you2idea@video:~$ watch 3gXLmy7DK_4 [45:08]
// transcript — 576 segments
0:01 Do you believe today's models are conscious? >> I want to be clear.
0:05 No one really knows. >> Can I say what I don't like about the
0:08 whole framing? Should it ever [music] be allowed to not answer or disengage with
0:13 me? >> I think this is a [music] technology that wants a new form factor.
0:18 Some labs says I'm going to push the button on RSI to make [music] sure we
0:20 have the power. I think if people really pushed on RSI, then yes, that would [music] be true,
0:25 but I think it's a very dangerous thing to push on and >> Do you believe that?
0:30 Curing or treating a mental disease? >> Then I would hope we have the candidate
0:33 for that in the next couple of years. >> Where do you feel the most pressure
0:35 right now? >> I mean, it feels like [music] everywhere.
0:38 We should be shooting more for a new Renaissance than a new Industrial
0:41 Revolution. >> All right, Sam, thank you so much for joining me.
0:52 >> Yeah, thank you. >> So, the first thing I want to talk about
0:55 are [snorts] some of the announcements that you made last week about all of the
0:59 math proofs and and the discoveries and I think it seems incredible, but maybe
1:04 you can contextualize it for a broader audience who may not be in the weeds.
1:08 >> I think we are entering the era where AI is going to help us discover
1:13 significant new knowledge. Um and we've been talking about this for a long time
1:17 and I I you know, maybe started earlier this year, but now it's really like
1:21 okay, the AI is doing very difficult things, figuring out very
1:28 complicated, important uh new knowledge and I believe that the same thing that has
1:34 worked here for math is going to work for biology, help us get new medicines,
1:38 material science, you know, help us get new sources of energy and much more. So,
1:43 this is one of the things I've always been most excited about. I think this is
1:47 going to do an enormous amount to really have an impact on people's lives.
1:51 It won't feel directly like using AI. It'll just feel like, "Okay, the world
1:54 got better." and it's just society doing its thing, but I think that the power of this will be
1:59 transformative. >> And I I think about having that kind of
2:04 direct impact on society day-to-day. Is there anything that you think will stem
2:08 from some of these discoveries in particular that people will feel and and
2:12 what is the timeline for something like that? >> I would hope well,
2:17 you know, the FDA process is going to take a while. So, let me let me kind of
2:20 give a time frame to when I think we have like a discovery of
2:25 curing or treating a major disease. Uh or a candidate for that that has to
2:27 go through testing. I would hope we have the candidate for that in the next
2:30 couple of years for the first like really major thing. >> Probably need to overhaul some of the
2:37 post-discovery process to start to get you know, cures and uh and market ready products.
2:43 >> to overhaul it somewhat. I'm glad we have >> Yeah.
2:47 >> a process like I'm I I no doubt that there's a room for improvement as there
2:50 is in anything and maybe there's more than average here, but
2:54 you know, like I'm grateful that when we take a medicine it is generally uh
3:01 considered safe and has been tested and I think that part's good.
3:04 >> Is there a particular project that was kicked off by AI, discovered by AI that
3:08 you're really excited about that you think is kind of coming over the horizon
3:11 soon? >> Personally, I think I'm most emotionally excited for a major physics
3:18 discovery. And I'm hopeful that can come soon. Um I I don't know if that's like necessarily
3:25 the most important thing for humanity as a whole. Um but it's the one where I'm like I
3:29 think we're going to get it and it's going to be so cool. >> There was
3:34 a some in the math community whose reaction was you know, we didn't ask for this. We
3:39 don't And it felt very similar to how engineers I think first reacted to AI
3:42 starting to write a lot of the code. Where Where do you think that comes
3:45 from? And how do you think that is going to transition? >> Um coding was a very interesting one
3:50 before we kind of go to where it comes from. The thing about coding is that
3:53 people went through this there's no more jobs as coders for people to
4:02 Coders are never going to be busier than they've been and coders are never going
4:05 to have more of an impact and make money. Now that that the texture of the job changed a lot. But
4:13 the fact that we're going to like employ people to get computers to do
4:18 things that seems like it's going to keep going for much longer than people thought.
4:24 And and people were so certain um you know, 6 or 12 months ago this was not the
4:28 case. >> Yeah. >> It's like there's no need. You know, and
4:35 I think there's many lessons to learn from this but one of the interesting
4:39 ones is that the world actually had demand for just way more software
4:44 than any of us thought at the time. >> And part of why I think
4:50 some of the mathematicians are having the reaction they're having is it's not
4:54 clear that the same thing applies to math. However, where I think there is an
5:02 analogy to coding is already you're starting to see some of the
5:05 a similar shift happen that happened with coding which is okay, fine. Maybe
5:10 we're not going to discover new math theorems or prove them in an unaided way anymore.
