The OpenAI Founders On Their Plan To Battle Elon, Compute And Everything Else

0:00123123.

0:03We're going to do all that personal

0:04drama in 90 minutes. Oh yeah, oh don't

0:06you worry. Don't you worry.

0:08Is this part all going to be in the

0:09podcast?

0:12Bear with me through my slightly silly

0:14intro, but uh

0:17it's going to be okay.

0:19>> [music]

0:30[music]

0:32>> Welcome to the fortress of finance, the

0:35capital of capital.

0:38This is not that podcast. This is core

0:39memory. I am Ashley Vance.

0:41>> And I'm Kylie Robison. And I think we

0:43have a hell of an episode for you.

0:46Normally we do an introduction of people

0:49at this point, but

0:50that's perhaps unnecessary today. We

0:53have Sam Altman

0:55and Greg Brockman, the co-founders of

0:57OpenAI. You may have heard of it. Thank

0:59you guys for being here.

1:01Thank you for having us.

1:02>> [laughter]

1:03>> Thank you so much.

1:04I think this is the first time you guys

1:06have ever done a podcast together. That

1:07is amazing, but I think that's true.

1:09Certainly in a long time. In a long

1:11time. Maybe the first one. And I will

1:13not fight you for doing it on our show,

1:16but you did buy a podcast. Is there We

1:18just lucked out. You lucked out.

1:20>> Yeah. I'll take it. I'll take it. I was

1:22curious why you guys bought a podcast.

1:24It's not really

1:25You don't have to go deep on it, but did

1:26you have quick thoughts on that?

1:29I think that the people who who do TBPN

1:32are incredible. I think they're just

1:34very creative thinkers, and I think that

1:36in this world that we're moving to of

1:38building these AI systems that are so

1:39useful for people and helping people

1:41understand why why that's valuable for

1:43them in their personal lives and work

1:45lives. Like these are the kinds of

1:47people that I think could help tell that

1:48message. I've seen you on it. Have you

1:50been on TBPN? I have. Yeah, okay. Okay.

1:53Um

1:53I don't watch all the episodes, but It's

1:55a fun pod. Um well I thought, you know

1:57since this is you guys haven't done this

2:00together at least in a while, we would I

2:02was going to go down nostalgia lane just

2:04for a little bit at the beginning.

2:06Um,

2:07Kylie and I have gotten to know both of

2:09you over the years. You know, I was just

2:11I we were reflecting as we were

2:13preparing for this.

2:14We're a little bit past the 10-year

2:16anniversary. You guys are two of the

2:18remaining co-founders. I think Wojciech

2:20is the third. So you you know, you're

2:22this constant line that's been running

2:24through the company.

2:25>> company you started as an underdog. You

2:27ended up as the the top dog. It's all

2:30through a lot of drama, undulations.

2:33Um,

2:34I was

2:35I was just I was genuinely curious of

2:38how your relationship

2:40through all this has changed and how you

2:42guys have played off each other and if

2:44it's morphed over time.

2:46>> It is extremely nice in

2:49Look, we always wish there were less

2:50drama. We wish we just got to focus on

2:52the tech.

2:53But in a world of so much chaos and

2:55drama and you know, tension and fighting

2:58and power struggles, it has been

3:00unbelievably nice to have a relationship

3:03with someone that's got the full

3:04context. We have all this history and to

3:07really depend on each other in amazing

3:09times, very tough times.

3:11Uh, it has been one of the nicest things

3:12about OpenAI. You know, in many ways the

3:15very first

3:16moment of OpenAI was right after this

3:19dinner that we did in July of 2015 and

3:23Sam and I were driving back to the city

3:25together

3:26and we looked at each other and we were

3:29like, we have to do this. Right? There'd

3:30been all this conversation of is it too

3:32late to start a lab that could go after

3:34AGI and have have a positive impact.

3:36>> so ridiculous now that we were so

3:37worried about that. Yeah, it was too

3:39late.

3:39>> Yeah, that's the You missed it. You

3:41missed it.

3:42>> feeling like that when you guys started

3:43though. I was like, no. DeepMind's going

3:45to run away with this, you know, yeah.

3:46Yeah. And you know, the conclusion of

3:48dinner was it wasn't obviously

3:49impossible. And I think both of us just

3:51felt like, okay, this is just so

3:52important. We just have to do it. Yeah.

3:54And I think that spirit continues. And a

3:57lot of how we operate in the early days,

3:58like I remember I was unemployed at the

4:00time, so I was full-time on it the next

4:01day. Sam actually had a day job, um but

4:03we were constantly on the phone, like

4:05probably like five times a day. Yeah,

4:07exactly.

4:07>> Were you Were you guys already Were you

4:09like close friends already at that point

4:11or or not? We had known each other for a

4:13super long time. Or I felt like a super

4:15I don't actually Yeah. What What year

4:17was that? 2010? 2011? Whenever you

4:19started at Stripe. That's right. Yes. So

4:21we met through the Collisons.

4:22>> Yeah.

4:23And so we'd been kind of casual social

4:26friends.

4:26>> it wasn't as long as I thought. Yeah,

4:27maybe it was 2010 and this was now 2015.

4:30So five years. Time compresses. Yeah,

4:32and and and obviously I mean

4:35the being in the

4:37pressure cooker of all this doing this

4:39work, I mean I would imagine it only you

4:41guys have only got closer over time.

4:44Yeah, you you know, people use the word

4:45trauma bonding. I hate that. I've I've

4:47like I like the other things about like

4:49the people you're in the fox holes with

4:50the fox hole with, but

4:52the the

4:53one of the nice things about hard work,

4:55no matter what, but certainly hard work

4:57in um

4:58stressful times is you really like forge

5:00these relationships that

5:01I at least have not seen get formed any

5:03other way. Yeah. And I do I do think the

5:05way that Sam and I work and relate is

5:08maybe different from what you'd expect

5:10from a typical co-founder relationship.

5:12Like I think that we are just in

5:14constant contact, that five calls a day,

5:16two minutes, five minutes each, that

5:18kind of spirit remains. Like I think

5:19we're just in constant sync. And we

5:21don't always agree on everything, right?

5:23It's not like we come at the world from

5:24exactly the same point of view, but

5:26that's why

5:27we're so strong together, right? Is I

5:28think we have very complementary

5:30approaches, that Sam will say, "Here's

5:32an idea." I'll think about, "Well, maybe

5:34we could do this other way." Or what

5:36about if we approached it with this

5:38angle? Or how does this relate to this

5:40other thing that we're thinking about?

5:41And one thing I deeply appreciate about

5:42Sam is that I think he always sees these

5:45connections between different ideas, or

5:48just like keeps focused on here's the

5:50big picture that we need to get to and

5:52then together we figure out well how do

5:54we actually do it and I think connecting

5:56the grand ambition with the execution

5:59like that is what has always

6:00distinguished Open AI. Yeah, what are

6:01some of the points in these 10 years

6:04where you felt like it was really

6:05important that you guys diverged? Do you

6:06remember any key moments for you guys?

6:10I

6:11I think one of the things that Greg has

6:12done

6:13the best which is not my instinct is uh

6:16really just pushed to

6:20focus on the most important thing in his

6:22own work and also in what the company is

6:24going to do. Um so there have been times

6:26where I have wanted to do more things

6:30uh and Greg has just said, you know, is

6:32this the most important thing? Let's

6:33really just do this. Let's get the

6:34company focused.

6:35Uh and we've diverged on that and that's

6:38been like a very helpful spirit of

6:40Greg's throughout the company.

6:41Yeah, and I would also add I think even

6:44for example thinking about compute and

6:46just constantly raising the ambition.

6:48And sometimes I feel like okay, I kind

6:51of logically know that yes, like we're

6:53moving to this compute powered economy

6:55and yes, that demand is always going to

6:57outstrip supply, but like we've got all

6:59this like hard work to do and we already

7:00have all these big computers and we're

7:02operationalizing them and you know, you

7:04still have all of this like just

7:05physical infrastructure to build and you

7:07feel already swamped in it and swamped

7:09in it and Sam is like no, we need even

7:11more. And I think that that actually has

7:14been a very important thing to really

7:16not like sometimes it's easy to lose

7:18sight of the higher order bit of just

7:20the fact of this is going to be so

7:22important for

7:24not just the next 6 months, but this is

7:25what's important for the next 2 years,

7:27the next 5 years, 10 years and I think

7:29that the keep

7:31like you need this balance of sometimes

7:33swimming in the details, but you can't

7:34be swamped in the details and I think

7:37that that balance is something that I

7:38think again is what really has

7:40contributed to what Open AI is and where

7:42we are going to go.

7:43Is there There must be one

7:46product or

7:48strategy you guys have What's the one

7:49that you've disagreed about the most

7:51vehemently?

