Ben Goertzel: Questions after the talk From Here to Human-Level AGI in 4 Simple Steps

and everyone come to the human level my outcome depends on what kind of research we're doing a lot of kinds of research right so I mean I think for many aspects of Ag our research ahead is enough if you think about it if you're dealing with natural language dialogue you can deal with a lot of grounded natural language dialogue with the head that can see in here because you can you can ground what someone says with perceived objects you can connect sight with hearing certain cross summary sensory interaction you're not manipulating stuff but the head is moving so it's action is affecting what it was it's seeing and of course it can tell someone to go away they might go away so I think actually the robot head with language vision and hearing is enough to get a long way in terms of general intelligence research on the other hand like planning planning in the in that context it's hard to do without a high level of semantic understanding I mean you can plan a conversation but that is only really meaningful once you understand a lot of what's going on the conversation where its movement lets you do research on planning in a much simpler and and more elegant way and planning is also important right so I think for for some aspects of what you want to do for a GI having a moving body is pretty valuable and I also think for understanding different perspectives on things conceptually having different perspectives on things physically is useful like here I'm in this room when I go out of the building like where's the room friend it's it's some little corner I can see from outside there as a young child you have this sort of experience all the time of seeing the same thing from different views and that that probably by abductive reasoning affects you more general understanding of perspectivism so I think I think there is a lot to be gotten by moving around but but on the other hand we swap a lot to learn from the robot head like the robot head is not having a human level dialogue yet I I mean those particular legs are expensive and use hydraulics and I mean there are applications for that but I think if you get a walking robot body that's lower cost and can be mass manufactured at a lower price point that also lets you sell home service robots and sales robots all around the world and that has indirect benefits for AGI right because then you're gathering a huge amount of data and you have a lot of case studies for the AI to do in terms of interacting with people around around the world everyone here knows everything about AGI will create a copy of myself and upload the copy I don't know it depends on how cheap the copy machine is right yeah yeah I think yeah this is what I often have said once once we've created a gi's much smarter than people essentially there's two alternatives each of us will have like what one is you you basically upload yourself into your ears Samsung phone version 25 and then and then you will be uploaded into some trans human mind cloud thing is once you're there the odds of remaining in human form are very low it just won't be appealing there'll be some of the other things to do once you're an uploaded super mind you will probably rapidly lose your sense of being an individual self and illusions like self and free will and identity will probably be forgotten so in essence you'll be fusing into some greater distributed in mind you want to do it slowly enough so you can savor the experience I guess but on the other hand hopefully assuming the age as we create our ethical we'll have the option of remaining in a human body like watched over in the people zoo by the machines of love and grace or however you want to put it heard the maybe the National Park is a better metaphor the the zoo like we we keep the squirrels and ants in the National Park for aesthetic reasons basically and probably the super a is that we create will see a lot of aesthetic value in keeping their creators around right and so I think that option should probably exist also but it doesn't have to be either/or well I mean once you have a nice like femto tech powered 3d printer you can print a bunch of copies of yourself mind upload some of them keep keep other ones in in in in human form so that's I mean that there definitely are are other scenarios than this relatively rosy one but that's what I would see is a at least moderately likely and positive outcome I mean something to worry about is the the National Park metaphor because like if you build a house you will build those the ants and the mice living under that house and I don't want you don't worry about it too much you don't go there and really look at each ant to a new dwelling or something alright now we did so we care about the ants and mice have species we want to protect them in national parks but we don't care that much about each individual guy right so if a super human a I viewed humans the way we view these other species and that's what it would like it would might keep a few billion of us in the National Park but if a few billion others happened to get bulldozed I mean it's still keeping the species I mean people get old and die all the time anyway right so we want the AGI to not just be humanely ethical we want it to be nicer to humans than we are to ourselves or to other species right and so that's that's a significant challenge in the number of regards because you also can't just tell it be nicer than we are like formalizing ethics in logic or even natural language doesn't work the only way to get human ethics across to a machine is by entering into them in shared social and emotional context but then it's getting like the whole mess of human ethics which is not all about ethical ism by its own standards and it's self contradictory in many ways so you want to get the whole mess of naturalistic human ethics into the AI by shared social emotional interaction then you want to by sort of instruction and declarative interaction you want to nudge the AGI to be on the really really ethical side of of the mess of human ethics right and that's that's an interesting challenge and that's where I think the political aspect comes in because if it's like the National Security Agency or the Chinese Ministry of