0001: What is Intelligence?

This is Brandon. This is Ted. And we are doing Concerning AI as the episode one. Episode one on intelligence.

What is intelligence? Such a great question. And a very intelligent question. Well, I don't know that.

We would, it means to be seen. Yes. But we're trying to answer this question for ourselves. We're trying to have, I believe, a shared understanding between the two of us of what is intelligence for the future of Concerning AI so we can even talk about what artificial intelligence is.

Well, are we going to get a precise, absolute, universally correct answer to this question? I hope so. I think it's undeniable that that's where we're going to end up. Well, actually that was one of the first things that we ran into is that there's seems to be that the humanity doesn't actually know the answer to this question.

No. I mean, there's been a tense of this where they get 100 or so experts on a boat and try to hash it out and only 50 of them end up signing the statement at the end of the thing. That kind of result. Yeah.

So it's some of the examples where of other topics that we knew operationally what it meant enough to make progress, but that it took hundreds of years to actually formally understand what the, what's an example was numbers. What's a number? That's crazy. We've been using numbers for thousands of years.

But it wasn't until recently when we had actually a mad except theory that we came up with a kind of the current modern understanding of what a number is. Yeah. And then the other example that really I really liked what it resonated with me was the idea of a self-driving car. The word self-driving car is relatively new, but people have been trying to have cars driving themselves for a very long time.

What if you had a vehicle that was pulled along by a string under the road, you know, like on a track, like a train for it. I mean, is that a self-driving car? You know, nobody's had a good answer for that. But there's been lots of different efforts.

What if you program in where the turns are and you're on a closed course? Is that a self-driving car? I think there's been progress in the last few years of, you know, now Google has these cars that are actually driving on regular streets. Hundreds of thousands of miles without a crash.

But is that a self-driving car? Yeah, and they still struggle with certain parts of the driving experience. No. No.

No emerging and they get that. So fascinating. We don't really have a good answer. And yet it's clearly happening.

Something's going on here. We have a good enough answer and the people who are working on have a good enough answer to make forward progress and in the process comes to a better answer. Yeah. So I think that's what we're seeking here is some kind of rough approximation that is good enough for right now.

All right. I'm okay with approximate. Okay. All right.

Sounds good. So, do we talk about some examples of answers that people have given? Would that be useful? Sure.

Sure. You want to come up? Why don't you bring what you got from the Quora person? Yeah.

So I don't remember which of my notes here are from where. Obviously I have a couple of examples of Go for. So, intelligence could be a set of abilities required to acquire, process, synthesize and put to effective use of knowledge and abilities. Too long.

Yeah. Clearly. But it's the idea that we're going to get some information, knowledge, and we're going to then be able to put to use. That's kind of a sense.

So learning is a critical part of that definition. Yeah. Intelligence is the ability to purposefully process and relate information. Learning again.

Very important. That's right. So that person goes on to say, in other words, it is the curious ability to want something and to detect and match patterns which will increase changes for that outcome to be obtained. So, maybe I was excited in learning a little prematurely.

Well, yeah. I think that this idea of there is a state A and there is a desired state B and how do we make changes in order to advance ourselves from state A to state B? If we can to successfully do that, that would be one minute of intelligence. You know the phrase I love that points at that that I got from Ella Yeezer.

Is future steering. Future steering. So that future steering is to create a future that does not exist in the present. There's some difference between the present and the future.

I would say it's less creating. Like steering to me has less about creating something than about changing the dynamics of the situation such that we're going to get to the future. But let's get to the future we want. Not the future we don't want.

So we're just going to steer this massive momentum thing, which is the course of time progressing. And we're going to steer it where we want it to go. Okay. So wanting.

What does it need to want? Well when I meditate, I notice that I want a lot of things. So I wonder how much wanting is a critical part of intelligence. Right.

I mean, is this so hard to think about? Like those are dog want things? I mean, it seems like a dog gets hungry and then it wants food. Is that intelligent?