7:52>> I was just thinking when Greg was

7:53talking about this. This is not a

7:55a product, but it was this thing that

7:57came to mind is that I was going to say

7:58it before you asked that. Uh

8:01we used to talk a lot about how to talk

8:02about safety. We never disagreed on

8:06the extreme importance of safety

8:08and what it will mean to get this right

8:11or get this wrong, but the the field has

8:13had a strange relationship with how

8:16we've talked about safety, how we've

8:17used safety, and how much that becomes

8:19about power versus actually keeping

8:21things

8:22safe. And I I think earlier in our

8:25history, I got swept up more in the we

8:28got to really

8:29talk about this in a particular frame,

8:31and Greg was very disciplined

8:33about we're not going to fall into the

8:35traditional frame. We can't talk about

8:37that way. Now, even then, I think we

8:39have

8:40because this is so important,

8:43we I would ever have fallen into the

8:45trap of still talking more in the wrong

8:48frame than we should, but I think one of

8:49Open AI's greatest contributions to date

8:53has been

8:55finding a different way to talk about

8:56safety, not just in how we build the

8:58products,

9:00and how we talk about society needs to

9:01do, but like what

9:03how we deploy them, the idea whole idea

9:05of iterative deployment,

9:07and actually getting to

9:14maybe not the

9:16actually getting to a world where we're

9:17figuring out how to deploy products that

9:18get increasingly safe as the stakes go

9:20up, and Greg really held a line there

9:23that I think has been quite important to

9:25the company against extreme pressure not

9:28to do that.

9:30And I think it's been like quite quite

9:32important to our whole strategy, not

9:34just how we talk about things, but how

9:35we ship and build products. Yeah, and if

9:37you look at for example

9:39the OpenAI Foundation

9:41which is the nonprofit that governs

9:43OpenAI and has a very large chunk of of

9:46equity

9:48one of its pillars is AI resilience. And

9:51what that really means is thinking about

9:53how do we make AI be something positive

9:55for the world? And the answer is not any

9:57one intervention, right? It's not you

9:59have

10:00chain of thought monitoring and now

10:02you've achieved the mission. It's really

10:04a whole

10:06deep sequence of different ways that

10:09society should orient around this

10:11technology. And I think that this

10:12perspective of you're not going to solve

10:15AGI going well for the world in a paper.

10:18It has to be a worldwide effort from

10:21contributions from society, from many

10:24different people, from many different

10:26ways of really understanding what this

10:27technology is, how it will affect

10:29people, how it will affect the world.

10:31And this this is something that was not,

10:33I think, at all appreciated or

10:35understood when we were starting out 10

10:37years ago because it's very easy just to

10:38fix it on, you know, we're

10:40technologists. We're, you know, building

10:41technology. That's the only problem we

10:43need to solve. I'm not saying anyone

10:44explicitly said it that way, but I think

10:46sometimes you can fall into a mental

10:48trap of thinking about it that way. And

10:50so a lot of what I think we have spent

10:51time on is being first principles

10:53thinkers and really thinking about how

10:55do you operationally

10:57deliver transformative technology to the

10:59world in a way that is going to actually

11:01help people in their daily lives. And

11:03one thing you realize is, well if you

11:05have a very powerful piece of technology

11:06that will change things, probably it's

11:09going to go better if you've had a less

11:10powerful piece of technology you've

11:12already helped change things in a

11:13positive way. And so if you just think

11:16about it that way, you start to really

11:17be pulled down this road of thinking

11:19about resilience, thinking about

11:21iterative deployment. And again, I think

11:23this is a lot of the dynamic within

11:25OpenAI with between the two of us that

11:27we're always thinking about these

11:28questions of how do we actually achieve

11:30this mission and make it go better.

11:31Yeah, I I feels like just yesterday I

11:34saw you at South by 2022 2023 and it was

11:37a completely different discussion about

11:39AI than what we hear today and it feels

11:41like that's something that's changed in

11:4310 years of like how you discuss safety

11:45and alignment and I'm wondering how you

11:47reflect on that now like what would you

11:50have changed in those panels and those

11:52news hits and like what have you learned

11:54about talking about safety?

11:58Even before we get to safety, uh I think

12:00we have fallen in the frame as tech

12:02nerds of talking about we're going to

12:04build

12:05superintelligence and

12:07dot dot dot it's going to be great for

12:09you. And we've not filled in enough of

12:11the dot dot dot. Like we talk about

12:13we're building this amazing technology,

12:15it's going to do all these wonderful

12:16things

12:17and

12:19there is like a sense in the world now

12:21of okay

12:23looks like you were right, you are going

12:24to build this thing. Um why like

12:28why do we want that? What's that going

12:29to do for us? And the thing that a lot

12:32of the field has said of oh it's going

12:34to you know cure cancer and you'll be so

12:35happy or like

12:37that's clearly not quite resonating. A

12:38lot of people are like sure cure cancer

12:39that'd be wonderful. Um

12:42I think what people really want is

12:46prosperity

12:47agency that they're going to continue to

12:49have meaningful work to do. Um there I

12:52saw an incredible post the other day

12:53that really stuck with me which was like

12:55a a right to adversity.

12:57People people actually you want some

12:59challenges in life. You don't want every

13:00day to be perfect and everything done

13:01for you and there there's a fear with AI

13:03that like mm let's say you're right,

13:05let's say you build it, let's say it

13:06like makes all this money and does all

13:07the work and whatever like what do I do?

13:09What's my kid going to do? What's life

13:10going to be like? Where's the growth

13:11going to come from? What's the what are

13:13people going to strive for? Um

13:15and

13:19I think we've talked as a field and as

13:21OpenAI and as Greg and I we've talked a

13:24lot about the amazing technology and

13:26what it can do

13:27and what it like the technological

13:29marvel

13:30and we have not connected the dots

13:32enough on

13:34here's

13:36what the future's going to be like. Here

13:38is, you know,

13:40when when

13:42when I talk to parents with kind of

13:44school-age kids the most common question

13:46is like what what should my kids study?

13:48What's the future going to be like? What

13:49what will still have economic value and

13:52I I

13:53and I realize that that's not quite the

13:55question they're they're really trying

13:56to ask. It's like how is my kid going to

13:57have a fulfilling life in this new

13:58world? You know, you can answer it

14:00economically if you want, but

14:02that doesn't actually settle people and

14:03I think it's because there's there's a

14:04deeper thing here.

14:06I was just going to say and I think I

14:08think we actually have a lot of

14:09perspective on this and I think that

14:12for example

14:14within ChatGPT, we have so many people

14:17who say that my life or the life of a

14:19love loved one was saved through

14:22information I got through ChatGPT,

14:24right? Whether it's there's someone who

14:26had a kid who had

14:28blinding headaches. They were denied an

14:31MRI. They used ChatGPT to try to

14:33research the symptoms and used it to

14:35argue to be able to get insurance for

14:37the MRI. Turned out there was a brain

14:39tumor. They were able to intervene and

14:40save his life. And just navigating that

14:43experience that that family says, "We

14:46have no idea how we would do this

14:47without ChatGPT." And that's just one

14:49story. There are so many of those and I

14:53think people are really understanding

14:54that this technology can help not just

14:56abstractly society, but can help them,

14:59right? It can help them make money.

15:01Right now we're starting to see this

15:02wave of entrepreneurship. I think that's

15:04going to be a huge theme throughout this

15:05year. And I think we have a lot of

15:07perspective on well

15:10we're going to build these AIs that are

15:11able to make rather than you contort

15:14yourself to computer, right? You think

15:15about the way we do work. Yeah. It's not

15:17natural, right? It's not It's not kind

15:18of what we were designed to do. Um

15:21and instead the computer's going to do

15:22work for you. And the question of well,

15:26what is the work? What is

15:28good? Like, what is the thing that you

15:29actually want to be done? That's going

15:32to be something very deeply human,

15:34right? This agency, this empowerment.

15:36And so, I think there is this very

15:37optimistic, and not blindly optimistic,

15:40but I think this very positive change

15:42that this technology is bringing right

15:43now, but it's so much easier to just

15:46notice what's going to go away. Like,

15:48what is the thing that you thought was

15:50solid and stable that's going to change?

15:52But, it's much harder to see well,

15:54what's coming? Like, what's the new

15:55thing that you get? What is the the

15:57benefit that that comes as a part of

15:59this transformation? And I think that's

16:00something that we're increasingly

16:02realizing that we also have to really

16:03spell out in addition to spelling out

16:05the other sides of it. All right. What

16:07do we do at Core Memory? We cover

16:09innovative, fast-moving,

16:11forward-thinking companies, which is why

16:14Core Memory is sponsored by Brex,

16:16because Brex

16:18is the intelligent finance platform for

16:21many of these companies. 30,000

16:23companies from startups to the world's

16:25largest corporations

16:28rely on Brex's technology for their

16:30finances. They've got smart corporate

16:32cards, high-yield business banking, and

16:36expense automation tools that are

16:38fantastic. I hate doing my expenses, and

16:41Brex's AIs and software run right

16:44through those expenses, figure out where

16:46we're spending money, and take care of

16:48so much stuff for you, so you don't have

16:50to waste your time on it yourself. Go to

16:53brex.com/corememory

16:55to learn more, and just, you know, get

16:58with the program. Let's get going. Let's

16:59get out of this archaic finance software

17:02and move toward the future. Core Memory

17:05and Brex.

