So. The. Book is called, the, sentient machine, and. It. Really is a, varied. Book you, know it starts off with. Some. Philosophical. Ponderings. On what the advent, of AI really, means for us there. Are as, you know some. Existential. Concerns regarding. The. Advent, of more, and more powerful. AI, AGI. And then ASI, and lots of very. Worthy scholars, have written volumes. About, these for example Nick Bostrom super intelligence, which I'm sure many of you have heard of if not read. And. Then. Beyond that there's also discussion, around. Essentially. These two fears, that keep sort of rearing. Their head in in different. Ways. But. One that AI will take away all our jobs and it, might render, us useless when it gets to a certain level of complexity. And capability, and the. Other is that it might kill us and of course that has, many, different. Aspects. And and situations. Under which that, fear manifests. Itself but, in, a nutshell those are two real. Conversations. That are happening these, days and, for. This audience I'll also tell you that these. Are not just, hypothetical. Or philosophical. Quandaries, and questions, anymore they, are now. Being played out at the highest levels, of government. So, spark-ignition. Works, in three principal, areas national, security. Industry. And energy and finance, and as, a concept and I won't talk much about our own work this is this is a talk about the book this is not a talk about spark-ignition, or my work per se but. Because. Of that background, I end up meeting with and having. Some. Pretty interesting discussions. With, the. Senior-most military, leadership not just in this country but also for, example in Allied States in Europe. About two weeks ago I address. The NATO. Council on their adaptation. Report. That was about to come out and how believe. It or not AI will, play a great role in, the new adaptation report. That was just released. Based. On the writings that General. John Allen who's my collaborator. And also on the spark-ignition. Board, and I published. Earlier this year so. That's, just one example to. Say that artificial, intelligence is, becoming, real in, many, many. Ways and. Perhaps. In narrow domains, initially, but the capabilities, are widening, and, for. Some of these existential, concerns, that, people have expressed. For. Example, that it'll take our jobs away realize. That we don't need Commander, data you know a GI level, capability. For. Those, sorts, of things to be threats, Ani. Capability. Implemented. In many, different areas can. Lead to. 30%. 40%, god-knows-what. Percent unemployment, that, remains to be seen and. That's. For developed, countries on, the. Other hand for undeveloped, countries, or developing countries, they.
Have Invested. A lot in their. Burgeoning. What, they call their demographic. Dividend, you, know people that have been brought, out from. Conditions. Of sort. Of, you. Know under. Privileged and are now being educated, and are being made. Available to. Do complex, tasks, in the economy, well. Some of those complex tasks, might be subsumed, before. Those, generations. Get an opportunity to really make a mock so there's that. Emerging. Sense, of the fallacy, of the burgeoning. Middle class in, developing, countries, and whether they'll be able to play the role that, we once thought they would be able to play we. Don't know we'll see at what rate these technologies, progress, but my point here is that already. We. Are at, a point where the. Discussion, around artificial. Intelligence, is partly. Technology. But. It's also partly. Policy, and. In. My own case. I've tried to bring these things together and, in the book you see, the. Science the. Philosophy. As well as elements, of policy because ultimately we have to do something, about this. And I'll tell you later on as we get into this for. Example some of the discussions, around bans and, autonomous, weapons I've been quite, deeply involved in all of those debates and have, met. With a, lot. Of folks that that really do matter in that in that debate so, we, will go through that the. Way I'll structure, the talk is I'll, start off with a very brief reading and. You. Know the beginning of all of these sessions ends up being different just, based on what chapter you choose so, here we were talking about autonomous, weapons maybe I'll start with the. Beginning of a chapter called warfare, in AI and. Then we'll, talk about more broadly some, of the content in the book but, I've also structured, a presentation. That takes us a little bit beyond the book there are some. Concepts, here which for accomplished, computer scientists, that are in the audience might. Seem to be very basic and you. You may be familiar with them and I'll go over those quickly if they become boring, but. Then we can get into some of the other issues, and. Some of the problems that I think we still need to solve so, that's, how I'll go and of course I'm open to questions at any time and, comments, I would welcome that. Okay. So. For those of you who do have the book on page 87, I'll start, with a very brief reading, of this, chapter that, that. That is titled warfare, and AI. Join. Me for a thought experiment. Originally, published in, the u.s. naval Institute's, proceedings journal. And conceived. Off by my friend and collaborator, General. John Allen of the, United States Marine, Corps a four-star. General and, past, deputy, commander, of US, Central Command, it. Is January 2, 2018. And a, captain, is contemplating. Damage to his ship after a surprise, attack, this. However was, no ordinary attack he. Is about to discover that this was a massive, widespread. Strategic. Surprise, our. Captain. And his crew had not anticipated the. Incoming, swarm, because, neither he, nor his, ship, recognized. That their systems, were under, cyber attack. The. Undetected, cyber activity, not, only compromised. The sensors, but locked, out defensive. Systems leaving. The ship almost entirely. Helpless. The. Kinetic, strikes came in waves as a complex. Swarm of drones tore, into the ship it. Was attacked by a cloud, of autonomous, systems moving together with purpose. Yet. Also. Dynamically. Reacting. To one another and to. The ship. More. Than anything the, speed, of the attack stunned, and overwhelmed the. Sailors, though. The IT specialists.
