DeepMind: The Quest to Develop Artificial General Intelligence

i was i was also an investor in deep mind before google acquired it the thing that makes deepmind unique is that heatmine is absolutely focused on creating digital super intelligence an ai that is vastly smarter than any human on earth and ultimately smarter than all humans on earth can find the deep line system can win at any game can already beat all the original atari game is super human and plays the games at super speed in less than a minute we're rapidly headed towards digital super intelligence that far exceeds any human i think it's very obvious artificial intelligence yeah the thing that has something to do with computers being smart almost everyone knows that computers can beat us at games like checkers scrabble or even fairly complex ones like chess or even jeopardy however there's still some debate about whether or not computers will ever be able to outsmart humans altogether also could computers understand human emotion could they replace us at our jobs there's a company that you probably heard of that's developed systems that are getting better than us one task at a time and now they have a level of generality that's impressive or frightening i guess omar and i made this video to showcase the extraordinary accomplishments the team at deepmind has made and make you reconsider what our future might look like if computers really do surpass humans in every way [Music] in this premium episode we'll talk about one of the few companies in the world which is dedicated to the mission of solving intelligence the company is called deepmind and this alphabet subsidiary has been praised by elon musk as one of the most promising companies to keep an eye on in terms of artificial general intelligence development or agi we'll explore how demine is developing ai systems at a crazy rate that for example are able to beat professional players at what they excel at let's say chess go and even starcraft also we'll dig into how deepmind is using artificial intelligence to accurately predict 3d models of protein structures and has the potential to accelerate research in every field of biology in the grand timeline of human history these breakthroughs could unlock a new revolution in science by unlocking the true power of ai which is discovering things that humans cannot we could not only be surpassed by digital intelligence but also depend on it even more than we do today we'll start with the story of founder and ceo demis hasibis he's cultivated a company environment that he describes as something in the middle between the best of the startup world and a research lab but with alphabet funding as a way to accelerate progress try to keep in mind throughout this episode that the primary goal of deepmind is to solve intelligence and then use it to solve everything else demis hospice the founder of deepmind was born in north london in july 1976 he's the oldest of three children among a family of diverse origins his mom is chinese singaporean and his dad has a greek cypriot background here we have his father mr hasabis himself on bbc radio explaining how demis was at a very early age he was always very interested in everything that surrounded him even two to three years old he was already um asking questions and uh he was uh interested in how we thought he was fluent reader before he even went to school before reception class but he was always free to to pursue his own intellectual interests especially that way and other interests obviously yes he always seemed to know what he wanted to do next early on demis was capable far beyond his peers in his skills and ability to learn new things he quickly showed academic proficiency and interest in board games especially chess he was literally obsessed with it and his mom even homeschooled him so he could focus on it more and just at the age of 13 he achieved the rank of chess master at age 14 he was the second highest rated player in the world only behind the hungarian trust grandmaster and strongest female player in history judith polgar not everything was easy for demis though he shared some of the difficulties coming from a working class family and the struggles that his parents had to make to attend all the events he wanted to be in demis described in his own words in an interview for the times last year i fell in love with chess at four years old he says my parents are working class and my dad gave up a lot of his career to take me around chess tournaments i remember we would stay in hostels in bunk beds dad would be in the bottom bunk and this put pressure on us because you knew the money mattered if i won the tournament we would break even on the hostile fees so you were always on the breadline and that caused the little tension in the family every weekend we would drive in my dad's camper van to the tournament usually not the very posh parts of britain we would be inside a church hall somewhere sometimes we slept in the camper van over the engine in the sleeping bags at the back it was quite fun but stressful at age 12 demis hasibis had a eureka moment an epiphany while sitting staring at a chessboard during a marathon match against the danish adult champion that lasted more than 10 hours he later told his epiphany in an interview for the achievement academy in 2017. it was like 10 hour match against i think it was the danish champion at the time the adult champion and i lost the match after 10 12 hours i remember thinking was that really a good use of of all that brain power um and i went for a long walk in some beautiful field at least i remember that in my mind in the in in the mountains uh and i remember thinking you know maybe this isn't you know there's this whole uh room full of amazingly bright people and they're using their minds to you know basically compete each other and try and win and what was really what was the you know maybe they should all be using their minds to solve you know cancer or something and i was just sort of thinking maybe we could harden you know all of these bright minds maybe there's a better use of that time and energy uh that would be better for the world and so i made this decision actually at that moment that i would move away from chess and start exploring other areas and passions of