17:07I have a couple questions based off what

17:09both of you were saying. Um

17:11So, I just went to Oberlin College with

17:13my son, and we were in this room

17:17um

17:17with I don't know, 600-700 people and it

17:20was fascinating to me. I mean it's in

17:21Ohio and people had come from all over

17:24the country and

17:25the president was up there giving a talk

17:27and then opened it up to Q&A and so many

17:30of the questions were about AI but they

17:32were

17:33it reinforced for me that

17:36how much of a bubble we live in here

17:38because they were what I would consider

17:41I mean it's not trying to like put

17:43anyone down or they they were pretty

17:44basic question I I just my my

17:46overwhelming sense of the room was like

17:48wow you guys kind of like don't know

17:50what's unfolding and about to to come

17:53across

17:55all of you and I don't know it was I

17:56don't know it was like alarming it was

17:58just it was eye-opening for me you know

18:00and and they were asking um

18:04They were obviously asking stuff about

18:05like well how are you going to use AI in

18:07class how are you going to stop kids

18:08from just relying on it but then it was

18:10like you know they kind of advanced a

18:12little bit beyond there but I felt like

18:14this room

18:16of pretty smart people were not not

18:17really in touch with what's going on. So

18:19I mean if this is a problem that people

18:21don't know these examples of of what

18:24could be done and we don't even know

18:26what shape this could take exactly you

18:29just I don't see how we solve this in

18:31sort of

18:32um kind of prepare

18:34people in some way.

18:35>> Well I I do have one

18:37sort of positive take on this which is

18:39that I think that if you just read about

18:41AI

18:42it gives you one impression right it's

18:44again it's like very it just feels like

18:47you're trying to wrap your mind around

18:47this new technology what is it but when

18:49you use it it's so intuitive.

18:51Right and that is the purpose of AI in

18:53many ways it's like you think about back

18:55to

18:56how we design computers for the past you

18:58know 70 years it's really about the

19:01machine doesn't quite get you right it's

19:03like you have these goals in your head

19:04you have to break it down into the

19:05machine's language whether that's

19:07writing assembly code or okay fine now

19:09we get higher level languages now you

19:11can kind of talk to your computer, but

19:13even if you think about how ChatGPT

19:14works, you still have to really

19:16understand concepts of language models,

19:18right? The fact like why do you have to

19:19create new conversations, right? New new

19:22new tabs. Like, why can't it just be a

19:24thing you talk to that remembers

19:25everything, right? Why there's there's

19:27these limitations that are techno

19:29technological limitations, but we're

19:30improving them. And so, I think that

19:32what we're building is the most

19:33intuitive technology. We are building

19:35something that can torch the machine to

19:37you. And when people use it, they

19:39realize, "Wow, this is what I can do

19:41with it." Like, some of the stories that

19:43I love hearing are someone

19:45in the Midwest who I you know, one one

19:47of my friends

19:49is who his sister was telling him

19:52about this app that she wanted to to see

19:53someone she wished someone had created.

19:55She described it in a detailed way.

19:57While he was listening, he was like,

19:58"Uh-huh. Uh-huh." typing into Codex

20:00exactly what she was saying. Pushed

20:02enter.

20:03Few hours later, he shows her the

20:04result.

20:05App that's exactly what she described.

20:07And she's like, "What the hell is this?

20:10Who built this? Like, this is exactly

20:12what I wanted." And he said,

20:14"You built it."

20:16Right? And that is the kind of aha

20:17moment that I think that everyone is

20:19going to go through. And once you're

20:21once you experience what AI can do for

20:23you and the fact that

20:25you are now empowered, like you have a

20:27image in your head that you want to see

20:29in the world. If you look at some of the

20:31upcoming

20:32some of the new image models we have,

20:34absolutely incredible. Like, you can

20:35create in a way that was never possible

20:37before.

20:38And the fact that if you are

20:41like like another story that I think

20:42about. So, my grandmother had

20:45dementia Alzheimer's, and that

20:48it was extremely tough for for everyone.

20:52But, one thing that really stuck out to

20:53me was she was actually able to use her

20:54Alexa to play music, and she could still

20:57remember

20:59the lyrics and sing along to songs. And

21:01it was just a way that she was connected

21:03to who she was through technology. And

21:07that really stuck stuck out to me. It's

21:09like if you can build interfaces that

21:10are intuitive

21:11that

21:12anyone can connect to them in ways that

21:15right now it feels like we we have

21:17pieces of technology that are targeted

21:19to specific individuals that you have to

21:21build all the skill that's not really

21:23related to deeply what you want, but

21:25it's just somehow specific to the tech

21:26itself. Three things came up for me

21:29during

21:30what Greg was saying, um

21:32which I I agree with all of.

21:34One

21:35before we launched ChatGPT, we used to

21:37try to talk to people

21:39and say, "This AI thing is coming.

21:40You've got to pay attention. It's going

21:41to change everything. It's really It's

21:43It's really important." And it got kind

21:46of no attention. We wrote these

21:47beautiful blog posts. We sort of

21:49did all these incredibly impressive

21:51feats. We won, you know, video game

21:53competitions. We had a robot that could

21:54do a Rubik's Cube with a hand. We like

21:56did all this amazing stuff. And we kind

21:58of felt pretty cool and we kind of said,

22:01"Oh, you know,

22:02the New York Times wrote about this. We

22:03must be doing great." But really truly

22:05no one cared.

22:06It got no actual impact. And then we

22:09launched ChatGPT, which was by far not

22:12the most impressive technological thing

22:14we had done. By far by far.

22:16And

22:18as soon as people could feel it, they're

22:20like, "Okay, I understand it." I think

22:22that was the moment probably most that

22:24the world has said maybe this AI thing

22:25is real. Like collectively all at once.

22:28Because people can use it and they can

22:29get value out of it. They develop their

22:30own sense. It's very different than

22:32hearing about it. That's happened again,

22:34as Greg said, with coding models. Um but

22:36those have really been the two kind of

22:38like

22:40large significant moments where I think

22:42the world has updated and said, "Okay,

22:43there's this thing happening." There

22:44will be more in the future. But

22:47so far the fact that you can

22:49ask a computer anything and get an

22:50answer

22:52or

22:53have a computer do anything anything

22:55with code for you.

22:56Um

22:57Tho- those have been really powerful and

22:58I I think that's how the world's going

23:00to update. Us saying superintelligence

23:02is coming is going to change everything.

23:04You know, maybe the people who listen to

23:05this podcast will sort of say, "All

23:06right, sounds sounds reasonable.

23:08Probably should pay attention to that."

23:09But it won't it won't have that big

23:11impact on the world. Uh

23:13So, I think the most important thing we

23:15can do to help that audience of people

23:17think not just about like what does this

23:19mean for people not cheating in class,

23:20but like what is this going to mean for

23:21the world is to ship great delightful

23:24products that create a lot of value and

23:26are easy for people to use.

23:27And we will continue to do that.

23:30The The second thing is we have seen

23:32again and again in our history that when

23:35we put something out that's pretty good,

23:37people say, "I can't really imagine what

23:39happens if this gets better. It can't

23:40get much better."

23:42This image model that Greg was

23:43mentioning, which we'll launch soon, was

23:45a real example of that for me. I kind of

23:46roughly thought image generation was

23:48solved. I was like, "Yeah, it's really

23:49good. Like, I don't need it to get any

23:51better. There's a lot of" And then this

23:52new thing has been a real reminder of

23:54like, "Wow, this can go so much further

23:57and there's so much more I can do." What

23:59does it do?

24:00It makes ridiculously great images. Uh

24:02and Has it solved text? Because I tried

24:04to make core memory merch mock-ups Try

24:06again very soon.

24:07>> [laughter]

24:07>> Uh

24:08the the team really did a great job at

24:10this one. But even with ChatGPT, like

24:12when we put GPT-4 in ChatGPT, I remember

24:14a lot of people, like sophisticated

24:16friends of mine, saying, "This is it.

24:18This is AGI. Like, if the model got

24:19smarter, I don't care. Just make it

24:20cheaper. Like, I can't This is amazing."

24:23Um

24:24and if you go back and use that whatever

24:27it was, uh

24:29I guess like March of 2023 version of

24:32GPT-4, you'll be like, "This was

24:34terrible."

24:35But at the time people were like, "It's

24:36solved. It's solved. I mean, it's beat

24:38the Turing test. It's done. It can't get

24:39any, you know."

24:40And and then it gets better and better

24:43and it like you can just

24:45keep raising the expectations and

24:46possibility of what you can do, to say

24:48nothing of then you're going to do

24:50reasoning models and do code. I'm just

24:51talking about how much better ChatGPT

24:52got in that time period. So, I think we

24:55see this again and again, too, where the

24:57world is like, "Okay, you made this

24:59amazing thing. That's as good as it's

25:00going to get. That's as good as I need."