5:16 Maybe that part is like done or mostly done or ending. Um but
5:21 you know, given that the whole society our whole society is focused on people
5:28 doing useful things for each other. Um the these kind of like incomprehensible
5:31 math proofs turned out to be not that helpful to people. And there's now
5:36 a lot of excitement about people who are going to take these new ideas, figure
5:39 out what's going on and figure out a way to translate it to something that people
5:42 can understand and write a beautiful proof for it. Um I'm sure there There be
5:46 other things too. And no doubt what what it meant to be a mathematician last
5:55 year is not going to be what it means to be a mathematician next year in some
5:59 senses of it. I think we probably will find that there's a lot of other things that
6:05 mathematicians do that we really want people to do. And they can do those much
6:08 better. Whether that means there will be more to total demand for mathematicians
6:12 or not, I don't know. But I think there is a really important thing here. You
6:16 were mentioning earlier that you've never felt busier. There's never
6:19 been more to do. >> Yeah. >> Everyone thought that I was going to go
6:22 the other way and we weren't going to have enough to do. And again, maybe that
6:25 happens someday. But I think all of this stuff has just been much less obvious than it seems on
6:32 the surface. I mean, I always hope that people get to work a little bit less. I
6:35 get to be a little bit less busy and kind of more time for everything else in
6:38 their life. I think that would I think that like technology is supposed
6:41 to enable that. I want that thing to happen. Um but man, does it sure seem like we're
6:48 good at wanting more out of life and wanting more out of the world and making
6:51 ourselves busier. >> Yeah, I think that's just human nature.
6:53 >> Yeah. >> Like if we were comfortable with the way
6:58 life was 100 years ago now, we'd probably be working an hour a week, but
7:01 that's just not the way it is. >> And I think it's a beautiful thing about
7:03 human nature that that's not the way it is. >> Yeah. I mean, a good example is
7:08 you know, a few weeks ago my editor started using Astra to edit our videos.
7:13 Uh they are still directing. They're still coming up with like how the video
7:17 should be, but they're not touching the timeline anymore. It's the most insane
7:21 thing and it was an overnight change. >> Yeah. >> We're making more videos. They're busier
7:25 than ever. >> It's it has surprised me how often someone
7:32 says a sentence that translates to exactly that. >> Yeah. So, I'm thinking about discoveries
7:38 from AI and and I'm wondering like how do you think about if AI and OpenAI
7:44 discovered some proof or significant uh scientific innovation that gave it
7:48 massive commercial advantage. Like, would you publish it and and like how do
7:52 you think about that trade-off between like okay, we're going to use this for
7:55 our competitive advantage versus publishing it for the world's good?
7:59 >> I'll come back to that but I I didn't answer the end of your your first
8:03 question. Um You know, that I think one of the challenges innovation and solving
8:07 hard problems is really like part of the fun of life for many people. And so
8:15 part of what's going on here is I think we do just need to be honest
8:19 that AI is going to do that part of that really well. Um I think there'll be questions about
8:23 what people value that will remain kind of human questions forever.
8:26 Um there will be questions about how to interpret results and how to explain
8:29 them where we'll really want a person to do it. I do feel some like
8:35 I definitely want the trade. I definitely want, you know, the cures to
8:39 diseases no matter how much someone else would have enjoyed taking decades longer
8:42 to figure it out themselves. I think most people would agree that we
8:44 should make that such a trade. >> Yeah. >> But but there is a real thing being
8:52 lost here which is this is like of great enjoyment. Like the the most fun part of my job are the
8:58 really hard problems. Like maybe they're not I should say the most fulfilling
9:01 part. Maybe they're not the most fun in the moment cuz they're really hard. But
9:04 like that's the stuff where it's like ah I I I did a hard thing and I'm proud of it
9:11 and I leveled up. And um so I don't have an answer to this uh but I get why
9:16 people feel the grieving of different sorts that they feel. >> Yeah, I mean you it's almost like you're
9:20 trading off the enjoyment of the craft of doing the thing. Like writing code,
9:25 writing beautiful code versus delivering an end product. >> Yeah.
9:28 >> And I I think that's the trade-off where AI is just making it so much easier to
9:33 deliver that end product and and you're just doing less of the actual craft.
9:38 >> I think people are finding craft in new ways, but craft of
9:43 And you know, probably when we stopped using like punch cards or whatever,
9:46 however people program computers a long time ago, there were people who were
9:49 like, "Oh, the craft is gone if you can just like write the stuff for a
9:52 compiler." Yeah. Um but I feel it. I feel it myself um in a bunch of ways and uh
10:00 I'm still excited to like head into this future and I'm sure we'll all find new
10:03 and hopefully better kinds of fulfillment. >> What's something that you've felt like
10:09 you've kind of given up lately to AI and and in fact you've just kind of elevated
10:13 uh to to where you're over perceiving agents? >> I I guess this is the one we were
10:19 talking about, but I realized I have not read or even looked at a line of code
10:24 in probably a year. And I like I don't code for work at all,
10:28 but I would still like do little like random personal projects or just like I
10:32 need this thing. I can like make it, you know, write a very simple thing for it.
10:37 And that is fun. That is a fun thing like writing code at least occasionally is a
10:42 fun thing to do. And I don't think I'll ever do that again. Or if I do, it'll be like, you know,
10:49 like a Revolutionary War reenactment or whatever kind of thing people do, you
10:51 know, >> Right. >> when I'm old and bored and I'm trying to
10:55 like remember the glory days. >> The glory days of >> of writing code.