So then go to the food bowl to look for the food? Well, you know, one of the definitions of intelligence has to do with optimization, external measures of your ability to optimize the world to match up with your wanting. So I would say a dog that like tears through a food sack in order to get to their food. It's optimizing a fairly challenging problem to get to what it wants.

Yeah. So I would say, yeah, the dog wants. Yeah. And so we have a buddy rabbit at our house and that the dog really wants the bunny rabbit.

We have a dog. But the way rabbit will not add the metal plasps on the cage. And because I built the cage myself from those super metal things that you put together with plastic connectors, the bunny sometimes is able to push hard enough to be able to break the cage to be able to get out. But the bunny fixates not just on anywhere in the cage, but fixates on the weak parts of the cage.

That's where the bunny know is where the bunny pushes. The bunny always breaks the cage and I look at what it's doing. It's in the weak part. It's in the part where the hinges or the doors.

Huh, which is fascinating to me. So there must be some kind of desire there too. But some sort. But can a bunny actually be intelligent?

I mean, it's not human. I don't know. It doesn't appear intelligent to me. I mean, I don't relate to it as if it's intelligent, but it's clearly wanting to get out of that cage.

There's some want of some kind. Is that the program to want? So maybe we can agree that humans aren't the only things that want things, that there are other DNA based life forms. So even if the wanting is instinctual, well, so there is nothing in the DNA that says I'm in a cage, I want to get out, right?

Because there's no cages in the DNA. There's no concept of a cage. But there is some desire that we're exhibiting. Yeah.

So while this brings up a fork in the road, are only people intelligent? Or is there a spectrum of intelligence that we can go to? Because it seems like a lot of people have to purmorphize intelligence. In fact, I think that's the position that I naturally fall into unless I'm reading some other people and they open up my horizons a little bit.

So is there intelligence that's non-human? Well, my tendency is to answer that, yes, without doing a much thought. What would we mean if we said no to that question? What would we be saying?

What were the implications of that? Well, to me that brings up one of the questions or one of the surprising things that I encountered in reading in preparation for this was the idea that consciousness is not necessarily a required component of intelligence. So a B, for example, is a B conscious? Maybe we'll have to save that exact question for later, but certainly not at the level that we're conscious.

And yeah, it does a great job of optimizing a particular domain, for example, creating honey comb structures to store the nectar and pollen and stuff that it gathers. It's pretty good at optimizing that particular problem, right? Well, right. Is it the B or is it the hive that does the optimization?

Does that matter? Does it matter whether the individual is intelligent or whether somehow? I mean, I know that that's sort of opening up a larger question on the hive mind. Let's not go there for this episode.

Okay, well, that might be another episode. Okay, so I think we can agree that a B is taking actions to make the future sphere. There is a... For sure.

There is. There is to have the hive work. There is to be able to take nectar and make it in the honey to propagate the species if nothing else. It's not simply doing the same thing over in Oregon.

They have to go out into the world and find the resources it needs. So there's future steering going on. Yeah. Let's agree that based on the working definition that has future steering as an important component of it, bees have some degree of future steering in their domain.

And from the same conversation, I think that our buddy talk and our dog talk both would be future steering as well. So what about a bee driving a car or... Ah, yes, in fact, that brings up another example that I encountered in the reading of a wasp. That brings its food up to its little wasp house and sets it down and checks inside the house to make sure that there's nothing in there already that's going to take the food that brings in and go back out and picks up the food, brings it back into the house.

But it's really dumb because if you pick that food up and you move it by an inch, when it comes back out, it picks up the food and sets it back down by the entrance to the house and checks again. You can just keep it doing that over and over and over again. Yeah. So there's some kind of hard-coded quasi-intelligent but not generally not generally intelligent process going on there.

And it doesn't learn in that instance. It doesn't learn. It does not learn, right? So is that intelligent or not?

It's dumb. But what does that mean? Dump? So less or level of intelligence?

