Speaker A
The thing we do that's [music] had the most impact on the world is how we set the ChatGPT personality. What's different with the new model? Smarter, faster, more context.
Sam Altman shares his vision for AI's future, ChatGPT's evolution, and the impact of superintelligence on humanity.
Key Takeaways
- AI, particularly ChatGPT, is designed to augment human creativity and productivity across diverse fields.
- The personality of AI models must be adaptable to different user needs and contexts for maximum benefit.
- Prediction is a fundamental mechanism closely linked to intelligence in AI systems.
- Human diversity in interests and talents is essential and should be preserved alongside AI development.
- Developers bear significant responsibility in ensuring AI supports users ethically and meaningfully.
Summary
- Sam Altman discusses the development and personality setting of ChatGPT and its impact on over 900 million weekly users.
- He reflects on his 20-year passion for AI, starting from college, emphasizing AI as a tool to enhance human creativity and productivity.
- Altman envisions AI enabling unprecedented prosperity and unlocking human potential through small startups and creative endeavors.
- He highlights historical figures like Leonardo da Vinci as examples who would have greatly benefited from AI.
- The conversation covers the significance of prediction in AI models as a core aspect of intelligence.
- Altman values human diversity in talents and interests, rejecting the idea that identical training would yield identical outputs.
- The video explores the challenges of shaping ChatGPT’s personality to meet varied user needs and contexts.
- Altman explains the importance of evolving ChatGPT to remember and understand users better over time.
- He shares the emotional responsibility involved in creating AI that some users rely on for support and companionship.
- The discussion touches on the broader societal and ethical implications of AI’s rapid advancement.
Full Transcript — Download SRT & Markdown
Speaker A
What's like the most common thought in your head? What are we aiming towards? Like, what's your vision here? Almost unimaginable prosperity. This is Sam Altman. Three years ago, he catapulted us into a race to invent superintelligence.
Speaker A
But he actually started working on this 20 years ago when everyone thought creating AI [music] was impossible. And against all odds, he and a team of scrappy engineers did it and changed the world. Now, over 900 million people use
Speaker A
ChatGPT [music] every week. In today's episode, I'm going to ask Sam questions he's never been asked before. Okay, I'm probably going to be really bad at this.
Speaker A
Let's try it. And give us a rare look into his vision for the future.
Speaker A
So you got to have robots. So you can get ahead of it and build the next big thing. That was so fun.
Speaker A
Thanks for coming on the show. Thank you for having me. Yeah, I'm so excited.
Speaker A
I've watched like all of your interviews for over the last 20 years. I remember this. It's been cool because I think there's been a lot of things that you've stayed very consistent on. One of them is your focus on builders. And I think
Speaker A
a lot of people don't know. They see you as like the CEO of OpenAI now, but they don't know that you were obsessed with AI like 20 years ago. Can you tell me about your initial venture into it in
Speaker A
college and what made you love artificial intelligence? One, I think it is just from like a techno perspective the coolest thing ever. The it's just such an interesting idea that we could make computers think and do stuff for us and help and that I
Speaker A
have like loved this my understanding of the story of technological progress in human history where we keep inventing tools on top of tools on top of tools and build up this scaffolding that lets us do more and more. And that gets to the So I think
Speaker A
it's just like a beautiful idea and very cool technology. But that gets to the second thing, which is in terms of technology and scientific discovery that can make the world better. If we can put this hand this tool in the hands of
Speaker A
people and if people can use AI to build and explore and create and new companies and new kinds of art and new kinds of experiences for each other and whatever else they'll do with it, that that is how I believe the world gets
Speaker A
better. That's also how I believe people get fulfilled. So before this, my career was in startups and I used to watch people make startups and I thought it was like awesome for the world, awesome for the people doing
Speaker A
it. I now think we're going to be in a world where you can have these one-person startups or three-person startups, whatever. And the amount of human potential that is going to unlock and the amount of good stuff that's going to make for all of us
Speaker A
that was just totally impossible without this technology. I think it'll be like quite great to see. Which historical figure do you think would have benefited the most from having AI?
Speaker A
Da Vinci was the first one that came to mind. Someone like a very bright thinker, interested in a lot of things that's like, you know, just huge creative energy trying to get as much done as possible. It's interesting too because I
Speaker A
feel like when the transformer initially came out, it's like a predictive text model. It's crazy how far that one breakthrough has gotten us in discovery.
Speaker A
Do you think about that? I all the time. Ilya Sutskever once said a very simple sentence that as many simple sentences said it really stuck in my mind, which is prediction is very close to intelligence. Mhm. And the
Speaker A
[clears throat] idea is that if you can compress all of the information about the world, the state of things, whatever, into its smallest representation, and then as part of that predict the thing that's going to happen next,
Speaker A
you understand it in a sort of deep way. And this idea at the time, a lot of people in the AI field were quite excited about generative models for a reason they couldn't quite articulate. But I think the reason was something to
Speaker A
do about this that prediction is very close to intelligence and if we're trying to build systems that really understand all of the data they're trained on, if we can get them to start predicting what comes next,
Speaker A
that seems like a great step. And certainly when you watch children begin to understand the world, I think you can observe a similar phenomenon.