On Board the ship were able to release some defensive, systems from, the clutches of the cyber intrusion, the, rest of the crew simply, did not have enough decision-making. Time to, react, mere. Seconds. And. In these few seconds, some of the sailors ascertained, with, their limited situational, awareness that. The enemy's, autonomous. Cyber and kinetic, systems, were, collaborating. But. In a matter of minutes the, entire attack. Was over, the. Captain survived, and courageously. Remained. On the bridge but. He was badly wounded as was. Much of his crew, fires. Were burning out of control and, the ship was, already listing, badly, from flooding, because. Of the damage the. Captain was unable to communicate, with the damage control assistant. Who was herself, badly. Wounded, it. Appeared, that some of the autonomous platforms. Knew exactly, where, to strike the ship both, to maximize, the damage and reduce, the chances, of survivability. The. Captain's, ability to command his ship was now badly, compromised, and the, flooding was, out of control, after. Surveying. The entire situation he. Realizes, he. Must make a call that, no American, skipper, has, made for, generations. He. Issues, the order to. Abandon ship. Okay. So. Going. Through this. You. Might wonder, if this is fiction. It. Is, you. Might wonder if this is entirely. Imagined, fiction with no grounding and truth it is not. It. Was very interesting this, past year I had the opportunity to travel through the Middle East and as you know there. Are many numerous. Active. Conflicts, going on in the Middle East and many, of them are conflicts. That are taking an asymmetric. Sort. Of a tilt. Where. You've got somebody, like the Houthis, as, an example in Yemen or you've got. Isis. Terror organizations. In Syria and some parts of Iraq and. You've. Got then on the other side for. The most part well armed, military forces with, sophisticated. Radars, and, Patriot. Missiles, and so on and so forth and. Let. Me just tell you that I know, for. A fact that. Much, of what is described, here maybe not with this level of nation-state. Sophistication. You know because here what we describe is a swarm, of. Improvised. You calves coming, in with, pretty sophisticated vision. And other recognition, capabilities, going, for a well protected asset but. Similar. Scenarios, have actually, played. Out in. The current conflict in the Middle East the. DIY. AI. Improvised. Flying. Ie D is, already. Here. Two. Weeks ago the. Convention. For conventional. Weapons which is a little, awkwardly named but. CCW. At, the UN got, together I think for the third or fourth time over, three or four years and there. Are a hundred and seven mm, states of this organization, and the UN and.
Every, Year they've gotten together and expressed, their dire. Concern. Over, the, potential spread of autonomous, weapons and what they must do and of, course you may remember the famous what's, called the Elon Musk letter which, was really not Elon. Musk's letter it was the, most recent one was written by Professor, Toby Walsh, from. The. University. Of New South Wales in, Australia he's, an AI professor, there and of, course Elon Musk was a signatory, and then, the letter was presented, as a, desire. Or a, request for a ban which it. Was not it was a request for a discussion, and. At the end of this most recent session a hundred and seven countries couldn't. Even get together after, four years of debate and agree on what the definition of an autonomous weapon, is. In. The meanwhile the Kalashnikov Bureau, which many of you may have heard of which is a Russian weapons, manufacturer. Announced. That they were testing, a ugv, an unmanned, ground vehicle. Which, in field, tests, had already shown, better than human performance, now. You, may doubt these claims, you may think these are oversold. But, wait a year or two. Similarly. China announced. That they were fielding. A I powered. Cruise missiles, and the. MiG Bureau announced, that their new, next, gen MIG aircraft would. Have AI autopilot. Operating. To. Control, the. Flight envelope at hypersonic, speeds, so. This is just one vignette, this is just one side of AI which is it is, a technology that, brings. Or, distributed. Autonomy, at large scale, to the field of battle it, is, a strategic. Level. Up if you will in terms of capability, and no, significant. Player. No, significant. Military is, going, to ignore, this, and. Just, to give you further evidence, of this of the hundred and seven countries that, were at the CC, W. Session. There. Were only 22, that. Came out and. We are in, favor, of a ban of. All. The nucleus states there was only one, which supported, the ban and of, all significant. Militaries, in the world all states with significant, militaries in the world only, two supported. The ban the, largest, militaries, the, countries with the largest number of nuclear weapons all, argued. For further, discussion, let's push this off to next year so. That's. Where we are that's just one vignette but with that let, me start, talking about sort, of some of the things that we that. We cover in the book so. What's. Quite clear to me now is that we've made enough progress in, several. Areas where a, new, form of intelligence, really, is coming I mean it's no longer sort of the wizard of The, Wizard, of Oz a man, hiding behind a curtain it's, no longer just large, numbers, of if-then-else. Statements. And, while we keep uncovering every, now and then company, X and Company y outsourcing. Some activity, that they call intelligent. But really it's going. Out to you. Know the Amazon. Mechanical Turk, type, situation. Really. Aside, from all of those things intelligence. A, increasing. Level of intelligence, is being built and we've, had with deep learning in particular we've, had great impact, on perception. Tasks, you, know where we want to for example classify. We want to perceive something and extract.