25:03And then, month by month, or at least

25:05quarter by quarter, expectations and

25:06ability ratchet up in a huge way.

25:09And then, the third thing that Greg was

25:11saying there is

25:14these models are still

25:16quite dumb relative to what they will

25:18be, but more than that, they have quite

25:19limited awareness of your life.

25:22You are still having to like

25:25massage them and

25:27cajole them and try to get the thing

25:29that you want.

25:30Uh we are not no longer that far away

25:33from a model that just

25:36knows all of your context. It knows

25:38about you. It knows about your life. It

25:39knows what you're doing. It knows what

25:40you care about. It knows about the

25:41people in your life. It has access to

25:43your computer and your browser and, if

25:45you want, of course, in the ways you

25:46want, and it has access, maybe

25:48increasingly over time, to what's

25:49happening in in the real world around

25:50you.

25:52That

25:53is going to be a complete change to what

25:55it feels like to use a computer

25:57and what it feels like to use AI.

25:59And

26:01I am tremendously excited about that,

26:02but I don't think even we have a good

26:04intuition yet for what that's really

26:06going to feel like. Yeah. And to to that

26:07point, yeah, you think about how much

26:09time you spend right now just explaining

26:11to chat or whatever tool you're using

26:13what's going on. And if you think about

26:15how frustrating that is, like, if it's

26:17like you got a coworker you're

26:18constantly trying to explain to them,

26:19"No, this is kind of what I want. This

26:21is what's going on." I have projects I

26:22come back to and it's like that.

26:24>> [laughter]

26:24>> Exactly, right? Like, it's it's just not

26:26it's not really how you want to how you

26:27want these systems to behave. And and

26:30and just to say one more thing on on

26:31what Sam was saying is that I think that

26:33one of the biggest technical challenges

26:35we have at OpenAI is too much

26:37opportunity. Like, AI is this like field

26:40of boundless opportunity. No matter what

26:41dimension you expand on, there's going

26:43to be something new and unprecedented

26:44and amazing. And so, the important thing

26:46is having a vision for where can we

26:49really focus, where we're going to get

26:51the most returns and the most benefit

26:54from this you know having multiple

26:57different efforts that all add up to

26:58something. And I think that when it

27:00comes to we have these coding systems

27:02now, but I think it's going to clearly

27:04expand to just all computer work. It's

27:06and it's funny by the way that you know

27:08you think about just the degree to which

27:10you have all of this like the work that

27:13you do it's like very seamlessly

27:14interlinked between you're having

27:16in-person conversation, you're typing

27:17this on your computer and trying to

27:20really figure out how to get the context

27:21in there. It's going to be such an

27:22important problem. All these things

27:23matter, but then also in your personal

27:25life you really want this and we're

27:27starting to call where we want to go the

27:29personal AGI, right? This AI that really

27:32knows you has the context that you it

27:34you can trust it, right? That you ask it

27:35for you ask it questions related to

27:37finances just or health and it can give

27:39you trustworthy information. All these

27:41things are important and it needs that

27:42context as well. And so you start to

27:44really see that there's this blurry line

27:46between an AI that's used for this deep

27:48computer work and an AI that's used in

27:50your personal life at a technological

27:52level even if you want different systems

27:54cuz you want one that's really just

27:55knows your work context, one that knows

27:57your personal context. And so you can

27:59pursue those

28:01in parallel and build on the same

28:03technological foundations. And the thing

28:05that I find most amazing about the

28:06technology we're building on is it's all

28:08at the core

28:10one neural net, right? It's still deep

28:12learning. You're scaling it up. You're

28:14building one system and you're applying

28:15it to all of these different amazing

28:16applications.

28:18I I

28:19Can I ask a quick Uh I want to ask a

28:20question sort of along those lines. I

28:22mean I promised myself I wouldn't do

28:23this to you. Um but I mean I was kind of

28:26like I wasn't like a non-believer, but I

28:28like I could see AI doing incredible

28:30things, but I was I was kind of like

28:32skeptical until relatively recently on

28:35on

28:36>> What what updated you? Well, a couple

28:37things. I mean one was I started playing

28:39with agents a lot more and really

28:42shaping them to do what I wanted and

28:44kind of feeling what both of you were

28:46talking about. I was like, "Oh, this is

28:48actually saving me a lot of time. It's

28:50doing my bidding. It's doing it quite

28:52well." And the other is, you know, I

28:53cover biotech so much and

28:56just seeing some of the results that are

28:58coming out is is

29:00It strikes me that like coding and

29:02biotech seem like the most obvious

29:04places right now where it's really

29:06going. But then, you know, I was reading

29:07it I was talking to you the other day at

29:08dinner and reading some of the things

29:10that you've written about,

29:12you know, I see this clear path

29:14based off LLMs to going to what you

29:16describe. I still have this part of me

29:19though. I think it's because I deal in

29:21language so much where the writing is

29:23still

29:25not great where I

29:28This is nothing other than like

29:30anecdotal and intuitive. It's just It

29:32strikes me that

29:34the feeling I get back still is like,

29:37"No, this is not a super intelligence."

29:40We we are not there yet on personality.

29:43>> But like does LLM Right. But LLMs will

29:46get us there.

29:47I was just thinking of it as it's very

29:48jagged, right? And I mean, it's Okay,

29:51for example, just in the past couple

29:53days

29:54our AI solved this long-standing British

29:57problem.

29:58So, this mathematical mystery that's

30:00been, you know, of great interest for

30:02for a long time. A mathematician spent a

30:04lot of time working on this, many years

30:06thinking about this problem, posted

30:07about his his views of of what was what

30:10what the contribution was. Terence Tao

30:12also saying that this looks like there's

30:14like maybe connection between different

30:15fields of mathematics that this AI has

30:18discovered. And we're starting to see

30:20real beauty come out of these machines.

30:22Now, that's a particular domain, right?

30:23Mathematics is very different from

30:24creative writing. And I think that the

30:27fact of these AIs being able to have

30:28interesting insights and being able to

30:30help like it's you know, there's so much

30:33that these mathematicians are now

30:35saying, "Well, think about what more we

30:37can do." And so, I I think that we have

30:39a jagged frontier in terms of what these

30:41AIs are good at. We know how to keep

30:43pushing that frontier back, but I think

30:45that the the thing that we're really

30:46looking for, if you think about AlphaGo

30:48move 37, right, that was something that

30:50not just it was this deep insight, but

30:53it really changed people's understanding

30:54of the game of Go, and more people play

30:56Go now, right? So, that it really

30:58increased the meaning of what people

30:59were doing. And so, I think that what

31:00we're going to do is we're going to make

31:01something that is going to you're going

31:03to have fewer complaints about it, but I

31:05think you're also going to be just like

31:06way more able to do the thing that you

31:08want in writing than you can imagine

31:11today. I'm sorry to harp on this, but I

31:13was told GPT-5 was going to be really

31:15good at writing. And my like update when

31:17I became less cynical was reasoning

31:19models, and when Claude delivered like

31:22really good research, and then you guys

31:24had a research feature, that was for me

31:26like, "Oh my god, this is amazing."

31:28Saved me a lot of time. But, I was

31:31promised before that I would have good

31:32writing by now.

31:33>> The writing still there's no there's no

31:35soul, you know? There's there's

31:37something missing. Well, I'll I'll tell

31:39you. So,

31:41from a technical perspective, the

31:42technology we have is you can train a

31:45model

31:46in this unsupervised way, right? So, it

31:48really we look at all the publicly

31:50available data, and it learns to predict

31:51what comes next, and it's really about

31:53putting it in new situation, trying to

31:55figure out what's a reasonable thing to

31:56do. Then we do a reinforcement learning

31:58step, where it actually tries out ideas,

32:00and it gets rewards and punishments

32:02based on how well it did. You know, just

32:03positive signal, negative signal, not

32:05not You're not beating it. Exactly. Not

32:07at all.

32:07>> [laughter]

32:07>> Yeah, just just you know, this this

32:09signal. Um and then

32:12the hard part is, well, how do you

32:14judge? How do you decide if something

32:16was yep, thumbs up or thumbs down? And

32:19so, in math and science, much easier

32:21than in some of these more open-ended

32:23fields. But, we also have AI's that are

32:25getting much smarter and much more able

32:26to provide that kind of reward signal.

32:28And so, I think that the part of the

32:29challenge has been how do you expand the

32:32task the set of tasks that can be

32:33graded, and that's been a lot of the

32:35focus. And uh there's actually very

32:38interesting like a lot of this stuff

32:39dates back to the very beginning of Open

32:41AI. We kind of had a picture this is how

32:42it would go. But I would just say I

32:43think we're getting there. I think we

32:44definitely have a lot more progress to

32:46make. But yeah, hopefully you can keep

32:48giving us feedback and we'll we'll be

32:50able to improve it for you.