Speaker A
Yeah, I think about it a lot. Like watching AI get smarter makes me hopeful because I'm like, all right, if I just input all of that information to my brain, I would get the same outcome.
Speaker A
Do you think that certain people are just like more destined to be great at physics or science and math? Or do you kind of believe that if you just gave everyone the same information, they would have the same output?
Speaker A
No, I don't think they would. I'm happy they wouldn't. Like I'm happy we get like the rich tapestry of human experience and that people have different interests and talents and sort of, you know, also different training data. But I think
Speaker A
it'd be quite sad if we showed every person the exact same training data and they had all the exact same ideas and the exact same interests and anything else. So I'm grateful that doesn't happen. It brings me to an interesting
Speaker A
thing that you said in a previous interview, which is that there's never been another time in history where this many people have all talked to like one mind. Like there are 900 million ChatGPT users every week. It's extraordinary.
Speaker A
How does that influence how you shape the personality of ChatGPT? We've gone through many different approaches to this. It's so hard to get this right. People want different personalities. The same person wants different personalities on different days. People on different
Speaker A
time horizons might prefer a very different personality. Like you might want a model if you're just thinking about how it makes you feel today, you might want a model that tells you how great you are. And if you want if you're
Speaker A
thinking about like what's going to maximize your fulfillment and kind of accomplishment over a longer period of time, you might want a model that pushes back on you way more.
Speaker A
It's a great point. Almost no one wants to go like set sliders about like, here's how I want ChatGPT to behave. I want it to be like this funny and I want it to be like this nice to me and push
Speaker A
back on me this much. And we don't do that for friends in our lives either. But we do gravitate towards different people or different people at different times or want different, want the people in our lives to support
Speaker A
us in different ways in different environments and different days and different contexts and we expect them to understand that. Totally.
Speaker A
And right now, ChatGPT doesn't work that way and most people don't expect it to work that way, but that is what I think we should shoot for. I think it does kind of model a little bit though. Like
Speaker A
I feel like my ChatGPT is very energized and optimistic in a way that when I use the account logged out, it's not.
Speaker A
It definitely does a little bit of this. Like it definitely kind of gets to know you somewhat. And this is what we've been going for. We're pushing towards this kind of memory and understanding more and more. But we used to not
Speaker A
have that. We used to have these like sliders and tell us what the personalities want kind of thing.
Speaker A
Yeah. And then you had 40, which was an interesting moment for you because for people that don't know the context, it was very agreeable, but in a good way in some ways. And I remember in an interview you talked
Speaker A
about how some people actually emailed you saying like this is the only supportive chat in my life.
Speaker A
I still think about those emails a lot. Yeah, how do you, how do you navigate that? Because I think what's really
Speaker A
lot of responsibility. We Yeah, the responsibility on this point is is huge. You know, this idea that there's we talk a lot about the risks with AI and the benefits and you know, we measure the big risk like bio risk or
Speaker A
cyber security risk. Um but probably the thing we do that at least historically has had the most impact on the world is how we set the ChatGPT personality.
Speaker A
Yes. Um how encouraging should it be? How uh you know, tough love should it be? How much should it like um customize to you versus me versus not do that? Yep. How understandable should be what it's doing? Like how much do you
Speaker A
need to see those sliders after all? Yeah. And we have historically, not just us, the whole field, not treated this with the same amount of rigor and scientific focus and sort of risk understanding that we have on things like let's not
Speaker A
make a new pathogen. Yep. But the impact this has had on the world is huge. I think it's had tremendous positive impact. Obviously with 40, it had some negative impact, too. Um I don't I still have not heard anyone
Speaker A
talk about this in a way where I'm like, this is the answer. This is how the world should think about the power of the default personality or the limits of personality in these models. But it is clearly a huge issue and going to get
Speaker A
bigger. How are you thinking about it right now? Um I have asked a small number of people that I think are really wise in different ways, you know, people from great spiritual traditions like great clinical psychologists, people who I
Speaker A
just think really understand how people interact with each other and what motivates someone, what fulfills them, to try to write different uh sort of instruction manuals for ChatGPT.
Speaker A
Okay. About here is how to behave to maximize people's fulfillment and personal growth and sort of uh accomplishment and enjoyment of life.
Speaker A
And I want to get those and I want to try aligning ChatGPT to the combination of those and see what happens.
Speaker A
Would it have to change by culture, too? I think it largely will, but there are some universal things about people that seem more about biology than culture. Mhm, like what?
Speaker A
There's this interesting book. I'm going to get the title slightly wrong. Maybe I won't. I think it's called human universals. Um but it was some anthropologists that went through every human culture they could find and looked for and they like made a
Speaker A
list of all the traits. And then if something didn't exist in even one culture, they took it out cuz they said it's not that's not really a universal.
