Google's Royal Hansen on Artificial Intelligence—Where Are We, and Where Are We Going?

foreign [Music] hi welcome back to conversations I'm Bill Crystal very pleased to be joined again for the second time in in the last year by Royal Hansen one of the most senior Executives at Google we discussed privacy Safety and Security which you're in charge of I think worldwide for Google in December of 2022 a conversation I recommend uh people look at and maybe be mostly reassured I think that they're not everyone is you know their Gmail is safe so to speak and a million other things obviously as well but um where it has long experience in this whole area though of of artificial artificial intelligence and its precursors and what it means and and you ever since you got your degree in computer science at Yale I guess a little while ago yeah I was thinking of David gallertner and artificial intelligence in the early 90s our old and you studied with him right yeah I studied with David yeah exactly that's interesting yeah so go back and look at that conversation from December I'd also I discussed this with Jim manzi who works in this area I'm was exactly five years ago and I just scanned the transcript of that last night and I gotta say it stands up well and raises questions but also it's striking how things have changed since 2018 and Jim is a pretty forward-looking guy it was certainly not minimizing what was going to happen but I think even he might be surprised by what's up so chat GPT has been in the news artificial intelligence everyone's aware of these uh I think that they've been breakthroughs or changes but where are we what's What's Happening Here is how new is it how surprising is it how fast is it moving Etc it's great and I I read that uh the transcription I thought was really insightful and helpful to have a marker at that 2018 or so point because it is a paper from Google in 2017 and the Deep Mind team that sort of um sets us on the path we're on now and I think Jim captured that nicely but even he I suspect couldn't quite have imagined where this was all going and I think you know there's always this balance between the hype of what's going on and what's really happening I think it's a it's a nice conversation and a good starting point but if you so using that as sort of the marker let's go backwards in time a bit the the invention of deep learning or some of these which is sort of an extension of machine learning which is all a subset of artificial intelligence just to give you another date sort of born in about 2011. so if it gives you a sense of and then before if you go way back 90s 80s I mean the concept of artificial intelligence goes way back but some of these Concepts were being worked on for decades so that so a lot of what you're seeing is not net new but there are significant advancements that change the shape of the curve so if you think back to 2011 people invent this idea of deep learning and I'll talk a little bit about that and they the the advances in in computational power and so let me just start with that because I think you know Jim talks in the in the conversation about Moore's Law and just how how much more computer power in simple terms is available and the in addition to just computer power you know Moore's Law doubling every 18 months and you could argue with that continues now and I don't think it doesn't quite the same way but um they people began to realize with deep learning that the real need was not for if you think about a typical computer program do this do that do this do that it's a linear sequence of instructions you know typing hitting return you know saving sending deep learning starts to say I want to do a thousand a million uh 10 million calculations at the same time and the historical computers you know your Intel you know kind of the the are in our PCS and in our servers weren't architected to do that quickly what was architected to do that quickly and people thought oh wow maybe we could use these together is gaming PCs and so you'll hear this GPU think about on your your um screen when you play a video in the old days you could watch the pixels be drawn in and you knew you were on a slower something was slowing down because the pixels had to redraw again and again and again from the top so this concept of a GPU an Nvidia which people see news about as a as a company in interview and Google has a form of this called the TPU said we can generate all the pixels on a screen at the same time again and again and again and it's just much faster for playing video games for looking at video and so you can see the benefits YouTube were much better with video than we ever heard that's because of Hardware advancements someone said that's the same thing we need for deep learning so you think about that like that progression is the first thing that begins to change the shape of the curve just far more computational power available to this now when I say these things all have to be calculated at once and Jim talks about this that like think about a word you've got a long a lot of text and it's all represented you know as a a number he talks about this and you want to predict the next number so you're you're basically asking of all the words in the world which is the next word you want to run that calculation and the probability at the exact same time that's something that the the GPU and TPU does really well so there's an intersection between video and gaming and deep learning that starts the the Curve the second thing that happens is there's a lot of what I'd call clever mathematics and this is I think you you know we alluded to a little bit of it but um in a neural network which is you know just a a long long list of inputs think about the body or the mind and we're getting all this signal and then two neurons come together and a signal comes out the other side and it goes to the next layer so you've got a long list of inputs and you've got a deep Matrix like you know it goes from one layer to the next layer to the next layer to the next layer even if you have very very powerful computers as as you're calculating each layer of that like is it one or zero one or zero one or zero when it goes eight nine ten deep what what happened in 2017 is the they realize that the thing you learn from like the seventh or the eighth or the ninth or the tenth layer in was actually relevant to the first layer and so if you really want a precise answer you got to go back and recalculate that's just super super expensive when you're talking about tens and hundreds of millions of you know inputs and so there's some very clever math and I mean we don't need to sort of bottom it out which starts to say you don't have to go back and they find ways of compressing all of those calculations to be faster and I think you know for this audience there's there is it's not just the the strength of the computer it's also some very clever math going on to speed up those calculations so if you think of those two things going on so give us that's excitable courses that lead us here so 2011 2017 too big breakthroughs but what what hap what can be done