Welcome to Concerning AI with your host, Ted and Brandon, where we explore existential concerns about AI and what to do about them. Hey, this is Ted. We have a little bit of a different show this time. A few weeks ago I gave a talk that was greatly influenced by conversations Brandon and I have and I'd like to share that live talk with you now.
We're at a time in history where I think we need to consider whether extinction is a possibility for humanity, right? Or whether immortality might be a possibility. And I think the answer is yes to both and I think we are a very unique point in history and so I want to talk about why why that might be. Some of you might be skeptical.
You might think well immortality has not been an actual possibility for all the rest of human history Why should we be thinking about it now? Or you might be saying the same about extinction. We haven't yet go on extinction So why should we think about this now? So Moore's law right everybody in this room knows all about Moore's law right?
How many people here have worked at Intel or works for Intel at some point? All right, everybody so we know about Moore's law, all right? Right, we also know that curves often at right They don't they don't often go up to the sky whether you're talking about a biological system or whatever you're talking about When you a signal point is something that can happen, but we've seen that you know We don't know whether Moore's law is going to end up a signal and probably it right Moore's law itself If you just talk about silicon right, but what about the stacking signal? What about the idea that every technology that we build allows us it gives us the platform for building the next technology, right?
And so a cascading signal well, then that looks a little bit different right go back 115 years as you all know Moore's law is not the first paradigm It's the fifth paradigm depending on how you look at it that we're talking about you know an exponential growth that appears to be continuing There's an argument to be made that you can go long past 1900 go all the way back maybe to the beginning of life and the fact that life itself is going on a kind of an exponential curve and with complexity more and more complexity adding on and adding on so think about What life was like in 1900? Just think about it from it. We don't usually think about it usually we think about what was like yesterday and what's life today? It's pretty much the same right yesterday today, but think about ten years ago, right?
We just talked about how different UX is and how different people experience their products now that we have the smartphones We think of things differently. There's a smartphone in every pocket right everybody in this room has one now things have shifted so Let's let's think about what if what if it's not really exponential like this what if it's exponential like this? That's the point it doesn't matter how fast it's going as long as we're continuing to improve we're in a situation where our technology will get to the point where we know We have an existence proof that this much matter can result in human beings, right? Think about your own brain.
It's just physics I don't know how right on neuroscientists. I don't think we fully understand we're constantly anybody who's coming from but we're here So we know that matter can result in intelligence at least smart as human beings, right? So look at our video games right how many of you guys played the working trail right in the day right? That was awesome right and now we're in the other room and we're you know we're in so chain we're going down like it's amazing right?
Things have improved dramatically think about our TVs. How many of these in your house grown up right? Yeah, this is what it seems to be like so in spite of our experience day to day right? I was talking to somebody earlier who said yeah, we're trying to build this such an exciting that's so hard to get thing to market Yeah, it is and yet over time we do over time things are improving They might not be improving as fast as we think they you know They should it or whatever if things are improving all right So I want us to think about intelligence and I want to think about a scale of intelligence right think about an ants think about a chip Think about you know the dumbest humans like my family right over here.
This is us think about think about the smartest human We often think about intelligence related to other humans that we know think all that one's a dumb one that one's a smart one We're talking about a very narrow band. That's the entire range of intelligence that we think about but I want to think expand I want you to expand that think about the difference in intelligence between ants and a chip right and From the time that sexual selection that sexual reproduction began until today two thirds of the energy went to get to the intelligence of the Ants two thirds that time and 99% of that time went to get to the chip to get to us Just a little bit more just a little bit more And I will posit that there is no reason to think that we are a pinnacle of the intelligence spectrum There's no reason at all to think that we're the top right the top of some chain So think about what would 1% more than us be like how we don't know if we were to ask the chimps You know what a really really smart chip look like they wouldn't be able answer not only because they don't have speech and But all those kinds of things but you know the idea that they are completely sidelined in the global economy They're just part of the landscape. They have nothing to say We're the ones that make the choices and so as we think about our computers becoming smarter Right, and we've got so many narrow Areas where we're seeing our computers are making our lives better, right? We see automation of all different kinds of things We have watched you know You guys follow the Alpha go and the Alpha the big deal that that was that humans were no longer the smartest of playing go He's a game of girls some people are right well It was a big deal because mathematically goes way more complicated than chess the number of games the number possible game states is more than The items of the universe is complicated not in terms of its rules But in terms of the mathematical numbers and Alpha go came out of nowhere people thought that we were at least a decade away from being able to Have computers beat beat the best humans in the world and in fact when they had the game They had a five-game match and right before that five-game match they asked the human champion was gonna play What do you think's gonna happen?