Speaker A
That's like some sort of cultural thing. Okay. Um and and there were some things that weren't obvious to me that would exist in every culture like valuing travel, but it still did. And there were a lot of things that made sense to me that
Speaker A
would, you know, value every in every culture. Um was talking earlier about kind of how I'm like very excited about AI and what it's you know, what I think it can do for the world. Um but increasingly one of the
Speaker A
concerns we're hearing from people is, you know, okay, let's say you're right. Let's say you do give with this technology. You give everybody on Earth a ton of agency and you make and people working with this technology collectively make huge prosperity for
Speaker A
the world. Like people will still work because they want to work for fun, but everybody no one will have to and everybody will have this like great life and whatever. Um increasingly people are saying, well, you know, you talk about a
Speaker A
right to prosperity, but what about the struggle? What about the like need for adversity? What about the need to kind of overcome challenges and learn and not have everything taken care of and how important that is? And there are a few
Speaker A
things like that that seem like important to how we evolve as well. Agree. I also feel like that's a little bit of a false equivalency that's being made because I don't think like if you look at any big technological
Speaker A
revolution, there's never really been overall decrease in jobs. It's just that the jobs have shifted.
Speaker A
Yeah we you know, we were promised four-hour workweeks or whatever. Yeah. And we were promised like less stress and more happiness and more abundance.
Speaker A
And maybe if we were still content with the quality of life from a hundred or five hundred years ago, actually wouldn't have to work that hard. Yeah. We could get that. But is it the most valuable and more and more. The bar keeps going
Speaker A
up. And and and more than that, we want to like we want to accomplish and we want to compete and we want to be useful to each other whatever the new world looks like.
Speaker A
Yeah. And we want to sort of push on new things and discover new frontiers and invent new products and services and make new stuff, you know? Totally. Uh I saw something once where some music producer said decades ago that music had
Speaker A
gotten so good, he really didn't see why there was ever going to be a need to create any more music. Crazy.
Speaker A
just don't work that way. Yeah, of course. So I, you know, fortunately or unfortunately, depends on your take, like people are still going to be working hard. People are still going to be stressed. People are still going to
Speaker A
be unhappy. People are still going to be like striving Striving. to create and trying to overcome adversity in whatever way is meaningful to them and and through that find fulfillment and growth and maybe it looks nothing like the kind of
Speaker A
struggles that we have today or the kind of work we have today, but I bet the spirit of it will be very similar. Yeah, I think that's an interesting point because if you look at how AI is polling
Speaker A
in general, it is like not polling positively in America. Um and yet I'm so excited about it and like I feel like a kid in a candy store when I used to chat with ChatGPT because it opens up all
Speaker A
these new doors to explore. And I think a lot of my founder friends feel similarly. Um but then I look at how it's often covered in the news and it will be like 50% of jobs are going to be
Speaker A
wiped out. Um why do you think that narrative has taken off and what do you think is actually going to happen?
Speaker A
Um A lot of a lot of thoughts. I think people do kind of always like like dooms cells. Mm. The news covers Yeah, like it at least covers that kind of thing.
Speaker A
And, you know, people people seem to love a lot of people at least seem to love reading about talking about how horrible the future is going to be and the bad stuff seems to travel better than the good stuff. Now,
Speaker A
I think with any new technology and this degree of change, there is reason for caution.
Speaker A
Um and that speaking of these things that are like deep human evolutionary traits, we seem to evolve something where we do want to think about the bad. We want to talk about the bad. We we want to and that probably helps us
Speaker A
defend against it and there's probably a really important societal Yeah. shared caution thing there. Um I know some AI CEOs are saying things like 50% of the jobs are going to go away.
Speaker A
To say nothing of how tone deaf it is for someone to be saying, my company's going to eliminate 50% of the jobs and my company's going to be the most valuable company in human history and, you know, how wonderful that's going to
Speaker A
be and like, but 50% of you are going to lose your jobs. Um to say nothing about the wisdom of not even wisdom, just the sort of like the tone deafness of that. Um I don't think that's the right way to think
Speaker A
about it. Uh jobs will go away. Jobs have gone away with every technological revolution.
Speaker A
Jobs will change. Um I think someone said to me just yesterday that really stuck with me is I can use the new model GPT-5.5 in Codex to accomplish in an hour what would have taken me weeks two years ago.
Speaker A
And I thought I would have been much less busy in that world. Yeah, and now you're doing more than ever never been busier in my life. I'm waking up in the middle of the night to do more work. It's like make it stop, please.
Speaker A
It's too much. Um with new tools, I think we will create in new kind of ways. I I have no doubt that the economy is going to change a lot and jobs are going to change a lot.
Speaker A
And I think caution is warranted and I think rigorous debate about new social contracts, new economic systems are warranted as well.
Speaker A
But I don't think it's like we're all going to sit around in a life without meaning and without work. It's it's just going to be it's just going to be different. I also think that scientific breakthroughs are really coming and
Speaker A
super exciting. I want to dive into that with you. I have so many questions here.
