Scale Transform 2021 - The Future of AI Research with Sam Altman of Open AI

[Music] hey everyone really excited to jump into our next fireside chat with sam altman sam is the ceo and co-founder of open ai and the former president of y combinator sam and i first met when scale was actually going through the y combinator program back in 2016. sam is one of the people in technology who works on the most interesting set of problems in both technology and hard tech between ai nuclear energy and synthetic biology always excited to chat with sam on the future of ai sam thanks for coming thanks alex one thing i wanted to start on is i you know we mentioned this list of uh diverse and wide-ranging topics that you work on from ai to nuclear energy to synthetic biology what inspires you to work on these problems i mean i do like trying to be useful and i wish i could give like some like really sort of selfless answer but honestly i sort of like to have an interesting life and i think these are the most interesting problems in the world i think ai is um like the ai even gets close to where we think if it really is this sort of technological revolution on a scale of the agricultural revolution the industrial revolution the computer revolution um those don't come across every lifetime and so the ability to work on that is uh i feel super lucky and it's amazing to work on open ai has had some incredible breakthroughs in research over the past few years and it's been truly incredible to watch from the outside what do you think sets openai apart from other ai research organizations i think we have a maybe a unique or at least a rare combination of we are we are good at research engineering and the sort of safety policy planning thing that think tanks usually do and we have all of those in one small organization that on a per head count basis is extremely well resourced and extremely focused so we're willing to concentrate a lot of resources into a small number of bets and bring these sort of three different pieces together to do that um you know we sort of we have a plan that we think goes from here to agi i'm sure it'll be wrong in important ways but because we're because we have such a plan and because we're trying to think about how the pieces fit together and we're willing to make high conviction bets behind them uh that has let us make i think certainly relative to our size and capital outsized progress super interesting you know one question i have is how how intentional was this you know magical mix of of uh multi-disciplinary uh interests on your team as well as the strategy or is it sort of emergent from assembling a group of very smart people who you enjoyed working with i mean i said both like we intentionally thought that to do this well you would need to put everything together and then when we looked at the landscape out in the world before open ai most of the groups were really strong in one of the area maybe one and a half but no one in all three and so we very consciously like we call those the three clans of open ai i've always wanted to be good at all three but then the other thing is i just like i think really talented people that are focused together and one not only one vision but one plan to get there is like that is the rarest commodity in the world and uh like the best people are are you know steve jobs used to say this any more eloquently than anyone else but the best people are so much better than the pretty good people that if you can get a lot of them together in a super talent dense environment to sort of surprise yourself on on the upside the the sort of the central learning of my career so far has been the exponential curves are just super powerful always under almost always underestimated and usually keep going and so in some sense like by the time we started opening the eye it was clear that these curves were already going um it was clear that i think the biggest miracle needed for all of the ai which was an algorithm that could learn was behind us we can have better algorithms that can learn we can learn more efficiency efficiently but once you can learn it all once a computer can learn it all then if you sort of think about that from first principles a lot of things are going to happen and so that like the miracle was already behind us when we started um and it then became a process about doing uh like a really good job of just executing on the engineering sort of figuring out the remaining research breakthroughs and then thinking about how it all comes together in a way that is good for the world hopefully sam you're you're such a good student of history especially in terms of uh you know the the situations the events that led to many incredible innovations in the past like the internet or gps or even the computer originally what what lessons do you learn from those the histories of these incredible technologies and how do you try to apply those into your work at open ai okay i have a non-consensus answer here i study all of those things they're all like super interesting i do love reading about history um but i think most people sort of over rotate on that it's like always tempting to try to learn too much um it's always tempting to say like what did the atomic bomb people do what can we learn about climate change and i think every there are themes um you know the stuff there's some similarities and you would be very stupid not to try to take those into account um but i think the most interesting learnings and the most interesting way to think about it is like what about the shape of this new technology what about the way that this is likely to go is going to be super different than what's come before and how do we solve this with all the benefit we've had in the past but really not trying to apply that um really trying to think about the world today the quirks of this particular technology and what's going to happen is going to be different um i think ai will be super different than nuclear i think it'll be super different from climate i think it'll be super different from bell labs and uh i think most people that try to do something like this take too much inspiration uh from efforts in the past not enough yeah super interesting so how do you pick the um you mentioned that openai you know part of the strategy has been to be relatively concentrated pick small number of bats yeah that that you guys have high conviction on how do you go about picking these bats and and what what represents a good bet versus a bad bet do more of what works is part of the answer and i think weirdly most research labs have a do