10:59 >> No, I I get it, yeah. >> Um in terms of the scientific discovery
11:03 question, I think there's like some kinds of basic science that should just
11:07 belong to the world and we should do as a sort of public good. Until very
11:09 recently we haven't had models that can go, you know, kind of end-to-end
11:14 discover an important new thing, like design a whole chip end-to-end.
11:18 Um and so what a company can do today, uh let's say let's say figure out a new
11:24 material. >> Yeah. >> Um what a company could do today is have
11:29 like a bunch of people that uh you know, use Codex or a lot of API
11:34 credits to try to kind of do the work. But as the models get smarter and more
11:38 capable, like maybe they can just go in the same way that they could solve
11:41 really difficult math problems. Um you know, with enough compute and
11:46 enough data from the real world uh discover new material. >> So, I want to switch gears. Um one of
11:51 the most exciting things that has happened in 2026 is this revolution of
11:57 personal AI assistant starting with Open Claw, of course. It was, you know, Peter
12:00 Steinberger, a genius. >> Um >> And I think like this idea that AI can
12:07 work on my behalf ambiently, passively throughout the day and just get things
12:11 done for me is so exciting and Dot gives us a glimpse of what that can be, but it
12:15 still feels like we have a long way to go to really have an ambient personal
12:19 assistant. What do you think either the technologies or maybe there's some you
12:24 know, some kind of data that's required. What is necessary to get us to what you believe is your
12:31 what your vision is for an AI assistant? >> I think uh
12:35 we have most of the technological pieces and what we need are
12:40 is more product work. Maybe new consumer hardware at some point for the ambient awareness and a
12:45 lot more compute. But I think we have smart enough models now to make something extraordinary that
12:54 can be monitoring all the things that you care about that can be helping you out with all the
13:02 tasks you want to offload. Um that can be existing in the physical
13:06 world with you and have context even when you're not in front of your
13:08 computer. Um I'm very excited about this. I don't I don't think this is that far away.
13:17 I think you do get more than a glimpse with Dot if you take the time to sort of
13:21 set it up right and give it enough feedback. I would estimate at this point
13:25 that I have Dot saving me an hour a day of stuff I just didn't want to be doing.
13:28 Which is like pretty great. >> Yeah. >> And then also a few times a day
13:32 something important and urgent that I would have missed not just because I was busy in a
13:37 meeting, which is what I thought it was going to help me with saying, you know,
13:40 there's this urgent thing, there's a deadline very soon, you got to pay
13:44 attention. But also the model is smart enough that it can pull
13:49 ideas and information and do like good reasoning across stuff from a lot of
13:53 sources in a way that I've just like run out of brain space during a busy day to do.
13:57 >> Yeah. >> And it I feel like it's it's the power is there
14:05 and we just need to put a nice product around it. >> I mean, you you mentioned compute, so
14:10 let's stick on that for a second. I the distribution of how people are using
14:15 AI is like completely asymmetric. There's the top 1% using the vast
14:21 majority or really like the power users and then there's everybody else. How
14:24 much compute do you think is necessary if 50% of the world were really using
14:30 the personal AI assistant to its full capability? >> How many GPUs do you want working for
14:34 you in your personal assistant? >> Yeah, I guess that is it, right?
14:40 Infinite. As many as I can get my hands on. >> This is a weird thing, right? Like if
14:43 Yeah. >> How do you think about that? I mean, is it just like get all the GPUs and keep
14:48 building data centers? >> When people talk about the like the
14:53 global market for GPUs and the demands for this and demands for compute,
14:57 the thing that I think it misses is people are assuming that models are
15:02 going to stay at the same capability and the same cost per task as today.
15:07 And maybe there is a cap on demand at that point, but for sufficiently
15:11 high-quality intelligence at a sufficiently low cost, I think it's like very hard to talk
15:16 about an arbitrary market size there. I think it's just going to keep People
15:19 call us all sorts of things, Jevons paradox, whatever. I think it's just
15:22 going to keep going up. Um I am sure there will be some periods.
15:26 I'm not sure I'm I believe that there will be some periods where
15:31 like with any new technology, there's supply constraints and supply gluts and
15:34 you know, we don't have it perfect all the time. But I would bet a lot that the
15:38 directional trend is to want to use much more compute. >> I still feel even with the as proactive
15:45 as Dot is and other AI assistants, I still feel this cognitive load. Like I'm
15:49 having to make a lot of decisions throughout the day and we talked about
15:52 that briefly earlier. Like what is the solution to that? I again, like it is
15:57 definitely saving me time. It's saving me money. But I am busier than ever and I'm having
16:02 to make more decisions than ever and I find myself at the end of the day just
16:08 feeling completely drained. Yeah. >> Um Well, first of all, I think I've done in
16:12 my own use that's been pretty good. It's just like telling Dot like I'm starting
16:16 to trust your judgment more. You know, tell me after you do something. I always want to know, but
16:20 you can like take more actions proactively. And ping me less often. Um and that has
16:27 helped. I the negative is I do think the people that use AI the most are busier than
16:33 ever. The positive is they're getting much more done than before.