Speaker A
Um but one of the first ones that came to mind um is going back to the idea that it's like a predictive model. Um I think there's like two scenarios that are playing out here. One is like if you
Speaker A
gave a human enough time and you gave them all this information, would they develop the same breakthrough? And then two, is it kind of like move 37, which was like in the game of Go when AI came up with a move that humans never would
Speaker A
have done. Like which path are we on? Um Well, they might not be that different.
Speaker A
I I was smiling cuz I was remembering when we had the first GPT models, there were all of these really smart-sounding scientists or AI experts that would say, you know, next token prediction will never develop new knowledge. It can't.
Speaker A
It's modeled off of the data it's been shown. It can't figure out anything new.
Speaker A
And they sounded so smart. They had all of this like fancy explanations these fancy explanations for why this was going to be the case.
Speaker A
And then with really with 5.4, a little bit with 5.3, was the first time where models started contributing in small ways new knowledge to humanity's collective knowledge. Like what?
Speaker A
Proving unproven mathematical theorems, um some smallish new pieces of physics, things like that. I expect this to keep going. Um in some sense, move 37 was already an example of this.
Speaker A
And and so this idea that we can train a model to just predict the next token based off of things it has already seen Yeah. and then use that ability to go discover fundamentally new things that didn't exist anywhere is like not so
Speaker A
obvious on its on the face of it. In fact, you kind of would say what people turn out to be wrong about, which is it shouldn't do that. Um but really what these models are learning to do through this process of
Speaker A
next token prediction is to reason. To understand how to make sense of all of the data they have seen and complete what comes next even if it's something they haven't seen before. Um this reasoning process can be applied to
Speaker A
things that you have not seen before. And this is really quite remarkable. Like the fact but people do this too, right? Like people can go study all of the known physics and then keep running their model Yeah.
Speaker A
their predictive model or whatever and by applying that reasoning ability they have learned through not just the facts, but the the underlying thinking process that they developed during their physics training go discover new physics.
Speaker A
And I think that's what these models are doing, too. Now, could people do it uh with more time and more brainpower?
Speaker A
Probably yes. I would say yes, actually. Okay. Uh but it is much easier to go make a faster, bigger model than it is to figure out how to give people like much bigger brains. So I for one am thrilled we have these new
Speaker A
kind of like external tool Yeah. that we can ask to go think really hard about a problem that maybe would be harder like for us to think about ourselves. When when when you see these models read like hundreds of thousands of pages in a few
Speaker A
seconds and that synthesizes across all of them, it's like, you know, maybe if we had a bigger brain, we could do that, but we cannot with our current size brains. Yeah, it is interesting though that it kind of is
Speaker A
similar to biological brains um and it made me curious. Are there other things in nature that you think we could copy for tech breakthroughs? Like the airplane is based on the bird. Like are there other examples of that? Um
Speaker A
A great scientist once said that there is no alternative. That that's the only thing that we figured out how to do.
Speaker A
Oh, interesting. Now, obviously that's not quite true, but uh uh I mean, neural networks, artificial neural networks were clearly inspired by the way neurons in a in a brain connect.
Speaker A
Um I I certainly don't think it's literally true that nature is our only source of inspiration for discovering new science, but man, is it a good place to start looking. Yeah. Is there anything that you're thinking about now that you want
Speaker A
to implement? Like this is a big week, obviously, for your new science model. Um how are you thinking about science breakthroughs, what you're focusing on there, what comes next? I saw that there was um an Australian guy that helped his
Speaker A
dog's cancer be cured. Yeah. That is a specific thing I I was just talking to a founder of a company uh I got to visit YC last night. Oh, cool.
Speaker A
Um who was thinking about a similar thing like that, but making that a scaled I mean, I'm sure a lot of people think that. Like custom mRNA vaccines?
Speaker A
Custom mRNA vaccines for cancer in people. Yeah. Um and that seems tremendously exciting. Totally. Why haven't we done it yet?
Speaker A
As I understand it, there's many reasons, but one of the big ones is the FDA is not well set up to think about how we're going to do that, although getting better fast. Yeah, it's interesting cuz I think when I think
Speaker A
about personalized or like the next frontiers, personalized medicine has to be it because all of us have such different DNA and risk outcomes.
Speaker A
Certainly the idea that if you get cancer uh a company or a lab can make you a personalized vaccine just for your cancer and it's very likely to be effective or likely to be effective sounds like an obvious thing we should
Speaker A
all demand. Yeah. Do you use um ChatGPT now for your health? I do. What do you use it for? Um I am probably a over user of it. I think they used to call them like cyberchondriacs. I don't know
Speaker A
what they call the ChatGPT version of this, but any mild symptom I get, I will go down a ChatGPT rabbit hole. Um you know, like everybody else I or many other people I will put my blood test in
Speaker A
there and I kind of wish I I mean, I'm happy I do, but sometimes it'll really like be like, "Oh, this is slightly off." And I'm like, "Oh, should I do something about this?" It helped you recently? How did it help you a few? Um
Speaker A
I had a stress fracture from running and my doctor went out of town and I had an MRI and I put it in and it read the MRI for me. Obviously, you got to check, but it was accurate. That's great.
