking future rewards and propagating them backward through these uncertain actions and states to improve your ability to make snap judgments about what's going to move you towards a reward. In your Hawaii example, is this where you learn whether buying a plane ticket was a good idea or a bad idea or not buying a plane ticket was a good idea or a bad idea? Otherwise, how would you know that buying a plane ticket was a good idea that helped you reach your goal of sitting on the beach in Hawaii?
Yeah, so that's an important, that's like that would be one of a major learning that you would probably get in the process of creating an artificial general intelligence that could get itself to Hawaii. That would be something that it would have to learn. Yeah, and it would have to learn that swimming there probably is not the optimal path. Their plane is better.
Right, and it would have to learn also that things can change. Like your seat can get canceled or that you need to show up early to the airport and that sometimes traffic is a problem getting to the airport and that you need your passport or your license or something to be able to prove who you are and not to bring razor blades through the straight razors through the security checkpoint. And it's just a myriad of things that can derail that goal and a myriad of things that can sort of incrementally improve you towards it. And okay, here's an important part.
You can't make a perfect plan in advance because there are all sorts of external things that are affecting the state of the world. It's not just your plan that affects the state of the world. Right, your flight canceled. Yeah, your flight being canceled.
You didn't plan that. It's not in your plan. Right, but based on then the new state of the world, the state of the world of I'm at the airport and the flight I was planning on taking is not happening. So now what do I do?
Yeah. Yeah, so it's so reinforcement learning feels or something in that area feels like a critical piece of a solution. The ability to deal with long-term rewards. All right, so can I ask my question now?
Can we ready for that? Yeah. Okay, so as we think about AlphaGo or Watson, maybe these are, you put these both in the same bucket of what they're showing us in terms of advances? I wouldn't.
No. Okay, so let's look with AlphaGo. So AlphaGo has given us some breakthroughs in the world of Go and we know that they're retiring from playing Go. They're going to start using the same concepts on other problems.
So what other problems could, I mean, you know, getting us to know why is one that doesn't feel like a real world problem because we as humans really do well at that. How is that going to help us to have an AI work on that? I don't know. Can we think about what other problems we might work on or does that matter?
Or is the idea that we know we're just going to pick up incrementally, we'll pick up more and more complex problems that algorithms can't help us with and we'll start solving those with machines and that'll help us. What is this here? Yeah, well, that's a good question, you know, because every time we build a narrow AI, like even a self-driving car, like a self-driving car would be an example of something where these two abilities could mostly, I think you probably solve that problem with these two types of learning, you know, hybrid system that uses those two things. But still, how close are you then generalizing?
I don't think you're close at all. It's a tool. It's a tool that gets better and better. Yeah, so it does feel like there's something missing there.
Okay, you know what I think is missing? What? I always come back to this, always come back to this. Being grounded in the world.
I was going to say something different. What's your answer? My answer is self-improvement. What does that mean though?
I mean self-improvement. What does that mean? Well, the ability to examine the algorithm having the ability to examine its own process and get better at achieving whatever the goal is. Although maybe these other things already do that.
Maybe that's what reinforcement learning does. Yeah, I mean, it's just an optimization problem, right? You've got some, you're just optimizing some cost function in either case. Yeah, maybe the embodiment is actually the key distinction because AlphaGo doesn't move the pieces on the board itself.
Although it could, right? We could build robotic extension. We could build a robot that AlphaGo could manipulate the pieces on the board. But that wouldn't lead to general intelligence either.
Yeah, no. It would be like what I think would lead towards general intelligence is bootstrapping semantic concepts. So bootstrapping, some sort of qualia, I guess, is one of the terms that gets thrown around that's often used by philosophers to point at these sort of magical ideas of the soul, that leads towards a soul or some non-mechanical sort of magical property. But that's not what I mean by it.
I mean by qualia, I mean the canonical example of qualia that's often used is if you had a woman who lives in a completely black and white world, like it's been artificially created to be completely black and white. And within the confines of this bunker world that she lives in, maybe she even has glasses that only let certain frequencies light through. So she's only seen these things. But she has access to a lot of educational material and she can learn all about the neuroscience of color perception.
