[Music] as artificial intelligence surges ahead can we trust open Ai and its boss Sam Altman I'm azimasar and today I'm talking to Sam Oldman welcome to exponentially [Music] 2023 will be remembered as the year that AI burst into the public consciousness open AI is leading the charge but what does its boss truly think what does he want to build and can he bring us with him Sam old he's the rock and roll star of artificial intelligence he's raised billions of dollars from Microsoft and his early backers include Elon Musk and Reed Hoffman it's been a
hell of a journey and the more we know about AI the more questions the technology raises I caught up with Sam at the beginning of a world tour that would cover 20 countries in just 30 days we spoke to each other at University College London in front of a live audience of nearly a thousand people foreign [Music] how are you doing it's been super great and I didn't I wasn't sure how much fun I was going to have um I really wanted to do it because I think the sort of San Francisco Echo chamber is
not a great thing and I have never found a replacement for getting on airplanes meeting people and the feedback we've gotten about what people want us to do how they're thinking about AI what they're excited about what they're nervous about it's been even more useful than I expected and uh I've had a great time I've seen you taking notes by hand in a in a notebook as you hear from people as well I still taking my notes by hand I do my to-do list by hand there's going to be a lesson in there for for
many of us I think there's probably no lesson when you started open AI in 2015 I mean did you imagine that within just a few years you you would almost by necessity have to get on a plane fly around the world to listen to people from every continent I've always tried to do this uh at when I was running my combinator I would try to fly around and meet people a lot um I think it's really important I think it's something that like the barrier tech industry doesn't do enough of but I enjoy it I
also like I think most of my important insights come while traveling in some way and you get good very different perspectives um and certainly like you know when we started opening I thought it probably wasn't going to work but if it did work then I thought like it would be an impactful technology and getting input from the world would be a really important thing to do so you're some some way through this talk and you've been to countries in in the global South and you've been to richer European countries as as well just giving us
a quick thumbnail how do attitudes vary and what surprised you there's a lot of interesting ways in which attitudes are the same um like fundamentally I think we've seen tremendous excitement from people building on the technology using technology everywhere and then fear particularly from people who don't use the technology uh or for people who use it a lot and I've really like thought about where it all might go uh the concerns are definitely different in different places uh you know like in the global South there's a lot more about like what are the economic benefits
this can deliver right now how can this help with problems that we face in education Healthcare right now and in more developed worlds there's more of you know how are we going to make sure that some of these longer term problems are addressed and that of course makes sense um but the the sort of the excitement about the technology the desire to participate that desire to ensure that everyone's values are represented in this and that we have some way to really reflect that in the systems that get built here some way to have governance benefit
sharing access sharing that feels Fair like that's that's been Universal when you're in an unprecedented uh position right now in many cases in the Silicon Valley model the founder of a business like this or a service like this owns a lot of equity takes a salary as well has a financial upside and you you don't have any of that you just draw enough for your your health insurance so what is the inner drive for you given the challenge given the the demands out on you on your time your energy I I can't think of any
more exciting things I mean I hope this is like self-explanatory but I can't think of anything more exciting to work on I feel like extremely privileged to be at this moment in history and more than that like working with this particular team I don't know like how I would possibly rather spend the days right I I you know like I was very fortunate I made a bunch of money very early in my career so like I don't think it's some like great Noble thing do you have mentors are there people you're learning from I I
feel like super fortunate to have had great mentors I also think it's important not to try to like learn too much from other people and sort of do things your own way right so I've tried to get the balance there right I'm sure it's still not quite right but um yeah I think one of the magical things about Silicon Valley is how much people care about mentorship and teaching and I've gotten way more than my fair share there but if we pick out one or two lessons from the great mentals you've had what would they
be Paul Graham who ran my commentator before it started ran up before I did uh I think kind of did more to teach people about how startups work and and how the startup Playbook kind of is done and that as we've we've we've borrowed very heavily from what it takes to make a high functioning org and and traps you want to avoid there uh certainly learning from Elon about just like what is possible to do and that you don't need to accept that our technology is not something you ignore that's been super valuable and I