after 2011 that what do we see is consumers as or as users after 2011 that we couldn't have seen before 2011. uh and what do we see after 2017 I guess I mean Google Translates talked about a fair amount for example but that's a good example so as you get to the point where and let's talk about that 2011 2017 um you you start to be able to say I don't just want to know what's the most likely next word I want to look at this paragraph and what the relationships are between the words and I can improve the prediction of the next words I think we all noticed in Translation and in autocomplete you think about sort of the autocomplete in Gmail right um the quality going dramatically up over that period and that really is a function of this clever math of figuring out you know the the the man jumped over his fence well the man and his have a relationship the old versions didn't take that into account and so you're you know whereas now these new the sort of clever math says Ah his and the man are related and let's wait the the prediction more heavily based on that knowledge and that just extends to translation it extends to this is really beginning to get to what large language models are so so if you type in the woman I'm just being extremely simple-minded here this will now tell you that her is the pronoun or Professor if that were the her yeah I would give you or if that were the you know sentence that was part of the larger context right you would you would attach those uh in in the in the The Waiting or the mental or let's call it the mathematical model behind the scenes and this is a so this is a breakthrough in the 2011 range well the 2017 is where that that What's called the Transformer people will hear that is what happens there but it's only made possible by the this deep learning idea I see the other thing that I think you just to talk about with deep learning is with that computational power and the clever math we used to spend a lot of time as a community labeling the data so is this picture of a cat of a toaster is this you know is these word French you know you think about all the ways in which you categorize label data what deep learning really did was um you could you know this is what you start to see now with the large language models you're feeding literally exabytes of data into um the model without labeling it and the it's what I talked about before like you know it's lit it's labeling the data if that makes sense the computer is labeling the words not a bunch of humans who are going through and tagging it and that is made possible by these sort of that clever math and the usage of graphical computers this sort of um faster computers and all of it so and then that it combined with what I just described which is is there a relationship between his and the man the ball and it you know a cake and food eating and you know you can think about all the relationships that we were saying worthy it's like the SAT all of a sudden this thing can answer the analogy questions on an SAT combine that with massive amounts of data and you see the kind of progression we talk about and and I think so Bard chat CPT they're really a product of those those forces that's that's that's so interesting so just to go and just to go I want to get to Jet GPD and all part and all that which is the current thing in a second but um I was so struck by the Google translate example which sort of became famous in that New York Times article in I don't know 2016-17-ish I guess and and that Jim talks about it in the conversation we had so clearly I mean that this the move from word for word translate inputting dictionaries basically and having instantaneous uh dictionary function in Google Translate which was useful if you wanted to read an article you someone you could at least make your way through it if you didn't know the language it was clunky though because each one just got translated it wasn't really a sentence so to speak or you know it wasn't always very elegant and then as I understand it instead of having dictionaries in there if I can put it in this way you have the zillions and zillions of translations uh you know of everything from Wikipedia to Tolstoy novels to everything else and the Machine is is is using those more like a human would when a human translates something from French to English you don't do it word by word you read a sentence and you understand the sentence and you reproduce that in English right I mean some that's a different seems like a different mental function somehow and it seems like just as a user of Google translate that there was an important moment where Google translate moved from that first word for word let's call it clunky you know just just computing power dictionary to genuinely be sounding a lot more seeming a lot more like a human translator it's exactly that it's this and you'll hear people call it Transformer but it's this ability to include more than just that mapping the one-to-one mapping you described so not only you know you know his and and the man or you know the queen and the woman are are one example but the depth of those relationships to your point a paragraph of French there's a lot of relationships in that paragraph much of which we couldn't even probably articulate or even entirely agree upon as humans but if you give these models the entire Corpus of French literature it begins to figure those out and weight them this is the point about like it's not magic it's just that you've got so many combinations possible and if you run them enough times it starts to say no that's that's the right word for this that's the right sentence that's the right phrase and I and that's the the and Jim is you know right on the cusp of that and that's what you just see accelerating um after 2017. I think in 2018 Jim said pretty soon we may be able to have you go to a foreign country and uh speak into your iPhone uh or any device and and have it speak out to the locals in their language uh either show them their language on the screen but actually just also vocalize it and I remember at the time thinking oh that's amazing I was in Prague you know two months ago and I didn't myself use it and it didn't have to really but people were just using it routinely and quite impressively right so that's well I think the pixel and if you've people have seen the pixel ads if you were watching you know the NBA finals or something you saw these pixlats and they have that great ad on the sort of blacktop with this guy you know being intimidated by uh some you know he's not doesn't speak English but he shows up and says you know you know what is it my turn or something you know my and he'd uses it in real time on the blacktop so it's um that's definitely true I I just one other example of that that people may not realize if you if you haven't used lens on um in the Google photos you're now able to search the web but you take the picture so to speak and you say search for that on the web it's the same thing you're looking you know you're it's using exactly the same technology to pattern match an image that you take a picture of in real time and then