She can learn all about the physics of color, how, you know, what leads to different wavelengths of light, all that stuff, you know. So she could become like the world's foremost expert in sort of mechanics of color perception. But as soon as she goes outside, takes off the glasses and looks at a red apple that she will have gained some new insight into color perception that she did not have before through the experience of red. And so she will then have this, this qualia, this internal experience of a thing that is a kind of a meaning of the thing, of this real meaning of the thing.
So coming back to go, an AlphaGo for a moment, AlphaGo is not going to talk about a tiger's mouth. AlphaGo is not going to talk about a bamboo joint. AlphaGo is not going to talk about pace of a move, whether it's slow or fast, light or heavy. These are concepts of a general intelligence being applied to a specific problem.
They're embodied qualia, if you will, maybe qualia is still not the right term, but they are these chunks of meaning. It makes the question whether they actually are meaning or not. Like constellations of stars in the sky were also human imposed meaning on the randomness of the nighttime sky. We added meaning where it wasn't actually there.
And I wonder whether these things like the blue joints and tigers mouth are not the same. Oh, no, I think there's, yeah, okay, keep going with this because it's interesting. They're not there in any sort of objective sense that there's, yeah, I guess the only, maybe this is a utilitarian argument that you're proposing, that things are only really there if they're useful somehow. And constellations were useful because it let us track seasons and time of night and directions and other things like that, that were useful.
Yeah, it was useful to be able to refer to particular stars, certainly. But the actual animals and all the mythology that went with it was not useful. And I wonder, the idea that this is a bamboo joint that there are two stones on this side and then a space and then two stones opposing it. Does the name bamboo joint or the concept bamboo joint or what bamboo is in the world does that actually help us understand go?
Oh, it doesn't, it doesn't help us beat AlphaGo, right? But it helps us beat another general intelligence that doesn't have the concept. Oh, it doesn't. It hasn't, it hasn't made that happen.
That theory is interesting. I don't know whether that's true or not. That's interesting. If AlphaGo does then have some embodiment and know what bamboo is and what a bamboo joint is and have that label for that configuration of stones, does that by its nature give it more insight and the ability to beat the previous version of AlphaGo?
They didn't have that. Well, let's start. Let's take AlphaGo and say, you know, what can AlphaGo learn about driving a car based on what it's learned from playing go? Like it has the ability to say, I like this.
This is, this is, this is, this is in AlphaGo's sort of, my mental space. This is a beautiful move. This is a very effective sort of thing that's going on and go. And here is a situation in driving a car where that beautiful, effective thing I learned is also beautiful and effective.
Like how do you, that's the, that is the critical piece that I think we need in being able to generalize is having a type of knowledge that can be, that can be mapped onto many different situations and humans do it through sort of metaphorical thinking. Like this reminds me of a, like what bamboo looks like. And boy, bamboo is strong, right? So this bamboo joint is uncuttable.
This is a very strong thing. You know, tigers are very aggressive. This tiger mouth like this is a very powerful, aggressive shape that just radiates out in this direction and it looks like the mouth of a tiger. And everybody knows how strong those are.
Let's, let's, let's make more tiger mouths. I'm not sure that this is, is fruitful in the go case because I think in the go case, these concepts are shorthand that allow us to get beyond the complexity where AlphaGo already has more ability to get beyond the complexity. It feels like AlphaGo doesn't need these short hands. Does it need these?
Well, AlphaGo doesn't have to get beyond the complexity, right? AlphaGo embraces the complexity because it's focused and specialized on this tiny, tiny little domain. But which is go and it's using highly specialized reasoning to deal with the complexity there versus finding its way through the complexity through generalization. Yeah.
And that's what I mean. The idea of generalizing is not going to help AlphaGo become a better go player. The idea that AlphaGo knows about bamboo now, not going to help AlphaGo become a better go player. But not what this concept of being embodied and therefore having these physical concepts that you can use as analogies or use to generalize, that feels like really fruitful.
Yeah. So, so one of the things people can do is we can come to a situation called that we've never been in before, a game, a sport, a sort of mental task, and we can use our general knowledge to bootstrap a decent, a decent model of how to make effective in that situation that is, you know, it's decent. And we'd have very few training examples to build it off of because we have all of this, you know, we have all of these other training that's been embodied in these other concepts that we know. So, you mean, if I'm dropped down in a country where I don't speak the language, I have my general intelligence that there is language, language exists.