less of what works approach um you know there's this like thing of like oh like once once you once it works like it's no longer innovative we're gonna go look for something different we just wanna build agi and figure out how to then safely deploy it for maximum benefit but if something's working even if it then is like kind of like a little bit more boring you have to like just put in a lot of like grunt work to scale it up and make it work better we're really excited to do that um we don't we don't take the approach that personally makes very little sense to me but seems to me what most research labs in most fields not just a i do of do less of what works um so we have like some thoughts we may turn out to be wrong but so far we've been right more than we've been wrong about what it takes to make general purpose intelligence and we try to pursue a portfolio of research vets that advance us to those when we have scaled something up as much as we can when we have gotten the data to be as good as we can get it and we still can't do something that's like a really good area to go do novel research in but again the goal is to build and deploy safe agi share the benefits maximally well and whatever tactics it takes to get there we're excited about and sometimes it's surprising like sometimes you can just really scale things a lot sometimes when you thought you would work needs a really new idea um but we keep finding the new ideas and we keep finding how to make bigger computers and get better data yes so you know one of the the super interesting intellectual questions that you know many people engaged in ai have kind of all pondered on is um you know agi is this thing that at least theoretically is certainly possible because you know it's to accomplish it through our brains and and there's this interesting question of like what is the what is the technological path to arrive at agi actually look like and obviously this is a um this is almost a philosophical question more than a sort of technical a real technical question but based on what you know today all the research you all have done it at opening what you've learned through that research what do you think is the most likely path from here to something that represents agi either creating agi in a computer is a certain possibility either physics all works like we think and we're living in the physical universe and consciousness is the intelligence is sort of this emergent property of energy flowing through you know in your case a biological network in the computer's silicon but but it's going to happen or um we are living in some sort of weird simulation or we're like a dream in universal consciousness or nothing is like what we think and in any case you should live as if it's certainly going to happen and so i i still find it odd that people talk about like maybe as possible it's like one i think you should certainly live as if it's possible um and do do your work as if it's possible in terms of what we need um you know we don't like talk too much about unannounced research but but certainly i think most of the world ourselves included have been surprised by the power of these large unsupervised models and how far that can go and i think you can imagine combining one of those that sort of understands the whole world and all of human knowledge um with enough other ideas that can get that model to learn to do useful things for someone that would feel super general purpose yeah you you in in that answer you brought up one thing which i think is uh that i i'm always very impressed by with you which is kind of this thought which is you might as well believe that the technology is possible because if it is it kind of changes everything yeah there's like this old philosophical debate which is either like sort of saying descartes was right and you can either you can say that like i have this certain knowledge that i i am i exist my own subjective experience is happening and you can't be certain of anything else so like maybe this is all like you're in a virtual reality game you're dreaming it's like some apparition of a god whatever or like it really is just like physics as consensus understanding is but in that case like it's totally possible to recreate whatever the subjective experience of of self-awareness is and so it's like either you believe that physics is physics or not but in the or not case then like something else is very strange so who cares right yeah yeah well the other part i was going to mention is that it's uh part of it part of i think what is what you mentioned to find your career is like this belief in exponentials and i think you have a strong belief in exponentials then the question for these great technologies is never like yes or no it's usually when yeah right for sure um that that is another i think uh if you can like train yourself to overcome one cognitive biased bias to sort of like maximize value creation in your own life this is like the one right it's understanding these exponential curves for whatever reason evolution didn't prioritize this we're very bad at it it takes some work to overcome but if you can do it yeah it's super valuable it makes you look like a visionary there's actually a lot of uh of neuroscience research that shows that you know people uh in their brains they have circuits to do all sorts of uh you know mental operations like addition like subtraction etc we're very bad at exponentials if we can like catch a ball or throw an arrow or something yeah it's not an expert we can do the crowd apparently parabolas yep that's where it ends okay um and this is another you know in some sense also philosophical question but you know how obvious do you think it'll be when we develop our first system that represents agi you know there's a few beliefs there's one belief where you know it may kind of be this emergent uh circuit in the middle of this like giant soup of yeah of stuff and we might not even understand that it exists when it when it emerges i don't think we'll understand it when it emerges i also think that it won't this is like pure speculation i think it won't be this sort of like single moment in time it'll just be this exponential curve that keeps going um but there will be something that emerges that's quite powerful it takes us a little while to really understand do you have uh do you have a personal sort of turing test of something where if it happened it would be sort of evidence that that uh we've achieved it in terms of something that's like people always use this term slightly differently something that's like self-aware or something