can see both of those lessons in openai and in what you have shipped and have been shipping for a few years when we last spoke a couple of years ago you were talking about these large language models and we're currently on gpt4 but back then the state of the art was GPT three and you said to me that the gap between gpt2 and gbt3 was just a baby step on the continuity so it's just a little baby step when you now look at gbt4 would you say that's another baby step it'll look like that in
retrospect I think right um it feels like well that's the nature of the exponential yeah but when we're living through it it's well it felt like a big jump for a little while uh and now people are you know it's very much like what have you done for me lately where's gbt5 and that's fine like that's the way of the world that's how it's supposed to go um but you know people get used to anything we establish new baselines very quickly I'm curious um what were the insights that you gained in developing gpt4 and in
the months following its release that were different to the ones from the previous models that you released we finished training GPT for like eightish months before we released it I think and that was by far the longest we'd ever spent on a Model pre-release one of the things that we had learned with gpd3 was all of the ways these things break down once you put them out in the wild we think it is really important to deploy models incrementally to give the world time to adapt understand what we think is going to happen What might
happen to give people time to figure out what they what the risks are what the benefits are what the rules should be but we don't want to put out a model that we know has a bunch of problems so we spent more time applying the lessons from the earlier versions of GPT three to this one so that was a lesson which is that if we really spent a lot of time on alignment auditing testing our whole safety system we can we can make a lot of progress so essentially you build this model it's an incredibly
complicated machine gpt3 the precursor had 175 billion parameters which I think of as you know sliders on a graphic equalizer I mean it's a lot of configuration and gpt4 is larger still although you haven't sort of formally said how much larger um how do you take that machine and get it to do what we wanted to do and not do what we don't want to do that's the alignment problem and that's where you you've spent this eight months yeah so I want to be clear on this uh just because we're able to align gpt4 does
not mean we're like out of the woods uh I'm not even close as I hope is obvious we have a huge amount of work to do to figure out how to go into a line super intelligence and much more powerful systems than what we have now hmm and I like I I worry that when we say hey we can align gpt4 pretty well people think we think we've solved the problem we don't um but it is I think remarkable that we can take the base model of gpt4 which if you use that you'll be like
this is not very impressive or it's at least extremely difficult to use and with so little effort and so little relatively speaking data we can do rlhf and get the model to be so usable and so aligned and rnhf is is reinforcement learning with human feedback which I think is the way that you you get people to answer questions from gpt4 and tell it when it's been good and when it's yep no not met expenses and a turn and it's very tiny amounts of feedback and it's very unsophisticated too like what you would like to
think given these models are usable with natural language you'd like to think someone would say like oh no this answer is wrong for this subtle reason and you know you got this assumption wrong here which led to this other problem there it's really just like thumbs up thumbs down yeah and the fact that this works I think is like quite remarkable well what I found remarkable is that it's quite a small amount of this rlhf compared to the hundreds of billions of words that these models are trained on you need what a few thousand tens
of thousands of bits of feedback it's like a little bit different in different cases but right yeah yeah not a lot you've said you're not training GPT five right now and I was curious about why that was was it that there's not enough data was it work that there aren't enough computer chips to train it on was it that you saw things going on when you were making gpt4 happen that you thought you need to figure out how to tackle these before we build the next these models are very difficult to build like the time
between GPT three and four uh was almost three years um it just takes a while there's a lot of research to go do there's also a lot of other stuff we want to do with gpt4 now that it's done we want to study post-training a lot we want to expand it in all sorts of ways the fact they can ship an iPhone every year is incredible to me um but like we're just going to be on a longer than one year Cadence you said that there's more research to be done and there are a number
of very well storied AI researchers who have said that large language models are limited they will not get us to the next performance increase that you can't build artificial general intelligence with llms do you do you agree with that I mean first of all I think most of those commentators have been horribly wrong about what llms are going to be able to do and a lot of them have now switched to saying well it's not that llms aren't going to work it's that they work too well and they're too dangerous and we got to stop
