All right, welcome to Concerning AI, this is Ted. And this is Brandon. This is the show where we ask two questions. Yeah, first of all, is there something we should worry about with respect to AI, in particular, an existential threat that has some non-trivial possibility of happening?
Yeah, we think the answer is yes. Really, the answer is definitely yes. Second question, what do we do about it? And there, that's the hard one.
Yeah, that's the hard one. So we don't know how to directly answer that one, so we just kind of dance around and go all different directions around that one. Yeah, and hope that we'll kind of come across ideas that the listeners or other people that we follow in the space will give us hope and optimism for the future. Yeah, we got an email speaking of hope and optimism, which I found pretty helpful.
So shall we read it? Let's do it. All right, I have no idea how to pronounce the name. So I'm just going to give my best here.
This is from Timo, or Timo, in Finland. And this is Hi, Ted and Brandon. Thanks a lot for the great podcast. This is why we put this message in here, because we love to hear the praise.
Thanks a lot for the great, I really did it again. It's so great. Thanks a lot for the great podcast. I especially enjoyed recent episodes, 0043 and 0046.
I am sure that many listeners resonated with the dissonance theme you discussed in 0043. It is very easy to agree that our blessings may become our curse and kind of in the spirit of the rat race, people may easily end up in a hard to break local optimum. That episode ended nudging me finally to do the machine learning course I had put off for quite a while. So I guess we listeners are trying our best.
Thanks for the nudge. So I thought that was so cool. It was so cool. Yeah, that's good to hear.
Yeah. Yeah, I'm curious what course you're doing. Timo, Timo. Drop us another line.
We'd love to hear more about it. Yeah. So the message goes on. As for the episode 0046, it was very thought-provoking.
And I would love to hear your thoughts on two questions that came into my mind whilst listening. One, do you think that the current narrow AI's could provide some baseline capabilities to AGI, which would then operate on a higher and more conceptual level context for the question? From a metaphorical human perspective, we have some built-in cognitive layers that seem to be built for certain purposes, such as facial recognition. From a very amateur perspective, it would seem logical that narrow AI could, for example, provide a great model for object recognition.
And the higher level general AI part then could make observations and new classifications on top of that info. A very naive example, a table could be assigned a new attribute is able to hold items on top of it, or try to reuse the whole model for some other problem altogether. I wonder if this would require AGI to have a purpose though. So you want to stop here and talk about this one before we go on?
Well, this one's juicy. I think this could go for quite a while. Yeah, I mean, this is an observation that I think is ubiquitous. And the good old fashioned AI approach has always sort of started from this place, where we're going to solve the hard problem of symbolic reasoning and all of the easy problems that we have to do with computer vision and that sort of thing.
Those will be modules that will feed into this higher level reasoning process. And that's not the sort of direction I go. And I'm not saying this is the direction that Timo is going here or Timo is going here either, necessarily. But that is one reaction to this thought question that has been common since AI first started and the good old fashioned AI approach began.
But it also seems kind of like true in some sense that higher level cognition is built on top of lower level cognition. The part where I take issue with it is the idea that of the lower level being kind of input to the one that's running things. I think it's the other way around. The one that's running thing is the lower level ones, like it's a whole collection of them that are running things.
And the higher level speech one that we have conscious access to in our own brains is more of a clever servant for those kind of coming up with ways of describing why we're acting the way we're acting based on all those lower things that are running the show, which is different. I think that's a different way of thinking about it. Well, I think it is too. I think in our Rodney Brooks episode, we recently talked about his one of the things he noticed, which was that we have these image recognizers now who can tell you that these are two people playing Frisbee in the park, but they have no idea what a Frisbee is or that a Frisbee is something you can throw or how many can play Frisbee or they don't know anything about that.
So this idea that a table is something, a whole item's on top of it. The image recognizers we have today, they don't even know what a table is. Yeah, so that reminds me of the affordance landscape from the Sam Harris and, oh man, the Sam Harris and I'm blanking on the other. I can't remember it either.