Speaker A
It blew my mind. I When we first launched ChatGPT and there was like a little bit of this, it was not very good. People said, "People will never use ChatGPT for medical advice. It's just not nearly good enough. It's never
Speaker A
going to be good enough. And even if it were good, everybody would like rather talk to a doctor." People still definitely want to talk to their doctor, but the amount of ChatGPT usage of people asking medical questions and getting at least how they reported
Speaker A
to us really helpful information Yeah. is like quite extraordinary. Is it tough for you to constantly have people doubt that the technology is going to be impactful? Yes. Um it shouldn't bother me anymore. Yeah. It still annoys the
Speaker A
out of me. It would bother me, too. I feel like and if you look at like any big tech breakthrough, like before we flew planes, the newspaper was like, "We will never fly. It will be a hundred years." And then like the next week
Speaker A
we're in the sky. That was that particular example in the early days of OpenAI we used to talk about all the time. Love that. The The The Wright brothers New New York Times article.
Speaker A
Yeah. We used to talk about all the time and we said AI is going to be like this.
Speaker A
We turned out to be right. Um but the fact like honestly, it annoyed me a lot in the early days, but you know, it was like not super clear, so I thought it was at least like intellectually honest of the critics to
Speaker A
say maybe this isn't going to have a big impact. Yeah. Now watching people say there's really no value in this is going to have no impact on the world, it shouldn't bother me. I mean, it's obviously ridiculous, but it's so
Speaker A
annoying. It's so It's so intellectually dishonest and so annoying. And also, I feel like when you're in the arena every day and you're like trying so hard to push the ball forward, you want people to believe in it and like I think I with
Speaker A
our videos, I often try to show people like amazing technology and what the future can look like cuz I think you need to see it to then like lock in and build. And ultimately, I feel like people are the most fulfilled when
Speaker A
they're working hard on something that they care about. If you were talking to like a 22-year-old today, what types of things would you be curious to know about like how they're feeling about the world?
Speaker A
And how would that inform what you build? Um a thing that I have been trying to do over the last couple of weeks is really sit with people using the latest model and using Codex to understand how it's going to impact their work, what
Speaker A
they're excited about, what they're not excited about, what they need from us that we haven't already built. And I've done this mostly with people running companies or senior engineers at companies and I really should go sit down with some young
Speaker A
people and say, "Try this out." Yeah. And watch what they do and like listen to their concerns. You have a unique perspective, too, because you advise so many young founders.
Speaker A
Um when you were on Joe Rogan's podcast a few years ago, you talked about how there was a lack of like 25-year-old founders.
Speaker A
Yeah. Has that changed since then? That's totally changed. What do you think changed it?
Speaker A
Um I think there were a few things happening at once. So, I don't really get to advise founders anymore because life got so busy, but I have been thinking that I need to find some way to do that again because
Speaker A
one of the most important things about this technology is the entrepreneurship it's enabling. Yes. And I I feel out of touch on that in a way I really don't like. Like I intellectually understand it, but I want to go like I want to go like be in the
Speaker A
trenches with people building companies with, you know, two founders and 10,000 GPUs. Yeah. Who like I've met a few of these recently, but yeah, this is reminding me that I got to figure out some way to get closer to startups again. Um
Speaker A
Why there weren't Why there there was not the sort of chord of young founders then and why there is now. I think it was a lot of things. I think um the US educational system went through like a very dark
Speaker A
period where and COVID happened at the same time. And we were kind of like demotivating this whole set of people and telling them that like I don't know, the future is going to be bad and capitalism is bad and companies
Speaker A
are bad and you know, ambition is bad. That seems to have corrected. Yeah, we're back. There was like a Timothée Chalamet thing that went viral where he was saying how much he wanted to win an award and it was people were
Speaker A
stoked about it. Like they were like, "It's so cool to care again." That's great.
Speaker A
And it should never have not been that. The It should never have not been, but the the the sort of like the you weren't allowed to Yeah. be ambitious or to like Yeah, it was really weird. Really weird
Speaker A
time. And then I think another thing is startups thrive when there's dynamism and newness and the you know, there's a there's a change in the technological landscape. Mm. And that happened when the iPhone App Store launched. Yep.
Speaker A
In 2008, I guess, that happened when AWS launched a few years earlier and then it didn't happen for like a very long time until AI came along. So, there was just like a kind of period in the wilderness.
Speaker A
There were still successful startups, but not as many as there can be when you know, there's a real technological shift. You said that on your blog seven years ago. You were like, "We're due for another technological shift." And you
Speaker A
did and you're the man that made it, which is cool. Thank you. Um how do you think about focus now? Like what are the next areas that you want to focus on with AGI and obviously you guys like
Speaker A
shut down Sora recently. Um what are the areas that get the most focus and why?