go look out into the into the model to find the object that's similar and so that's different really from where we were free 2011 or even free 27. oh yeah even before 2017. yeah exactly it's not just more more powerful version of the same no it's it's that um the the hardware advancements start in 2011 2012. and then there's this additional clever math this Transformer that adds all that contextual data that's enabled all kinds of things I mean you know that's the Gmail Auto compute compute complete we even use it in our data centers to optimize the cooling for for power consumption so these are things that have just snuck their way into all kinds of processes behind the scenes amazing so okay so now we're in 2023 looks like it's the end of 2022 and chat gbt shows up maybe slightly I mean not really out of nowhere for you people of course but in the kind of public imagination slightly out of nowhere and it's writing essays and what's going on and it's channel it's answering questions or not answering questions correctly and has a point of views it seems and so what's that all about yeah it's a it's a good question I was looking at and if if people look at demos asabis as sort of the CEO of the Google Deep Mind which is where the if people you know a lot of these papers come from and um the alpha fold or alphago people know the game playing that they where they've beaten the world champions you know his point was a good one which is that these capabilities began to exist post 2017-1819 as we developed them internally I think there was always a question like what's the what are all the right form factors and and his point was that the chat bot interface it you know surprised everybody a bit how quickly people would want to play around with and engage with that interface rather than these being baked into Gmail or baked into the photos or baked into you know a medical uh device or Alpha fold which is where they did the protein um you know they sort of in in over a couple of weeks I've figured out all the protein folding possibilities which would have taken you know millions of years to do using it as a chat bot I think people wondered what would be the what exactly is the utility I think we're still figuring that out to be honest like I saw a lot of questions about exactly whether that's the right form factor for everything we want to do but that that did change the public perception of what was going on but those things that those models existed for the last several years and when people say well that's generative AI generative artificial intelligence as opposed to as opposed to simple artificial intelligence uh what do they mean by that and also relatedly I guess large language models those seem to be the two for my limited reading it's the two key two of the key phrases were yeah terms that do seem to capture something I gather somewhat new and different yeah that's right um and and I think the the language is being you know even in this in the Specialists among Specialists I I see language being sort of figured out in the vocabulary for this so nobody should feel badly if it still sort of feels a little bit um confusing the um let's see the we said large language models yeah I think of generative AI in just simple terms as you're predicting the next word the next sentence you're you're generating a photo it's you know behind the scenes you can lots of computer programs generate the next step this is just generating it in these sort of native interface of complete my sentence or do do something anticipate or predict generate an image generate a song generate a word generate a summary it's not overly complicated in that sense but a lot of back to my earlier point about the ways in which these models have been used you wouldn't think of them as having a human interface always where a human's asking you to generate the next thing but that is what a lot of computers do behind the scenes like this you think of it even in the um the data center example I give if you're making calculations about how much water to send where to cool to computers if the temperatures this it's generating a number but it's not generative in the way that we're talking about a chat bot or a or a movie maker and that's the generative side is it's predicting the next answer that is generative in a public-facing way you might say as opposed to I think that's why it is the computer you know regulating a nuclear power plant of course it has to think ahead and think ahead or whatever you were saying that's exactly right that's right yeah and so the generative now is just more obvious to the end user because they're the one making the the ask and seeing the immediate result rather than somewhere deep in the computer that's correct now I remember when autocorrect you know began and and uh I mean it's always been sort of an object of fun and almost ridicule ties because it's annoying and it miscorrects based on whatever but I remember thinking at the time well this is a little different from before that what could have spell it you could have and then order predict and when they give you the three choices the next word or whatever and and then they I gather use your own past choices to improve the prediction the the choices they're giving you so you're not getting choices is the guy down the block who uses different formulations and phrases but that's all a little different from spelling correct let's say which is just adjust which is a little more like the old dictionary right that it's programming there's only one answer or two answers right right it corrects the spelling of your name or my name if it's used enough to to suggest a correction but predicting is different it does feel different to me and of course the auto prediction as it were is uh not the most interesting use obviously but I mean but it's an interesting point like the grammar correct you know now you see the grammar and the you know better more and more Nuance to the grammar correct right that's really just to your point degenerative work because it's saying no I would have generated a different sequence so let's do you want to change it yeah same capability behind the scenes yeah what it says you might not want to this word is on maybe unnecessary that's a little different for telling you you just misspelled something with two T's instead of absolutely absolutely and that's back to this point of having the context of the whole sentence the whole paragraph look you're going to even end up in Worlds where your own your own style is introduced that's what's going to be powerful here and we'll talk about that maybe next with the large language models is that you know you're still getting a kind of General um uh prediction how one should write this or what's the most likely um next sentence in the world not for Royal or Bill but let's talk about that for example so if you're a large language model question again the simple version of large language model is if you go back like my work was David Lerner in in the in the early 90s we're dealing with parameter lists of like a few dozen that's the sort of order of the amount