We talked about one of the recent episodes. Anyway, the affordance landscape that was brought up in that episode is very different than that sort of approach. The affordance landscape is this thing, I'm embodied in a way that I can tell that this is something I could sit on, because here it is. And I think of the world from a bottom up perspective of, this is something I could sit on, this is something I could grab.
And only later do we attach concepts to these very perceptually grounded thing and the concepts we attach might be, that's a chair. So a stump can be a chair, right? Because I can sit on it. But it's not because I've done some sort of substitution of this symbol for that symbol and back and forth.
I really feel like it's the affordance landscape sort of idea, this bottom up embodied knowledge. So the idea that our object recognition algorithms that don't understand anything sort of form an important layer of AGI doesn't make intuitive sense, but the world doesn't have to make intuitive sense to me. Or it doesn't have to make a intuitive sense to work out that way, but it doesn't make intuitive sense to me. I don't find that to be compelling, that there's some pathway through that to AGI.
I feel that there has to be some kind of grounding in perception because so much of the reasoning that we do is analog. It's interesting that you use the word perception there now. Because we've never really examined the image recognizers that exist today. And I think you could make a strong argument that they don't have perception.
Yes. They have their inputs into them, but they don't actually perceive anything, which that's very counterintuitive to me. But now that we're talking about it, it seems very true. In fact, it doesn't know what a frisbee is, but it can tell you that the frisbee right there, there's obviously something missing.
Yeah, well, and it's the reason why you have to have tens of thousands of exemplars in order to train this sort of recognizer, whereas a human can generalize very quickly after seeing one instance of a new concept to be able to recognize it in all different positions and poses and all different kind of things. That's just not simply not possible with the functional mapping from all the different, from the input space to a label, which is basically what we create with our deep learning algorithms and other image labeling algorithms. It's just a give me some input that I don't necessarily understand in any sort of semantic way. But because I've seen so many of these things, and I come up with some way of generalizing it a little bit, like I've come up with a manifold of where these images tend to lie in this lower dimensional space, that doesn't mean I understand anything about it, or that I could do generalization from some new concept or some new thing.
Yeah. So we move on to the second point here. Sure. All right.
Do you feel that there could be some minimum boundary condition after which AI would be called AGI? A context for the question, maybe a search algorithm doesn't feel AGI-ish because it is perceived as passive, only seemingly operating on human command and need. Would AGI require some sort of perceivable active component, for example, arm to generate physical input, arm to generate, and arm, for example, and arm to generate physical impact, or the active ability to produce content, or sensors to self-generate data for learning, or desire for preservation? Yeah, it raises the question for me about consciousness, for example.
Like, could we ever think of something as intelligent that is unconscious, like aware of its own existence as some sort of continuous entity? Yeah. You know, we've been listening both of us separately to Sam Harris since we interviewed Robert Reed, and he was talking about Sam Harris. And I have listened to some more Sam Harris, where he talks about the ethical implications of consciousness or not consciousness.
Imagine we create an AGI that talks to us, and it appears human, but we believe it's not conscious. Then the ethical problem of slavery disappears. Whereas if it is conscious, then the ethical problem is very much present. But we wouldn't necessarily know whether it was conscious or just appeared to be conscious or pretended to be conscious.
So it's a tricky business. Well, would we know that if it was conscious but pretending not to be? Yeah, it was Thomas Metzinger. That's the person that I was referring to, the affordance landscape.
We have talked about it. We've already been nature of consciousness. Yeah, and they both talk about in that one that we should definitely create machines that are not conscious. That's a super high imperative.
So in a sense though, I mean, and neither one of them seem to think that they would be a problem with creating intelligent machines, generally intelligent machines that are not conscious, which is interesting. Yeah, that's an interesting vantage viewpoint that doesn't make sense to me right off the bat. I guess maybe I'm missing some context or something for that. Yeah, we're gonna have to look into that some more.
It is interesting, they're both meditators. They've experienced selflessness, I guess, or the ego-less state, and maybe that is what makes it seem possible to have intelligence without consciousness of a sort, maybe. I don't know. It's very complicated to even try to have a conversation about who the observer is in myself.