Complex, Patterns, and even, patterns. Across temporal. Boundaries, we've been able to do that very well with deep learning and now, we're sort of running forward, with reinforcement. Learning with, lots of new innovations. And the. The the key thing to take away there is we're moving from the domain of perception, to, the domain of action and even. Within reinforcement. Learning now we have the ability to. Train. Systems, up maybe, in simulated, environments, and with some of the break breakthroughs. That are taking place and transfer, learning we. Take the. Learning. That's done in a simulator, and and translate, that to the real world so, this. Is not, a talk, about the. Current. State of the art in all three of these areas but just a couple of minutes to say that this is becoming quite real and indeed. A new, form of intelligence. If not sentience. Is is coming sentient I think is far away. So. With this just. A quick background. One, of the things that you. Know I think I do. Differently. In the book is simply. A consequence, of my own background so, I'm, a serial entrepreneur I've, been, based in Austin and I went to school at UT Austin computer. Science and have, done a number of software companies since, then and spark-ignition. Was. A company I founded back in mid, 2013, the. Company focuses, as I mentioned, on national. Security finance. In industry, in fact we're a Google partner on a number of different things and and. The company's grown really really fast in fact it's the fastest growing company, in Austin now. With that being said that's the business side of things right and how you actually, take this technology and make it work and make it solve, problems, for the largest companies in the world but. The other side of, this is that I come to this not just purely from a business background I am a computer scientist by training I, love, computer, science I live computer, science and I. Serve, on the Board of Advisors of UT, CS which, is one of the great, really. Pleasures. Where. I'm when I'm able to spend much time there. So. It brings sort of the business aspect, the practicality. Of making these things work with. The science, and attempting, to advance the science and then finally, the, center of a new American security, is one of the premier. Think. Tanks in DC, and. I, serve on their. Advisory. Board for artificial. Intelligence in fact about. Two or three weeks ago I was in DC we were we had a scene as conference, on AI and what this would mean for autonomous. Weapons and there were lots of generals, and serving. And otherwise and and many. Policymakers. In the audience but, we also had Eric Schmidt there and I. Had an interesting discussion with with, Eric and one of the topics that came up was, well. How. Long given, that China just announced their 2030, AI plan. Which many, of you may have seen it's. An investment of a hundred and fifty billion dollars, of government, spending. Over. The next five years in, 2015. The US government, spent 1.1. Billion dollars, on AI and in 2016, we, spent a whopping. 1.2. Billion. Dollars on AI. Again. The Chinese government, has committed just, governmental. Spending, 150, billion dollars over five years and if, you read the 2030, AI report, it says by 2030, we will be, the dominant AI player.
In. Addition to that they also talk about all the applications, of AI and a big chunk of that is military application, so here we were in DC at this conference, and I asked Eric I said Eric you. Know I have a view but what's your view how soon do you think China will be, able to overtake, the u.s. in core, AI capability. And. It's on video and then there will a lot lot of articles written about it but he said five years I don't. Disagree with him it sounds, very. Aggressive, but. The rate at which progress, is being made the rate at which you. Know just if you look at face plus plus and the rate at which they're improving their vision, algorithms. And a few years ago I remember people. Used to cite well you know AI papers, in China sure they're publishing a lot of AI papers but they're not as good as good as us well. Now. That's. Really not the case anymore and it sort of reminds me of the, whole. The. The reaction, that we've had to many, other countries that have been catching up it. Probably just copied it it's kind, of like a fake you, know it's, sort of kind of there but it isn't, the same thing and then, suddenly very, quickly you realize, that that, folks are catching up and where, we are in our current, situation. As. A country, is that, we're doing things like. Preventing. The, spouses, of h-1b, immigrants. From working which means that fewer smart people will be able to come in we're. Trying to ban entry, from a number of trees and, limiting. The number of smart people that will be able to bring into the, US so, while you, have an air pure competitor, putting a hundred and fifty billion to your 1.2, billion you're. Also then, strangling. Some. Of the the core, elements. Of innovation, that, have historically, been so, useful. Useful. For you that, I think is bad. Timing. So. Moving on to another element. In a lot of these stocks people asked just the simple question, of well what does it mean for a machine to think this is obviously a very complex, question and there's many, many different ways in which a machine can think and we, tend to describe, things in the context, of a neural network where you take a neural network you give it a lot of data and then you can ask it a question about, what you've trained it on and it'll either classify, or in a regression sense, give you some sort of an answer but I figured, that from a visual perspective you know there's many ways machines, can think and AI, isn't just machine learning there's many other things in AI as well, for. Example search, based optimization so. Here one way that you can think in. A problem domain is let's, just take the simplest example, of tic-tac-toe. Given. Just a couple of rules you, can pre, generate, all the, possible outcomes, and then, what. Is perceived, by the human player, to, be a smart. Move is simply. A goal, seeking behavior, where I know what, a win looks like and I've generated, the tree or the graph and I'm trying to traverse the graph to find the most efficient, path to what I know to, be a win and that's. One way, in which you can make. Machines, appear. To think but, we also know that not, every, problem has, such a small state space so, there are problems, I mean even games take another game like pac-man where.