mine and and use that those meta skills as my foundations for getting to these other things but i just felt although i love chess and i still love chess now i felt was too narrow a thing to dedicate your entire life to alongside chess demis also had a fascination with technology he once used his prize money from a chess match to buy a zx spectrum computer which he soon upgraded to a commodore amiga he took it apart reprogrammed it and created a program that could play chess by itself he tried it out on his younger brother and to his amusement the program was victorious he describes why computers were so compelling in the following video and i remember thinking this is amazing that you can create a program that you know you can go to sleep afterwards and it can carry on number crunching and then you can wake up the next day and it will solve something for you that you you know you hadn't you know you you wanted them to know the answer to when you went to sleep so it seemed from again i guess i was even then obsessed with efficiency is that that seems incredible like uh a way to enhance your mind right so suddenly you can outsource problems to this computer and it can solve certain types of things for you to me that just seemed like magic and a magical extension of the mind and from then i was sort of hooked on computers and programming and and then ultimately ai demis didn't dismiss his education he finished his studies two years ahead of everyone else in his class he applied to cambridge university but they wouldn't let him start because he was just 16 years old so he took a gap year and spent his time coding a simulation game called theme park at uk studio bullfrog productions it was a great success then in 1997 he graduated with a computer science degree with honors from the university of cambridge you might imagine demis as kind of a nerdy guy who wasn't very popular but in fact he was very popular among his classmates at university here is a brief extract from a podcast on bbc radio in which his peers and ex-professors described demis during his first years at university so when he got to cambridge much to the surprise of all his other students around him he was driving a porsche it was certainly a fancy car and nobody had a car ben coppin was in the year above him was his designated buddy in those first weeks though if anything he seemed in awe of him he was almost like a minor celebrity right from the beginning you know everybody knew who he was and everyone knew about his background asabi's told radio 4's desert island disc program he parted big time in his first year at university i was working literally from the day after my a-level exam so the day i got to cambridge so i was determined at cambridge to have you know a great social time it was never like a super wild person but he was hugely popular like everybody liked him if he ever had a glass or two of wine or something he you could tell because his cheeks weren't bright red needless to say he still found time for chess his head of studies robin walker impressed i saw him in the college bar playing a game of speed chess and it was just unbelievable almost frightening to see the speed with which they could move pieces on the board after finishing his studies he founded the video games company elixir studios producing award-winning video games for global publishers such as vivendi universal and after a decade of working experience leading successful technology startups demis returned to academia to complete a phd in cognitive neuroscience at university college london followed by post-docs at mit and harvard his thesis from his doctorate was related to the links between memory and imagination he also co-authored several influential papers in the field here we have a clip of demis explaining why he studied neuroscience and why it's really the only field worth studying so from a young age at school i kind of came to this realization that in some fundamental sense there's only two subjects really worth studying physics and neuroscience and and for physics of course is all about explaining um the external world so um you know the external world out there including of course the entire universe and neuroscience and psychology is really about conversely explaining about what's inside here our internal world so this is where ai comes in because to the ultimate expression of understanding something is being able to recreate it and as richard feynman said one of my scientific heroes what i cannot build i do not truly understand and that's one of the things that i'm excited about with artificial intelligence i think ultimately will help us understand our own minds better what i cannot create i do not understand said physicist richard feynman the winner of the nobel prize in physics in 1965. hospice took this approach to heart building small bits of intelligence that grow in their level of generality this is the key insight that is catapulted deepmind to what it is today a blend of company slash research lab whose sole purpose is to build intelligent tools and then use those ai tools to solve difficult problems that would be impossible to solve without the use of vastly smarter machines it's just impossible for a human to find the next optimal move in board games such as chess every single time now imagine all the complexity and open-endedness of real-world scientific problems that open-endedness is what makes this all so fascinating the entire industry every single major tech company and particularly deep mind is working on problems that have increasing degrees of freedom now going back to demis he has long been determined to build intelligent machines he pivoted from an early interest in board games to game design then computer science and ai then he started his own entrepreneurial endeavors for 10 years before going back to university to get a phd in neuroscience demis explains this in the video why neuroscience was the last piece in the jigsaw puzzle for him to create an ai company with the purpose of solving intelligence and then the final piece of the jigsaw for me was after doing this for 10 years i sold my games company and i went back to university to do a phd in neuroscience to study how the brain itself solves some of these hard problems and um i decided to