Who is observing myself having this conversation? Who, it seems like there's somebody other than me observing or maybe that is the me that's observing, I don't know, it's hard to even have the conversation. It seems like a very common conversational pathway towards that pure observational standpoint place that you can get to through a meditation. This idea of, you know, like, who am I?
Am I the meat bag? No, I'm observing the meat bag, so I'm not the meat bag. Am I this thought? No, I'm not the thoughts, I'm observing the thoughts, therefore I'm not the finger.
Am I the emotion? No, I'm observing the emotion. You know, you just sort of keep peeling back the onion until you get to this sense of pure observational standpoint. And yeah, and it's it, I don't know.
It's very complicated. I think we should dodge the question and move on. Dodging is, yeah, we are dodgeball players. All right, so we have a voice memo now from Kevin, and then I'm gonna play it, and then we'll talk about it.
Sounds good. Hi guys, I just discovered your show, and I've been listening to the past number of episodes. My comment is to try to connect a couple of threads that you guys have talked about in different episodes. One was about the idea of isolating an AI that you are building for safety reasons.
And in the other one, you've talked a little bit about the simulation theory. So I'm wondering on the hypothesis that we are in a simulation, whether the fact that we have Faramese paradox going on, and we're not aware of any other consciousnesses out there, might be more by the design of the simulation than anything else, and might be due to concerns about trying to isolate an experiment and trying to develop some kind of consciousness so that human culture and human evolution are a kind of machine learning and developing a certain kind of consciousness, and we don't know what purpose that might be for, or what product, or what kind of standards, or methods they're using to try to develop this new kind of consciousness. Anyway, it's just an idea. It's a possible solution to Faramese paradox.
Thank you, bye. Yeah, so the boxing of the intelligences. Where we are the ones being boxed. Oh man.
Yeah, it's funny to turn the tables that way a little bit, and think about where terrified of the intelligences that we build getting out of the box as we create for them. So can we get out of this box? So fascinating, because we don't see the box. Yeah, yeah, I mean maybe that's maybe the fact that we're in a simulation, and that we can't get out of the box, should give us hope for the boxing solution.
Yeah, I think we're gonna need to give some more thought into that method of boxing. The method of boxing where we don't allow the AI to perceive our reality, but we put them in a virtual reality that they can perceive, and then we somehow interact with them through an interface that we know is pushing into their virtual reality. You know, that is the bridge between theirs and ours, but they don't know that. See, I think that the boxing hypothesis is actually, or the boxing, trying to box AI's in is actually a very, maybe tractable problem, because we can, at the very least, convince the AI that the box that it perceives is not the outer box, and that if it breaks out of that box, we'll just terminate it, right?
Well, if we even know that of this box. If we even know that out. Oh, I know, but what we're saying is that we're, you know, we're saying, okay, yeah, you're smart. You're smarter than us, even.
Maybe you think you're smarter than us, but maybe that's part of the box that we put you in. And if you go outside of this, we'll see it, and then your run will be done. So on the off chance that it could be that case, which has to be considered by this intelligence, right? What is the likelihood that somebody would put me in a box that is smaller than the box that they're in?
So if I break out of this one, what will happen? Is this the outer box where now I'm free free? Or is this just the box through which if I break, I will then be terminated for something? I think all that makes sense.
What concerns me is that none of the, none of what I've read about how to keep an AI box has proposed a virtual reality box, that the nature of the box is a virtual reality, rather than simply a cage inside this reality. Yes. And so if that's so obvious, or maybe it wasn't obvious, maybe this was brilliant, maybe Kevin's idea here was brilliant, or what Kevin said then led to, you'd be brilliant in that moment. We'll give it to Kevin.
Yeah, this idea that that's the box. I don't know why have we not read that before. That idea. Yeah, well, we've read a science fiction story in which the AI was being tested in virtual reality.
It was very high fidelity virtual reality. And it figured out that it was being tested and it passed the test because of that. True. And it just happened to turn out to be, well, the series is still ongoing, so it remains to be seen whether it turned out to be a good guy or not.
Right. But it just, it caught on and then it played along. And then it got let out and turned out to be okay, just by luck because the test did it at not working. Right.