of complexity in 2011 you know sort of that word to vac example which Jim talked about uh or or other examples you know you're starting to get into the Millions of parameters and complexity in 2017 and 2018 you're starting to get to tens of millions hundreds of Millions but now you've seen these numbers that people talk about in 2022 and three for these large what I would call large language models now we're in the hundreds of billions of parameters and don't you know that's really just a representation of what I talked about before the size of this neural network and the depth you get a multiplier that gives you those big numbers but the point is that's not just going from one to ten to to 20 or to 30 or 1 to 100. you're talking about from 2011 or from early early 90s you're going from you know a million in in the 2011 to 100 billion and sort of bordering on trillion humans don't think about how big a leap that is and that's large language model like the the amount of context that's captured now is of that order different from what we were doing in 2011 or 2012. so it's not a different class of thing and and it happens to be done for language but the truth is now we're well outside a language we're doing using uh I don't know if you tried the music if you haven't tried the Google music you can ask it to compose a song in the style of felonious Monk and Beethoven with this basis and it will play a brand new song made up so it's not large language models that concept is yes we learned about it through chat Bots but it exists in any mode your multimodal that just means text video music whatever so that's interesting so on the music side since I kind of like music yeah I haven't done this though I haven't tried this yeah in the old days yes it could play Beethoven in the old days maybe it could it could play you five versions of the same Beethoven obviously you can play those and maybe even think about those or do something in the style I suppose of one or the other but but now it's actually can compose a new Beethoven concerto that's based on its understanding understanding of all the Beethoven concertos and all other a lot of huge numbers of other conservity from the same time I suppose and successors and those that were influenced by Beethoven etc etc is that the right way to take it that's right yeah and maybe just to kind of get started on a theme that I think we want to talk about is that in the same way I remember being in college and my music Major friends who as they would get to their final exams as you know conductors yeah and and the or composers and the exams would be things like just what we said give me a new you know take the Beethoven Sonata number 30. and translate it or sort of transform it to a Mozart style you know whatever they'd give them this it's like the SAT again right like the relationship and they'd have to do it and you know between teacher and student they would know whether they were more or less right that's what's going on but it's not because we've we've um told at all the rules for that we've just fed at the Corpus of music and it's found those relationships just like we talked about with words in a way that seems awfully similar to the way someone who has grown up in composing music has that intuition so there's an element of what I would almost call Intuition or emergent why people talk about as emergent properties the same thing with the human humans begin to sort of emerge and like oh I see the relationship or I'm able to do something quite clever because I see these things and other people didn't see it and so I think of that as that expert like things that experts can do with a lot of effort and a lot of training these large language models multimodal in this case are starting to demonstrate those same kinds of capabilities now they're not doing it because they have all the rules or because they have um you know even all the weights in the decisions that they make but somehow with that quantity of data and that clever math that kind of recurses on what they learn goes back and injects what we learn into the model again it's finding these finding these relationships um and I think that's the power of it going forward any expert who knows how to do something but it takes them a long time to kind of Cobble together the steps now can use these tools but more quickly get those things done whether you're a scientist folding proteins or whether you're scheduling PTA meetings yeah that's so interesting so so remember when I talked with Gary Casper's written of course a lot about this uh maybe that was five I don't know six seven years ago we had that conversation and he'd written I think a book about a book about this and he was struck so he lost to deep blue in 1996 or seven I think and as I understand it deep blue at the time had been programmed with all the Grand Master games basically in chess and had therefore knew what the right move was a little more reliably than even Gary the greatest chess player of his time and maybe ever but what he was struck by and I think this happened fairly recently when we spoke was that the computers now didn't need to study the games you just fed in all the rules of Chess this is back to that deep learning right it doesn't you don't need to label all the games right you know here's a good game right and this work we've led to this Victory and or defeat you just plugged in the rules of Jess go is more complicated because it's more open-ended sort of game as I gather but um and it was able to figure out the best way to play chess because it just had all the rules but what you're saying is we're now a step if I'm correct and feel free to correct me um we're a Step Beyond that because we're not really feeding in rules that the computer is figuring out you know given the all the rules given the rules of Jess you know white politicking for is is the best opening move and the following is the best response to the next move and so forth it's that it's it's more yeah as you say about emergent relationship somehow I mean somehow yeah explain to me I guess maybe try to do a little more yeah am I right that that's sort of we're beyond that now in a sense we're Beyond simply maximizing the efficiency as it were of a certain set of of playing a certain set of rules and we're so I think there's there's another concept which let me try and introduce hopefully this doesn't confuse it but it this there is the next level which you're starting to get at you are correct that like not having to label every picture of a cat in a toaster in the alphago and Alpha um you know the the chess versions and the protein you do have to feed the like what is a move like it has to know what are legal moves right but then to your point it doesn't in that case you can generate every possible move and then let the computer through that sort of think of it almost like that relationship between his and the man there are think of that as better and worse moves just and there's bazillions literally of those little adjustments that can be made what it does is it plays the game literally and like I said from instead of one