focus i chose my topics imagination and memory and an area of the brain called the hippocampus which is responsible for imagination and memory because these are capabilities too of the capabilities that we don't know how to do very well in ai and i wanted to see and get some inspiration for how the brain actually solves these problems so after a couple of postdocs mit and and harvard i then decided that i had all the ingredients and the components to um start uh deep mind and actually attack the ai problem head-on so all these experiences then culminated in 2010 and me co-founding deepmind and um and the idea behind deepmind was really to create a kind of apollo program uh mission for ai and now at deepmind we have over 100 research scientists 100 phds um top people in their machine learning fields and neuroscience fields uh working on solving ai and the type of ai we we work on is this neuroscience inspired ai so inspired by how the brain works at a very high level systems level so one way we articulate our mission at deep mind it's very easy to articulate actually obviously it's hard to do is a kind of two-step process so step one solve intelligence and then step two use it to solve everything else deepmind technologies was co-founded by demas hospis and two others one was shane legg a machine learning researcher from new zealand and two was a childhood friend mustafa suleiman an oxford dropout their company website reads like the hubble telescope that help us look deeper into space these tools are already expanding human knowledge and making a positive global impact our long-term aim is to solve intelligence developing more general and capable problem-solving systems known as artificial general intelligence deepmind is a hybrid between a startup and a research lab and the reason for that was explained by dumas in this clip from the deepmind podcast hosted by anna fry what i try to do is basically take the best from startup world the kind of focus and energy and pace that you get in the best startups um say in silicon valley and i wanted to combine that with the best from academia which is you know blue sky thinking incredibly bright people working on you know long-term big research questions and stepping into the unknown all the time but i felt if you were kind of smart about it you could extract the best of both those worlds and combine them into some kind of hybrid organization and i feel that's what deep mind is and i don't think many people have ever done that and and so that's why it seems quite strange probably as a as a organization i think we've shown with our scientific output even if you measure it by normal measures nature science papers this kind of thing that normal academic labs have measured themselves by we've been very successful the deepmind team is comprised of a diverse group of people the team is not just engineers and scientists who study computer science they collaborate and work with people from many other fields in the scientific community they also work alongside artists and other creative areas of course the team has its fair share of experts in areas such as engineering or artificial intelligence but it's worth noting that one of the most common fields of study is actually neuroscience the rationale behind that decision is likely related to dumbest's background in the field but also because the brain is the best example of general intelligence that we currently have neuroscience is essentially our human intelligence while this ai work is digital intelligence by having these team numbers it's more likely the alternate form of intelligence will develop in sites they didn't know of before this all creates a nice cycle understanding more about human intelligence can improve digital intelligence and as that improves we can learn more about and improve our human intelligence here's matt but vinnick director of neuroscience research at deepmind to explain why neuroscience is central to its mission the brain is arguably still the only example that we have of something that approximates the kind of intelligence we're trying to build and so it seems like a good idea to pay some attention to how it works at d mine the way neuroscience informs ai research i think it works at several different levels so there are cases where we're interested in very specific findings from neuroscience research like memory replay for example or the way that dopamine works in the brain another tool that we use every day at deep mind which is something called reinforcement learning which is about how you can learn to behave effectively given only rewards and punishments that also came originally out of studies of animal learning and psychology we look for opportunities to take what we've learned about those very specific aspects of neural function and then translate them into things that will work for ai in my experience so far i've i've realized that there's a subtler and maybe more important way in which neuroscience informs ai research which is that when you have neuroscientists in the building interacting with engineers and computer scientists and experts in machine learning you get a wider variety of intuitions and viewpoints and associations which ends up paying off for ai research the way i think about it is back in the days when people were trying to build machines that could fly they might have discovered the solution without worrying about how birds are able to fly and in the end the solution for machines ended up looking different from uh the solution that birds use but at the same time it seems like it would have been silly for people trying to solve flight to completely ignore birds and that's i think what we would be doing in ai research if we didn't attend to neuroscience just one year after the company was founded in 2010 deepmind raised an undisclosed amount from founders fund and horizon ventures there are some rumors according to forbes that even facebook was in conversations to acquire the company but the acquisition eventually failed in 2013 just three years after being incorporated and with a company size of around 75 employees deepmind released one of the first demonstrations of a general ai agent they showed an algorithm that was able to play some atari 2600 games at a superhuman level that was one