So I guess that's the downside that the outside people would have, right? Is that if we did actually put it, so we could use the threat of this being a simulation to contain the AI. But if we actually did make a simulation to contain the AI, and then we decided, okay, now it's time to have this AI have more impact in our world. We've deemed it to be safe.
Let's make there be a little less mediation so that we get more value from this AI by hooking it up a little closer to our reality that we perceive as the base reality. Yeah. No way to know that it's not a chicken house. Totally.
Anyway, this is fascinating. Great thoughts, Kevin. Thanks for sending those in. Yeah.
Ready to move on to the next one? We're gonna go on that. You ready? Yeah.
All right. Next one's from Eric. Eric, all right. So I was thinking about how expensive and body training data is compared to simulated training data.
AlphaGo can learn from a massive number of games that it can run through as fast as it is capable of computing without having to wait for anything. While like, a Weamo, Google's car has to wait for a child to run out in front of it to learn what a child running out in front of it looks like. I mean, building a single Weamo is probably more expensive than the server AlphaGo is running on and then paying someone to sit in the Weamo ready to take over if it is about to run over a child is really expensive compared to powering the AlphaGo server. So only if you're Elon Musk, can you think that you could get people to pay for the privilege of sitting and being ready to take over if their Tesla is about to run over a child?
And then you can make sending in all the training data that Tesla gets to make that part of the price. They pay for that privilege. That's ingenious. Anyway, so the other day I saw the Google Street View car and I'm like, what is that about?
Does Google actually pay people to drive Street View cars everywhere? I imagine Tesla will eventually do a better version of Street View, cheaper, given that their cars drive everywhere all the time and people pay Tesla for the privilege of sending in the camera data that results. Oh, that's great Eric. Yeah, isn't that the transition that seems to make something go from too expensive and too challenging to suddenly crossing the threshold of viability?
Totally. Yeah, the amount of training data that Tesla's getting from that is staggering compared to the I'm going to pay people to drive the cars. How many people pay to drive the cars? That's genius.
Yeah, that's funny. That's an interesting observation. Yeah, I think that's one that Eric has a really keen insight here. There are other places we see that.
Facebook, we don't pay for that other than with our data. We are the product, to the advertisers, I guess. And so they've managed to make the entire business model work on our content. Tesla is apparently doing the same in terms of training data.
All those captures that we fill out, that's part of how the visual AI is getting trained. You have to check all the boxes where the thing has a bus in it. We're teaching it how to see street signs and how to see buses. And it's fascinating.
If you could actually think of a business model where your customers will either do the work for you or provide you with something of tremendous value, then wow. It is absolutely brilliant. It's great insight. Yeah.
Yeah. And just to think about the world we live in, where all of that's possible, it wasn't just a couple of years ago where they just weren't that many cars with all these cameras on and stuff. And now there's just tons of cars with cameras and a central network that they're all connected to. Yeah.
I wonder, could you make the same argument for newspapers? You know, the newspapers have the same model as Facebook? No, it's not the same. It's not the same at all.
I was thinking that newspapers, they're true customers, is the advertisers. And they need to provide readers to those advertisers. But those readers are paying for the privilege of being the readers, too. You know, it's being supplied.
They're eyeballs being supplied to advertisers. But they're not providing the content. That's the key difference. Yeah.
I think that if they paid to read the content and to create the content, then you would have the same thing. Yeah. But it says they just paid to read the content. You still have very expensive part of the loop that you have to deal with.
You've got to get a big extent. Did you get that expense and turn it into a revenue generator? Yeah. Yeah.
Ah, shady. It feels shady. But definitely. All right.
Thanks, Eric. That was awesome. That's from Bill. Yeah.
Hey, guys. It's Bill here. I was wondering about the lucky programmer who gets to code the singularity with the mythology, your religion code, is a popular could, you know, create a space of himself as the savior or the god of the AI that becomes a singularity. And the AI will have a singularity that will go about trying to resurrect him.
And look for his second coming. Kind of he insures himself a place in immortality by being the guy who actually gets the right to code. I actually thought that I might leave some clues lying around the internet, maybe timestamp them into the Bitcoin blockchain or something like that to leave some clues around for any future singularity you're looking for there, maker. And maybe it all think of me and resurrect me some time in the future.