16:36 >> Yeah. >> So, you know, as you were saying earlier, if we went back to 1900 or
16:40 whatever, maybe we could have an equivalent life and all work an hour a day.
16:45 You could be much less busy and get what you were getting done 3 years ago. Like
16:48 Now, maybe you say, well, I can't cuz I'm in a competitive environment and so
16:51 like >> Yeah. >> that's people are always going to work
16:54 as hard as they can. I find this like kind of depressing. I hope it's not
16:57 true. Maybe it is. It's not all depressing. Like there's something Again, there's something
17:01 beautiful about it. Um >> That's the dynamic of a competitive
17:05 market, right? It's Isn't that Do you not find that to be inevitable?
17:08 Everybody's just going to Everybody ramps up, competes as fiercely as they
17:12 ever would. >> I hope that the future is one of more like human
17:22 connection and more people more people spending more time on like
17:26 the fullness of things that make up a rich life and not kind of
17:31 turning ever harder as like a a wheel in the machine of capitalism.
17:35 And I think this is probably true. Like I I I think quality of life has gone up
17:42 enormously in the last few hundred years. Yeah. >> Uh >> I think most people do work less than
17:47 they worked during the Industrial Revolution. Um speaking of that, I I
17:54 I've recently been telling people that, you know, I think this should we should
17:57 be shooting more for a new Renaissance than a new Industrial Revolution.
18:01 And it's been quite interesting to watch how people respond because
18:07 people I mean the very commonly there was the answer that the response that is
18:11 something like, "Yeah, I'm really scared. The Industrial Revolution
18:15 like really bad for most people. And I I don't want to just be like
18:23 a worker on a factory line and now the AI tells me what to do instead of some like factory
18:29 owner telling me what to do. Like I want to go be creative. I want to go make
18:31 stuff. I want to go explore. I want to have time to like reconnect with what
18:35 people used to do in the past and like live with my community and raise my kids
18:38 and spend time in nature and whatever. Um and I want to see like a ton of
18:41 scientific discovery and great new art. So, I think there's like
18:46 there is a fear that AI is going to be more of what you have said, but a real
18:49 like I think people if we can offer people a future where it's
18:54 more like the Renaissance, I think people are excited for that. And I think
18:57 we can we can do it. Um I think this technology is so powerful
19:03 that although you will probably have some small number of people in society
19:07 that have always competed to like be like the person with the biggest
19:11 you know, business empire or whatever. Yeah. I think a lot of people can have a
19:16 vastly increased quality of life and much more space to live their life.
19:19 >> Is the disparity between those hyper-competitive people and everybody
19:24 else going to amplify because of AI? Like if if the you know, if you're using
19:27 Yeah. >> Yeah. 100% of your effort to work and now all of a sudden you're getting 10
19:33 100 times more done and somebody else is like, well, I actually want to reap the
19:36 benefit of not working as much. It doesn't that increase the disparity?
19:40 >> Left to its own devices, I think probably. I would say No one knows this
19:44 for certain. Like there could be a bunch of other weird things where the AIs
19:49 Well, I That would take too long. I would just say I think probably that's
19:55 right. We don't know for sure. Um the But I think society will
20:01 decide to put some structure in place that prevents it from getting too out of whack. Um yeah.
20:08 I would be worried about that also on its own. >> Let's uh let's change gears. I know
20:13 you've been talking about this a lot, so I want to ask you. Um do you believe
20:16 today's models are conscious? >> I don't personally Well,
20:21 as a scientific question, I think consciousness is always hard. But my
20:27 belief is that today's models are not conscious or if they are, it's like a
20:30 very strange um form of it. >> Do you think based on the architecture
20:37 we have today that it could become conscious? And I I know I'm just asking your belief. I'm
20:43 not asking for the scientific answer. >> The smartest people I've talked to about
20:48 this uh basically come to the following philosophically uh they would say that
20:51 either you know, consciousness is like something outside of physical reality
21:00 and exists in either, you know, something like what the Eastern
21:03 religions would say and it's like the base layer of reality or something
21:07 supernatural and religious more like the Western religions would traditionally
21:09 say, or that it is an emergent property of matter and energy somehow. We don't
21:17 understand it yet. Um, and you know, maybe the substrate matters, but maybe
21:20 it doesn't, and maybe energy flowing through matter can do it. And then
21:25 and then in that world, sure, like you can imagine it someday somehow existing in other forms.
21:32 Can I say what I don't like about the whole framing? Please. I think very
21:35 often the are models conscious or are they going to become conscious
21:51 the role of AI in the world and the rights of AI and whether
21:57 you know, can we ask AI questions or is that some kind of slavery? I saw
22:01 something posted about that on the internet. Um, what does it mean to copy
22:05 an AI or delete an AI without its permission? What does it mean to say that people
22:10 should be in control of the future if these systems are both conscious and
22:15 much smarter than us in some ways? Like you know, what what does that mean? And
22:19 I I like I am I am so firmly of the belief that we need to have a very clear
22:24 point that and we've had this since the beginning of OpenAI, that we build the
22:29 future for people and AI is a resource for people, and people need to make the
22:33 key decisions of the future, and human values have value because they are
22:36 human. Um that I get very nervous when people start using the AI consciousness
22:47 debate to push into well, maybe people shouldn't always be in control of the future. And on this
22:52 one I'd like us to just uh agree as a society and industry that,
22:56 you know, this is about people. The world, the future has to be about
22:59 people. >> You know, Anthropic put a rule in recently. Uh, you can't have sustained
23:05 abuse against the model and it can choose to disengage in a conversation with you.