I have language myself. And now I'm in a place where I don't know the language and I can use my general intelligence to start to figure things out. And at first I'm going to make a lot of mistakes, because I don't know what the heck people are saying. But eventually I'm going to hear some things that sound similar or I'm going to use my pointing and saying this is a mango and they're going to say this is a whatever, they're going to give me their, you know, I'm going to start to build up new knowledge based.
Is that what you're talking about? Is that the concept here? Well, maybe I guess maybe a different example would be, you know, like opening a door. There are lots of different ways to secure a door, different kinds of latches, that sort of maybe if I've been through a few doors, I kind of have this idea of what it's trying to accomplish, which is a stay shut and then go from shut to open in a very clear transition and not be ambiguously in one state or the other unless I want it to be, but really to be either fully open or fully closed and not going to accidentally switch between those two states.
So I can take that, I've walked through doors and I can use that same thing to actually figure out how to get the oil because the same general sort of principle of this two state where it's either in one, clearly in one or clearly in the other with no accidental transition is a kind of generalizable knowledge. Okay, so from our tool based development of AIs that we're doing now, AlphaGo is designed to solve Go and then maybe we can use the same infrastructure architecture to solve self-driving cars potentially. It could be transferable, not the knowledge but the methods. Right.
So from that point of view, having a door opening AI, it can learn about all kinds of doors. I actually ran into a door, a latch the other day that I couldn't figure out for quite a while, but that's neither here nor there. It was designed to keep kids inside the playground, but let adults open the door for the little ones. It actually had to pull up on this on this pole that went up in the air.
It was very strange. Anyway, that's neither here nor there. Being able to generalize that your door opening AI into taking the oil pan off your car in the tool building world that we were in today, it doesn't make sense. Why would you want that?
Why would you want your oil pan AI to have any knowledge from doors? You can just train it on oil pans because it's a specific, you know, it's narrow, but you're trying to go beyond the narrow. Now, what, at what point do we start to think about solving problems that it hasn't seen at all before rather than just train it on narrow things? The narrow thing is that we want to be able to give it a goal that's abstract and have it be able to figure out how to solve it.
Like your goal of this abstract is, I want to get out of the playground. And so then, you know, like it all depends on the playground that you just landed on and all sorts of things, but because you know how that this is a gate, because you know what's going on in the world, you know, keeping the little ones in, the big ones out. So instead of carrying my camera and my flashlight and my compass, I now carry my iPhone. It's all there.
I don't need to carry my door opening AI and my oil pan taking off AI. It's all one. Is that the idea? So that's one of the ideas.
And another part of the idea is that the world is full of these novel problems that we haven't taken two months and a billion generations to train up an AI on. Like AlphaGo got to this like, this luminous go playing position through billions of iterations and you know, months, years of developing itself. So, you know, how could I build a machine intelligence that could go from zero on go to like, at least following the rules and playing with some modicum of strategy, you know, within an hour, like a human would and two examples, not a billion examples or something like that. Yeah.
Yeah. And not by teaching it the way we teach a human. I mean, AlphaGo, I assume to drive things from just putting teachers on the board, like the rules themselves, the first principles didn't learn the way we learned. Yeah.
It's funny. AlphaGo probably doesn't actually know the rules to go. Maybe it does. Maybe it has built into a pruning part that will keep it from making illegal moves.
But there's only like one or two pretty simple cases where it's on legal move and go. Well, it begs the question of whether we're teaching, whether we teach correctly, right? Or whether we teach in the best way for humans, should we be sticking more to first principles and letting people derive, you know, the axioms or whatever, you know, derive what's good and bad rather than saying, well, this is good and this is bad, you know? Yeah.
No. No. That would be incredibly inefficient. No human would learn Go if we did that.
We can't do a billion iterations to learn Go. No, I know. That is not available to us, right? Right.
If a human were to discover a ladder rather than be taught what a ladder is, you know, would that be different for them? Yeah. Why do we call it a ladder? Well, does it matter what we call it?
Does AlphaGo call it a ladder? Certainly not. But can AlphaGo choose good moves around a board space in which a ladder is a imminent possibility? Presumably.
Sure. But it's not going to think of it as a ladder. No. In fact, it probably won't even like put down its finger on the board and trace out this jig-jag.
This is where it hits a ladder doesn't work. Or, oh, hey, the ladder works. You know, like, no, it doesn't do any of that. That is all applying our embodied knowledge of moving our bodies through the world to this slightly more abstract space.