You. Know where, is miss pac-man quite. At this moment and how many of the golden. Nuggets have been consumed and what's the direction of each one of the foes and did, you eat the berry or not and does a lot, of variability, in, that state. Space so that's the that you wouldn't want to just encode, in this way. So. There now we try. To do things like reinforcement. Learning where this, starter, often start. Playing the game and you end up and you die pretty quick but maybe you made 50 points and what, you are trying to remember is what sequence, of tasks, got you to those 50 points the, first ones, that you took are pretty much worth the 50 points, because they led to you getting the 50 points but as you move further along in that stack, of moves you realize. That the closer you get to 50. The. Less valuable. Each one of those more recent, moves were because goddamnit. The last one got you killed so, that can't be very valuable and so, you, can have this sense of well let, me few. Others again many many different, approaches. To this but you can have an element of randomness. To, where let. Me try and get the maximum reward but when the level of reward goes below a certain threshold I'm going to try different things and maybe, I find something that's more interesting so this is sort of like a self. Pruned, search, you, know in reinforcement, learning in just layman's. Terms it's, sort of like a self pruned search where, you start off with something and you don't abandon that but you just look for improvements, where you can you can find them and we've. Seen great progress with this now. Another. Thing that I will point out is that even in this idea. Where you are generating, entire, states by the way, stuff. That's not fashionable, people, stop thinking about but there's a lot of problems that can. Be. Smartly. Pruned, to. Where these, sorts of solutions are, still pretty good solutions, I mean you, don't hear a star search much now but, if you apply data, properly, to a star search and you come up with clever heuristics. On how to prune what, gets generated and, what gets searched there's a lot of problems that you can solve pretty. Cleverly with a star search but, anyway, in, this particular case where you see the full tree one, thing to note is that generating. These states sometimes. Can be incredibly. Simple, okay, and that's one concept, that I'll build on here. So here what you had to know to generate, this whole tree is that, for, every progression. You can only change, or add one symbol at a time right, so if you've got knots and crosses either, it can be you. Know you can add one knot or one cross you can't add two knots in one go, you. Need to know what the winning state is and that sort of we all know a diagonal, or a line or a horizontal bar, and then. With, with every, step here as you step through the tree in. Every layer you are alternating. Symbols, so first. You get a knot, then you get across then you get a knot and so on. That's. All that's, all that you need to know in, order to generate, something, like this and so, is. That you know sort of this mind-altering, fact, in the context, of knots, and crosses not. Really but, it does go to say that, very, very simple, things. When. Iterated. Upon when. When. Dealt. With with recursion, can create, tremendous. Complexity. That can be useful. So. Seed. Of. Specification. Can can, create. Something. That is. Very. Very large very, very useful and sometimes, a little unexpected these, are concepts that at least in two different places and we know we talk about the game of life briefly, in the, sentient machine but. What. I cover is just, the the basic introduction, Stephen. Wolfram who's the creator of Mathematica. And somebody, that I followed for many years, very, interesting, thinker in his, book a new kind of science he. Spends, you know almost, 200, pages just. Going, over different forms, different variations. Of the game of life and these, again if you're from how many of you are familiar with the game of life, all. Of you okay almost, and. So again these are really really very simple rules and what wolfram, shows is that you can have these.
Incredible. Levels of complex, non repeating, patterns, that come from very, very basic rules. Another sort, of more and in lines, of sort of continuous. Mathematics is, you. Know this this notion of fractals. And there's two, things that I want to quickly say so, one. Since. You are familiar with the game of life I'm sure you've seen simulations. Like this but to me every time I see stuff like this I'm amazed you, know it's, that's. These creatures. With like distinct. Behavior. That, that evolve every time and they, some, of them find stability and, others oscillate. Between two, states and then. You have some movement you have this you, know artifact, called a glider that just sort of walks across diagonally. Usually. You, have. These. Blobs that can combine and, then what, comes out of that is at least not visually, immediately, predictive, and and these can be very complex, sort of behaviors, and, looking at that you know it does seem to be like there's something going on here and of course we know what's, going on here is just very simple, three simple rules but. The. Manifest. Complexity. Is far. More than those three simple rules an. Initial reading of those three simple rules would, imply. The. Same is the case really when you start, thinking about fractals. Because, I, mean, this one is the, Mandelbrot, fractal and, Benoit Mandelbrot came. Up with an expression which is. Geylang, I mean that's it right that's the the, amount, of math that's, being generated, into, the, structure, and the, reason why we're going into this is that I mean again for, those of you who've read Ender's Game the book, not, the movie. You, you. You. Realize, or, recognize or. Remember, that, there's one, comment in there that was made where ender, was at this training facility, and he was given access to a large computer and somebody. Pulls him away and says what are you doing and he says I'm traveling. Through the fractal, and now. The fractal. Is larger. Than the known universe. So. When, I read that as a teenager, it sort of stuck, in my head I don't think that when the book was written. That, computers, had actually generated, a fractal that was larger than the known universe but, now there. Are many examples of, this so, to think, that this much math this, much specification. This, much code can. Generate, something, with, unending. Complexity. That, is. Unpredictable. Actually. And. Unique, at so many different levels. To. Me is pretty amazing and the two ingredients of that of course are the specification. And the. Iteration, and the recursion and that, gets me to one of the points that I make in the book also about the universe, being computable. In a different way so you've heard. Elon. Musk recently, talk about how the universe might be a simulator, I actually, came across that concept. In. My youth, gentleman. By the name of Ed fredkin, had. Written extensively on, this and then when I started digging, deeper, I realized, that konrad zuse a-- even. Back in the 40s, had, talked about these concepts, and ed fredkin, you. Know this article my father gave it to me it was published in a magazine it. Was called is the universe a computer, and the. Idea there was not so much Elon Musk idea which is that we're all living inside a simulation which. Is one type, of simulator, inquiry, but, the other idea was is the universe fundamentally, computable. Like everything, we see is it a consequence, of computation. Many. Years later my sister, became. A string, theorist and I. Tried. To get. At least a. Workable. Understanding, of string theory and my many conversations with her and one day she sort of lost their patience with me and said listen all, of what we write in in, words, in English, it's, just to sort of point you roughly. In the right direction if you want to understand, any of what we're really saying you have to work through the math none, of this really translates, in. In. In, language and, the. One, thing of course from string theory which. Is very interesting is the. The rediscovery. Potentially, of what, the greeks called the, atom you know uh Tom, that, which cannot, be cut, they. Were in, search of that final. Particle, that was truly indivisible, and, perhaps. With the Planck length it's not so much the particle, but.