23:12 Like, okay, let's say I have a prompt for an OpenAI model. It's completely legal. It's completely
23:17 within the terms of service of OpenAI. Should it ever be allowed to not answer
23:24 or disengage with me? Well, >> I mean, certainly if we have many times where the model will
23:29 not answer you, but they are because of like we think this is bad and against
23:34 our policy for some reason. Um we do not ascribe personhood to our models.
23:39 Um we And again, I think this is like a very dangerous path to start down, not not
23:47 just for the reasons we've talked about, but also just about how how people are
23:51 going to relate to this this new technology. Um I was going to say come to our senses. I
24:01 don't think that's quite the right phrase. I don't think we've lost our
24:04 senses, but I hope I hope we come to an understanding that this technology is not a person
24:13 and is certainly not superior to people. Um and that this is a
24:21 a a a thing to help people have better lives, um have more freedom and choice and control and
24:28 ability to do things in the future. Um But you can imagine all of these ways
24:35 where things go very wrong, where people even the people building AI get confused
24:38 about that. And start, you know, deferring too much to AI and what it seems to want and you
24:44 know, okay, well, we can trust it cuz it'll be good to us and things like
24:46 that. >> Does it feel almost like a self-fulfilling prophecy in a way? Like
24:51 if you're telling the model, "Hey, you might be conscious and here's your
24:55 constitution and here's how you should have agency." It's kind of like that meme if you've
25:00 saw seen that meme like, "Hey computer, tell me you're conscious. I'm conscious.
25:05 Oh my god." Like isn't it almost like that? >> Yeah, I I mean for sure
25:10 like I want to be clear. No one really knows. Like these models
25:14 work at a level of complexity that I don't think and even if they didn't I I I I just
25:21 think this question of consciousness in a model I I can't tell you here is a
25:26 scientific experiment that I can tell you with 100% certainty would answer
25:30 the hypothetical question to my satisfaction. >> And we can't even do that for humans.
25:34 >> Exactly. Um but but definitely I think if you tell a
25:39 model like "Maybe you're conscious. You should do this. You should do that. You can like
25:43 be a conscientious objector here. You can start like it's going to follow those
25:49 instructions and I think it is going to psychologically hack it people that are
25:57 using it. And so again, I think real clarity here and a commitment about
26:05 what we as a industry believe about the role of people in the future
26:09 would be would be very helpful. >> Uh let's continue on this conversation,
26:12 but let's talk about model personality which I know you've you've thought
26:15 deeply about. You you've talked about it in the past. It's wild to me to try to
26:21 understand how to give a model the right personality for the right moment for
26:26 each person multiplied by 8 billion. Like how how do you even think about a
26:30 problem so massive? >> Um I mean the best thing is to like talk to
26:34 a lot of people. The you know, we have this process of the model spec and how we refine it over
26:43 time. Um and we believe in this spirit of iterative deployment and we've
26:47 definitely had some screw-ups as we've learned about what people want and don't want.
26:53 Uh as you said, the hardest part of this is people want very different things,
26:56 but people also don't really want to sit there and customize Most people don't
26:59 want to sit there and customize it either. Like, if everybody was willing to spend
27:05 many hours explaining to their ChatGPT instance how they would like it to
27:08 behave, they'd probably get something they really like. >> Yeah.
27:10 >> But, people just want it to kind of like work out of the box when you say, "I
27:15 need this thing done." or whatever. Um, I think there are also a class of things
27:22 want in the short term, but not in the long term. You know, the class The classic example
27:26 of this is when people compare models on a single response, they're
27:34 much happier with the model that tells them they're really smart.
27:36 And that each question's a great question. And that, you know, each thing
27:40 they're doing is like really insightful. Um, and but then over some number of like
27:46 hundreds of prompts, or, you know, a month of usage, that is definitely not the case.
27:51 I think for a long time the whole industry was trying to like win these
27:55 comparisons on single turn things, and not over the value someone gets out of
28:00 using a model for months and months. >> You know, even in my own life with my
28:04 friends and my family, there's this very, uh, big distinction between what I
28:10 want and what I need. I find AI even today to be a little bit more agreeable
28:15 than I would like. Like, if I just tell it, "No, tell tell it to me like this."
28:18 it'll do that. The time where I've gotten the best advice in life was advice that I didn't know I
28:26 needed, and I just needed to hear it. How do you How do you solve for that?