intelligence it creates the first AGI I mean these agencies do a lot of good in the world therefore a lot of really nasty plots so I don't think they're bad in themselves but I don't think their goal systems are really what you want the first AGI is goal system to be nudge to and I mean you know Google Baidu intense and these are all great companies I mean I love Google's products right on the other hand in the end their purpose is to maximize shareholder value right and and their motivation now is to do that by advertising this isn't really what you want the first AG eyes motivational system to be to be nudged in in terms of right so you're gonna want a more participatory way of nudging and guiding the motivational system of the of the first AGI and that that has to do unfortunately with like politics and economics and industry structure and government policy as much as with technology right I mean there are these things are will be acting and interoperating with each other in some way yeah whoa so um so roger penrose um I take it seriously but I think a lot of what he said is wrong and doesn't make sense so man I used a new penrose a long time ago but I mean he first of all his differentiating hypothesis is more than what you just said right his differentiating hypothesis in this area is that a the human mind when it carries out creative acts like mathematical discovery is carrying out transferring hyper computation and be that this transferring Harper computation is achieved by the human brain carrying out some quantum gravity computing that goes beyond just the standard model of physics and and interoperates general relativity with quantum mechanics now the thing is I don't really believe that conglomeration of things like none of none of the current approaches to unifying quantum theory and gravity are hyper computing based including his own twister theory which is very discrete or string theory loop quantum gravity I mean they're all basically they're all basically quantum computer ball rather than hyper computable right and so on that since he's making an appeal to some physics theory that as a physicist he doesn't even have any reason to believe exists right and he's a great physicist on the other hand the idea that human mathematical creativity is generally hyper computing base seems just nonsensical to me I mean I think with the work that Joseph's group here is doing and so many others around the world were we're gonna solve mathematical creativity in the computer using Turing machines it's not going to need some some quantum theoretical due to do math better than people do so so in that sense I think what he's saying it's just is wrong it's like wishful thinking that he wishes it was he wishes that there was some amazing supercomputer thing in the trance computable thing in his brain and it feels like it like it feels like he's thinking by divine inspiration but I mean that doesn't matter what we feel like we're doing often bears no resemblance to what we're doing there's loads of examples in cognitive neuroscience of that now on the other hand if you set aside the nonsense there there's interesting evidence that there is macroscopic quantum dynamics there's Colin biology going on in the human brain which is which is quite interesting right and and we don't know how important how important it is like we know like the bird in the bird's brain there's quantum biology like macroscopic quantum coherence used on the level of many protein molecules that helps a bird navigate in the Earth's magnetic field and that there's weak evidence that like water mega molecules in between neurons in the brain have some sort of quantum coherence it's notion of weak measurement in quantum theory so there could be there's some weird clown stuff in the human brain I mean if so that doesn't necessarily mean that it makes us more intelligent than a than a classical computer of course because quantum computers it's a very broad class right and they could be stupid or smart and the brain could use it for some things and and not not for other things right so they sometimes I feel like Penrose and Hameroff from these guys actually distract attention from more interesting research into how my quantum biology impact human cognition because he you could have the same computations done by a classical machine or a quantum machine in different ways and it's quite interesting to to study how how they could be combined no I mean about trans turing hyper computation again it might well play a role in this universe that we live in we really don't know like I mean science science has explored only a very small amount of the of the universe and there's I'm in the experience of consciousness is something we understand very poorly so I mean I don't rule out that hyper computing could be important on the other hand I don't see how you get it from putting gravity in the standard nozzle together and I don't see that that's important for I don't see why you would think it's important for mathematical discovery which which Penrose says yeah I mean it's if you view his theory as sort of a placeholder for other totally different theories in the same vague direction maybe it's interesting but it's a bad theory [Music] yeah we've been paying a lot of attention to that so this is this is an issue that more people should be raising and people in the blockchain space are not paying enough attention to it so actually the the first scalable version of singularity net in essence is going to heavily exploit existing cloud computing platforms so the singularity net agents can run anywhere they could run on your phone or your pacemaker your own server or whatever but on the other hand if your singular if there happened to be a hundred thousand singularity net agents all sitting on Amazon Cloud or $0.10 cloud well they're gonna cooperate together really fast right because you can make it special infrastructure abstraction abstraction there and so that that's why the bootstrap things because right right now there is no decentralized infrastructure that does everything like AWS is API calls do or that it doesn't exist so at this moment and I'm sure by the end of 2018 running a bunch of AI agents in the scattered way all over interacting I mean it's so much just about blockchain