It's, The fact that we know how small a particle, can be we have if, the universe is minecraft. We know the smallest size, of the. Pixel of the. Block right. Within which there, can be nothing, else other, than just one symbol contained, and what, is that symbol that symbol can be a configuration. Of a string so, in that sense I started thinking well if that's the case then, you. Essentially, can model the universe as, a. Data. Structure, that, has these fixed size cells that are Planck length sized cells which, have a number, of these. Symbols. In them and in, that sense it's computable, right, so, who. Knows, lots. Of different people are thinking about this in different ways max. Tegmark has, his book. He. Talks about some of this of. Course edie fredkin, and Konrad Zuse alike, I said have been thinking about this for many decades and. Stephen. Wolfram has his take on it but, there's, something here there's something here about the, fact that, computational. Constructs, on very, very small. Recipes. Can, create, this sort of useful. Emergent. Complexity, and. Now. As we know you know computers can create computational. Realities, I can, today basically. Build my own world I start the book off by saying that I got into computers, because at the age of four I ran, into a Commodore, 64, I saw, hangman, playing on the screen and it, blew my mind because I had never seen a TV screen, play. Out what, I wanted it to play out and yet, there was a keyboard, which I could touch and suddenly everything was fungible, this notion of programmability, for a four-year-old mind was, was completely, mind-blowing and from, there it went to well what can I not create and, now of course we know we can pretty much create what we want and even, the physical, dimensions. Of all of this are not gated. In any way by by, reality, these are some just, fundamental. Things from computer, science that we ought. To realize, that. The. The basic constructs, of computer, science the magnification. All constructs, the iteration, the recursion and so on and so forth applied, to very basic, specifications. Can, yield a lot and then. If you start thinking about the mind of a machine which. Any. Such, thinking. Is partial, because we don't know we you know there's a lot to be done there's a lot, of questions to be answered but. Think. About things, just in terms of differences with us well one thing we know pretty, well is that, you, know our brain. Fits. Into, a relatively. Small cranium, and consumes, about 20, watts it's, very very efficient, it has a very large number of neurons it has a very large number of connections but. For. All of its efficiency and size etc it's not really substantially, going to consume more than 20 watts in, its present form. While. We don't have a computer. That's as efficient, as a brain yet we do know that our computers, can, consume. Much more than 20 watts they don't have to sit inside a physical, cranium, that's the size, of ours, perfect. Recall which is that and. This is interesting. Because. You. Know we tend to sort of live, in the moment we get what we need from that experience. And we tend to forget and in many ways this is good for us it's good for us actually, because it avoids overload, and this is common with another thing that we do which is very aggressive, pruning, so, we tend, to go down certain.
Solutions, And we tend to discard, things that sound ridiculous, to us so. The, reason why that move that, Lisa Dole and others, found, so. Magical. Even. All the commentators, that were great, practitioners, of go was, because it, just, was, one of those things that they were willing to discard, nobody's, done this before I've never done this before in this situation, why the hell would you ever do this before it, becomes sort of common sense it, becomes the kind of thing that is, system. One thinking, you, know if you go back to Kahneman and his theory about how we think Thinking, Fast and, Slow and, we, start pruning things but, a machine intelligence can actually be, in a place can take everything, in can. Learn what it can at the time but. The original experience, is entirely preserved, so if it's ability, to extract more, knowledge from that experience, improves, over time the. Original data and full, fidelity is still available in, fact, my. Colleague professor, Bruce Porter, who's the chairman of the UT computer science department, is working on a, long-running. Project, that, does something like this which. Is that he's developing, machine so. Natural, language understanding, software. And he's, been working in that area for 35 plus years and his, approach, is that what. I can't understand, with my current algorithms. In the corpus, will get tagged in a special way and every. Iteration of the algorithm will go back and look at what the previous iteration was not able to understand, so. This sort of constant, learning with, the availability, of the fully preserved, information. That's not how we usually. Think about things we filter out a lot of stuff and then of course there's other things like being, disembodied, I mean there's no need to protect a physical, body. At all there's no need. To you. Know comply with a size limitation, and of course we know about the faster processing so you. Know the question here is if human. Beings had evolved, with, an eye also on the back of our head would. We be fundamentally. Different. Probably. Probably. I mean a lot of the amygdala. Driven response, that we have now in situations of fear where you, know we. Are. Aware that there, could be something, pretty close behind us that's got to keep us that, could get us and and therefore we have to keep at a high. State of readiness well, maybe. That we, would be less. Neurotic. You know with an eye at the back of our head who knows. So, these are all the kinds of things that of course. Machine. Intelligence we. Get to experiment with. One. Other element very, practical, element of what's happening right now, is. You. Know so we're, talking about AI of the future and the mind of a machine and so on but what's happening, right now, you. Know and on the valley everybody, knows Marc Andreessen said a while ago that software, is eating the world and that's. Sort of a euphemistic. Thing but for me it's a it's a very real, physical, thing, so, this is a you. Know, traditional. Combustion. Engine and you've got an electric motor on to the right hand side and you've got valves. And and. Spark. Plugs and EF eyes and carburetors. And you've, got a block, and you've got all sorts of things going on here and each, one of them has a specific function in each one of these mechanical, elements, performs, a specific function and then, you've got an electric motor where most of the capability, of the mechanical, elements has, been transformed, into software, the EFI for example, how you gate. Energy, is now, all software, so the, reduced number of physical, components. In that, picture. On the right is. When. You subtract, that from the picture on the left that.