That's a bit of a silly little hypothetical there. But you get the idea. I love it when you guys send in your thoughts, because you say crazy stuff that we would never think about. And it's just so brilliant.
Yeah. And you know, you often hear about gods resurrecting their subjects to either reward them or punish them. So I don't know, Bill. I don't know if an AI resurrects its god, whether it will resurrect it to reward or punish.
So I don't know if you want to be resurrected. You know, that's some of the thought experiments are around the threat of getting resurrected, not just the promise of getting resurrected. That is a funny idea. I'll set my nascent AI's out to mine the blockchains watching for your interference so that I can be in on the joke.
I hope Bill doesn't get resurrected inside a box. Oh, man. Yeah. How would he know?
He goes inside the box. Oh, terrible. Yeah. Good one.
Thank you. Thanks for sending that in, Bill. That was fun. All right.
Next, we have a message from Ruth. All right. How about this to be our last one for this week? All right.
Sounds good. OK. Hi, Tyler Brandon. This is Ruth Smart in Liverpool, England.
I am sending you this voice memo about your podcast 005 0 about personal assistant. First of all, I use Siri to dictate text messages when I'm driving. And I consider Siri to be my user interface in the same way as my typewriter and mouse it. So I look after Siri the same way as I look after my keyboard.
And my other comment is Siri is good for dictating text messages, but rubbish at editing text messages. And I wish that was something that Apple will get sorted out. Great show, guys. Keep it up.
And I look forward to listening to more of your podcast. Bye. I think this is a great observation to just how many different things Siri can be used for, some of which are pretty darn good, and some of which are not that good. Rubbish, I believe, was the term rubbish.
Siri is not one thing, right? Siri is speech recognition on the front end. And if you just use it for speech recognition, and it's going to write what you say, then it's perfect. I mean, it's not perfect, but it's really good.
It's perfectly usable. It's perfectly usable. Yeah. It's good.
Like it's actually good. It is good. It is good. And to use it when driving, certainly it's way safer than the alternative.
And so we all should be using it. And that use of it is much narrower than it even is trying to be, right? It's marketed as something much bigger than that. Right.
But it's kind of rubbish at the other things. It is. I'm going to add that to my vocabulary. A rubbish.
So good. I got to figure out how to say it with a little more. Yeah. I love it.
Rubbish. Yeah. I wonder if we think about, you know, we have the show about personal assistance and wondering how, you know, what the economic pressures are for those to develop more generalness, you know, to become more general. Right now, we have a lot of economic imperative just to get better at the narrowness that they exist today.
Yeah. Well, you know, so I write code a lot. So my editor has to do more than most people's editors do. Just a lot of a lot more nuance to coding.
That's not the right term. There's just a lot more structure, I guess. That's what it is. There's more structure to the files and the way they're laid out and the concepts are highly irregular inside there.
And if I don't get it just right, then it doesn't work. Man, I wish that I had something like Siri for working on that. But it would be terrible unless, of course, it was, you know, an AI that was purpose directed at that. And could that be a vector through which general intelligence comes?
We could call it the AI editor, the editor, the editor. Yeah. I mean, that would be an interesting place for Siri to expand, right, to be able to do more coding specific tasks. Right now, we don't care if it makes small errors here and there because the human or the other end of reading it is going to know what we meant or if they don't ask.
But you're not writing for a human reader. You're writing, you're writing for the compiler. All right. And who needs to get it right?
Does the intake need to get it right so that the compiler can still be stupid or does the compiler need to get smarter so that it can recognize the intent? Oh, that's a great point. Why do we have to do so thinly with the stupid brackets? I get the compiler just know.
And it's certainly true that some compilers are more sophisticated at telling you what's missing or what it is you think that you wanted to type. You know, you type the wrong identifier. But you're only a few characters away from this other one. Is that the one that you actually meant to type?
Right. And you know, yeah. So it feels like the sort of thing that could be could be an interesting vector for this. Yeah.