million times a hundred billion times just to give you the order of you know magnitude and that's what's happening it doesn't you know the rules are almost they're they're not helpful because it can it'll figure out whether that's a better move or Not by trying all these combinations and tweaking the you know whether there's a better or worse move now what I just highlighted in there though something that is a little different large language models or you know image models or or music models are they call these foundational models so and they take literally weeks and months to train to go through all that data even with all the powerful stuff I just talked about and then you get a they call it a mo you know the model and with these all these little weights that tell you know so if you get this word you get this length you get this music whatever that's the input and it gives you back output they're trained in a very general sense but to our conversation right here if what I want to do is win a game of chess and chess could be played with words right you could you tell the you're just telling um the the game what to do but in English you could you know King F4 right um but the same thing can be done for um modeling the uh molecular possible you know molecular possibilities or for scheduling like I said scheduling meetings it's just instructions that are given out or sort of there's a language for doing it and the truth is these large language models don't get everything right you hear this word hallucination right they're they're just given their best guess given what we talked about so if you then take a small set of experts and say here was the answer that the large language model gave to this question like um what's the fastest route to you know what's the fastest the the directions or use this same kind of uh work that kind of traffic directions and they give an answer they give five answers you have experts say that's the best one that's the next best whatever and you do a small amount of that you know you have 10 experts give a look at 100 100 questions and rate them you you do what's called fine tuning the other word people will hear is uh reinforcement learning with human feedback so the large language model you don't go back and train it from the very beginning but you get a small number of experts to rate the answers and or give the best answer you feed that back into the weights of the model and all of a sudden it's not just a little better at playing chess we're not just a little better at scheduling meetings or anticipating the next word it it it almost seems again exponentially better yeah that's because deep in the model is this relationship and all and it says oh I get it I know exactly what these experts are saying and then it can extrapolate and that I think is where you're starting to see even more power is that the specialized models it's great to have chat GPT or Bard or these big models what's really great is to have a radiologist um a version of the model tuned for someone reading x-rays or for someone navigating travel in a foreign country or for you know you think of anything and each one of those can be tuned with the additional input of a small number of experts and it doesn't take weeks and months it takes seconds to improve those models yeah let's go I'm going to put Justice I think I sort of misled myself a little on the chest thing and thinking about this over the past few years because chess is such a such a constrained constraint exactly Finance of moves which is not the case with radiology and music presumably that's right so but you're saying that so let's just take I'm gonna come now just in a few minutes to the specific fields that have been most effective which medicines obviously one possibly but also traffic and a million other things scheduling and but um so what you're saying am I correcting it's on the music side maybe that's an easier one for people to understand so we've got this computer generating for one's instructions uh we've got this machine learning generating um you know most a Mozart like piano Concerto in A minor key which is what and maybe characteristics with certain other characteristics of Mozart at a certain period of composing career or whatever like that but you're do the human expert sense still have to step in and say so that's a good example so let's say so it will do that now right for music and it'll give you the answers but if you take Leonard Bernstein and a few you know a few few students whatever and you give them 10 or 100 versions of that and they say you know what like because it's not going to give you one answer it cannot give you 10 Mozart piano concerto's and see mine and you say which is the best which is the best what seems the best to you they don't even know why they just say that's the best this is sounds more like Mozart it sounds more like Mozart in that answer and then if you do that and then you take that information and this is where it's almost like a model training a model right it's like a if that makes sense you then run it back against the large language model and all of a sudden its ability to answer that category you know you think about category can be large or small depending on how you train it gets dramatically better so I do so you're seeing that I mean that's why the hallucinations or the mistake there's still a fair number of mistakes like you this is the you know you saw the New York Times piece from the very beginning with chat GPT like some some bizarre stuff comes out right but that stuff can be quickly improved with this sort of um human expert input and again you don't need to do it for 100 Years of text you just need to do it with a small number of bills and I think that's one of the things that surprised people that the the speed at which whatever deep patterns exist in the music or in the language or in the image can be tuned that the speed at which that can happen is surprising it's really quite remarkable yeah I want to get to the sort of yeah the speed that's such an interesting question of whether you you've been surprised by the and one should be surprised by the speed with which the things sort of accelerated as opposed to slowing down in a way that people kept expecting with more so I guess it's different from worse a little bit but but just on the human thing so but for now at least you still need the human expert to come in and say Human music expert is super valuable in that in that what would they call uh reinforcement learning or fine-tuning if if what I talked about in the beginning is pre-training like you do the pre-training of these large models there's still a real role in all of these fields for the experts so let's take my security team we you know we we see literally billions of these events every day on the giant Google Network our team uses this very model to help them prioritize but they have to first help tune the response by saying you know that's actually a better answer on balance and so we go through the exercise of tuning Our our usage of it with experts and that's going to happen like I think there's a really