Is The amount of physical stuff that software just ate okay, and there are similar pictures. Like this across a whole host of areas we. Work very closely with Boeing. I. Can, tell you that the future of aviation even, though it's not going to be very. Imminent. But Boeing's, invested, in a company that's doing. Electrical. Engines for, proper. Commuter. Aircraft you. Know not the size of a triple. Seven yet but, you, know multi seed commuter, aircraft short-haul. Electric. That, changes, a lot of things if you, look at what companies. Like velour, and hang, are doing the Chinese company hang with autonomous, drones the. City, of. Actually. The country of UAE appointed, an AI minister, recently and they've, expressed, their their, desire, to, build the world's first autonomous, flying. Taxi. Service and they, in fact even signed a contract, with a, Chinese, company and they're now looking to move that to somebody else but, they are committed to doing that, so these things even. In the narrow context. Are happening. And. They're. Certainly, coming. Too to bear. The. Last point here I'll make and then we will kind of stop for questions. Important. Thing is not, so much. Whether. A I will do everything, the. Important, thing I think is, what. All can a and I do in, a given period of time and, if. You look at the right there that's a study that was done recently, a, poll, many. AI experts, and you, may agree or disagree with some of those and some of them may be optimistic, and some of them may be pessimistic but for, example the ability to assemble, any, lego. Right. In the, next 10 years or so now, we know that when you can assemble any Lego, you're not just assembling Lego it's a fairly general purpose capability, that you have. Manufacturing. Robots, have been increasing, in. Huge. Ways if, you look at warehouse management, just five years ago and what we have now in terms of warehouse management capability, it's tremendous, and, nobody's. Going to want to discuss, you know the, poetry of Robert Frost or Rumi with a warehouse management, robot, but it is going to take it. Is going to have an impact on on jobs and so, part. Of what my. Push. Really has been particularly. On the policy side has been look. We, hear all the platitudes, that, you. Know people say the machines are coming the machines are coming but really it's, like any other revolution, it's like any other technological. Area. Of progress, era, of progress rather where, there'll be new jobs and all these people that are displaced here, will find things. To go do there and I think that's just complete, hogwash. I think that's nonsense, I think, the two, things that we've done one, we've replicated human. Muscle with the steam engine, which. Basically we cut ourselves out, of every, job that requires, muscle, and. Now by, replicating, not the entire mind but even parts of the mind slivers, of the mind that are good enough to form to, perform a function we're, at a point where we can automate much, of what we do in many many professions, and that's. All we are we. Are muscle. And mind, I mean, that's what that's what it is so there's, no third thing to go and replicate, and with. This the impact. May not be us in, barcaloungers. You, know ala Wally but. It, might be 30 40, percent unemployment and who. Needs to start, thinking about this the, people that need to start thinking about this are the people who make policies, and there. Are some countries where, for example they've gone and they've started experimenting, with themes things, like you. Know minimum, wage. Essentially. Minimum guaranteed, income. There. Are countries like France which are progressively. Reducing, the number of work hours so, that automation. Can pick up the slack and people. Can get time back I don't, know Bill Gates has proposed. A tax on robots, I am. Not proposing a specific solution but, what I am saying is that the level, at which this discussion, is happening with. The. Inevitability. Of the impact, right. Ahead. Of us I think that level of discussion, is. Insignificant. And, insufficient. And I, started, this talk about. You. Know talking about autonomous, weapons and sharing, with you the rate at which the. CCW. UN is making progress on even defining, what an autonomous, weapon is while, at the same time autonomous. Weapons are being deployed, and. Here. Again we, might find ourselves in a situation where more. And more automation. Make it into, factories. Into, retail spaces. We're, working on an engagement with a large company, in Dubai to, do. Concierge. Intelligence, based on natural language processing in, retail. Stores, it's, gonna start off being. Interesting.