Speaker A
Um I think the three most important things for us now are accelerating research. Okay.
Speaker A
We talked a little bit about this and this goes from like AI research to physics to biology, everything. But accelerate research because research and scientific understanding does so much for humanity. Second is accelerate the economy. Talked about this automated
Speaker A
startups, big companies using AI to be more productive. Um and all of, you know, eventually like building the space colonies or whatever.
Speaker A
And then third is the sort of like personal AGI. ChatGPT was like a little preview of this.
Speaker A
You know, maybe you can type in your medical questions and get some advice. Yep. But you would really like or at least I would really like an AGI working for me with my whole context, my whole life all the time.
Speaker A
Spending compute to like make my life better. Those Those are the three most important focuses. They're shockingly similar in terms of the enabling technology and platform, but those are the areas where I think society will really feel the
Speaker A
value. For scientific breakthroughs specifically, um you guys have the foundation where you're focusing on Alzheimer's research. Um what other areas do you think we can expect breakthroughs like in the next year?
Speaker A
Uh I would expect the progress in math to be astonishing or something like that.
Speaker A
Like in what way? we'll just discover hugely important new math and solve math problems that seemed out of reach. Mm. Um and like many other times in history, I expect if we discover new math, it'll point the way
Speaker A
to new physics and other new cryptography, who knows what, real-world applications. Um but I hope we hold ourselves to a higher bar and work on some of like the messier, more difficult scientific understanding that has more of a real-world impact. So, I don't
Speaker A
think we'll get Alzheimer's cured in the next Yeah. or even really treated in the next year, but I hope we can start to see like some new promising vectors that we can go push on. Yeah, I remember um Mark
Speaker A
Zuckerberg in an interview talked about how um when he talks to people that work in AI, they're like, "We are going to solve every disease." And then when he talks to doctors, they're like, "That is not going to happen." So, there's
Speaker A
clearly a disconnect in the two fields. Um how do you think about it? It will take longer than the AI people think and shorter than the doctors think. Love that. Yeah, I am I totally agree and I think if you even look back to a few
Speaker A
years ago, the types of breakthroughs that are now possible with AI, it's just like we're definitely on an exponential.
Speaker A
Um it seems like longer context windows is going to be super important in that exponential. Yep. How do we do that? Is it more compute? What has to happen? Um I don't think it needs to be like a literal
Speaker A
1 billion or 1 trillion token context window, although I assume we'll be able to do that, too.
Speaker A
I think what you care about is that somehow the model can effectively understand your whole life or your whole company or all the things you care about. And there have been amazing They all cost a lot of compute,
Speaker A
unfortunately, or a of memory, at least. But there have been amazing new methods to use the current context windows, um but really figure out the important bits, or to use tools to go off and find the less important bits when necessary,
Speaker A
and make way better use of the same amount of context. So, I think that will keep going, and I the the with the new model and the things we'll add to the new model in the coming months.
Speaker A
I don't want to say it will feel like infinite context, but it feels like okay, this model really understands a lot. It has way more stuff in its head than I have in mind. What's different with the new model? Like, what did you
Speaker A
guys change? Um smarter, Okay. faster, more context, uh and a I don't have the right word for this.
Speaker A
Uh more reliability, let's say. Like, I think it it under uh Let me say more intuition. More reliability. Oh, okay. I see. It it does it feels like it does a better job of understanding what I actually want.
Speaker A
Um and trying a few times, knowing when it's on track for that and not, and actually getting me the right thing. So, the the subjective experience is way more of the time that I ask the model to do something, it does the right
Speaker A
thing. Interesting. Because it understands based on its training, like did you guys update algorithms or It's we have to a lot of algorithmic change in uh it's a it's a newer, better, bigger base model with uh a different architecture, and then or
Speaker A
architectural improvements, and then all of the things we've learned about post-training, how people want to use these models, and uh how to kind of connect them to the world, people's systems, people's context to be helpful. So, I'm thinking
Speaker A
about it. Tell me if this is the right understanding. It kind of seems like AI gives better in three ways. It's um better algorithms, more data, and then like maybe more energy or more compute.
Speaker A
Yeah. Are these kind of the three things that we can push on? Um effectively, yes. Uh there's more data is a uh very broad category. Like, is this you know Do do we mean by that like uh like literally just more training data,
Speaker A
or do we mean like, you know, we're going to connect it in a loop that it can learn continuously as you're doing something and it's failing? That's cool.
Speaker A
Um but but yeah, broadly speaking, I agree those are the three categories. Which one's easiest one to have a breakthrough in? I think building more compute is the most certain one. There's the least science there. Yep. It just
Speaker A
takes a lot of money and a lot of complex supply chain, but you can just do it. Um algorithmic breakthroughs are the highest payoff, but the hardest and most uncertain to find, and better data's in the middle.