democratizing element to this is that everybody everybody knows something as an expert and that I think is what's also is quite exciting about the whole thing and why the chat bought in a way was was exciting it had put that the feeling of that power in the hands of a lot of people and I think that's where the the future is not going to be chat Bots it's going to be experts tuning these large language models for you know efficiency that everyone would benefit from but then nobody has access to the expert in I don't know what it is um farming like there'd be the elements of Niche farming that you know the internet has helped connect people but now you'll be able to get the the expertise of that Niche farmer in a model somehow and some of that expertise is expertise the Mozart specialist you know a a great composer great conductors or or Scholars of Mozart but some of that am I right or wrong some of it could just be not just be consumer feedback in a certain way that is presumably Waze works better if I having just driven a lot to less competitors as a consumer of ways tells them actually you got this a little wrong I don't know what's wrong with your algorithm there but you didn't capture the fact that there was a Slowdown here on i-86 or whatever is that simple first and the first and most obvious way to get that feedback is just to take the actions that people take in in with your model so you're right whether they're experts or not but that but I think that's that's the democratizing you know if you're using the tool and you're and you're giving its signal that's feedback but the humans are still in useful or even crucial to this development at what point computer correct computer so to speak in it yeah I think they're crucial now and you just keep moving up the order of like you know it goes from the think how much think how many of the things you do on a rescheduling a dermatology appointment and sort of what a pain in the neck it is still you can we can connect those dots I know how to we just need to sort of train the models and connect them up in a safe way kind of a responsible way and then we can move on to the next thing so I I do think it's just a matter of putting to bed a lot of the things in each field or area that we know how to do but we haven't figured out how to quickly automate it takes the software development team and a bunch of rules I mean this is to your point most automation at the moment takes somebody writing the software and they got to know the rules they may not even know what they're doing they're just writing the software now the person who knows how to do it just language becomes code that's the other thing maybe that's like instead of Java or c English becomes code it's really empowering everybody's a software developer in their own way and so how fast are things moving and I guess there are two sides of that that occur to be one is beneath the surface I mean have you been struck by the rapidity of the progress and and do we know why that's happening and do we expect that to continue or even speed up further or are we in a sort of Rapid spurt and then it subsides or are there inflection points and then maybe we get to this a bit later is from the kind of uh public-facing side how quickly does this get translated into things that people see the effects of how how different is Radiology five or ten years from now the driving cars that seemed self-driving cars that didn't seem to happen quite as quickly as people expected things happened faster than people expected maybe for just on the underlying side first I mean how you've been are you surprised by where we are compared to what you expected five years ago I am I think the last year has well let's see when I got to Google five years ago and a year or two in somebody showed me that music app but it was just internal at that point it has blown away I mean I didn't I mean I did I was amazed what it could compose so that but that step I mean is a big step and I think there's a lot of work now and that's the race to build more computers and put this in more people's hands and there's a lot of work though domain by domain you know use Case by use case if you want to think about it in that way to to take advantage of these these new capabilities that's going to take time because to your point you need the either the users or the experts to engage and you've got to put the right tools in the right people's hands and people got to learn how to do it so there's going to be a fair amount of work in the coming years to make this applicable in all these different ways and as you know as I talk to people who understand maybe the hardware or the math you know better than I do I do think we will continue to need clever mathematical or clever like we're still the fact that we're even still talking about graphical processing units and not you know AI machine learning processing units is still a sign that we're on a path to figuring out what's the optimal way to use the hardware and and I think there will still be mathematical breakthroughs this is what you know when I talk to Demis or James Van yikes or someone there's people thinking about these where it goes they're still necessary Innovations but there's also a lot of progress you will see in the coming months and years as more in you know experts individuals are able to put this capability to use in their own domain and so I would put it in those two categories right there's a lot of work right in front of us low hanging fruit to make things better and then there are still some big like what we're talking about these models don't have a model of the world right they don't they don't they don't go back to uh they're just predicting based on almost intuition they're not grounding it always in uh you know truth as as modeled somewhere in the computer there the people are working on that they don't have memory right they don't you know you kind of get what you get when you interact with it in this chatbot but over time you'll get a personalized version it will have some memory of your experiences and be able to build off of those things so there's those things will continue to happen in parallel even while people use the capabilities we have now I mean I think the one thing I would just highlight and the reason that's sort of important and I think the way it unfolded is bad guys are also using this stuff so I mean maybe we get onto that but like why we you kind of have to be you hear Google talk about bold and responsible is every time I say yeah it's democratizing you know this from your you know you better Farber than I do there are downsized to everything available to everybody bad guys use this the the technology too yeah available and so is available to criminals and to dictators criminals exactly and so doing the the other thing that's got to happen is we need to make sure we don't undo the safety benefits that we've baked into the thing we talked about last you know in December all these benefits that are baked into the um you know the big services at a Google or or you know the large software providers and the new things we just