And Sort, of like you, know how do you get people in into the shop and you attract them but something that's new. And shiny but it'll get pretty good pretty quickly so, that's. Kind of where we are and my hope. Is that we, can with, all of the work, that we're doing and all the conversations, we're having in Brussels, and DC, and elsewhere is to, get the, leaders of our nation and frankly. The leaders of the Western world to, take notice and to. Truly. Focus. Themselves in, developing. Policies, that, can sustain this AI powered, world of the future so. With that I'll, stop. See. If there any questions. Thank. You for your your talk. My. Question was mostly about like you know the the ending, the last thing, you talked about right I. Often. Hear about the. Need, for you, know more discussion, policy, policy. Changes right but, it really alarms, me because when you see the kind of discourse that our, elected representatives often. Have about even like really basic, technological issues you realize they have no idea what they're talking about right, and like, it. Really, worries me when they. Might be having discussions about things like this which are potentially, very complex and nuanced like, is there a solution to this like you know you mentioned, even that like you, know they're having discussions, about defining, things that are already happening right, like, what. Can we do about this. That's. A very difficult. Question, I think one thing that we should, do is to not give up so, what I'm personally, doing is. That, I try. To. The best of my ability, to insert, myself into, every, forum. Where I can, impact. Policymakers. Where. We can go and share the story and explain to them the quantum, of impact, that is coming, two. Weeks ago I was speaking at Texas, CEO summit which is an economic summit so they talked about how the state's growing and what the future of work will be but, you have leaders. State. Leadership economic, leaders and so on present, there and I. Thought that what I said may have been surprising, but, it was well-received people, were willing to listen, I've. Briefed pretty, much everyone, at the Pentagon and. I. Told, you I met with Eric Schmidt you, know that was also in connection with a cena's, event so, that's. The one thing I know how to do which is to be, out there and keep repeating the, message over and over and then as a, consequence. Of doing that you find allies you find kindred, spirits you find, in an exchange I didn't know how Eric was going to answer that question but, he happened, to answer that question in a way that supported, the, the basic thrust of what we were talking about and that.
Became 25, media articles, you, know General, Allen I recently wrote, in foreign policy that. Got reprinted. In the national, newspaper of the UAE, it showed up in Canada, showed up everywhere, so again. We. Can't force, change. Physically. But, what we can do is influence. Huge. Numbers, of minds, and in doing that we, can find allies you know this is a mind shift so. I don't expect people that have. No grounding, and no interest in this area to suddenly see, the light but. I do hope. That we'll, be able to find around, them influencers. And shapers, that, can at least carry, the day and move them forward in the direction that the, country the world needs, them to move in it's. It's not an easy process, and it's not a direct answer, but that's, the best I know how to do. Thank. You thank. You all. Right, the title. Of the book in the talk is the sentient machine and we haven't talked much about sentient, yeah, right here but, that is a goal of many of the teams working towards this self-aware, machines, with, morality and ethics of their own real. Viewpoints, and perspectives now, that might be five years away 40, years away but. It, would be a good idea to have some kind of idea of how to deal with that when it gets here before, it gets here so in the circles, that you talk in is there any discussion, about ways to manage, regulate. And interact with AI other than, as property. That's. A very, good question and by the way I completely, concur with your original, observation, that the title, of the book is the sentient machine in this talk we didn't really get into what is sentient and so on the book does cover that my view of it and just, to give you a very. Quick. Sort. Of response to that in my, view you, know intelligence, and a lot of people agree with this but not, all. Intelligence. Is about goal, directed behavior and, the. Larger, your goals usually, the more intelligent we, assess, that entity, to be. And to. Me sentience, is a combination, of intelligence and, self-awareness, so.
What, I talked about in the book is sort, of I think therefore I am that, school of thought of the, principle. Truth. Even, to myself that, I exist, is that, I can. Externalize. My, my, myself, and then observe. Myself thinking. And then say aha. I must, be and, then from there we, go we go on but. That being said you. Know, your question really is. Whether. In, these. Practical. You. Know domains, whether. There's. A realization. That. So. The, thrust of your question really is managing, sentience, or, not. Not. Creating, sentience, in these machines so there is nothing to know so the government, manages, and interacts, with people right, but the government does not manage us as though we are property, right all, of the talks that I've heard are about managing AI as property, are there any alternative, talks going on that you've observed yes, I mean I'm part of that order that have taught myself what. I say, in this book is that even if you go back to our religious. Traditions not. To take them literally but if you go back to our religious traditions what, made Adam, great was the fact that he could take action, on, his own and up. Until the creation of Adam in our ABE, Romaic system, of religions. The. Angels etc could do just what they were told to do and. The fact that Adam could do what he what, he wanted to do was, what. Made him great if we, now are, after. Millennia. And millennia, of evolution poised. At the juncture, where we can be the kind of creator that can create something, that has, its own will. To whatever. Limited degree because, we haven't figured out how much free will we have but, whatever, limited, degree I, don't, think that that's an automatic. You. Know dampener. On this, process and I don't think we should stop I also think there's a long, ways to go a lot, to learn there's a lot to do with ethical, systems and safe AI and so on and so forth so, a. New. Form of life if it, truly is sentient, should. We treat it as property, no. Thank. You yeah I. Have. Another question. In. Your book. You. Talk a lot about. Opportunities. And dangers, of AI in. All sorts of different aspects. And again I want to encourage everyone, to read the book because it covers so much more subject, matter than this. Dog need one. Of the things I found missing, was when, you write. About mind hacking. Both. On like, on a personal level and, on a national level. Right. Is, there any like, solution, that that you see like. How to defend against, that like, sort of like an anti-virus, that will say your. Mind is being hijacked or your democracy is being hijacked, yes so. In the book there's a section, called AI shields. And that. Talks about how you. Would want to use AI. To. To, fend off that kind of AI, because. Basically what's happening now is that and we will know once this investigation. Into the Russian. Involvement. In our, election is fully revealed but. Tens. And tens of thousands, of BOTS were, using. Very, simplistic, nlg, technology, to, not. Just retweet, but come up with messages, that were targeted and, the. Intent, was to shift the, prevailing, sentiment, in the in the election so. We've. Developed systems, others have developed systems that can look at that kind of generated, activity, and. Identified. As distinct, from what, somebody actually wrote and even. Otherwise, looking, at the pattern of post behaviors, and profiles, and so on to detect. Is might, be a very sophisticated bot, but ultimately this is sort of a. You. Know sort of a cycle. They, build a better bot then you've got to find out ways to detect that better bot and so on and so forth, but I think that, is very critical the other thing which you. May like or not like given, that this is Google is. In. My view there's, too much control, of algorithms. In the cloud I think. That it's. Just to give you the very simple, basic, example, is that, I want. To control, what I see, I've made a conscious, decision now, that it is not good for a. Social media service and I won't name any specific one, all of them do this for. A social media service to, decide, what, they want to show me regardless. Of how good their machine learning algorithms, are and regardless, of how good the collaborative, filtering is, and regardless, of what they think my cousin likes or what my younger, brother, you know clicked on yesterday, I, want. To be, an active, participant, in the, filtering. Of my feed and in fact what I would like is I, would like all of, my. The. Posts from my network, totally. Raw feed and on, my end I get to decide what I want to see and what I don't want to see if.