it's just about having a decentralized infrastructure in general running it purely decentralized is going to be dramatically slower than running a bunch of agents on some big companies cloud and that's going to be true for the next few years I would imagine I mean over time everything that AWS or $0.10 clouds API gives you someone will make it available in the purely decentralized way and that that will that will be awesome alright but it's but yeah it's it's not the case right now and that was yeah like in in in this slide so on the bottom part you could have a special way of using ethereum blockchain on AWS or $0.10 cloud that made it really efficient with a bunch of agents we're running on there or if you had a bunch of agents running distributed all over in different people's houses that's a different infrastructure abstraction but it's the same thing to that to the ai-ai-ai-ai algorithm I mean Gollum is a cool idea but they don't have anything that that we can actually use to do what we need right now and even is Yasser IO there are all these other projects I mean so there's so many cool projects making scalable blockchain infrastructure and we have tried there code right like none of them has none that they do have code that works and does stuff but none of the massive platform that you could use to make a scalable singular you know that at this moment so what's uncertain is how far we will have to go in the infrastructure direction like right now we're not focusing on it because there's so many other things to do one possibility is some other project will just make the internet into a better AWS right another option is we need we need to do it and I'd say in China there's a real lot of people who own mining farms and we're now looking to repurpose their mining farms into AI farms because u s-- needs to be able to do like but used to be like a 30 mining with GPUs but now there's gonna be a six and do it better so what do you do with your GPU mining farms well becomes the deep learning mining farm or an open or an open cog learning farm or something right so there's a lot of people owning infrastructure who are interested in customizing it just for cool blockchain based projects because they have to do something with it so there's yeah there's a lot of energy going to that right now and uh yeah I've been talking to the guys in deep brain chain about collaborating on some of these things also I mean that I have a feeling the Turing test will be passed by a narrow AI actually I mean in terms of fooling and fooling the average person in a chat for maybe 10 or 15 minutes that you're a human I don't know but I'm guessing someone's gonna deepen their on that because I mean the average level of human conversations is not that subtle and there's just so much training data you could almost I mean what you can almost piece together that by cutting and pasting from all the conversations that that Googler Tencent hasn't has in their database right so I I don't necessarily trust that that test I mean a lot has advanced since Turing I mean there's no exact test for it but what one one test that would mean more to me than the Turing test is a and this one has flaws also but so there's the same robot college student test what if you could make a robot to go to this University with the same interactions that a person does and graduate right because there's many common-sense interactions there there's physical movement there's visual perception there's there's hearing you have to take the test but you also have to understand that taking a test is what you're supposed to do at that moment right so that that and that involves learning stuff about a whole bunch of different areas and the tests are there similar to things that were online but then again the robot it's not allowed to cheat during the test and look things up on the internet so I think if I wanted to make a pragmatic test I like that better than than the Turing test cuz that and that that's also part of human life but it's a part of human life that's designed that you have to be intelligent to pass it to graduate college right where's an average conversation is it is a different a different sort of game on the other hand that's obviously not a perfect test because you I mean you can imagine you could imagine an AGI that was super humanly intelligent in various ways but just at some particular deficit that it up in terms of getting through college because it's very making a rigorous test is a harder problem than you think because if you're not restricted to the human cognitive architecture there's always many clever ways to game different tests that are that are provided and in the end I don't know how important having a test is like what in a more qualitative way once we have a eyes that can do every human job better than the humans I'll be happy we've gotten to human level AGI because there's many human jobs that require a lot of generalization and transfer learning and creativity and and and many that don't but yeah that's because think about IQ test is very over fitted to human beings right I mean cuz to make an AGI the pass the IQ test doesn't mean it's intelligent in the same way as of a human passes the IQ test I mean the IQ test has all these like visual matching puzzles but clearly if you just train a neural net on the right answers to a shitload of visual matching puzzles it'll be able to do really really well on it so that's I mean you to make a test that's generic with regard to the cognitive architecture of the mind doing the test is a very interesting with quite quite difficult problem and in some ways a different problem than making the AI in the first place in my dog that they I could be about consciousness about that sympathizing consciousness oh yeah wandering various consciousness um I'm a philosophical zombie so I don't know what you guys are talking about yeah so yeah my own my own philosophy of consciousness on a person level is somewhat irrelevant to these projects I mean I I tend to be sort of pent psychist and if people are interested in philosophy of consciousness there's an interesting book by the philosopher Galen Strawson called physicalism entails pense psychism so he gives a very detailed analytical philosophy style argument that if he believed the