32:52One embedded challenge in that is that

32:54the the writing you want is very

32:56different than the writing that most

32:58people want. And right now we have to

33:00you know make a model that about a

33:01billion people use and all kind of kind

33:04of like. And what we'd like to do is to

33:06get the model so good at personalization

33:08that you think it's a great writer.

33:10And some other person that has a very

33:12different kind of life and set of needs

33:14than you also thinks it's a great writer

33:15for what they want. And those are very

33:17different things.

33:18>> I think that's what I'm kind of asking

33:19though. Like for the personal AGI

33:22you would want it to have that sparkle,

33:23that soul, that sort of that magic,

33:26right? And and I just I

33:29if it's not

33:31And I saw you know obviously Jan LeCun's

33:33been arguing against this for a long

33:34time. I I caught this Dennis thing very

33:36quick where he seemed to be kind of

33:38aligned with you guys but you know that

33:40LLMs will get us there but he seemed to

33:41be saying we still need a couple more

33:43ticks maybe of something special to to

33:46but you guys seem so confident.

33:48>> can solve we haven't tried to make it be

33:51you know a what Ashley Vance will think

33:53is a great writer. We have tried to make

33:54it you know solve open math problems

33:56that the smartest mathematicians in the

33:58world can't solve. And I don't want to

34:00say anything about your intelligence

34:01relative to these mathematicians but I I

34:03think

34:03>> Writers Writers are smarter

34:06But I was going to say like you know

34:07solving math is also hard. So if we can

34:09do that

34:11I'm pretty confident this approach can

34:12also learn what you think is great

34:15writing and figure out to do it for you.

34:16But but but by the way we also have some

34:19new new models coming on on this

34:20dimension as well. So after this podcast

34:23comes out

34:23>> wait to judge it. Let us know. Let us

34:25know if it looks better. writing I think

34:27personality and I think

34:29we're always improving like all the

34:30capabilities is moving up the jagged

34:31frontier. and so I think in this field

34:33it's always really important to judge

34:35the not the current position but what

34:37has that slope been and also fit it to

34:40an exponential. And so if you think

34:41about how does the writing stay compared

34:43to a year ago? Sorry if you're

34:44disappointed with 2.5. Don't worry we're

34:46we're we're very motivated to you know

34:48to to to really deliver. But I think

34:51that I think we really have a line of

34:53sight for how to improve it for for

34:54every application that people want.

34:56Since day one the core memory podcast

34:58has been supported by the fine people at

35:01E1 Ventures. They are a young and

35:03ambitious VC firm in Silicon Valley

35:06investing in young and ambitious

35:08companies and people. Thank you so much

35:11to E1 Ventures for all your support. In

35:13the world that you guys have outlined um

35:18this technology okay, you know, say

35:20everything is working great and this you

35:22know, this is uh it's a quite optimistic

35:24scenario

35:26um

35:27diseases are being cured um resources

35:30are becoming more abundant, problems are

35:32being solved. So so humanity as a whole

35:34is being lifted up. It's I still don't

35:36see I mean it seems like very very some

35:38of the world's smartest people are

35:39developing this technology. It strikes

35:42me it will disproportionately benefit

35:44them

35:46in almost all scenarios even though like

35:50everyone might get lifted up a bit that

35:53um you know, things really get more

35:56extreme because every time

35:58um

35:58what you guys are talking about about

36:00how you would manipulate these tools use

36:01these I I just feel like

36:04I feel like this

36:05things will go even more extreme and the

36:07have the haves will be

36:09>> the permanent underclass and that

36:11sentiment of like seeing these really

36:12powerful tools and feeling like oh my

36:14gosh, what what am I even here for then

36:16and feeling very disenfranchised.

36:18>> you get have all this time to do fun

36:19creative things. I just feel like other

36:20people will be manipulating this in such

36:23extreme fantastic ways that

36:24>> This This is at the core of the OpenAI

36:27mission. Like this is really why we

36:29started this place.

36:31Because you see this powerful technology

36:33coming. It's going to be the most

36:34important technology ever created. How

36:36do you ensure it benefits everyone? Like

36:39truly all of humanity. Like that is in

36:41our mission. Ensure AGI benefits all of

36:42humanity. And we really mean it. And if

36:45you look at our corporate structure,

36:47we've tried multiple different

36:48iterations at trying to in the structure

36:51codify some of our values of how do we

36:53ensure that this is something that

36:54really does benefit people. You see it

36:56in the product choices we make, right?

36:58We decided to launch ChatGPT because we

37:01really believe this is technology that

37:03we need to be able to put into people's

37:04hands. And that was very controversial

37:06by the way, that there was a different

37:07school of thought saying that the way to

37:09do it is you have to build it in secret,

37:11that you can't get people access, that

37:13you need to do it this other way. And I

37:16think that when we look at the how do we

37:17actually make society resilient, how do

37:19we actually benefit people, it all kind

37:20of points to the direction that we we

37:22have been going. And there's a lot of

37:23nuance to it. Um but I think that the

37:25way that we're we're headed is very much

37:28that I think the floor for everyone is

37:30not going to go up just a bit. Like I

37:32think we are going to head to this world

37:33where

37:34I mean you think about even having

37:36a doctor in your pocket that's better

37:38than the best medical team that anyone

37:41in the world could get today,

37:43accessible to anyone who has a

37:44smartphone. Like that is coming and

37:46that'll be free.

37:47Right? That is a wild wild fact. That's

37:49not a small It's going to be a little

37:51bit better. That is fundamentally

37:53raising the floor in this massive way.

37:55Now I think the question of also raising

37:57the ceiling, I think that's fair too,

37:58right? I think that the question of

37:59exactly how the distribution goes. And

38:01we think about this a lot. You look at

38:03the OpenAI Foundation. That's That has,

38:05you know, somewhere between, you know,

38:0625%, 30%, somewhere around there of

38:09OpenAI equity, right? That's like more

38:11than 150 billion dollars. If it is the

38:13case that OpenAI becomes very

38:15successful, that all of that value

38:16locked up in the nonprofit, like that is

38:18something that will be for really

38:20benefiting the world and we think that's

38:22a good thing. Well, we were at dinner

38:23the other day, It's real quick, and I

38:25was listening to you talk about what you

38:27do with agents, and I've known you I

38:29mean, look, you're smart as

38:30and

38:31I was just like, "Jesus, man, Greg, with

38:34this

38:35machine

38:36I like even though I feel like I'm doing

38:38interesting things that help me out when

38:41I heard you describing your life and

38:43what you were doing, I was just like,

38:44"Oh my god, man. I mean, there's just

38:46going to be super people running

38:47around."

38:49I don't know, it's like kind of

38:50intimidating, really. I mean, it was

38:51cool, but it was also yeah, quite

38:53intimidating. I'm like, "I don't know

38:55how to use this like you do at all." Can

38:58I Can I try a less sanitized version of

39:00Yes, please.

39:04I hope I don't get in trouble for this.

39:05Uh

39:06I I can see three

39:08futures of the world. Um

39:11I can see one where, as Greg was saying,

39:14the floor comes way up. You know,

39:16everybody gets subjectively like 10

39:18times richer, just the sort of

39:20materially

39:21material abundance, the prosperity is

39:23crazy huge. People are like, "Man,

39:26relative to my life a decade ago, I'm

39:28doing great." But also in that world,

39:30these people who really learn how to use

39:32agents and get a lot of compute together

39:33and whatever, we have some

39:34trillionaires, you know, maybe 10

39:36trillionaires, whatever. So, like, the

39:39floor comes way up,

39:41like dramatically up, but because this

39:43is a lever that people can really use,

39:45uh the

39:47uh the the sort of like most capable,

39:49most ambitious that people who already

39:50started rich and have access to a lot of

39:52compute, inequality gets worse.

39:54So, that's that's one world I can see.

39:57Um I can see another world where we,

40:00through

40:01many different kinds of things that

40:02could happen, um the the floor doesn't

40:06come up as much, we don't generate as

40:07much total prosperity, but also there's

40:09like less inequality. So, maybe people

40:11in a decade feel like

40:13twice as rich, but inequality has come

40:15down.

40:16Um

40:18and

40:20actually, I'll just like stop and talk

40:22to those two because I think this is

40:23like the real crux of the issue.

40:25>> Is the third much scarier?

40:26>> No, but it it's like a distraction from

40:28the point I want to make. I think a lot

40:30of people in the world

40:32uh like we think it's obvious. Greg and

40:34I think it's obvious that people should

40:35prefer

40:37the first. I assume you also agree with

40:38that.

40:39Um but emotionally, it's not where a lot

40:41of people are.

40:43Wait. Well, I was going to say I don't I

40:45don't want to claim what what is obvious

40:47or not. I but I think it's I think for

40:50me personally, I just see so much

40:52potential in this technology and I think

40:53it's just so important from a societal

40:56level, just thinking about even American

40:58competitiveness.