of the very first times that an algorithm showed a wider level of generality learning just by seeing the pixels from the screen can you imagine that the algorithm is called deep q learning and here's demis explaining its working principle so what i'm going to show you is the ai playing these atari games but the only sys the only thing it gets the system gets is the raw pixels as inputs so it's just like a human looking at the screen seeing uh all the pixels on the screen so there's about 30 000 numbers per frame because the screen is 200 by 150 pixels in size and the goal here is to simply maximize the score everything is learned from scratch and we we insist that one the same system plays all the different atari games the hundreds of different atari games so i'm just going to run this video now this uh one minute video this is space invaders the most iconic game probably um on atari and this first minute this is before the first time the ai has ever seen this data stream so don't forget it doesn't know what it's playing it doesn't know what it's controlling and you can see it's actually losing its three lives it's controlling the rocket here at the bottom of the screen and it's losing its three lives immediately because it doesn't know what it's doing but after you leave it playing overnight on a single gpu machine you come back the next day and now it's superhuman at the game it's learnt for itself through experience how to play so you can see now every single shot it fires hits something it can't be killed anymore it's worked out that the pink mothership that comes across the top of the screen in a second is worth the most number of points and it does these amazingly accurate shots to do that and you can if those of you remember space invaders as there's less of them on the screen they go faster and just watch the last shot um that it does the the rocket does this is predictive shot for the to hit the last space invader so you can see how um perfectly it sort of modeled the uh the the game world and that data stream and so accurately it can predict ahead of time what is going to happen just from the pixels on the screen impressive eh later the algorithm comes again with unexpected results here demis features a strategy that the algorithm figured out by itself just by learning to play the game these kinds of insights are indicative of simple and also complex counter-intuitive things that ai could help us with in the long run in this example humans for the most part understand conceptually how to maximize their score however ai is not limited to this type of learning if we consider how well designed programs can learn about more complex games we realized that computers could learn things that we don't understand moreover computers could identify strategies that we never considered so it's a second video it's my favorite video actually this is a game of breakout and there's more gradations here of the agent getting better the system getting better so this is after 100 games so just 100 games and you can see again here the system is pretty terrible but you can probably convince yourself that maybe it's starting to get the hang of the fact that it should move the bat towards the ball um now this is after 300 games so it's now i'm hitting the ball back uh pretty consistently and it's almost never missing so it's about as good as the best humans can be at this game um and then we thought that's pretty cool what would happen if we just left the the machine playing the game for a couple more hundred games and this amazing thing happened what happened was it discovered the optimal strategy was to dig a tunnel around the left-hand side here and then send the ball you know with this unbelievable accuracy round the back so so that's really cool because um actually the the brilliant programmers and researchers are on this program um they're brilliant at programming and coming up with algorithms but they're not so good at playing atari so um so they didn't actually know that strategy for themselves so this is something that their own creation taught them these progress updates helped jumpstart an increase in the number of mentions in book publications they released a paper titled playing atari with deep reinforcement learning in and shortly after deepmind was acquired by google in early 2014 this 650 million dollar acquisition was alphabets or google's largest acquisition to date surely the guys at google understood deepmind's long-term potential and value plus acquiring a whole bunch of top-notch team members like the individuals at deepmind could be a bargain for that price the value of the company today could easily be at least 10 times the price they paid early on the depth and breadth of impact these ai systems will have could dwarf many other initiatives the company in the early years was so promising due to its early results and exceptional team that even elon musk himself invested in the company early on before google acquired it here's a quote of elon speaking about ai and confirming his personal investments in deepmind i was i was also an investor in deep mind before google acquired it it's not from the standpoint of actually trying to make any investment return it's it's um it's really i'd like to keep an eye on what's going on with artificial intelligence i think there's potentially potentially a dangerous outcome there and we need to dangerous potentially yes i mean there's been movies about this you know like terminator movies even if that's the case what do you do about it i mean what dangers do you see that you can actually do something about so you want to make sure that technology is used for good and not terminator like evil yeah i mean i don't think i mean if the in the movie terminator they didn't create ai to you know what i mean they didn't expect uh you know uh some sort of terminator-like outcome it's sort of you know like that monty python thing nobody expects this vanishing position it's just like it's you know but you got to be careful so you want you yeah you want to make sure that here's the irony i mean the man who's responsible for some of the most advanced technology in this country is worried about the advances in technology that you're aware of i mean yeah i guess that's why i keep asking what so