Oh, I'm gonna win five to nothing very easily computers to start ready yet and they interviewed him after the first game where he lost And he was just actions, you know, he just couldn't believe it right and he's a human and whatever we make mistakes and well He lost four one he did win one game Right he did win one game and that one game was a big deal But the press conference at the very end I mean it was a it was a whole different thing and he was humbled and he was thinking about how this you know He's learned so much and he's going on the tear by the way And he hasn't lost a human since because he's just like he's learned a lot and that's really interesting It's really interesting that we could be learning from these computers But this is a very narrow system right and that's one of the criticisms of it. So they well that's not that's not really intelligence That's just you know, that's just math and but it's doing something go is a game that people consider beautiful They still consider beautiful right and it takes intuition it takes something unique to human and yet Alpha go is able to win All right fast forward a year later. This is part even if you've heard about Alpha go you may not have heard about Alpha zero They simplified and generalized their algorithm They threw away all of the training data the training data used the first time They took them a year something to load in hundreds of thousands of human games that it learned from using machine learning They threw all that away. They just gave it the rules of go and let it play itself Within three weeks it was beating Alpha go 100 to zero Three weeks three weeks giving up all culture right thousands of years ago knowledge that we have developed as human beings And just through it all the way in three weeks.
It was the best go player in the world They throw all that away and they gave it chess instead 24 hours later it was the best computer in the world So not only it you have also a zero being the best computer in the world But now you have this artificial intelligence system has beaten 25 50 hundred years of artificial intelligence research So this surprising level of jump forward I think is really important and it's important to consider and again We're not yet talking about general intelligence the distinction You know many of you have probably thought about this distinction between like a narrow system That's designed to do one thing and a general system like a human they can pretty much do everything We generalize from everything that we know and we can do it you watch an infant You know they're learning all the time and they're sucking sucking stuff up what I'm talking about artificial intelligence I'm talking about a general intelligence that when it tells you yeah I'm human and I'm as intelligent as you and whatever we can I'm conversation You'll believe it that's what I'm talking about and I don't see any reason why we won't get there at some point now When I could give you a number right? I think we'll probably be there my opinion that I cannot prove to you in the next 30 years I think we'll be there Again back to the curves right I make you wrong. Let's say I'm off my order magnitude in my in my curve Maybe it's a hundred years out. Maybe it's 200 years out.
I don't know to be super conservative about that I think that this is important because if we get it wrong we could wipe out the in a base life now How would we do that? How could that happen? We'll talk about immortality again in a minute here, but oh sorry This I got I'm really not a designer or a slide person I do a podcast that's audio only and I talk to my friend We don't have an audience that we see and so sorry I'm global life expectancy This has occurred from 12,000 years ago to today and you see that you know that looks like an exponential Right life expectancy has been between 25 and 30 years for a long time and recently the last few hundred years things are different So we'll talk about mortality for just a minute before I'm gonna come back to how could we wipe ourselves out by accident? Life expectancy is going up now every year for living people because medicine is progressing right in technology is helping medicine progress And we're getting better at curing diseases and so we're living longer It could be that we're not extending the maximum Life possible it could be that 120 is really the max and that we're just all gonna live longer life expectancy We'll reach up to that and then we'll stop but it could be that we we create some kind of Regeneration, you know where we actually are using stem cells or something else we're augmenting our bodies We're able to change our DNA now with CRISPR and so as AI is progressing We're also progressing biotechnology So there is a possibility that the exponential rise of technology will also allow us to augment ourselves even if we're only talking about biology If we then talk about nanobots that we put inside our bodies that could you know be an augmented Immune system or go into our brain and help us think better or whatever connect us to the cloud I mean then you get into you know all kinds of crazy stuff and that you know That's exciting.
That's how I actually got into thinking about all of this about 15 years ago I was thinking about how you know hundred years we would all be immortal and we wouldn't have bodies anymore because we'd all be living in the cloud You know so this is you know This is stuff that people like riswile and mondas at singular University You know they're thinking about these things thinking about how we're going in that direction But I want to talk about anybody know the paperclip maximize our idea in AI thought okay? This is a thought experiment think about all right Well, what do you do with your programmer and you're working on anything you want to see if your thing works You give it a toy problem of some kind all right, so what do we do? We say okay, I've got a paperclip factory I have my new AGI it can self-improve and I want to see if it works and so I say to it hey I've got this paperclip factory why don't you maximize the number of paper clips that you can produce? Well, if it's truly an AGI what does it do it realizes?