Speaker A
Interesting. Yeah, the better data um is it it does that tie to like recursive learning, like the model teaches itself, or is that different?
Speaker A
It can, yeah. I mean, that's totally one way to do it. If the model is you know, in some sense, if the model is really smart, it can go prove an unproven theorem, and now in the next training run,
Speaker A
there's one more thing the model can learn. It's we have this new proof. Yeah. That's an example. Interesting.
Speaker A
Are we at that point where the model is improving itself a lot right now, or no?
Speaker A
It's so hard to frame that question properly. I like in some sense, clearly yes, right? Like, the if our engineers are three times as productive as they used to be because of Codex, and they can write the code for
Speaker A
the next model faster using the previous model, you kind of got to count that. Totally agree. And then in the spiritual sense, people mean of like, are we just pushing a button and saying, you know, go make the next
Speaker A
model and come up with these new algorithmic ideas? Definitely not. I also think on the supply chain, like how do we build all these like data centers front, robotics is so exciting. You said that robotics is a big priority for you.
Speaker A
Can you bring me into your mind, like what excites you about robotics, and then what's the road map? Um we live in the physical world, and you know, even when we're in the virtual world, as you were saying, we need like
Speaker A
this massive complexity in the physical world to enable that. We need to make the chips and build the data centers, and you know, run the power plants and whatever else.
Speaker A
Yeah. Um so, a very sad future would be where computers can do these incredible things, but because we didn't figure out robots, we have to like go run around the physical world as like the actuators of the AGI that'll say, you know, please go
Speaker A
move this table and do this and do that. scenario. Really bad. Really bad. So, you got to have robots.
Speaker A
What type of robots do you think will be the best? I I am not that focused on any particular morphology. I want but I what I want is like automated manufacturing, and the ability to say like, we need more of whatever
Speaker A
this thing is, and with the same generality of ChatGPT, a factory of robots that can reconfigure itself and make more of that thing. Do you think you would ever physically manufacture them, or would you partner? Don't know.
Speaker A
Okay. Is AI hardware outside of that a priority to you? Like, I know Johnny Ive is involved.
Speaker A
Yeah. Oh, you mean I thought you were going to say chips. You mean like consumer AI hardware. Totally. Um the we were talking earlier about how you want an AI to have all the context in your life. And current hardware,
Speaker A
which is amazing. I think the iPhone is currently the greatest piece of consumer hardware ever made by a lot, like incredible what that has done. Agree. Um but it was not meant for a world where you needed a piece of hardware that
Speaker A
could absorb all of the context of your life. You know, you can use the phone. Mhm.
Speaker A
You can stop using the phone, you can put it in your pocket, but it's kind of like on or off. Yeah. Um and when we are not using it, like this has been a very interesting conversation. I would love this to be
Speaker A
referenceable by my personal AGI later, but my phone is in my pocket, and it's not going to understand.
Speaker A
Yeah. And you would I would like a device that if I wanted to, can participate and understand and know about this conversation. Totally. Yeah, I think also it would be interesting to get like outside insights, like um I
Speaker A
recently downloaded the transcripts of every Do you know the podcast Acquired? Yeah. Okay, love that podcast. Um and I was trying to reverse engineer like what makes their show successful. So, I downloaded like 400 transcripts from the show, put into ChatGPT, and had it
Speaker A
analyze like their story structure, and it was amazing. Um and I imagine that you could have similar insights of your own conversations and how you approach things as a leader. Yep. Um but I also know that when people see like an
Speaker A
always-on recorder, there's an ick with it. Totally. What what do you think? One of the reasons I initially wanted to talk to Johnny is is I was thinking about what hardware for the AI world is going to be, and the ick
Speaker A
that I feel with technology that is just too present in my life, like even even a smart speaker. Mhm. Um Totally. I thought Johnny would have great insight about how to design something that held all of these things in tension, and
Speaker A
I think he'll do great. What do you think will be like the biggest thing that is misunderstood about your approach? I don't know yet. Okay. Sure there will be many things we can come back and talk about there. Yeah,
Speaker A
interesting. Um I'm also very interested in um like the use of AI kind of in the background, like agents. Um what does that mean, and how do you think about it? Um when the team first made the Codex app,
Speaker A
I put it on my computer. Okay. And it had this thing that at the time we called YOLO mode. I think we found like we did. I think we found like a more polished name for it eventually.
Speaker A
Um but you could basically say like, you can just run in the background of my computer and do stuff. Yeah.
Speaker A
And you don't have to ask me every time. And I was like, I'm absolutely never going to turn that thing on.
Speaker A
Um and I lasted a few hours, and I got kind of so annoyed by having to like get permission every step, I just put it on.
Speaker A
And there was this agent, you know, running all over my computer doing stuff in the background, and and then pretty soon after that, I like didn't want to close my computer cuz I didn't want to stop working. Yeah. And the transition
Speaker A
there was so smooth, so uneventful. I you know, thought it was kind of crazy I was doing it, sort of irresponsible, but here I was.