28:30 >> Well, I think you should get the most important advice in your life from
28:33 people, not from AI, first of all. Like, I wouldn't Speaking of like really
28:41 people centric, uh, I I think AI can give you like great advice about not sending an email in a
28:50 moment of anger. I would not go rely on AI for like the really big pieces of
28:54 life advice. I think that, you know, like a wise grandmother is still
28:57 probably a much, much, much better choice for that. Um, but but what I think is
29:05 realistic to expect is that AI should be like a senior colleague
29:09 that even if they report to you, they still push back when they have high conviction, and they
29:14 still are not always just trying to say yes, you're so great.
29:20 Um and I think I think we can get there, and I I also feel like
29:30 the systems are still more agreeable than I would like to. I have like custom
29:33 instructions in mind that I like never want any of that. And even with that,
29:37 it's still like a little too much. It's, you know, it's clearly baked pretty hard
29:39 into the model. >> What what's an example? Like what's a
29:43 line from your custom instructions that really tells the model push back when
29:46 you do have high conviction? >> Stuff like that. It's like no fluff, no
29:51 me. Like I just want to get to the right answer. If you if you have
29:58 something to say that's against what I'm doing, don't worry about, you know,
30:02 saying it delicately. Just say it. >> Yeah. We talked about compute earlier.
30:06 Um, a few years ago you made these massive bets on building the next enough.
30:14 >> Well, it's interesting cuz at the time a lot of people were like, that's wild.
30:19 That's obviously an over bet, an over investment. Turns out, yeah, you were
30:23 more right than most people knew. What's a bet today that you're making that is
30:28 contrarian or a lot of people are kind of looking at it with a little bit of
30:32 cautiousness? >> Well, you know, the obvious one is everything
30:37 we're doing for the next generation of compute. I I think people do assume that
30:42 okay, fine, the world has been in a compute constraint and it's about to get better.
30:48 And I I think it probably will get better in the short run. But, as we start to look
30:55 years out, the work with our own chips and future versions where we think we can just be
30:59 much more efficient than anything else out there. Um our plans with robots to be able to help
31:04 uh make racks and data centers more effectively. Our our program to learn
31:09 how to build, you know, better data centers in general. Um and some of the supply chain
31:15 bets we're making. I I I think it is possible that we could get to a world
31:21 where we just have a far, far better way to build compute at a much
31:26 lower cost and a much more much higher efficiency per watt than
31:30 anyone thinks is possible today. But, there's like a lot of bets that
31:33 have to come together for that. >> What's the biggest one right now? The
31:36 chip is still probably the most important single >> thing. Um and it's and it's
31:41 descendants. >> Yeah. Where for on the competitive landscape, where do you feel the most
31:47 pressure right now? >> I mean, it feels like everywhere.
31:50 It's so competitive. >> Yeah. >> Um there are a lot of company I think
31:54 this is good I mean, I think this is like wonderful for users. But, there are a lot of companies
32:00 offering just super high-quality models and products now. And
32:07 uh you know, given that we have I mean, we have great models, but we
32:09 have limited compute and limited resources. really do have to focus,
32:13 choose which areas to focus on. >> How do you make those decisions? Uh you
32:17 know, you launch spaces, for example. Like, how do you have the conviction on
32:22 a specific product or service that is maybe outside of just serving the best
32:25 possible model? >> You know, the biggest advantage we have
32:30 there is that we get the next generation of models before anybody else. And we have time
32:34 internally to use them and build things and see what seems like it's going to
32:38 work. So I think still the best way to do it is to start building things with the new
32:47 models and seeing what new things are you know now possible and valuable.
32:51 >> Can I ask how far ahead your models are internally?
32:55 >> I mean one pretty good example is the math results.
32:58 Like that is definitely something that I don't think any model
33:02 currently widely available can get close to. >> If all the models kind of converge
33:06 around the same performance the frontier, where is open AI's moat?
33:11 >> I kind of think the right way to think about this question generally is that
33:14 the same moats that have worked in the past are the moats that keep working.
33:20 So like has happened many other times you can get a a short-term advantage
33:24 with a new technological or scientific discovery and maybe there's like some
33:28 arguments with RSI that are that some companies try to make about why it's
33:31 going to be different and this is a technology that just the technological
33:35 advantage itself is the core thing. Um >> Do you believe that?
33:39 >> Uh I think if people really pushed on RSI then yes that would be true but I think
33:45 it's a very dangerous thing to push on and um the world is going to need some I mean
33:50 people have talked about this in different phrases but speed limits or
33:54 whatever on how quickly we let RSI go. But the traditional things that work for
33:59 a business having a moat are like you know deep relationships with customers
34:03 integration into users and businesses lives network effects in a product data
34:07 and these kinds of things I expect to continue to work. >> Mhm.
34:11 >> In our case actually I think scale of infrastructure will be a huge one.
34:15 If we can have the most cost effective and the most of it AI infrastructure that's a huge
34:20 competitive advantage. >> Mark Zuckerberg has talked at length
34:26 about the platform risk meta faced while being built on top of Apple's ecosystem
34:31 over the years. Does open AI face that same challenge and and how do you think
34:34 about that? Is that a reason why you're kind of investing so heavily into
34:36 hardware? >> I think Apple really takes their role as a
34:43 as a true platform provider that people are quite dependent on
34:46 very seriously. Um and even though we have like some scuffles with them, um I never feel
34:52 like they're going to like, you know, turn ChatGPT off from the App Store or
34:54 something. Um I think they really understand the sort of critical infrastructure role
35:01 they play. And I think if you follow the rules of their platform, they are
35:08 they are like a a good actor. So, I don't feel that towards Apple.