it's just like really generally intelligent that can learn fast what do you mean by it uh generally intelligent can learn fast capable of doing like given enough um given enough education sort of can do anything that humans could do yeah i think that's actually like not a super hard test for for what you were just saying like there's like a lot of that would be so economically valuable to the world that it will show up that way and so like once it can start doing some significant percentage of human labor really well that would pass the test for me yep very cool one topic that's really come up a lot um especially recently is this topic of a responsible and ethical ai yeah i think any powerful technology has the ability to that that will change the world has the ability to be uh you know responsible ethical good for the world overall or bad for the world overall how do you think about ensuring that the the benefits of ai are are equally distributed at open eye yeah um two principles here uh one i think the people that are going to be most affected by technology should have the most say in how it's used and what the governance for it is um i think that this is something that some of the existing big tech platforms have have gotten wrong and i believe most of the time if people understand the technology they can express their preferences for how they like that to impact their lives how they'd like that to be used how to maximize benefits minimize harms but in the moment people don't know me included people don't always have like the self-discipline to not get led astray so i can certainly say that i my best life is not like scrolling all night on my phone like reading instagram or whatever but then like on any given night i have a hard time resisting it and so i think if we ask people like show people like here's this technology how would you like it to be used you know what what do you want the rules of this advanced system to be um that's pretty good uh and then this and i think we'll get pretty good answers and the second is i really believe in some form and there's a lot of asterisks to be placed here but in democratic governance um if we if we collectively make an agi uh i think everyone in the world deserves some say into what that system will do and not doing how it's used how we share the benefits how we make decisions about kind of where the red lines are um i think it would like be bad for the world um it would lead to like it would be unfair and it would lead to a not very good outcome if a few hundred people sitting in the bay area got to sort of determine the value system of an agi along on the other side like sometimes in the heat of the moment democracy doesn't make very good decisions so figuring out how to balance these seems seems really important um but what i would like is sort of a global conversation where we decide how we're going to use these technologies and then you know my current best idea and maybe there's a better one is some form of a universal basic income or basic wealth sharing or or something where people get to sort of we we share the benefits of this as widely as we can definitely um you know one one thing that's super interesting about ai is just it's it's a very different paradigm from a lot of technologies yeah before that's hard that i think that always makes it hard that's hard with any new technology um but for me it seems maybe everyone thinks this in their own era but it seems particularly hard with this one to reason about because it's so different yeah and one of the you know one of the super interesting things that's this has really come up in a lot of um recent instances of ai is this problem of of bias that arises from the data sets right and and you know if you talk to some folks like andre karpathy's been very public about this you know the there's a belief that you know data really does sort of 80 90 of the programming of these the true quote programming of these systems how much scrutiny do you think um we should put as a community into the data sets and the the code and algorithms you know sort of relatively in the development of responsible systems um i mean i think what we care about is that we have responsible systems that behave in the way we like them to and if we can accom and again back to this sort of like do more of whatever works if we can get there with data alone that'd be fantastic if it requires some intersection of data and algorithms which i suspect will be the case plus sort of real-time human correction and user correction um that's fine too so i think we should have a design goal of responsible systems that are as fair as possible sort of do what the user wants as often as possible and i think it will take all of the above certainly i think there's a very long way to go with better data and that has been if you sort of think about the holy trinity here is data computing algorithms i'd say it's still been the most neglected yeah and i think that was a really um out of opening eye there was this amazing paper scaling laws for large language models and i think that was the the understanding that the sort of scientific understanding of how that holy trinity interacts i think was it i'm also you know someday we're going to get to models that can tell us the kind of data they need and what data they're missing and uh and then i think things can get better very quickly um you know one of the things that we uh we spoke to with one of our other uh fireside chats was drago angulov who's the head of research at waymo and one of the things that he discussed was there's this natural um almost misalignment where you know neural networks are very good at optimizing for average loss functions that's what they're just incredible at that that's what they're naturally good at and that you know uh the loss function is not representative of what your design goal is you know as you mentioned and so how do you think about how do you all think at open eye about this kind of um this misalignment is created by how the technology is developed between what you actually want the system to do and what what your loss function tells the system to do and uh and how do you think about aligning those over time yeah i mean this touches on the earlier question about bias this is why i think answering this is why i think it's not only about this is a group one example of why i think it's not only about data sets i think this is like a sort of an interesting example for how really how we dis precisely how we design these systems and what we optimize them for uh has a bunch of effects that are