what can you do in other words this stuff is almost inexorable isn't it how if you see that there's these brain-like developments out there can you really do anything to stop it i i don't know well what should ai be used for what's its best value i don't know but there's some scary outcomes and we should try to make sure that the outcomes are good and not bad um yeah or escape to mars if there's no other option the ai will chase us there pretty quickly it may seem odd to some folks that elon made his investment not as a way to make a profit but as a way to keep tabs on the development of ai through this investment he now had more exposure to the capabilities of the team and now with google funding and resources the potential of the company increased even more one of the key limitations of artificial intelligence methods is the access to huge amounts of computing and data and google was and probably still is one of the best positioned companies in that regard and on google's end they'll reap the benefits of the algorithms and innovations developed by the deepmind team one of the early applications that they got was the optimization of the energy usage in data centers for example the team at dmine was able to reduce google's data center cooling bill by 40 percent and with examples like that elon was further intrigued he made comments in a documentary about ai safety about demine and the dangers associated with their applications if they are left unchecked google acquired deepmind several years ago deepmind operates as a semi-independent subsidiary of google the thing that makes deepmind unique is that deepmind is absolutely focused on creating digital super intelligence an ai that is vastly smarter than any human on earth and ultimately smarter than all humans on earth combine the deep line system can win at any game you can already beat all the original atari games is super human and plays the games at super speed in less than a minute deepmind's ai has administrator level access to google's servers to optimize energy usage at the data centers however this could be an unintentional trojan horse the mind has to have complete control of the data centers so with a little software update that ai could take complete control of the whole google system which means they can do anything they can look at all your data you can do anything based on elon musk's statements regarding artificial intelligence it's no surprise that neuralink exists deep mine and similar companies are working to create a super intelligence vastly smarter than all humans combined the problem of ai control is stated by elon here we're rapidly headed towards digital super intelligence that far exceeds any human i think it's very obvious ai doesn't have to be evil to destroy humanity if ai has a goal and humanity just happens to be in the way it will destroy humanity as a matter of course without even thinking about it no hard feelings it's just like if we're building a road and an ant hill happens to be in the way we don't hate ants we're just building a road and so goodbye ant hill we we have five years i think digital super intelligence will happen in my lifetime one 100 i think it's incredibly important that ai not the other it must be us and i could be wrong about what i'm saying i'm suddenly open to ideas if anybody can suggest a path that's better but i think we're really going to have to either merge with ai or be left behind the concerns are clear digital super intelligence is one of the sure things that will happen in the near future and companies such as d-mine are building these systems with a level of generality that is both exciting and frightening at the same time deepmind and the team continue to unveil key innovations on the road to solving intelligence one of the features of that was the creation of a project called neural turing machines in which an algorithm mimics the properties of the human brain's short-term working memory heavily inspired by neuroscience the algorithm was unveiled in 2014 and it was able to mimic some of the features of our human brain's memory and even program like a human to solve problems the alphabet subsidiary has no shortage of breakthroughs the global recognition started ramping up when deepmind beat the european world champion fun hui 5-0 in a go match against the system called alphago developed by deepmind that was the first time an artificial intelligence defeated a professional go player here we have the impressions of fun hui speaking about his experience playing against the machine i play with alphago ajay huang push the stone for alphago and of course i think i will win the game with alphago because it's just a program the first game i make some mistake and i lose a game in that first match i think something clicked for him that this wasn't an ordinary go program we weren't just doing the the same as everyone else that something new was happening the second game i try to change my style but the problem is also i lose the second game and also for third game first game even the last game i lose five zero alpha go win all the game later using the same system they defeated the world champion lee sedol in a 4-1 match in favor of alphago there was even an award-winning documentary created about that historical moment only comparable to when gary kasparov the world chess champion was defeated by a super computer in 1997. here we have demis explaining the working principle behind alphago and also wired magazine's senior staff writer explaining its importance in a broader context the way we start off training alphago is by showing it a hundred thousand games that strong amateurs have played that we've downloaded from the internet and we first initially get alphago to mimic the human player and then through self-playing reinforcement learning it plays against different versions of itself many millions of times and learns from its errors these specific ideas that are driving alphago are going to drive our our future the technologies at the heart of alphago they're what are called deep neural networks which essentially mimic the web of neurons in the human brain it's a very old idea but recently due to increases in computing power these neural networks have become extremely powerful almost overnight even elon complimented