whole world is physical which I don't necessarily but it's an interesting argument then it follows from that that everything in the world is conscious and every other position is logically incoherent and they it gives a pretty good argument I mean in essence he decomposes dualism like consciousness is something else but then how do they talk as a physical world in consciousness or like what are the properties of that antenna interface have to be and to argue for some interface between those domains that that isn't on one side or the other it's not very coherent in the end he fails to find a lot of coherent formalism for dualism on the other hand the argument that you know there's no separate domain but what consciousness just doesn't exist so it's very hard to maintain when you look at phenomenology and our own conscious experience because I mean from each of our individual point of view if you go back to like Heidegger huh sorrow and so on I mean we're starting with our perception we have a bunch of sense perceptions and we're building the physical world as a model out of this so then to argue that the raw perception that's in the beginning of our experience out of which we build the model of the physical world doesn't exist is is also a hard position to maintain in a logically coherent way whereas the idea that everything in the physical world has at least some consciousness or protocol justice or something it's completely logically coherent the only argument against it is that it seems counterintuitive to a certain percentage of the Western population because it's intuitive to essentially everyone in the Orient and in Africa but it conflicts with certain aspects of Western culture somehow and if you frame it that way then the interesting puzzles of consciousness become different they become like what is this type of physical system have the species of conscious experience that it does which clearly is very different than the species of conscious experience that Iraq has or proto consciousness that Iraq has right so why is this type of consciousness come about in this type of system with properties of the human brain are needed for things like a self and reflexive consciousness and the feeling of having will which are not necessarily there and everything that does like rah proto experience right and then which which aspects of these could you put in a digital computer and so far so far I don't see any reason to think that any of the aspects of the brain that are associated with like the uniquely human or mammalian aspects of consciousness I don't see any reason to think that any of those properties of the brain need quantum computing or Chrome gravity computing or something I mean I think it may be that the brain uses quantum computing to solve the bonding problem and help with coherence of different things across the brain but I mean so what you could do that a different way in a computer right look at like just like the bird's brain apparently uses some funky macroscopic quantum coherence to tell the direction of the Earth's magnetic field and our plane does it a different way that doesn't require macroscopic quantum coherence and does it more accurately right with compasses so I mean but still there's a lot of mystery there that remains to be unraveled so one thought experiment I've had which will be cool to do involves what I call second-person science so when we have brain computer interfacing like it would be possible for me to in essence do something like take this wire connected to my head connected to your head and we could even do it Wireless for the be a worst visual right then then increase the bandwidth to the connection eventually I think I would feel your consciousness and the like I would feel like we're becoming one mind or something and then we could decrease the bandwidth and go back to being ourselves but a bit dizzy and disoriented now what if I were to do that for example to this microphone stand right I might get a sense of what it is to be a microphone stand I'm guessing it would be less dynamic and interesting the weather's to be yosef right now on the other hand whether if we had Sophia plus plus running OpenCog version three who passed the Turing test and I wired my brain into hers right then what do I feel you know does it feel like the microphone stand does it feel clearly as aware and vivid as you or is it like something else really weird and disorienting right so that that's that will be a doable type of experiment we're not that far off on the neuroscience side now if you're willing to cut up in the head and stick a bunch of electrodes in so I mean I think that sort of approach of trying to share second person experience with a eyes as they develop is going to be the way that issues of machine consciousness get solved and they're gonna that's not going to be just an academic experiment because we're going to be sub marking it right we're gonna be wanting to put put computers into into our heads and then will you be will you be sharing experience with that intelligent computer in your head or will it really just be like an inert inert peripheral right and or something in between that we don't know how vocabulary for that should be quite interesting to find out and anyway I mean I'm I'm volunteering to be in the first couple dozen but not number one of the of the subjects yeah we're we're taking names for our early experimental mind uploading experiments yeah just have to sign a waiver I didn't I didn't know about either [Music] that's from singularity University perhaps so I wasn't I used to be involved with that with the organization's singularity University they put on a bunch of conferences called singularity summit in the US and I spoke at and they offer like executive education on the events singularity related technology for CEOs and stuff so I I since I moved to Hong Kong seven years ago I've had I've had less to do with them but I think they've started doing if event events around the world Ray Kurzweil and Peter Diamandis co-founded that I mean way back in the prehistory my employee and friend Bruce Klein organized the meeting at which singularity University was founded but now they've gone they've gone the wrong way yeah like Xi Jinping the leader of China for example yeah it's very