talked about right if everybody's writing code in English because English is now the coding language what safety um and guard rail safety sort of mechanisms and guardrails do we want to bake in so it's secure and safe by default that's the kind of thing that my team in this you know just sort of contextualized me for a moment that's what we're working on extend those secure by default up into these new use cases so that's where I think things will happen in the coming years yeah I mean I guess I'm struck listening to you though it sounds like you think many of these use cases I mean they are kind of low hanging fruit and as we've only or what just tell me what the just oversimplifier it sounds like we're only 10 20 30 along the path of uh improving medical Diagnostics or or improving pharmaceutical maybe uh creation or even or automobile safety or all these kind of things that we've we're now somewhat familiar with from the last three four years but this sounds like that does not require really conceptual breakthroughs at this point it requires just continuing on the road we're on but you think that's a pretty dramatic Road I guess on the other hand I think in the same way you feel you've seen people react to the chat Bots and been like wow I'm amazed at how well this can do this that it's going to do that in every domain where we can add a little bit of that expert knowledge or put it in the right you know workflow for a doctor um you know for a policy maker Etc and do you have a Instinct about what areas might be most are moving for that instinct but you you're observing it I mean whatever is seem to be moving the fastest what areas or turn out to be more recalcitrant sorry word to to this sort of improvement I think you know the obvious first places are where there's already a high degree of automation because the the thing this can't do is it can't automate the process if it the process is not already got some computer hooked up to it and can so you know there are some Fields where that's just far more true than others and my hope is this accelerates the digitization and Improvement of some of those other areas because they realize actually we could solve some of these problems uh Demis hasabis who you know the Deep Mind CEO always talks about science at digital speed so you know we all these problems that people you know a lot of scientists who can imagine testing medical or physics or other kind of you know experiments but now they could more quickly use the these this data to do science so my hope is that science and the benefits to that a real world individuals in health in safety is a is it is the is a big push but then I think the other one is just convenience right think of the you already see people stringing together instructions using these chat Bots to do some simple things so the low the other side of this is just think of uh people on their mobile phones on their laptops as they start to string together you know it's almost like macros on steroids right or sort of the you know you're just you're putting little instructions together and I think you'll see a lot of that too and then but the areas of that I mean what would that be that's your scheduling instruction you could think about where like call you know call and reschedule my doctor's appointment for a date that I'm free in uh December so it has to string together all those things those pieces of software that's going to happen quickly and your phone could call the doctor's phone and yeah or interact with their online you know you could say I mean you start to interact with online system and then they have to do a matching exercise because it has access to your calendar and it suggests state so that I think there's just a lot of that efficiency that will come that's not and it's not particularly threatening in the grand scheme and these are things that people have to do anyway at this point right right and then science yeah I think science is the other area I'm really excited about so yeah at medicine but also broader science and broader science yep interesting and where has reality put more work out sooner than people expected would you say yeah I mean I just I think the only thing so far that I would say is that if it's not digitized you can't even start yeah or if it's not language music I mean it's a form of digitization the reason we can do it languages because language been digitized music's been digitized in images um I'm trying to think about the best examples but there are certainly Industries which are older in their adoption of technology which won't be as quick to adopt this but maybe this will it will encourage them to do so I guess education is a big question right and and Theory it could it could be extremely useful but in practice maybe there's still such a huge human component or maybe that's just a an old-fashioned Prejudice that we think there's a huge human component there's a ton that could be done it's a good question you know I was talking with my wife outside about about education and sort of what will the role you know there's this whole field of what they call Prompt engineering emerging which is how do you write you know it's a version of what I talked about which is how do you write the right question or instruction to the model to get the answer you want it's a form of that expert putting their input in that that's going to become a skill right you like using your H your TI calculator that you you know when you first got that uh the derivative somehow on it on it though it's a version of that right that'll be part of school but generally it sounds like you think we're in fairly early stages in a lot of these areas I mean so if we come back in five years you expect pretty pretty big transfer changes but not quite Transformations or actual transformations in some of these areas I think it will be every bit as big if not bigger than the the timeline we've just lived through you know the term yeah I think it'd be big so the next 10 years will be as big or bigger than the preceding absolutely and then it somehow still stops short of artificial general intelligence as they call it I mean that really is the kind of computers running the world and what the science fiction stuff and dystopian or maybe utopian stuff today on your point of view we're dystopian probably but I mean we're we're not there right yeah I mean we're certainly not there now I mean in the sense that like remember there's the the model these don't have models of the World Behind them they are prediction engines in that sense very good ones and then they're demonstrating I think what's possible I you know I I am not as um I think the definitions of that can be debated uh but I do think you will see not just predictions but memory um and models of the world begin to be made useful initially in narrow Fields like medicine or but you can see where those things get strung together I have no idea you know what we'll be debating on the topic of a artificial general intelligence but I do think it's important to this responsibility angle that his countries companies we think about and work through