40:59Like you look at robotics. Like I don't

41:00think that we are ahead in robotics at

41:02all. We are not. But you're not.

41:04With the software side, we do have a

41:07lead we do have this opportunity and I

41:08think that is something all of these

41:10factors together in my mind are are

41:12worth considering at once.

41:14Y- yeah. I mean, clearly I mean,

41:16actually, people who disagree on a lot

41:17of other things I think do agree that

41:18America is needs to be competitive on

41:21chips and robots and AI and everything

41:23else. But but there's going to be this

41:24huge question about how we organize

41:26society and the economy and do we push

41:29for maximum prosperity and accept the

41:32inequality that will come with that or

41:35do we say like, "Actually, we're going

41:37to constrain it

41:39because we're going to be more focused

41:41on the relative nature and the fear that

41:44well, if we don't do something, yeah,

41:46maybe the world will get more

41:47prosperous, but the people who are

41:48really good at using this are going to

41:49have all this power." And

41:51you know, emotionally, I think

41:53intellectually, it's kind of seems clear

41:55to me. Emotionally,

41:56I really get why it's not clear. And I

41:58really get the the fear that if we let

42:02inequality run here, the compounding

42:05nature of this tool is something that we

42:08don't yet understand. Now, the last

42:10thing I'll say on this

42:12is

42:13kind of no matter what

42:16I think everyone should want much more

42:18compute, much more infrastructure

42:20uh and the cheapest possible access to

42:21AI because otherwise I think you really

42:24exacerbate inequality if there's a

42:26limited amount of this and the price

42:27goes up because of supply and demand and

42:29only the rich people have it. Yeah, I

42:30want I wanted to to just build on that

42:32because I think that to some extent

42:35and this also plays back to to earlier

42:37just how Sam and I work together, how we

42:39think, right? Because I think that

42:41laying out these two options I think is

42:42actually really good point where it's

42:44like, well, are those the only two

42:46options, right? Is there something that

42:47we're not seeing in the space? And I

42:49think that the last thing that Sam said

42:50to me is an unlock in terms of thinking

42:52about so AI is really opportunity. AI is

42:56opportunity for everyone if you have

42:57access, right? If you have compute. If

42:59you don't have compute, you can't,

43:00right? No matter how good you are with

43:02agents, if you don't have the computer

43:03to run them, you're not going to be able

43:04to do much. And so I think a world where

43:07if everyone does have access to compute

43:09and

43:10you know, I think for for my generation

43:12growing up, we were much better at using

43:13computers than our parents, right? We

43:16grew up with it. We were, you know,

43:17native. And I think that the generation

43:20growing up now are going to be as good

43:22as I am at using agents, I think they're

43:23going to be 10 times better. It's wild

43:25to watch that. It is crazy. I mean, I

43:26wrote this is like the first massive

43:28story on core memory was that dude who

43:31built the nuclear fuser with with

43:33Claude. I just remember it was the the

43:35fuser didn't even sort of matter as much

43:36as like what I saw him doing with his

43:38computer. I'm like, you just use this

43:40thing different. And you know, he's like

43:4122 and he has eight AI applications up

43:44and Yeah. And if you can get compute to

43:47that kid, if you can get compute to

43:49every kid

43:50then

43:51it is really going to be the most

43:53extreme version of the American dream,

43:55right? Where it's really anyone who has

43:57the desire, the willpower to really play

43:59with this technology, to really lean in,

44:01to build with it, to get the most out of

44:02it, I think you're going to see this

44:04outperformance and I think you're going

44:05to see people have all sorts of

44:07mobility. And so, I think that to some

44:09extent the right way of answering this

44:11question of option one versus option two

44:13is to say

44:15Neither Exactly. And something else. And

44:17I think it's possible. I feel like we're

44:19doing some broad stuff and and we have a

44:21bunch of specific questions that I'm

44:22going to I'm going to be bad and ask one

44:24more broad one. It's okay. Just based on

44:27what we're talking I mean, you guys

44:28probably know, maybe you don't. You

44:29know, I just cover hardware so deeply

44:31and um I think I'm pretty

44:35I would venture to say I've probably

44:36been to more factories and hardware

44:37startups in the US than just about

44:39anyone and keep an eye on China. I mean,

44:41the thing that always plays through my

44:43head is that yeah, we're ahead on

44:44software. We're ahead on AI. Software's

44:47been this

44:48the story of the US for the last 40 or

44:5150 years of our our strength. But, you

44:53know,

44:53when I think about how

44:55this technology manifests itself in the

44:58physical world, I do not see a future

45:00where the US is competitive in any way,

45:03shape, or form. Not just in robotics,

45:05but you know, just all of the

45:06componentry that goes in to these

45:09hardware systems. I go to El Segundo.

45:11I see 30 startups. It's great. There's

45:14like this flourishing of ideas and

45:15people trying things, but it's like such

45:17small potatoes compared to what I see in

45:19China. Um

45:21it just it just strikes me that

45:23there obviously will be a physical

45:25manifestation of all this in the world

45:28and and the US just seems like

45:30extraordinarily disadvantaged to to win

45:32that to me.

45:34Well, I think of people who are working

45:36incredibly hard to try to change that.

45:37I'd say Sam is

45:39Like on robotics and well, I know

45:42through through all the investments in

45:43different hardware companies or

45:45No, no, we're trying to figure out how

45:50Robots are an obvious part of this and

45:52we're trying to figure out how to be

45:53very successful at robotics. The I think

45:56if you could pick one thing to make the

45:58US competitive at manufacturing and the

46:00world of atoms in general, you would say

46:02we need a lot of robots that can build a

46:04lot lot more robots. But like we can't

46:06even make a actuator. You know, like

46:09going to figure that out. We we will

46:11Open AI.

46:12Okay. Like through your robotics? It

46:14would have to. We have to.

46:16Uh the

46:18But there's other parts of this that

46:19really matter. If you think about like

46:21the number of gigawatts of power that

46:23the world is going to need just to

46:24support AI workloads, the number of

46:26chips we're going to have to fabricate,

46:28the the boring stuff about like how

46:30we're going to just get enough racks

46:32assembled and enough network cables

46:34strung and made in the first place. Uh

46:36the US is extremely behind here. I think

46:39robotics will be the solution if you

46:40just look at how fast this has to happen

46:42and what the US is currently good at and

46:45how much infrastructure we have to build

46:46up and how long that would take to do by

46:48hand. Um but

46:52you are right in the diagnosis and the

46:54criticality of it and

46:57I think luckily we have like a new

47:00uh

47:01like a new piece on the chessboard.

47:04Through Open AI or the United States?

47:08>> Okay, tell me I mean what do you mean?

47:10I I I have a different feeling. I I I

47:13feel like we're play acting

47:14hardware and we're going to get

47:16completely If host. Well, on the current

47:18trajectory, yes. No, I we totally agree

47:20with that.

47:21Uh

47:22the If we can make true general purpose

47:24robots, if we can have a

47:27Codex equivalent power thing for robots

47:31that you can say like go configure a

47:32factory this way and make me more of

47:34this kind of robot or go figure out a

47:36mine and refine this kind of thing.

47:38>> then the playing field changes. Okay,

47:40but you said the US and not Open AI

47:42specifically. But yeah, I feel like

47:44you're pointing at something that Did

47:45you know though? Well, I don't think the

47:47US has a credible plan other than this

47:49kind of AI plus robotics to catch up

47:51fast enough. I didn't know what the

47:53chess piece was. Did something You saw

47:55something recently.

47:56>> Oh, I just I mean I meant general

47:57purpose artificial intelligence.

47:59>> I feel like there's there's a real

48:00chicken and egg that's existed in

48:02robotics where

48:04if you don't have the robotic hardware,

48:06it's kind of hard to develop the

48:07software. And if you don't have the

48:08great brain, it's kind of hard to be

48:09motivated to develop the hardware. And

48:11we really saw this. We had a robotics

48:12project back in 2018. Do you remember

48:14the robotic hand? Yeah, yeah. Super

48:16cool.

48:17The thing about that hand to know, so we

48:18trained it through reinforcement

48:19learning. Actually, the exact same

48:20algorithm we used to solve competitive

48:22video games, right? Wild. Single piece

48:24of technology doing both. The one

48:26difference is that this hand would run

48:28for 20 hours before it had these strings

48:30as its tendons, and they would snap. And

48:32then you'd have downtime, mechanical

48:34engineer would come in and fix it, you

48:35know, maybe be wake them up, have them

48:37come into the office.

48:38And you realize that you cannot

48:40do ML that way. You just cannot. So,

48:43actually, we ended up canceling that

48:44project. That team went on to go work on

48:46what became GitHub Copilot, you know, to

48:48contribute to that effort. And so, you

48:49just realize how much faster software

48:51world has been moving. But I think we're

48:52at a point now where we have these

48:55amazing general-purpose algorithms that

48:56can be used in all sorts of different

48:57ways. And we see how if you start to be

48:59able to apply those to physical world,

49:02then I think you are going to have a

49:03very different story for how the

49:05hardware development and how the

49:07hardware tuning to the co-design is

49:10going to happen. So, I think that

49:11there's like real potential for change,

49:13but we as a nation need to have the

49:15willpower to really do this. But I think

49:17we really agree with you that without

49:19something like that, the current

49:20trajectory looks terrible. We're so

49:24He's convinced that it's really bad.