where these are these I'm not sure if Donald Trump has conceptualized it yet yeah we keep getting pressure to build the Trump robot and it's interesting because there's not enough actuators in the upper lip right like he he's got like a prehensile upper lip to make all kinds of expressions and none of the heads and robots can do it so the dialogue would probably be easy like a statistical a Markov model could probably get it right but the the the facial movements would be will be very challenging yeah it's a different world though I know when I mean when I organized the first a GI conference in 2006 and Bethesda Maryland I mean the what it had the feeling like 50 or maybe 70 of us Mavericks were like gathering I mean it wasn't really in the musty basement it was in Marriott hotel conference room but still like you're gathering off for the corner and you wouldn't you wouldn't you wouldn't expect the leaders of the AI field in academia to come there at that time and certainly it would be it would be hard to get get your university to cover going to the conference because it looks weird so we're like gathering as Mavericks scheming how they overturned the AI establishment and bring general intelligence to the world now yet now the now the zeitgeist is is totally different than general intelligence it's part of the marketing brochure for big companies basically yeah it is very hard to chart your role in this complex causal network like that I mean look as I mentioned Shayne leg who worked for me in his 20s I got him excited about general intelligence and he went on to co-found google deepmind right so like one human contact like that can have half a causal role and in an indirect way on the other hand maybe Shane would have done that anyway right okay he didn't know that much about AGI before he said working with me but many people were going in there in that direction right and then I'm in the AGI conference series pulled together a lot of people who then went on to found companies and then do different projects so yeah I think and actually I mean I helped singularity University organize the first few singularity summits I think I gave them the idea of inviting scientists to speak at their summits instead of just business and marketing people which was it was probably helpful so yeah that's I think a whole bunch of us who were in the AI field for a long time and had these ideas in the back of our minds were being unpopular Mavericks throughout the like 80s 90s and early aughts and making a lot of noise about general intelligence and the singularity and transhumanism and all these sorts of things they even neural network cyclin when I taught a class in neural networks at University of Western Australia in 1995 no one was teaching that then it was it was gutsy an unpopular thing to be teaching neural networks to students at that time so I mean there's a whole generation of people who were pushing these things for a long time and then all of a sudden the world started to started listening which is almost counterintuitive because and I was actually doing cool things way before that but no one gave it credit for it and it's interesting in Europe now like within our singularity and OpenCog projects most of the people we have doing really deep AGI research and the stuff that I personally think is most likely to be the core algorithms behind general intelligence most of those guys are are in Europe actually saying we have no guys swallow who's French guy living in Bulgaria and Alexei polivanov has a team in st. Petersburg's and a couple guys in and it's because Europe has been training people and sort of mathematics logic probability theory and formal methods I got more the deeper technologies underlying AI which are not necessarily that popular in in commerce now right so I think yeah what we're seeing is a lot of stuff that's been developing submerge for a long time bubbling up to the surface feelings um you know it's usually frustration like why did that motor break huh but yeah in in the right moment in an interaction with the robot you get a really uncanny feeling like there's some there there I had that feeling most intensely with the philip k dick robot actually like which was I was being I was having a sort of interview with the philip k dick robot and i knew a very close friend of mines to fund boo guy was in the back room and there was a chat about operating it which had helped to create and he was overriding the chat bot no and then when it was too stupid so I even knew like the software and and the guy doing the overrides on that chat system but still while I was like sitting in this living room talking to Phil for a few minutes there I really felt I was talking to Phil and I tried to convince myself his disembodied spirit had somehow possessed the robot or the robot operator so as to bring the soul of philip k dick into that interaction because it it felt so much like that so i can if i could get that subjective feeling i could easily see how someone who doesn't know all the software and hardware behind it will have that feeling very very very strongly so another thing about sofia by the way is there are many Sofia's so we've built 12 therefore that are in common operation now so actually you don't always know and the heads and bodies can be mixed and matched to some degree also so so Sofia is really it's a character which is manifested in various bodies in various software system so that was an amusing thing when when she was made a citizen of Saudi Arabia like was it just that one Sofia it was really a citizen or is it is it any Sofia robot right what if we make a million of them are they all citizens of Saudi Arabia so they fortunately they're like no Saudi Rebbe is not the rule of law driven country so there's no detailed legal code of Saudi Arabia the legal decisions are just made by certain Imams interpreting some hadith of Islam that were made at various points in history so they can deal with it they're more flexible than Western countries right they can deal with it however they want as it comes up it's very convenient all right I think yeah I should get ready to go to the airport [Applause]

Channel: CIIRC ČVUT

Views: 6,037

Keywords:

Id: VUSrwXZJ5Yg

Length: 37min 22sec (2242 seconds)

Published: Tue May 22 2018