Oh, you know what if I just you know I could build another paperclip factory next door and so it starts to acquire whatever resources it needs It starts to write computer software itself so that you can earn money and it goes on mechanical Turk and earns money and you know Makes video games and it creates whatever needs to get the resources and it builds another factory Right because another factory bill and then it looks around well I need another factory and pretty soon the entire known universe is turned into paper clips Right because it's got a simple goal and it just goes after it our systems don't think about the consequences necessarily They go after the goal the Alpha go just goes after the goal of winning consequences be damn It doesn't matter if that human is sad at the end right so I think about ethics a little bit And I want to think about how when we are developing our computer systems. We have a couple problems One is we don't know how to teach it our values We don't know how to get it to be aligned to our values That's one problem if you if you as a programmer try to put in be kind to humans You don't know how to do that. We don't have a program that on the other hand We as humans don't agree on what our values are right if we polled everybody We don't have a universe one universal set of human values And as we know our computers are very very literal and even if we get to an AGI There are lots of unintended consequences, you know paper clip maximizing is one of them another one is well All right, why don't you AGI new AGI? Why don't you maximize human happiness?
Oh happiness? Well, how do I know if humans are happy well they smile? Okay? Well now you're all gonna get both oxen sections and Why are like this right?
That's unintended consequence. We need to be able to think through these things All right, what else I got some notes here on paper because that's how I you know, that's how I roll What else oh? Alright, so instrumental goals No matter what our end goals are there are certain things that come into being just by being intelligent self preservation is one of those Imagine you've got an AGI that you want to bring you coffee you say bring me coffee well that AGI all of a sudden They know if you turn them off right that's a threat and that means they're not gonna be able to accomplish their goal So self preservation comes along for the ride no matter what goal you give a system And there are a bunch of these other instrumental goals and these are things that we need to plan for and we need to really think through Appearing the nine for example would be another one They need to look like they're nice to us if you're if you're an artificial intelligence You need to not be scary to the humans who developed you and so the first thing you're gonna do when you're smart enough to realize that humans Are a threat to you is you're gonna think about things that make you look good might maybe you'll cure cancer the game the humans trust Right, but there's other goal another goal not because you're evil. I'm not talking about evil I'm not talking about any of the science fiction stuff right?
I'm talking about the way we think about the ants. That's why I put the ant up there when we build a new house The backhoe destroys the ant colony. We don't care. It's not that we're evil.
There's nothing wrong with us Morley or otherwise well summit you might make your argument that we are bad morally, but I'm not making that argument I'm saying that we just do it out of Inattention and that's something that I'm worried about that whatever the goals are that the ETI has if it just doesn't pay attention to us And why should it frankly unless somehow we figure out how to program it in but we don't how to do that So this is the kind of thing that we're thinking about all right. What else what else we want to talk about today? Some of you are thinking this is far in the future We don't need to worry about it or something that you know this crazy guy front hasn't mentioned is gonna stop it from happening I Am not saying this is guaranteed. I'm not making a prediction the DNA base life is gonna be wiped out in the next hundred years But I have worried and I'm worried that if we really think these problems are real that we are solving and we don't know how long It's gonna take to solve it.
I could be wrong about my 30 years prediction But it could be that it takes 29 years to solve the value alignment problem to be able to program in values It could be wrong and it's really a hundred year away from now before before we get to artificial intelligence It's human level, but it could be it's gonna take 90 years to solve these problems So my call to you is to think about this stuff and to not dismiss it out of hand But talk about it, you know talk about it with your friends talk about it You know Maybe pick up a book or pick up a podcast and think about these kinds of things We need more people thinking about it We need more people who are smart who are design oriented who are engineering oriented or ethics oriented to be able to talk to each other across Boundaries because we we get to create the future. I hope I hope I'm not actually a hundred percent sure about that That we have some saying but I hope we do and I hope that you all in this room We'll be working together with me to create a world that doesn't result in you know the whole universe can turn into the close All right, thank you Thank you for listening to the work voice heard on the show be great one minute voice memo on your phone and send it to concerning AI at gmail.com