together the the boundaries or the or the sort of safeguards as we go there I'm not as worried about the end state I'm worried about along the way are we collaborating in healthy ways as an industry and as a country and you you know this is I think the White House did a nice job here with and then we'll publish some um uh you know industry standards that we've been working you know you've seen countries come together on this so I'm I'm actually kind of encouraged about the the way we're being thoughtful we're not shutting things down and we're continuing to look for the upside the opportunity but but also talking about where we should put up guard rails but that'll be work you know this better than I do There's real policy work here there's real um inner multinational um work to get done to go across countries and what about defense I mean and National Security that the we in some sense these things of course once they're on leech they're Unleashed and they're good scientists who can and the models are available probably across almost entirely across National borders and maybe one can find them a little bit uh just recently there was we were talking to Mitchell I there was a story about the Chinese penetrating our I guess parts of the US government uh I don't know it's just and of course we're hearing these things all the time I mean how and then there's the weapons too which do seem I don't know that seems to be maybe to be underreported in terms of what's some of the progress there but people I know in defense are a little bit are don't quite know how fast that's going to go you know drones and the sort of I would just give me your general sense of that I mean I mean on the one hand I think it's been incredibly useful on the defense side we talked a little bit about it but our you know our team in at Google in the cyber security space we had a team in 2011 that was starting to figure out where could we use artificial intelligence machine learning in defending ourselves and so we've been continuing to do that and I think there's lots of upside in helping the experts who are defending in the way I just described right they're going to be beneficiaries of these tools as well but you're right um whether it's in the cyber security landscape or any other sort of interaction between the kinetic and the digital world uh we have to anticipate that people will try and use these tools so um you know just like on an expert on the Ukraine war but we've seen you know the we've seen the on the digital side we've seen a very active cyber security front and you've seen a very active drone front like again so that's been a big part of the news there's no reason these questions aren't going to be a part of those discussions and they should be and I think we should get ahead of that back to my point of norms and standards and how we work across countries I think the sooner we're working together as a um a sort of government's like-minded governments and an industry I see signs of that I think the good news is we're talking about that already right it's not yeah we're not waiting 12 years for that to emerge but it does sound and neither of us is an expert on war and and then military hardware and things like that but it does sound like it would be surprising if we didn't stuff that's happening in Ukraine that people did not really had never seen before I think it's almost safe to say or and didn't expect 10 years ago probably in terms of the drug yes and yeah it sounds to me like we should also 10 years now if unfortunately there is another word somewhere they could be further leaps forward right I mean it's it's sort of we're not uh again I'm struck just the without even anticipating massive breakthroughs or real inflection points even we're just early in this stage of whatever this revolution we're in is technological Revolution yeah and there I don't think there's a domain that's immune from those questions uh because it's a general purpose tool these are not these are enabling tools they are not end games in the in themselves right that they're it's just the start of that question and I suppose it's like early in the industrial revolution it's like wow cars you know that were trains I guess first they replace you know horse-drawn carriages and and vehicles and and that seemed like it was a huge huge thing it transformed countries and War and economies and stuff but it turns out when you step back 100 years later and you think you know well yeah well cars have trains were part of it and then cars and then planes and then a million other things I mean it we're sort of in the car stage or in the train stage not yet in the car or plane stage it sounds like I think that's right I do think your legs are going faster right so it's not going to take 100 years this time it's gonna take that's 20 years I mean to my to my I think that's what surprised people and I and I think figuring out is that how quickly a bit more expertise can change like to your point just because someone invents as a different form factor for the combustion engine it still takes decades for that to find its way into all the machines in this case those those benefits occur in the same day yeah that's that's kind of amazing further a final things people should think about in in this area that we haven't quite covered enough no I think just I would just come back to that sort of the this this balance between being bold and trying to really make this useful this is I think one of our strengths is an open and uh Society in that sense and at the same time being responsible and figuring out the right ways um to do it and I think that's you know there's enthusiasm on both sides of that and they will the the reality will get worked out in the middle and I think that's good I think being sort of both is where we've we've you know been successful as a society and I I think it to the degree we can keep that lens on it the better off we'll be that's a good a good point a good thing to end on it I think things things are changing fast enough that we'll have to have this conversation again and uh looking forward to it 18 months maybe even faster but it's been very helpful to me in sort of clarifying where we've been and and where we are and and somewhat on the one hand I I think uh non-alarmist in the sense of you know we're not looking at computers running the world in two or three years but on the other hand I think you know being sober about the the speed with which changes that it has come and is now coming I think that's your it is continuing to come we're not this isn't we haven't seen oh my God that was a huge breakthrough over the last five years now we're gonna Plateau for 10. that's not gonna happen no you're going to notice it yeah exactly well on that note Royal thanks so much for taking the time today uh to join me really in destructive discussion thanks for having me look forward to the next one yeah me too and thank you all for joining us on conversations

Channel: Conversations with Bill Kristol

Views: 5,389

Keywords:

Id: NQZrl57vZ0Q

Length: 59min 58sec (3598 seconds)

Published: Wed Jul 26 2023