We Don't do that we're in trouble and if, we don't do that then you. Know us engineers, and builders etc. Should go do that I. Think, having this. Notion of your own AI, shield, and your own AI filter, that's very very important. Having. The convenience, of multiple, data centers all over the world and having the convenience, of not having to buy and configure, computers, and having, the convenience, of being able to get to them from any point you, know on the world is. One thing but, then also not. Control, not even knowing what the today. I can't even tell, what the hell the raw feed, is that I'm supposed to get there's just no way to get to that it's, such a glaring, omission, that to, me is ridiculous. And I think that's another area, where the. Algorithms. Need to be controlled more by us even. If the. Cloud. Platforms. Provide the data and now to say oh you know don't worry you don't have the compute power to filter your own feed nonsense, come on that's. How many how many posts. Will I get in my feed ten. Thousand a day I could. Probably do that on a Raspberry Pi you know so. All of us can do that filtering so the technical, arguments, no longer apply it's a control, argument. And it's very important, to let people. Manage. Their information. The way they want to manage it. So. In the last slide wouldn't we talk about a, lot of tea the, loss of jobs and those, jobs potentially, not, replaceable. As has been in the past. So. There there serves that the muscle, where you're you're talking about labor. Moving, to steam engines, and those, innovations at the, factory innovation that came out of it but. Not is the first. Level. Of AI implementation. That we are looking at like self-driving cars so. In some ways they seem sort, of muscle plus plus there is a little bit of intelligence on top where. Do you see that heading. To in the next ten to twenty years where, that. Muscle, plus plus really, becomes as capable. As as an, infant's, mind for example and. What will be the impact of that on unemployment. So there's if, you can, if. You can look into the future five, years ten years and maybe even 50 years and see where we end up at so in AI, research, directions I'll answer that question in two ways one of the things that I'm actually very very curious, about and have been very curious about for a long time is this, whole idea of intrinsic, motivation. Hierarchical. Reinforcement. Learning and intrinsic, motivation, there are many challenges, with this because, I, mean ultimately if you if you if you. If. You say this simply. It it sounds simple it's, not very simple to go do but, the idea is well you can have. Reinforcement. Learning pick up one task and then you can have it pick up the task and then, one can be the sub task for the other and so you can have this hierarchical, tree and if you keep building all these tasks then you'll have a pretty big coverage but. The the, challenges, with that is whether. You can really do that and whether you can have these independent. Tasks and so on and second its intrinsic. Motivation. That's another topic that I do cover in the book and, Roberto, wrote about this father of reinforcement, learning Andrew Barton wrote about this as well but, that, is so where does that flame, come from where. Does that so when you talk about a three-year-old, child to. Me it's not so much that we can't build a machine that does what a three-year-old, child does lifts, the amount of weight or can be driven to where a three-year-old child can be driven in that very mechanistic, kind, of sense the, thing about a three-year-old, child is that that.
Three-year-old Child is born with a flame and. That. Flame, of intrinsic motivation is. Something, that we need to figure out actually, some, people have hypothesized things. Like well you know what that flame really is is. Emergent. You know I talked about that in the in the in the book as well emergent. Purpose I went. Looking in philosophy, for, what, that purpose is, somewhere. Along my life I thought I knew, what I was supposed to do and then one day I felt, what, the hell am I supposed to do now and so I started reading philosophy and. You. Know at that point I realized that you know even, even, people like Camus, Albert. Camus the, French philosopher and. Writer he, said things that were, you. Know the existential. Nihilist movement, he said things like well that. Leap. Of faith is. Philosophical. Suicide. Intellectual. Suicide because. You you. Accept, that there, came a time when. Your. Knowledge. Couldn't, take you over the hump so you just assumed, in other words you did away with that gift which was your biggest gift so. I don't know that what I whether on that extreme, but. I do think that that that, internal, flame that intrinsic, motivation, would still be something that's missing I think, narrow, systems, that, have. Not. A flame, but a set of criteria that, they are going to go and and optimize. You'll. See that, much. Of I mean all. Of our system one type stuff you. Can pretty much mechanize, and now. If you look at your day and you say well how much system to type stuff do I do, depending, on who you are it's going to be variable. But, I think for. A lot of jobs the, job itself is a lot of system one stuff and that, then goes away three-year-old, child everything. Other than that intrinsic, motivation I don't think we have an answer to that intrinsic, motivation. Thank. You, thank. You so much Amir thank, you. You. You.
2018-02-08