sort of not super obvious and depending on what you want the system behavior to be the algorithmic choices you make are really important yeah so you know there's been a number of incredible breakthrough uh results in the ai research community over the past few years uh many of them coming out of open ai like gp3 and cliff and dolly one of the one of the trends has been that these uh breakthroughs require more and more data more and more compute and more and more concentrated engineering resources and so it seems to be the case that the sort of effort required to do great research is increasing by quite a bit and the resources required increasing how do you think about this impacting kind of the the landscape of of research for ai i don't think i entirely agree with that i would say to build the most impressive sort of useful systems that does require huge amounts of data in compute so to like make gpt three or four or whatever that requires a large and complicated and high amount of various types of expertise effort and there's only a handful of the companies in the world that can do that but the fundamental research ideas that make those systems possible can still be done by a person or a handful of people and a little bit of compute in fact most of open ai's most impressive results have started that way and only then scaled up so uh it sort of makes me sad to hear researchers saying while i can't do research without a ten thousand gpu cluster i really don't think that's true um but but the part that is true is to sort of scale it up into the maximally impressive system that someone else can use uh that is going to be a narrower and narrower slice that can do it yeah so i think it's interesting i think it's um a i do think it's empowering for you know there's many researchers and prospective researchers in the audience today to believe that um you should you should certainly believe that it is possible to do great research for sure independent of all these resources but given what you just mentioned which is that to create the the most advanced systems with with the uh with the the highest performance for other organizations user requires lots of resources how do you think that uh what do you think that means for what the right collaboration between you know the research community industry and government needs to be for maximal benefit of ai technology i don't think we really know that yet i mean there's going to clearly need to be some but like collaboration is always tough right it's always like a little bit slower and a little bit more difficult to get to work than it seems like it should be and so uh what i am most optimistic about is that there will be organizations like openai that will sort of be at the forefront of creating these super powerful systems and then we'll work with this government other governments um sort of experts in other fields to figure out how do we answer these hard questions what should we do with this system how do we decide how we all get to use it so my guess is it'll be something like that yeah you one question that i think is is interesting to ask regarding ai is what are the what are the bottlenecks that you all have experienced in scaling up openai um and and do you think there's a reflective of the bottlenecks you'll continue seeing like scaling up the organization itself uh the organization and the research and results all together honestly very standard boring like there are these things that work in a 20-person organization that don't work in 150 person organization and you sort of just have to accept somewhat more process and planning and slowness in exchange for dramatically more firepower um but i don't think there's like a deep thing unique to open ai here yeah one thing that is uh you mentioned before how how openi works and and why you think it's it's been so successful but oftentimes the north stars of these organizations are are so important for how they'll develop over over decades how would you describe the mission and sort of the overall north star of open ai i think i've said it a couple of times without meaning to but our mission is to build and deploy save agi and maximally spread the benefit and it's like simple it's easy to understand it's like really hard to figure out how to build safe agi but like at least it's clear what we're going for um and if we miss it won't be because of a vague mission i really do believe that like good missions like fit in a sentence and are pretty easy to understand and i think ours is and that's like very clarifying whenever um whenever we need to make a decision the thing that i think we could do better at that i think many or almost all organizations could do better at even people who get the mission right is the tactics i recently heard of the ceo who wore a t-shirt to the office every day with his like top three or five priorities printed on it and i was like that is a good idea not only like should the ceo do that but everybody the company should wear a t-shirt with that every day so everyone's looking at it when they're talking to somebody else that's the thing that i think people don't get quite as right it's uh i i know that ceo and there's a there's a uh work remote version of that which you set your zoom background to be your top field interesting opening i just crossed it's it's five year anniversary yeah uh yeah i think so um and so it's founded a little more than five years ago i think uh you know my assumption would be that in the past five years accomplished a lot more than what you'd expected how you know maybe when you started it what were your expectations for what would be possible within this time frame and and how how have you done with respect to those i mean this probably speaks to just like absolutely delusional level of self-confidence but basically i thought it would go exactly like this exactly like this i thought uh gp3 would would uh i mean not like i didn't know it was gonna be called gpt3 but i thought we would be like about here by now and i'm thanks to like a lot of incredibly hard work from a lot of very smart people here we are so where do we get to in five years from now i don't like to make public predictions with timelines on them where do we get to next vaguely i think if you i i think like one very fair criticism of gbt3 is that it makes a lot of like stupid errors or sort of it makes it's like not very reliable and i think it's pretty smart um but the reliability is a big problem and so maybe we get to something that is like subjectively a hundred times smarter than gpt three but ten thousand times more