Okay. But back to how this is going to happen, right? Desensively, that's our topic today. There is a business model where a company can make money with a program that is a narrow AI.
There's a specific problem in the world. There's a specific thing. And we're going to try to solve it. And then we can sell that thing because people want that problem solved.
Medical diagnosis might be a really good example. Yeah. This is awesome. Keep going down this road.
So narrow AI's are awesome because I have a nail and I need to put it in the board. What do I need? I need a tool. I need a hammer.
It's a hammer's a lot better than a rock. And so now I've got a hammer. And somebody can sell me that hammer. And so that's great.
There's a sick people. We want to try to prevent those illnesses or we want to try to treat them. And so that's a narrow AI is called for in this time because I know I can make money by selling this thing. So the marketplace has strong incentives that so the dynamics of the marketplace need to mini narrow AI's.
Right. Right. So in our I-Zone example, is there an incentive to take those narrow AI's and generalize them and put them in a package that can solve many problems? Is there a wider market because I can sell the same AI into lots of markets or is it more efficient to build specific tools for each problem?
See, I would argue the market. Yeah. The short term markets, it feels like a low capture by a local optima for narrow AI's. Right.
Like we want to solve this problem. What's the fastest way from point A to point B? It's narrow AI that's focused on that problem. No doubt.
I think that's that feels like not particularly debatable even. Okay. So it seems like a foregone conclusion. When the iPhone was introduced, it was not the best camera in the world, right?
But it was an adequate camera and it or maybe it wasn't even adequate. But it was it was introducing something where now you can make the argument that the iPhone is one of the best cameras in the world. So does the iPhone example, does it change when we think 10 years from now? If there's a local, if we're stuck in a local optima, does that change just by each each of the tools getting better or somehow they can then be generalized later?
I still think at some point in time we're going to, okay, so this is making me think of something. So this is making me think that the AGI, so there's many different pathways, right? One is whole brain emulation. So just start from a human brain that we know is already generally intelligent and bolt things onto that.
That's fine. But we still don't know how it works and we still didn't get there from first principles. But getting there from kind of from nothing bootstrapping our way up, I think is going to require a similar process to what led to the human general intelligence and that definitely requires this embodied meaning to the ability to sort of have this almost visceral understanding of the world and what it means to succeed and fail in getting around the world and stuff like that from a body perspective. One, and two, I think the dynamics of that might be that we are from market dynamics that could lead to that might be that we really want to be able to talk to our digital assistants as if they, you know, in our language.
Yeah, hold on a second. Before you go any further, I wonder, we're pretty long on this show and we did a pretty deep dive on tools, specialized tools getting better. And does that seem like a path to AI? We didn't cover any of these other things.
This idea of the personal assistance is that the path to general intelligence or the whole brain emulation as the path to general intelligence. Should we pause here and go back to these at another show? Maybe each of these needs its own, you know, needs its own exploration. I feel like we did a pretty good job thinking about tools.
Right. And that feels kind of a little bit like a dead end, I think is our conclusion there. I think so. It's not clear how the tools, the individual specific tools that are now rewarded by the marketplace, how that jumps to a GI.
Yeah, it's maybe possible, but the pathway is not clear from here. Yeah. So maybe we ought to talk about personal assistance next time. I think that would be fun.
Yeah. Or soon, let's say soon so that we're not, you know, maybe something else will come up there. We want to talk about next time. Yeah.
But soon, soon, we should talk about personal assistance and maybe whole brain emulation is another one. I don't know if those are different or they seem different. Those are so completely different. So we have it.
So there's two possible topics there. And maybe there's others that we haven't thought of. Yeah. One is one is a problem domain, like the personal assistant.
And the other is a pathway to general, you know, a technical pathway to general intelligence. Yeah. They could solve the personal assistant thing, right? Yeah.
Or any other problem. And we haven't really gone deep in this embodiment issue either. Like robots, if we go with what's the name of Boston Dynamics, like is Boston Dynamics a path towards AGI, because they've got the embodiment built in and they're making the brains better. Maybe he's also like really good topics for future.
But all right. So if you want to get involved in this conversation, if you want to give us some feedback, record an audio message on your phone and email it to concerningai at gmail.com or visit us at Facebook. We're in the concerning AI Facebook group. And until next time, keep your Wi-Fi up.
Keep it up.