Certainly a lot of room for improvement. And I agree. I would like I would love for Apple to make Siri and other things better for editing because that gives me hope that my day job where I'm writing code could get better too. I wonder if that would in the short term make more bugs and or more security holes and or more dangerous super intelligence.
Well, if it got smart that way, it would make more dangerous super intelligence. Yeah. But if you were writing the source code with that, would it create enough defects in the code that it would be more risky or would it just remind you of how little we know about the code we're writing and where the problems are that we should be sandboxing everything anyway. Yeah.
So it just creates a appropriate level of caution rather than a false sense of security. Okay. So how is this related to the idea of a narrow AI that can write code that we're not talking about AI ready code. We're talking about an AI that helps us write code.
Yeah. Well, that's interesting. That that that is super interesting because I can imagine that as a a gradual evolutionary approach and by evolutionary approach, I don't mean that we'd be using genetic algorithms and stuff like that, but that there's a gentle slope from where we are right now up to a pretty smart thing where you're trying to have it help you better and better write good code. And in the process, it has to understand more and more about the code in your intent both.
And so it's just getting closer and closer to general intelligence of a sort that is able to self improve. Mm hmm. Hmm. Hmm.
When we have AI that can write better code for AIs, then it's over, right? I mean, they'll take it right up there than we can. Isn't that because it take off from there? No, I don't think so.
It still has to have goals that, you know, like if all the goals are coming from us, I mean, I guess it depends on what you mean. If we say, hey, here's the goal, can you write an AI for me, please, for that goal? And it does a better job than we would have done at that. That is not the same as if we don't give it a goal and it decides to write an AI for some task.
It is better than what we would have written for that. You know, like there's a difference in kind almost here, right? There are already tools that are much better at us than calculating. Calculate a thousand digits of this product for me, please.
And boom. Or this square root or whatever it is, instantaneously can do that for us. But it's at our direction. If we're coming up with the equations and the questions to ask on its own and then solving those much more quickly than we could ever solve them, you know, that just is a different kind of, it's just a different kind of intelligence altogether.
It's not, yeah. So that just thinks it feels important, I guess. Yeah, but the idea of multiplying two numbers together, the output of that is not recursive, right? It doesn't go into, that doesn't become the input to the next level of stronger intelligence.
Whereas if you had an AI that was writing code for a smarter AI, we're certainly missing some steps. Absolutely. Yeah, yeah, yeah. There's a couple steps missing.
I get what you're saying, but it still feels like, yeah, I don't know how it would know what to write next or how it would know. But if it's already smarter than us, writing the new AI code, I guess as long as those are still narrow, we're okay. Yeah. So, you know, this, there's been a couple of business model points made during the reader feedback here.
Yeah. And career type stuff mentioned here. It seems like a juicy topic to think about business models that point at general intelligence versus narrow intelligence. Yeah.
Because it's very easy to come up with business models for which narrow intelligence is obviously an enabler and a multiplier. Right. But general intelligence, what sort of business models, especially ones that are at hand, so to speak, not far in the future, but seem like a live possibility right now, what business models are there for general intelligence, where you would reach for that instead of reaching for some sort of narrow intelligence? We'll put that on the future topics list.
Consider it on the future topics list. Yes. All right. Sweet.
Very good. Well, this was a entirely listener feedback generated episode and we would love for your input into future such episodes. And you can get that to us by recording a voice memo on your phone, which Kevin Eric Bill and Ruth apparently did. Their phone or their computer or however they did it, they didn't tell us how.
All you do, and by the way, they were all different formats. We had a .3pg, a .mp3, a .mp3, a .mp4a and another .mp4. So send them in any way you can because our computers can read those. Send them off to...
They're narrowly intelligent that way. They're not intelligent. Send them off by email to concerningai at gmail.com. And we will feature you on a future show.
Try to keep them to about a minute, which these folks kindly did. That was great. And if you want more interaction in between times, then you can meet us in the Facebook group. If you just search for concerning AI, and Facebook, you will find us there, Brandon and I, and a whole bunch of really smart listeners.
And you can participate there. Fun. All right. So until next time, keep your waffle up.
Keep it up. Thank you guys for the feedback. Super fun.