Well, that would be not intelligent. So where I'm going with this is suggesting that we add an element of cross-domain to our future steering. Can you generalize across domains? Are you very narrowly able to deal with situations that match your very narrow sort of world you can work in?

All right. So if we take our money and we put it in a different kind of cage that has a different kind of clasp or a different kind of weakness, would we see that that's still how it tries to get out is by finding the weakness and by acting on... So I would argue that the bunny is actually very intelligent compared to the wasp. Yeah.

Right? So I think the bunny is going to find the weaknesses pretty well. Because it's looking for them. The wasp is not looking to change anything.

Yeah. The wasp is a pretty hard-coded, not a lot of learning gen generalization built into it. Hey, there was a funny example of this in something by Elle Ezer that was about deep blue beating, maybe big blue. Anyway, the IPM computer...

The chest plane could beat Kasparov. And as long as they stayed within the domain of that computer, it could be Kasparov. But if you let Kasparov have, say, hey, just beat the computer, can you just beat the computer, then Kasparov can knock the computer off the table, can do all kinds of things that quickly get outside the very narrow domain that that computer can play in. And it's hopeless after that, right?

So that is another example of very, very narrow intelligence. The domain across which that intelligence can operate, very, very narrow. Okay. So then the computer that played Jeopardy had to be a little bit wider, because it had to hear the question being answered.

It had to do whatever processing needed to do. It had to speak the answer out. It had to be more broad than the chest plane computer. The chest plane computer didn't move the pieces itself, right?

It just gave the move number, and the human had to move the pieces. So there's a narrowness and then there's a little bit wider. Yeah, for sure. You could imagine taking that original chest plane program and giving it some embodiment with a robot that can move around and pick up and move the pieces itself.

Now there's, if Kasparov knocks the robot over, maybe he can pick itself up, reorient itself to the board, and now it's not quite such a narrow intelligence. And even more than that, the robot or the computer might say, please don't knock me over again. Please do not knock me over. It hurts my feelings.

Okay, so the cross domain, that makes sense. So you have a chest plane computer that can beat the best chest-playing person in the world, but that's very, very narrow. Very narrow. So that's, for the purposes of our conversation, what we're thinking about general intelligence, general intelligence, it certainly doesn't match.

So the most important part, I think, of definition is future, of intelligence is future steering. But as soon as you want to get outside of really narrow domains, you've got to have this notion of cross domain. So let's call that another, like a kind of a more specific unpacking of the word general. Cross domain is, can be substituted for general.

Okay, fair. And intelligence is future steering. How does that sound? Does that make sense?

And when we talk about cross domain, it makes me think of the different kinds of intelligences that people talk about in the world. There's this guy, Gardner, who talks about, I lost my notes here, here, there. So interpersonal, linguistic, logical, mathematical, naturalist, spatial, body-kinesthetic, musical or interpersonal. I start with interpersonal.

Interpersonal. I don't know if these are actual intelligences or not. But that's how... And there are tests that have been come up with this sort of externally major, some of these intelligences, right?

Right. And the reason we need tests like that is because people measure intelligence just with one test, IQ test, right? For a while. But I don't know that either one of these makes sense for our purposes.

Yeah, so I think that all of these assume and are coming from the standpoint of that general intelligence exists. So let's try to quantify that general intelligence by looking at specific domains where we can do the quantification. Okay, so the assumption that a general intelligence exists, is that because humans are intelligent? Is that why we make that assumption?

Just about every human that would ever be taking one of these tests, I think would qualify as generally intelligent. So we don't even have to deal with the idea of whether it's generally intelligent or not. We're just trying to quantify, in particular domains, how intelligent is this particular instance of human? So the fact that we're having this conversation, we're not having IQ conversation, we're not having a guard right conversation about these different things because the context of our conversation is about non-human intelligences, right?

I mean, isn't that why? Is that why? Yeah, I think that's part of it. We're trying to understand what it would take to look at a machine and say that machine is intelligent.