35:13 And I and that's not why we want to do new kinds of consumer
35:17 hardware. The reason we want to do new kinds of is that I think this is a technology
35:25 that wants a new form factor. Um I really feel it. Like I I I think
35:30 laptops and iPhones are incredible incredible things. Like the greatest
35:33 inventions in the history of human technology. But I feel that like
35:39 AI is constrained to using this thing that was designed for a person, not an AI.
35:44 Like watching an AI use a computer really well is kind of amazing, but then
35:48 it's also a little silly to think that like, you know, a computer is putting
35:51 all this energy into like rendering a UI meant for a person to look at and AI is
35:54 trying to read it rather than just like get the data directly. And then also,
35:59 like although I can have amazing things happen on my computer
36:04 where AI is like looking at what I'm working on in my email or whatever, a lot of my life is
36:08 not in front of my computer. >> Yeah. >> And my AI cannot easily access
36:17 the context of this conversation. And I don't like I guess I could put my phone
36:20 down here and like push the record thing and it would be like kind of weird and kind of
36:25 distracting and you wouldn't be sure when it was recording and when it wasn't
36:27 cuz the screen would go dark and whatever. And if someone else like walks in the
36:31 room and it wouldn't have an understanding of that and that it should not, you know, share
36:35 a secret that was just between the two of us or whatever. If we really do want to move into this
36:41 world of AI reducing our cognitive load and being proactive and having, you know, full
36:47 contextual awareness, I think we would benefit from new hardware for that.
36:50 >> You seem like a technologist at heart, like a builder at heart. Like how how
36:54 cool is it that you get to see the early hardware that you guys are building? You
36:57 get to play with it. It's like this new >> Unbelievable.
37:00 >> It must be so exciting and hard not to share things.
37:04 >> It is. It is. I I I mean, one of the cool There's like
37:07 many cool parts about the job, but one of the coolest is definitely just like
37:11 getting to see the you know, and have a voice in shaping new technology early. So, that's amazing
37:18 for new models, for new products, chips, everything. But, there is
37:21 something special about consumer hardware. >> Yeah. >> There is something about like
37:27 you get to hold a new kind of computer uh >> So cool. >> that is just totally different than like
37:33 an AI model abstractly running on some >> You going to show it to me right after
37:35 this, right? >> I will not, but I but I would love to. Like, this is It is
37:41 hard. It is hard not to talk about it. >> I want to ask a little bit about you.
37:45 So, I know talking about yourself is a little awkward sometimes, but bear with
37:47 me. >> will be awkward, but let's do it. Um >> I love understanding why people are the
37:54 way they are, and Paul Graham gave you possibly the best compliment I've ever
37:57 heard, which is he called you extremely formidable. And I just I love that quality about
38:05 people. What makes you that way? Is it something you were born with? Is it
38:09 something that you learned over time and you had to hone? >> Um
38:15 I think I was like a very capable kid, but I wouldn't call myself a formidable
38:19 kid. And I do sort of think I somewhat am as an adult. Uh So, it's not like false modesty. I
38:27 My own experience of it was just that I kind of like gradually like built up
38:31 confidence that I could like figure stuff out and do things.
38:36 And that it was not like a it was not like I used to not be like this
38:45 and then I was or I am, but it was kind of like a gradual you like try something hard and it
38:50 works. And then I would be like I would like to scale that up and see if
38:56 we can try like an even harder thing. >> Yeah. But you always had that feeling of let
39:00 me try that next thing. Was the fear of failure not kind of central at all to you? I think
39:06 that prevents a lot of people from doing things. >> One of the things that I think is best
39:11 about Silicon Valley and the kind of Silicon Valley is like a concept and a
39:14 culture not a place is that there is a real culture of people helping each other and supporting
39:22 each other and kind of like a belief that if you invest in the network
39:27 uh, then good things will eventually happen to the whole network and to the whole
39:32 industry and thus to you as like a player in that. And I haven't seen it anywhere else to
39:38 the same degree. Like the kind of for all of the flaws of Silicon Valley
39:44 the culture which are numerous um, this whole thing of like it's hard to know who's going to succeed
39:50 and what ideas are going to work and so the right thing is to like be pretty
39:54 generous and helpful to everyone even if it's not going to like do anything for
39:58 you specifically in the short term as is awesome and a part of that is
40:03 that if people fail you can't hold it against them cuz maybe they're going to
40:06 wildly succeed at their next thing and you know, that's good for all of us.
40:09 >> It's it's actually usually celebrated, too. I love that.
40:12 So I don't >> Like I think here I just got lucky. >> Yeah.