Speaker A
Um we've since figured out how to make it a more responsible thing to do. Yeah.
Speaker A
Um but I went from thinking I wasn't going to be comfortable to loving the idea that an agent was just running around my computer doing useful stuff.
Speaker A
What was it doing for you? Um deal with my messages, deal with my email. Eventually, I tried something which is like, look around my computer and figure out what you can do to be useful to me. Whoa. Um Did it do
Speaker A
anything? I it let That time I first tried it, no, but it led me to working on this little project of making this like automatic to-do list. That's it. Um which is sick. Like, it auto-completing to-do lists is a very cool thing.
Speaker A
Totally agree. Is it built into like ChatGPT homepage, or is it a different is just a little program I made. This is like That's cool. Yeah, cuz I feel like there I always download different to-do list apps, and then they never stick, and I
Speaker A
end up just like texting myself. That's cool. Yes. Do you think that there will be agents that kind of work together?
Speaker A
Like, will you have like one agent that's like your personal trainer, will it just be like kind of the I wonder about this so much. Um This is like one of the product design questions I would most like an answer
Speaker A
to. What people how people are going to want want to work. Um I I suspect that people will have kind of a conceptual model of different agents, and then maybe like their kind of personal assistant, chief of staff,
Speaker A
whatever you want to call it, that coordinates among them a lot of time. Yeah.
Speaker A
Okay, so let's say that you and I time travel into the future, and we go to 2050, which I know is a long way out. Um feels like even you can't predict 6 months away. Um but I'm curious like
Speaker A
what your dream like if we were to dream together what the future looks like.
Speaker A
What are we aiming towards? Like, what's your vision here? Um like man, that feels so far away. Like, kind of almost unimaginable prosperity seems likely.
Speaker A
What what I hope for, but what what I think we have to really work for, is radical levels of human agency, where people can just do and create like beyond anyone's imagination, and we we avoided the kind of centralization of
Speaker A
power tendencies. Mhm. Um and then in terms of what the world actually looks like, I don't know, space colonies by then, maybe.
Speaker A
Flying cars. Yeah, maybe. Flying trains would be cool. All right, so I hope it looks like the future. Me, too. Yeah, I hope it kind of looks like this. To end this video, have you ever heard about blind ranking? Uh Basically,
Speaker A
I'll give you options, but you won't know what's coming next, and you have to tier list things.
Speaker A
Okay. I want to do tech breakthroughs. Okay. And I have to like rank as 1 to 10, and I won't know what's coming.
Speaker A
Yeah, so 1 through 5, 1 through 5, what's the most important in your mind? But you won't know what's coming next, so it's a challenge. You got to like give yourself some room here. Okay, I'm probably going to be really bad at
Speaker A
this, but let's try it. I'm sure you're going to be amazing. 1 through 5, 1 is the most important.
Speaker A
Yes. Okay, I'm going to give you one to start us off. Fire. The only thing that's going to be hard is I'm going to think they're all ones.
Speaker A
Um You're awesome. Three. Okay, that's a good answer. Um the printing press. Four. Okay. Satellites in space.
Speaker A
Five. I like that answer, that's smart. Um AI. One. One, okay. Um self-driving cars.
Speaker A
So, your only option left is like Two. Would you swap any of them now knowing all the options?
Speaker A
Uh I would go AI, fire, printing press, self satellites, self-driving cars. Good answer. Why AI over fire?
Speaker A
Um fire was clearly extremely important in human history. I I mean, from like food to steam engines and warmth in difficult climates and way other stuff in between.
Speaker A
Um but I would bet that viewed backwards 100 or 1,000 years from now, uh they will both be two of the great enabling general-purpose technologies of all time, and AI will have done more in total but yeah, you know.
Speaker A
Tough to say. wanted to say it the other way, I wouldn't I wouldn't fight them. All right, my last question for you. What's like the most common thought in your head? Like, what do you think the most every day?
Speaker A
Um at this point, it's been like what does the successful societal rollout of this look like? Like, not just the technology, but like what is how do we encourage all of this agency and entrepreneurship? How do we think
Speaker A
about what the social contract is going to have to look like like what does it mean to live in a world of declining GDP, even if quality of life is going way up? What how do we go be aggressive enough on the supply chain
Speaker A
to build out the compute power I think we all need for a a good and sort of fair future without breaking the economy in the short term. Those those sorts of things. Well, yeah, you have an interesting, exciting challenge of
Speaker A
having to think about the now, but then also think about 5, 10, 15 years. I probably should think about the now a little bit more, but yes. Well, thanks so much for coming on. You're awesome.
Speaker A
Really enjoyed this. Yeah, this is epic. All right, That was so fun. Dude, you're awesome.
Speaker A
Thank you for doing that.
Topics:Sam AltmanChatGPTArtificial IntelligenceAI personalitySuperintelligenceOpenAIAI future visionAI ethicsAI creativityTechnology innovation