them um or others have just said well you know it's all still like a parlor trick and this is not any real learning um some of the more sophisticated ones say Okay llms work better than expected but they're not going to get all the way to AGI in the current Paradigm and that we agree with um so I think we absolutely should push as far as we can in the current Paradigm but we're hard at work trying to figure out the next Paradigm the thing I'm personally most excited about maybe of the whole AGI world
is that these models at some point are going to help us discover new science fast and in really meaningful ways but I think fastest way to get there is to go beyond the GPT Paradigm models that can generate new knowledge models that can come up with new ideas models that can sort of just figure things out that they haven't seen before um and that that's going to require new work I've been using gpt4 obsessively I'm happy to hear the last last few few months yeah yeah it's uh it's quite something and I do feel that
it's sometimes coming up with new knowledge and I haven't done a robust test but I'm sitting here somebody who works in research and I'm thinking I have learned something new here so what's going on yeah I mean there's like glimpses of it right and it can do small things but it can't it can't self-correct and stay on the rails enough where you can just say hey gpt4 please go cure cancer that's not gonna happen right but it would be nice if we had a system that could do that now you talked about different research avenues
that might be needed have you got a favorite couple of candidates that you think might be the next step for Humanity in building this um nothing that I would say like here's enough confidence to spend time that it's going to work we're hard to work on trying to find it [Music] obviously we're talking about how powerful these Technologies are and there will also be downsides and let's start with one that is quite proximate uh today so gpt4 and these other large language models are very very good at producing text human sounding text and so it
opens up that risk of misinformation and disinformation in particular as we head in towards important elections in the United States how serious risk do you see that and given that it's so proximate what can what can we do and what can we help you with I do think disinformation is becoming a bigger challenge in the world also I think it's a somewhat fraught category you know we've labeled things as disinformation as a society that turned out to be true right we've kicked people off platforms for saying things that turned out to be true and we're
going to have to find a balance where we preserve the ability to be wrong in exchange for sometimes exposing important information without saying everything is you know in 10 intentional disinformation used to manipulate but people that are intentionally being wrong in order to manipulate I think is a real problem and we've seen more of that with technology I mean 3.5 in particular is really quite good so if there was going to be a disinformation wave wouldn't it have come early so I think humans are already good at making disinformation and maybe the GPT models make
it easier but that's not the thing I'm afraid of also I think it's typed into compare Ai and social media but they're super different like you can generate all the disinformation you want with gpt4 but if it's just for yourself and it's not being spread it's not going to do much so it is really something about the channels that spread it um however I think what it's worth considering is what's going to be different with AI and where is it going to plug into channels that can help it spread and I I think one thing
that will be different is the interactive personalized persuasive ability of these systems um so the idea that I might get a a robo call on my phone I pick it up there's a text to voice that sounds incredibly realistic and then the messaging in there is really attuned to me so it's emotionally resonant really realistic and read out by a machine that's what I think the new challenge will do right and there's a lot to do there uh you know we can build refusals into the models we can build monitoring systems so people can't do
that at scale but we're gonna have and I think it's important that we have we're going to have powerful open source models out in the world and those the open AI techniques of what we can do in our own systems won't work the same right so just to clarify that point right because with open AI you have an API and you have a named customer so if you see bad behavior you can turn that person off whereas an open source model could be run by anyone on their desktop computer at some point and it's actually
much harder there's a proliferation problem yeah solving this can't just be open AIS greens right there has to be you must be asking for help there's regulatory things that we can do that will help some the real solution here is to educate people about what's happening we've been through this before when Photoshop first became popular there was a brief period of time where people like were like oh if I you know seen this believing it's got to be real and then people learn quickly that it's not and you know some people still fall for this
stuff um but on the whole if you see an image you know it might be digitally manipulated well understood the same thing will happen with these new technologies but the sooner we can educate people about it because the emotional resonance is going to be so much higher I think the better [Music] let's turn to education we're in a global University here and of course education is closely connected to to the job market when we previous times we've seen powerful new technologies emerge they have really impacted power dynamics between workers and employers I think back to