49:25It's pretty bad.

49:26It's really bad.

49:27>> Taking a huge step back, you know,

49:28robotics, models, these are things that

49:31you guys are interested in building and

49:32really care about, but you guys have

49:34recently, it's been reported that you

49:35consolidated and focused, and I'm really

49:38curious, I'm sure like the audience

49:39wants to know, like, what's on the table

49:42now? What got cut? What do you guys care

49:44about? And why did you make those cuts?

49:45Why was it important? So, before I let

49:48Greg answer, before he does, Greg has

49:49taken over really figuring out what our

49:51cohesive product offering and the

49:53research to support that is going to be.

49:55Um and it's been amazingly uh joyful

49:57inside of the company. Like I it's going

49:59to take a little bit longer for all the

50:00stuff to ship. He's only been in the

50:01role for a few weeks, maybe, something

50:03like that. But uh it the the energy and

50:06excitement and the sort of like

50:07enthusiasm about what Greg is doing here

50:09is unbelievable. So, you can say what

50:11it's going to be. Dude, like can you

50:12expand on so like a few weeks ago you

50:14came in and because some of these cuts

50:16were already taking place, right? I

50:17mean, so but you've come in and like

50:19assessed everything. Uh so I've I've

50:22always been very involved behind the

50:23scenes with, you know, many parts of

50:25OpenAI and so I think that kind of

50:27taking a a foreground uh role is

50:30relatively recent uh in this particular

50:32area. Although, fun fact is I actually

50:34built the very first version of the API.

50:36So, like I've been doing products since

50:38product existed at OpenAI and always

50:40deeply cared about it. And

50:43there's a bunch of things I could say

50:43there. Like I think that how core it is

50:45to our mission uh was something we

50:46didn't appreciate before we started

50:48building products and afterwards you

50:49realize how important it is. So, that's

50:50why I've always been so close to it.

50:53So, the

50:55place that we're at now is that we are

50:57clearly at a moment of transition to

51:00agents.

51:02No question, right? People in software

51:03engineering you've been feeling this

51:05for, let's say, the past 6 months.

51:07And the you know, at the course of 2025

51:10I think there was a transition from

51:11yeah, it's like kind of autocomplete to

51:13okay, yeah, you have kind of a sidebar

51:14in your editor and then you'll start

51:16talking mostly over there, but you're

51:17still mostly doing the same kind of

51:19software development you were doing

51:20before to now it's like actually you

51:22want a tool like Codex that is really an

51:25agent management platform. And the

51:26agents are going to take care of all the

51:27details. The agents are going to do all

51:29the the nitty-gritty work. And there's

51:31still probably 20% of how you put things

51:33together and the the way that you will

51:35structure your code and exactly some of

51:37these these higher-level things that you

51:39would normally put in an architecture

51:40document that the human still really

51:41cares about and wants to manage, but the

51:43details of exactly the code, like nope,

51:45that's what agents are for.

51:47And so, the

51:49question that we had is, first of all,

51:50how do we really rise to this moment?

51:52Cuz it's not just software, right? We

51:54see line of sight for every single

51:56vertical, right? For law, for finance.

52:01Some of the mechanical skills there of

52:02writing, you know, creating spreadsheets

52:03and presentations. How do we make our

52:05models extremely good at those, right?

52:07It's like you work with domain experts,

52:08you produce evaluations, you produce

52:10training data, you have the AI actually

52:12take its great domain knowledge and

52:14apply it in these verticals and gets

52:16experience and get those, you know, yep,

52:18you did a good job in order to figure

52:19out what what good looks like. So, we

52:22have the mechanical the the vision of

52:25exactly how to do this, but we need to

52:27make sure that we're building the right

52:28product surface to unlock all of it. And

52:31one thing that we have found is that the

52:33models have shifted from being the

52:35product to being a part of the product,

52:37right? They used to have these very thin

52:38layers of software on top of them, and

52:40you didn't have to like think that hard

52:41about how it was architected. But now,

52:43it's a very fat layer, right? That you

52:45have things like skills, connectors, you

52:47have exactly how you hook up to computer

52:50use, how you manage context and memory,

52:51and all of these things. And so, there's

52:53just this deep layer of software, which

52:55is almost you could think of it as the

52:57AI. It's kind of like we have this brain

52:59in the form of the model, and now we're

53:00building the body. Both are hard. They

53:01have to be co-designed together. And so,

53:03a lot of what we're focusing on is,

53:04number one, getting together an amazing

53:06agentic platform. Like, that is the

53:09number one focus that we are delivering

53:10on. We have teams that are executing

53:12extremely well on this. I'm super

53:14excited about what we'll be releasing

53:15over upcoming weeks. Um the second is,

53:18where do you actually want to apply

53:19these agents? And that our priority

53:21there is really towards computer work,

53:23right? So, that And I I use that term

53:24very deliberately, by the way. People

53:26like to talk about knowledge work, but

53:27no one thinks of themselves as a

53:28knowledge worker, right? Like, that's

53:29not a thing that people do. It's kind of

53:31a term that's like almost removed from

53:33what the actual thing is. But the thing

53:35I like about computer work is it's like,

53:36I don't really want to do computer work.

53:38That doesn't sound like the thing that I

53:39want. But you realize how much of your

53:40time you do spend doing it, like chain

53:42behind your desk, you know, typing away,

53:44hunching over your shoulders, getting

53:45your carpal tunnel, all of those things.

53:47And so, we are focusing on that,

53:50bringing

53:51Codex that exists today, not just for

53:53software engineers, but really making

53:54Codex be for everyone. And that's

53:56something that's coming very quickly.

53:58We'll have some updates even coming

53:59today as of this podcast filming. And

54:02there's a lot of exciting things that

54:03are still in the pipeline for that

54:04direction. And then the third thing is

54:07really thinking about personal AGI,

54:08which is about Thinking about ChatGPT

54:11right now, used by a billion users. And

54:13every single person on the planet is

54:14going to want an AI that represents

54:17them, has their context, that they have

54:19built trust with, that it's not just

54:21something that you

54:22build, or it's not just something that

54:24you talk to one-on-one, but it can be

54:25out there doing things for you. For

54:27example, maybe it knows that you like a

54:28specific musician, and that musician's

54:30in town, and it knows it proactively,

54:32and tickets have just become available.

54:34There's some great tickets that are very

54:35cheap that you you get for you. It just

54:36goes and buys them, right? And maybe it

54:38knows it's built trust with you, so it

54:40kind of knows you're allowed to to do

54:41this without getting approval, or maybe

54:42it realizes, "I should probably I'm not

54:44quite sure. I should check." And so,

54:46we're building that as well.

54:49And if you think about these things,

54:51they all are kind of expressions of

54:53something that fits together into

54:54cohesive whole. Like in the end, you

54:56fast forward to where we're going, you

54:58really just want

55:00an AGI, right? You don't want a language

55:02model, you don't want threads, you don't

55:03want any of these like details. You just

55:05want something that is helping you, that

55:07is operating on your behalf, that is

55:09able to help you solve problems, that

55:10knows what your goals are and achieve

55:12those in work context, personal context.

55:14And so, this is what we are prioritizing

55:16and building. Yeah. And so, you know, to

55:18me one of the important questions, you

55:20know, people are asking, "How do we

55:22think about consumer? How do we think

55:23about enterprise?" And the answer is,

55:25if you take the definitions of these

55:27words as they exist today, we care a lot

55:29about consumer, we care a lot about

55:30enterprise. But I think that the meaning

55:32of these words will change and blur,

55:34because what we're doing is we are going

55:37to So, we're going to unlock this wave

55:38of entrepreneurship. Again, I I think

55:40we're seeing the leading edges of it.

55:41Small companies be able to get tons and

55:43tons of revenue that was not possible

55:44before. And that's been a trend for a

55:46while.

55:47It's just going to really accelerate. Is

55:49that enterprise?

55:50Is that consumer? It's kind of neither,

55:53right? And so, I think we are really

55:54focused on solving goals across all

55:57context. And that is the lens that we

55:59that we look at things through. And so,

56:01that has mean that has that has meant we

56:03need to deprioritize other things that

56:05are also amazing on their own.

56:07>> So, what got cut?

56:08>> Yeah.

56:09>> Well, So, Sora is the most Sora is the

56:10most obvious one.