And notions of EQ and IQ and these other kinds of things, I think are at a future sort of, at a future place, maybe. Or maybe they're actually the gateway. Just like the chess-playing computer is a domain, emotional intelligence is a domain. For an example of a recent movie, her is an example where the guy has an operating system on his phone that he never sees the embodied computer.

He just talks to that her and she is very emotionally intelligent, very intelligent in a particular set of domains. So I guess maybe we want to carry these along with us, but they might be decorations on top of our general intelligence definition. Cross-domain may sum these up. I think it does.

I think cross-domain sums it up. Because the implication there is we're talking about many domains and well, what are the domains? Well, people have done a lot of thinking about what are the domains that we can measure. Well, let's get one more concept in there and then call it an episode.

That is the idea of being computationally frugal or efficient. Alright, so let's unpack that for us. Well, here's an example of an incredibly inefficient sort of optimization process, evolution. We see the results, but we don't see the practically uncountable number of experiments that it took to get to those results.

One of the quotes I don't remember who this was attributed from, but was a graduate student can do in an hour what evolution would take a billion years to do. So evolution is pretty much the least intelligent optimization process you can imagine just about. But human optimization is way, way more efficient. But evolution has had a huge head start.

I think this is important. This notion of efficiency is important because it starts to get at time scales and the rates at which things change, which I think is going to be an important part of our podcast as we consider intelligence explosion and those kinds of things. Yeah, but from the time scale of a human being, boy, I need to do some more thinking about what computational efficient, what does that mean? Do we have examples of the time scale of a human being of what is inefficient and therefore maybe fails this test of intelligence?

It's not a major of failure or success. It's an important major of not impact, but like impact is I think is just how and intelligent it is. But given that you've got two intelligences that are at the same level, one that uses every computer in the world versus one you can carry around on your phone, there's a big difference between those two. Right?

Okay, but we have to say those are both intelligent. They're the same level of intelligence. One is just a lot more efficient than the other. And therefore way more interesting.

Okay, certainly. So evolution, we could say, is not interesting. It's not particularly interesting at the time scale of a human being. That's right.

So it may say that evolution actually does display the characteristics of intelligence, but at a much because it's so computationally inefficient, it's not even interesting for us to really look at. Yeah, and it's a intelligence conversation. And it remains to be seen, you know, how actually successful it is over the long term, given that now that there's human intelligence operating that it evolved, that was the game of evolution, you know, its entire optimization function could be thrown off by humanity. DNA based life might stop existing because evolution created humanity, which would be a big fail, right, for on evolution.

So how intelligent is it? Who knows? The key thing is that there's suddenly a vastly more efficient intelligence, that is the human intelligence compared to evolution. Because evolution has such a head start, we're now getting to be neck and neck.

And so things are getting very interesting. Yeah. Right. So what happens with machine intelligence and now there's these three big three in the future.

Yeah. Part of this podcast is looking. Well, it's exciting. So I like it computationally frugal, cross domain, future steering.

Yeah. This is from Elle Ezer. Yeah. Thank you, Elle Ezer.

We'll take it. Yeah, I think it works. But consciousness was such a shocking idea to me that consciousness is not necessary. Yeah.

Maybe we want to do a look at consciousness next time. What do you think? I think that's a really good idea. And we're probably going to need to come back to the idea of behavior versus process.

You know, like intelligence. I don't know if it matters, but it seems interesting to me the idea that intelligence can be externally validated just by watching behavior rather than knowing the internal process. Yes. Yes.

The mechanism versus the results. Exactly. And that way maybe we never know whether anything else is ever conscious or not because we're not inside of the thing. Maybe these are related.

Yeah. Well, let's take this offline. Maybe that's the next one we can figure out what our next one is offline. That's the next episode one of Concerning AI.

I'm Ted. And I'm Brandon. And thanks for being with us. See you next time.