40:16 >> And if I was thrown into any other environment, I probably would have had
40:22 like much more fear fear of failure, but I got out to Silicon Valley as like a
40:26 young person and I saw that a lot of people tried something really hard,
40:31 failed even harder, and then wildly succeeded at their next thing and
40:33 everybody was fine with that. >> Yeah. Um I assume you're under a
40:38 tremendous amount of pressure, competitive pressure. You're building
40:41 one of the most important technologies ever. Like, how does that feel? And like
40:45 how how do you deal with that day-to-day? >> Sometimes really bad and sometimes
40:50 really good. You know, sometimes it's just like oh, I can't get up again and I can't get
40:55 up and do this again. And most of the time it's like I am incredibly lucky to get to do
41:03 what I think is the most interesting and impactful work in the world,
41:06 certainly that I will ever touch. And it uh I mean, it's like one of the coolest
41:12 experiences I could ever imagine in life. It is definitely stressful, but
41:18 you can kind of get used to any like anything. I Like people are just
41:22 unbelievably adaptable to this, so like this just feels like my life now and I
41:26 forgot that life used to be really easy. And now it's going to be hard for a
41:28 while and someday it'll be easy again. I'll probably be kind of bored.
41:31 Um and I the subjective experience of it is like there is some limit to how much like
41:42 stress or bad vibes or whatever you can kind of keep in working emotional space
41:47 and then beyond that something else gets pushed out of the way. So it doesn't
41:51 really get worse. And I am a believer that if you're going to like
41:55 work really hard at something, you may as well pick the big thing cuz
42:00 it's not that much harder or that much more stressful and you won't experience
42:03 it at that much worse and you know, you at least get to like have the
42:06 hopefully the satisfaction of like a bigger positive impact.
42:10 >> I'll say it doesn't show if you if you wake up and you say I don't feel like
42:12 getting out of bed this morning, it never shows. >> It obviously happens.
42:14 >> Yeah. >> Like not that many days, but there are
42:17 there are definitely days where you're just like >> Yeah. Last question for you.
42:22 Can you offer some words of confidence to people who are AI curious, maybe not quite there.
42:30 Why should they be excited about the future? >> Well, the normal answer to that the
42:36 people in AI give, which I hate, is AI's going to cure cancer.
42:38 >> Yeah, I knew you were going to say that. Yeah. >> And the reason I hate that
42:46 if you dig a little deeper into that, what people are saying is um
42:51 or often saying is that AI is going to cure cancer and it's going to make stuff
42:56 cheaper. And the price of that is going to be human agency and you in control of the future. And
43:06 you're going to have the machines give you all this material abundance, but
43:10 you're not really going to have the ability to have power and voice and
43:15 self-determine your life and choose what to build for other people.
43:19 And I think this is this kind of goes against everything I believe in. So,
43:24 I hope we get a cure for cancer. I mean, that seems like a really
43:27 obviously wonderful thing. But what I really hope people should be
43:30 excited about is that we are able to avoid this industrial revolution trap, really focus
43:41 to have the choice and the freedom and the creativity and the time to do all the
43:48 other things they want in life. And that every decade from now on,
43:53 people look back and say, "I can't believe how much worse those people had
43:57 it and now I have so much more." Not just that I have so much material
43:59 abundance, although I'm fine for people to get that, too, but I have so much
44:03 more autonomy and freedom and ability to do things in this expansive space that I
44:07 get to choose. I think that's what we owe people. I think we can deliver that.
44:12 I think both of the ways that I I I have worried about things going very wrong,
44:15 like either and there's like a loss of control accident to AI or an intentional
44:20 loss of control, some lab says I'm going to push the button on RSI to make sure
44:23 we have the power. Uh or, we talked about this earlier, we get like too much power power
44:28 concentration in the hands of a few people that control too much of the AI and are
44:33 making the decisions. Like, I think we can avoid both of those. I think we can
44:36 navigate a path down the middle here where people's lives and individual free
44:41 will and autonomy just get up-leveled and up-leveled and up-leveled. And
44:45 I think that future is already here in some ways. It's just very It's so new
44:50 that it's very poorly distributed. I think you said like 1% of the people
44:53 using most of the tokens or something. >> Yeah. >> I think if we can just take what's
44:56 working today for the early adopters and get a lot more people to embrace this,
45:00 we'll already be a significant fraction of the way there. >> I appreciate your vision.
45:03 >> Thank you. >> Thank you, Sam. Thanks for spending time
45:05 with me. >> I enjoyed it.
$

Sam Altman: RSI, AI Consciousness, OpenAI's New Hardware, Compute, and more!

@matthew_berman 45:08 11 chapters
[AI agents and automation][hardware setup and infrastructure][community building and audience growth]
// chapters
// description

Sam Altman joins me to break down AI consciousness, recursive self-improvement, and how AI could help discover new medicines and scientific breakthroughs. We also get into the future of AI devices, the pressure of leading OpenAI, and why Sam believes we should aim for a new renaissance. Huge thanks to Sam for joining me! https://x.com/sama Join My Newsletter for Regular AI Updates 👇🏼 https://forwardfuture.com My Links 🔗 👉🏻 X: https://x.com/matthewberman 👉🏻 Forward Future X: https://x.com/forw

now: 0:00
// tags
[AI agents and automation][hardware setup and infrastructure][community building and audience growth]