deepmind's team on twitter for their unexpected accomplishment in saying congrats to deepmind many experts in the field thought ai was 10 years away from achieving this to which one of the founders mustafa suleiman replied nice one elon musk thanks for all your support over the years smiley face the winning match was not just a simple victory demis has always been striving to find unexpected insights from computer systems and that was embodied by move 37 in reference to the number of movements where alphago introduced a sort of machine based creativity into the game in order to defeat the world champion see the reaction of lisa doll and the commentators thinking that the alphago system made a mistake so i jab sees alphago plays the move 37 and aja puts the stone in the board value that's a very that's a very surprising move i thought i thought it was i thought it was a mistake when i see this move for me it's just the big shock what normally human we never play this one because it's bad it's just bad we don't know why it's bad but it's a little bit high yeah it's fifth line normally you don't make a shoulder here on the fifth line um so coming on top of a fourth line stone is really unusual yeah that's an exciting move i i think we've seen an original move here that's the kind of move uh that that you play go for hey interesting stuff this fifth line shoulder head i wasn't expecting that um i don't really know if it's a good or bad move at this point the professional commentators almost unanimously said that not a single human player would have chosen u37 so i actually had a poke around in alphago to see what alphago thought and alphago actually agreed with that assessment alphago said there was a one in 10 000 probability that move 37 would have been played by a human player so it knew that this was an extremely unlikely move it went beyond its human guide and it came up with something new and and creative and different not only was move 37 an unexpected strategy but it was also ultimately the move that caused alphago to win the match move 37 was representative of this larger ai transition taking place right under our noses this is what we referenced earlier it's expected that a sufficiently advanced ai system would do something unexpected that teaches humans something we didn't already know before these professional go players gained experience from the unusual movement made by ai in 2022 many experts felt humans fully understood the complexity and optimal strategy necessary to win this ancient game clearly that was not the case because the ai system was able to find creative strategies from unexplored movements against the world champion the developments didn't stop there for the deepmind team and like demessa said many times the goal is to create general-purpose algorithms that ultimately are flexible enough and don't depend on human input at all following these guidelines the team developed a new algorithm called alpha zero which doesn't need the data of human plays for training instead it learns by itself from scratch just by trial and error it was only told the rules of the game but not just that the new alpha zero system was now able to master not just the game of go but also shogi and chess using the same algorithm so the same algorithm was flexible enough to master three board games at a superhuman level here we have demis and david silver who leads the reinforcement learning research group at deepmind explaining the steps after beating the world champion and explaining the motivations for alpha zero there's been a big chain of events that followed on from all of the excitement of alphago when we played against lisa dole we actually had a system that had been trained on human data on all of the millions of games that have been played by human experts we eventually found a new algorithm a much more elegant approach to the whole system which actually stripped out all of the human knowledge and just started completely from scratch instead of learning from human data it learned from its own games and that became a project which we called alpha zero zero meaning having zero human knowledge in the loop alpha zero is a kind of experiment in how little knowledge can we put into these systems and how quickly and how efficiently can they learn the next stage was to make it more general so that it could play any two-player game not just go but things like chess and shogi which is japanese chess and in fact any kind of two-player perfect information game what we discovered was that actually this exceeded all of our expectations alpha zero could start in the morning playing completely randomly and then by t would be superhuman level by dinner it would be the strongest chess entity there's ever been after about eight or nine hours it was strong enough to be able to go out and defeat stockfish the incumbent world champion a program which was vastly stronger than deep blue the program which had previously defeated kasparov in the next iteration the team at dmine developed an improved version of alpha zero in which they were able to solve more tasks using the same algorithm this time with zero human data no domain knowledge and also without even knowing the rules of the game that means that the algorithm was really starting from scratch figuring out the rules and then mastering the game really by itself mu0 was more general purpose than previous algorithms because this time it was flexible enough to play the whole suite of the classic atari games alongside the board games such as go chess and shogi at an impressive level of mastery so the natural progression of the deepmind algorithms is to put the power in the algorithm instead of the human data or domain knowledge the benefit of this approach is that the algorithms are able to perform well in unexpected or applied environments one good example of this is the application of mu0 to optimize video compression in the open source vp9 codec in simple language this results in the same or even slightly better quality video streaming with less data usage at the scale of google it's simple to imagine how gigantic this cost reduction is collaborating with youtube the deepmind team was able to achieve similar quality in each video while reducing the bitrate required demonstrating