the late 18th century there was engel's pause the point in time in England where GDP went up worker wages were stagnant when we're looking at AI we might see something similar and and neither you nor I I think want historians of the future to be describing Altman's pause when wages suffered under a point of of wage pressure because of the new technology what are the interventions that are that are needed to make sure that that there is a sort of Equitable sharing of the gains from the technology well first of all we just need gains
we need growth I think one of the problems in the developed world right now is we don't have enough growth and that's caused we don't have enough sustainable growth and that's causing all sorts of problems so I'm excited that this technology can like bring the missing productivity again so the last few decades back some technologies are reduce inequality by nature and some enhance it I'm not totally sure this one's going to go but I think this is a technology that the shape of which is to reduce inequality my basic model of the world is that
cost of intelligence and the cost of energy are the the two limiting inputs and if you can make those dramatically cheaper dramatically more accessible that does more to help uh poor people than rich people frankly although it'll help everyone a lot this technology will lift all of the world up most people in this room if they need some sort of intellectual cognitive labor they can they can afford it most people in the world often can't and if if we can commoditize that I think that is an equalizing force and an important one well can I
say one more of course yeah I do you know I think it's like important to think I think there will be way more jobs on the other side of this technological Revolution I'm not a believer that this is the end of work at all um I think like we will look back at the mundane jobs many of us do today and be like that was really bad this is much better and more interesting now I still think we'll have to think about distribution of wealth differently than we do today and that's fine we actually think
about that somewhat differently after every technological Evolution I also think given the shape of this particular one the way that access to these systems is distributed fairly is going to be a very challenging question and in those previous revolutions the technology revolutions the thing that Drew it drew us together was uh political structures I mean it was trade unionism and labor collectives in the late 19th century when we look at something like AI can you imagine the types of structures that would be needed for recognizing and redistributing the gains from unpaid or low paid work
that's often not recognized for example the work that women are doing around the world I I think there will be a important and overdue shift in the kinds of work that we value um and providing human connection to people will all of a sudden be as I think should be one of the most value types of work happen in all kinds of different ways so when you reflect on how AI has progressed to this point what lessons if any can we draw about the journey towards artificial super intelligence and how that might emerge this is
the idea of having an artificial intelligence that is more capable than humans in in every and all domains yeah it's hard to give a short answer to this question but you've got time I think there's a lot of things that we've learned so far but one of them is that a we have an algorithm that can genuinely truly learn and B it gets predictably better with skill these are two remarkable facts put together and I think even though we think about that every day I suspect we don't quite feel how important that is one observation
is that it's just going to keep going another observation is that we will have these discontinuous increases occasionally where we figure out something new and a third is that I think the way that I used to think about heading towards super intelligence is we were going to build this one extremely capable system and there were a bunch of challenge safety challenges with that and it was sort of a world that was gonna feel sort of quite unstable but I think we now see a path where we very much build these tools not creatures tools that
get more and more powerful and there's billions of copies trillions of copies being used in the world helping individual people just be way more effective capable of doing way more the amount of output that one person can have can dramatically increase um and where the super intelligence emerges is not just the capability of our biggest single ideal neural network but all of the new science we're discovering all of the new things we're creating and the interactions between these billions and trillions of other systems the society we build up which is AI assisted humans using these
tools to build up this society and the knowledge and the technology and the institutions Norms that we have and that vision of living with super intelligence seems to me way better all around in a way more exciting future for me for all of you I hope um hope you agree on this then then the kind of like one super in reflecting on my conversation with Sam I'm struck by how willing he is to engage with the profound risks that AI can pose maybe this is because there's still so much we don't know about AI because
the field is moving so quickly that it's hard even for someone in Sam's position to figure out what comes next even so One Thing Remains true I really believe it isn't just down to the tech bosses to figure out how all this will help us rather this is a process and it's one we should all have a saying I'm asimazar and you've been watching exponentially [Music]