reliable and then that's a system where you can start imagining like way more utility uh also something that can sort of like learn and update and remember in this way that gpt3 doesn't you know the context is gone all of your sort of all of the background of you is gone if the system can really sort of remember its interactions with you i think that'll make it way more useful i think one thing that's kind of happened within the field of ai research is this incredible up leveling of of what it means to ai research so you know originally if you were to think maybe 20 years ago to build world-class ai systems it would involve a lot of hand feature engineering of you know and a lot of like manual parameters to do things correctly and then uh kind of with like more modern machine learning methods that kind of went away that was more made more about hyper parameter tuning and identifying the right architectures and that was where you know 80 90 of the work went and then with the recent breakthroughs with transformers also in the architectures are just copy paste and and so it's been this like leveling up leveling up leveling up where do you think this this goes what do you think are the the things that we do today that you know take up a lot of our time with machine learning research that you know in the future are going to be meaningfully automated we still have to write the code i mean when the ai is like writing the code for us or helping us to write the code do you think it'll happen soon again no time predictions i think it'll happen at some point and i think that will like meaningfully change people's workflows what are some of the the short-term use cases of ai that that you think are are sort of right around the corner that you believe are going to be very impactful for the world on the whole um that people you know maybe aren't thinking about or aren't expecting i don't think i have anything like deeply new or insightful to say here but um you know if you think about the things in the world that we just need a lot more access to high quality versions of um you know everybody should have access to incredible educators everybody should have access to incredible medical care um we can make a long list of these uh but i think we can get there pretty soon i think i i can imagine sort of like you know gpt seven doing these things incredibly well and and that having sort of like a really meaningful impact on quality of life so i think that's awesome yeah that's it it's really great and i think we can see glimpses of that in gp3 yeah for sure the power of it to understand super early glimpses but it's clearly there the the ability to just sort of distill human knowledge is really incredible today there's more and more people going to the field of ml and machine learning research than you know ever before and if you were to uh give a few words of sort of direction to this this community of people who are all coming into machine learning all all looking to do incredible work in the field what would be kind of um a few vectors of sort of direction that you give that you give this community to be maximally impactful to humanity i will pick only one piece of advice for a young researcher which is think for yourself i think that this is like pretty good advice for almost everyone but something about the way that the academic system works right now um sort of it feels like everyone should be doing this like really novel research but it like it seems so easy to sort of get sucked into working what everybody else is working on and sort of that's how the whole reward that's what the whole reward system optimizes right and and the best researchers i know in this field but really most others too are the ones that are sort of willing to kind of trust their own intuition to follow their own instincts about what's going to work and do work that may not be popular in the field right now just keep grinding at it until they get it to work and that's what i would encourage people to do yeah you know kind of to to wrap up with respect to to ai and everything that's that's been happening today you know i think there there's uh kind of as we discussed before there's very few mental models that people can have to actually understand uh how how to think about ai and how we'll change the world just because it's a new technology with new uh fundamental characteristics um one of the things that that one topic that really interests me is how what are the kind of uh the changes to the physics of economics that ai will will encourage you know i think when uh software first came out there's an interesting change where software required a lot of costs to develop but was uh was you know basically zero cost reduced that was an incredible thing what do you think are some of the the uh qualities of ai technology or some of the the the sort of characteristics of ai that you think are going to meaningfully change how we think about the like the physics of our economic systems well like you know the cost of replicated goods went to zero software i think the cost of labor for like many definitions of the word labor should go to zero and that like makes all the models very weird so if you had to pick like one input to the economic models that is not supposed to be zero kind of like my expectation is that's it and the you know there's been this long-standing push of too much power in my opinion shifting from labor to capital but my intuition is that should go way further and i think like most of the interesting economic theory to go figure out is how to counteract that um there's all these arguments about like is it going to be deflationary inflationary it seems like obvious it should be deflationary but i think there's these other like things that are weird like what it does the time value of money i actually don't know if he's certainly said this but i've always heard attributed to mark's the quote that uh when the interest rates go to zero the capitalists have run out of ideas which is sort of interesting in a world where we've been in zero rates for so long um but like maybe we get a lot more ideas really quickly once we have agi and maybe something crazy happens with interest rates i think all of that stuff is hard to think about yeah super cool thank you so much for uh for joining us for sure sam it's always interesting to talk to you about these ideas and uh we're very thankful for how awful you are about them thanks for having me

Channel: Scale AI

Views: 2,275

Rating: 5 out of 5

Keywords:

Id: rr8N5IOvTGk

Length: 31min 38sec (1898 seconds)

Published: Mon Jul 19 2021