on average four percent bitrate reduction across a large diverse set of live youtube videos it might seem like four percent bitrate reduction with almost the same quality is a negligible improvement but can you imagine the volume of content being uploaded to youtube or the whole internet here's jackson brosher product manager at dmi to explain its significance we're seeing a little over six percent improvement in the bit rate optimization so that directly correlates to videos that are six percent smaller being sent across the internet a six percent saving might not sound like a lot but scale that up to the whole of the internet and it's quite a significant saving it's incredible to see such improvements deployed in the wild we're already harnessing the power of ai in unexpected ways today another interesting application which is already familiar to many users without them realizing it is the technology behind google's personal assistant its voice was partly developed by deepmind through its algorithms called wavenet and wave rnn these generative models from 2016 produce vastly improved voices in comparison with the competition now they're behind google's best known services such as google assistant maps and search but not just that here's an excerpt from one of deepmind's blog posts wavenet is a general-purpose technology that has allowed us and teams at google to unlock a range of new applications from improving video calls on even the weakest connections to helping people who have lost their ability to speak regain their original voice but it's not just the embedding of this tech into google's products that's interesting here the technology is also helping to improve the lives of als patients by training custom models that help them express themselves effectively als sometimes called lou gehrig's disease is a progressive neurodegenerative disease that affects nerve cells in the brain and spinal cord over time some patients lose their ability to move certain muscles thereby making it difficult for them to speak here we have julie kaczyal project manager at google ai who talks about the use cases that this technology has in order to improve the lives of people with unfortunate disabilities my name is julie i'm a product manager here at google for the past year or so i've been working on project euphonia project euphonia has two different goals one is to improve speech recognition for people who have a variety of medical conditions the second goal is to give people their voice back which means actually recreating the way they used to sound before they were diagnosed if you think about communication it starts with understanding someone and then being understood and for a lot of people their voice is like their identity in the u.s alone roughly one in 10 people suffer acquired speech impairments which can be caused by anything from als to strokes to parkinson's to brain injuries solving it is a big challenge another gaming related ai initiative from deepmind is called alphastar it's an algorithm that was specifically designed to compete at starcraft at a professional level similar to how alpha zero did with chess and go this new challenge was tackled by deepmind and after some trials with other players they decided to have a match against mana one of the strongest players in the world he's been playing the game since he was 5 years old but like we saw with deepmind's experience in chess and go the story repeats here's a clip of mana having confidence in winning but alpha star beat him in a 5-0 match i'm hoping for a 5-0 not to lose any games but i think the realistic goal would be four and one in my favor i think he looks more confident and teal taylor was quite nervous before [Music] [Applause] the room was much more tense this time i really didn't know what to expect he's been playing starcraft pretty much since his fight [Music] i wasn't expecting the ai to be that good everything that he did was proper it was calculated and it was done well i thought i'm learning something it's much better than i expected it to be i would consider myself a good player right but i lost every single one of five games here are more of mena's reactions from a deep mind blog post i was impressed to see alpha star pull off advanced moves and different strategies across almost every game using a very human style of gameplay i wouldn't have expected he said i've realized how much my gameplay relies on forcing mistakes and being able to exploit human reactions so this has put the game in a whole new light for me we're all excited to see what comes next does that message sound familiar mana is echoing the idea that ai can teach us things we didn't already know this also shows that deepmind is making progress in ai in a much more dynamic environment such as real-time strategy games such as starcraft so we can see here that deepmind is not just working on developing these ai-based technologies but also applying them to solve huge problems and improve the quality of life of many citizens they're focused on solving huge problems in science in fact this is one of the main drivers for dumas hacibus providing tools to accelerate scientific discovery to solve the world's most urgent problems a major and recent breakthrough by the deepmind team was the creation of an ai system that can accurately predict the structure of proteins they followed the same successful strategy as previous algorithms such as alphago making use of human input data and domain knowledge in order to create an algorithm to solve a huge problem but first what is protein folding we've discovered more about the world than any other civilization before us but we have been stuck on this one problem how do proteins fold up how do proteins go from a string of amino acids to a compact shape that acts as a machine and drives life when you find out about proteins it is very exciting you could think of them as little biological nano machines they have essentially the fundamental building blocks that power everything living on this planet if we can reliably predict protein structures using ai that could change the way we understand the natural world [Music] this gigantic task is only comparable to the mapping of the human genome the implications of this breakthrough are difficult to predict but making this enormous data set available to the public to accelerate scientific discovery is one of the most remarkable things about deepmind they've not only shared countless great lectures on youtube about artificial intelligence but they've also open source many of their tools innovations and algorithms these kinds of practices are well catered to a positive future with ai in which the benefits are democratized and not just in the hands of a few powerful companies in 2020 dmine was recognized for solving the quote-unquote protein folding problem elon congratulated and praised the team for their huge achievement on twitter the scale of this major breakthrough is difficult to emphasize enough here's a chart of the relative progress they've shown with alpha fault the potential for this technology is truly revolutionary and the great benefits we'll get from it are just beginning to appear this blog post emphasizes its importance well as well as accelerating understanding of known diseases we're excited about the potential for these techniques to explore the hundreds of millions of proteins we don't currently have models for a vast terrain of unknown biology among the undetermined proteins may be some with new and exciting functions and just as a telescope helps us see deeper into the unknown universe techniques like alpha fold may help us find them continuing deep mind's great track record in may of 2022 the team released their latest agent that was able to master more than 500 different tasks a stepping stone for achieving general intelligence in the long term this is an excerpt from the abstract of the paper the same network with the same weights can play atari caption images chat stack blocks with a real robot arm and much more deciding based on its context whether to output text joint torques button presses or other tokens this work inspired by big language models such as gpt-3 was able to master a wide range of tasks even stacked blocks with a real robot arm the generality is increasing and also the level of mastery of these tasks is increasing here we have a graph that shows the level of mastery achieved per data set for example the rgb stacking simulator data set corresponds to tasks involving a robot arm in a simulated environment but they also performed well in a real one it's also important to note that there are only 1.2 billion parameters for gato this is nothing when compared to algorithms such as gbt3 which has 175 billion parameters we've shown some of the highlights of demine's progress towards artificial general intelligence the aspirations of the company are much much broader we hope these highlights not only give you an idea of the progress capabilities and motivations of the company but also open up your thinking with regards to the potential of ai and how our future could unfold based on their track record it's clear that deepmind will keep making progress towards a general-purpose intelligence system that may ultimately help us solve all problems but how long will it take here's a clip of demis hasibis answering this question at a conference for the future of life institute keep in mind he made these remarks in 2017. i think i think it partly depends on the architecture that ends up delivering human level ai so um it's the kind of neuroscience inspired fbi that we seem to be building at the moment that needs to be trained and have experience and other things to gain knowledge there may you know it may be on the order of a few years so possibly even a decade now fast forward five years to march 2022 and here's what he had to say so i think that the progress so far has been pretty phenomenal i think that it's coming relatively soon in the next you know i wouldn't be super surprised the next decade or two so it appears the timeline is tough to pin down and although he now thinks it could be 10 to 20 years from now there's no clear answer to when a truly artificial general system will come there are not many companies in the world that have the access to data computing and algorithms that deepmind has so it's exciting to watch the progress they're continuing to make we'll close this episode with some words from dumas hasibis imagining the day true agi emerges yeah i'd have dreamed about that for a very long time i think it would be more romantic in some sense if that happened where you you know one day you're coming in and then this lump of code is just executing then the next day you come in and it sort of feels sentient to you be quite amazing from what we've seen so far it will probably be more incremental and then a threshold will be crossed but i suspect it will start feeling interesting and strange in this middle zone as we start approaching that we're not there yet i don't think none of the systems that we interact with or built have that feeling of sentience or awareness any of those things they're just kind of programs that execute albeit they learn but i could imagine that one day that could happen you know there's a few things i look out for like perhaps coming up with a truly original idea creating something new a new theory in science that ends up holding maybe coming up with its own problem that it wants to solve these kinds of things would be sort of activities that i'd be looking for on the way to maybe that big day will agi ever be developed when let us know in the comments deepmind and openai are among the companies that are both working on general intelligence and one of my favorite episodes for neuropod happens to be about open ai if you missed that episode check it out here my name is ryan tanaka and this script was written by omar olivares and me thanks for watching [Music]

Channel: Neura Pod – Neuralink

Views: 200,092

Keywords: Neuralink, Elon Musk, Max Hodak, What is Neuralink?, Neura link, Neura Pod, Tesla, SpaceX, starlink, Nueralink, Nuralink, Brain computer interface, What does Neuralink do?, brain machine interface, Artificial Intelligence, Metaverse, Facebook, Neuralink News, Neuralink news 2021, Neuralink 2021, Neuralink Update, Neuralink Update 2021, Neuralink news and updates, neauralink update, Neuralink monkey, neuralink presentation 2021, neuralink pig, neuralink demos, Neuralink stock

Id: kFlLzFuslfQ

Length: 53min 29sec (3209 seconds)

Published: Fri Jul 22 2022