56:11>> And why? Well, because it's a different

56:14branch of the tech tree, right? So, if

56:15you look at the models that actually

56:17power Sora, that they're not unified

56:19with the core GPT series. And secondly,

56:22the use case is not quite unified

56:25either, right? It doesn't fall as far

56:27under this goal like there's creative

56:29expression. There's something very

56:30important there. I think it was like an

56:32incredible model. The team does

56:33incredible work. And I think that

56:35technology will live on for other

56:36applications. But what we were really

56:38focusing on was this the product suite

56:41that we want to be delivering. What do

56:42we want to be doing over the next 3

56:44months, 6 months, 12 months? And the

56:46thing I described, by the way, is just

56:48step one. Because we also see line of

56:50sight to much more powerful models.

56:52Like, you look at what we're doing right

56:54now in mathematics. It's Okay, it's

56:55actually kind of mind-blowing. Where

56:58this result that I just mentioned of

57:00solving the new Erdős problem that seems

57:02actually really significant.

57:04That was just someone using GPT-4 Pro.

57:07Like, 2 years ago

57:08>> Yeah, we used to like train our model We

57:10had a team of 20 people to try to train

57:12our models to go solve a computing

57:14Olympiad, and we got a bronze medal. A

57:17team of 20 people for like 2 weeks and

57:18lots of compute. And now it's just this

57:20model that we trained very casually.

57:21Someone is able to just point it at

57:23problems and get this kind of result.

57:24What if you point that at drug

57:27discovery? What if you do take that team

57:29of 20 people and all that compute and

57:30And really try to push it for scientific

57:32discovery. And that's something no one

57:34is pricing in right now. So, I think

57:36it's rising to the moment of agents

57:38really trying to make sure that the

57:40product investments we're making are

57:42sort of well-structured, that we think

57:44about how all these pieces fit together,

57:46that we have connectors that work really

57:47well, and each of these pieces can be

57:49composed, and really also build an

57:51ecosystem because it's not just about

57:53what we built, right? That we want to

57:54build some example agents. You think of

57:56Codex almost as an example agent, but it

57:58should be that if you're a developer, if

58:00you're someone with a creative idea, you

58:02can build your own agent, right? That

58:03you can build it for your application,

58:05for your purpose. You care about the

58:06specific math problem, you should be

58:07able to apply the agent to your that

58:09math problem. And so, we're we're

58:10enabling all of that. But you I mean,

58:12okay, just bear with me for 1 second

58:14just to go quick through a couple things

58:16around this. I mean, so we I mean, part

58:18of this was like you had to

58:20probably cut Sora for compute to get

58:22compute, right? So,

58:23that compute.

58:24>> reported that it was taking up a lot of

58:26compute. I mean, every everything in

58:28this field that is successful is going

58:29to take a compute. I mean, the two of

58:31you in particular, and I think Ilya and

58:33some others, you know, kind of famous

58:34for going all in on where you were going

58:37to direct the compute in the early days,

58:39and now you have to make these very

58:40difficult decisions. You have to serve

58:43all these customers that you have. You

58:44got to make some money cuz you're

58:46spending a lot of money, and you know,

58:48so you two are both wired to take the

58:51very biggest bet possible, but you're

58:54constrained now by the business into

58:57some degree. So,

58:59that just seems

59:01very difficult. I feel like it's not in

59:02your nature. Does it You're looking

59:04skeptical.

59:05It's a It's a strange phrasing because I

59:07don't feel constrained by the business.

59:09I feel enabled by the business because

59:11the business is what has really allowed

59:12us to say we can scale compute, right? I

59:15remember when we launched ChatGPT, and I

59:17remember we were talking right

59:18afterwards, and we're trying to figure

59:20out how much compute to buy. And

59:22I was just like, we got to buy it all.

59:24We just got to we got to do it cuz it's

59:25just so clear there's so much demand.

59:27And think that that has been a huge

59:28unlock in our ability to get lots of

59:30compute for mission critical

59:33>> this incredible revenue machine, we

59:34would not be able to

59:36convince anyone that we should get all

59:38this compute.

59:38>> Get get this. But then the impression

59:40obviously when you see the stories about

59:41Stargate and things like that is that

59:43you guys are somehow pulling back on

59:45infrastructure. I don't know where

59:46that's coming from. Like there you know,

59:48there will be like a site here and there

59:50we say okay, you know what this

59:51particular site maybe it only has air

59:52cooling and so this is not as valuable

59:54to us as this other site and that's like

59:56there's specifics and people really want

59:57to write the story of like pulling back.

59:59But very soon it will be again like open

1:00:01AI is so reckless. How can they be

1:00:02spending this crazy amount? So like the

1:00:05media will go

1:00:06will flip out either way just cuz you

1:00:08all need something to write about I

1:00:09guess. Um but we will keep building out

1:00:12as much compute as we possibly can. One

1:00:14one one one thing that I think is also

1:00:15worth thinking about is that compute for

1:00:19us is

1:00:21not a cost center. It's a profit center,

1:00:23right? When you deploy it in the

1:00:24product. And so in many ways our

1:00:25business is extremely simple, right? We

1:00:28rent or buy compute and then we resell

1:00:31it at a margin.

1:00:32And as long as we have some positive

1:00:33margin on it, then it's scalable, right?

1:00:36Because the demand is just unlimited. So

1:00:38okay, so data center hardware still

1:00:41all systems go?

1:00:42Or you mean like our own chip? Yeah, own

1:00:44chip, networking Yeah, very very excited

1:00:47about our our chip.

1:00:49We have an incredible team there.

1:00:51>> One Titan one? Uh no, we're not talking

1:00:54about timelines of course, but I'll just

1:00:56say that like I spend a lot of time with

1:00:58that team and I think it's so So they're

1:01:00doing great. So that's our all systems

1:01:01go. Robotics sounds like it's all

1:01:03systems

1:01:04>> to be a while till we have something to

1:01:05where you're like ah this is the chat

1:01:07should be team moment, but But but you

1:01:08haven't pulled back on that program.

1:01:10Social net social network

1:01:12Robots are not the current thing, but

1:01:15are going to be so clearly important in

1:01:17the future.

1:01:17>> Okay. And social network? Not doing

1:01:20that.

1:01:20>> Not not doing that right now.

1:01:21And then clearly like the super app, the

1:01:24browser, all that stuff still still

1:01:25going.

1:01:25>> Yeah, and one thing to realize about

1:01:27super app because I feel like it's one

1:01:28of these things where it's like a catchy

1:01:29word, but I thought you guys like came

1:01:32up with this. I mean but You guys used

1:01:34this like internal short Exactly. This

1:01:36is the thing to realize is that like

1:01:37sometimes we're communicating to to our

1:01:39team and then of course it ends up being

1:01:40communications the world and these are

1:01:42not intended to be the same thing. Super

1:01:44app in many ways is an iceberg, right?

1:01:46It's like yes, we're going to have a an

1:01:47app and you'll see updates to what the

1:01:49Codex app is today to make it so it's

1:01:51Codex is for everyone and I think that

1:01:53ultimately like what that becomes that

1:01:55we have we have a lot of steps to get to

1:01:56where we want to be, but it's really

1:01:58about saying we're building this unified

1:02:00identity infrastructure that I described

1:02:02and that is I think going to be a huge

1:02:04unlock for every single thing that

1:02:05people want to do. Okay. Um

1:02:08You and I have talked a little bit about

1:02:09yeah,

1:02:11you guys have had lots of drama. It's

1:02:12slowed you down at times. You still have

1:02:16drama sadly. Um there's a lawsuit

1:02:18coming. Uh It'll be fun. Yes, yes. Well,

1:02:21let's talk about that in a second. Um

1:02:24Like which company do you think has

1:02:25actually executed better over the last

1:02:29two years? Open AI or Anthropic?

1:02:33Look, I think that this is a hard thing

1:02:36to say from just the current moment,

1:02:39right? Because I think that the

1:02:42view in my mind is you got to step back

1:02:44and really think about what is it that

1:02:46we're all here to do, right? And I think

1:02:48that from the Open AI perspective, the

1:02:50things we've been talking about in some

1:02:51ways are about it's very clear that yes,

1:02:53selling to enterprises like figuring out

1:02:55how to really deliver these coding tools

1:02:57and it's been really it's been actually

1:03:00I think competition can really help

1:03:02elevate your own thinking and realize

1:03:04that hey, we need to focus on this. And

1:03:05one example in coding was that I think

1:03:07we got late to the game of not just

1:03:09building models that were good in the

1:03:11abstract of coding. Like we always had

1:03:13the best numbers on programming

1:03:14competitions, but you also need to apply

1:03:16them to messy repos, real-world data,

1:03:18those kinds of things. And that was

1:03:19something that I think we had

1:03:20appreciated later than Anthropic did.

1:03:22So, I think that's That's kudos to them,

1:03:25but also something that's helped elevate

1:03:26our own execution. And now,

1:03:28head-to-head, Codex versus Claude, I

1:03:30think that that that we get very

1:03:31favorable results.

1:03:33Um and I think we've done an incredible

1:03:35Our teams have done an incredible job

1:03:36across the whole company to really build

1:03:37a product that's not just competitive,

1:03:39but actually ahead in many, many ways.

1:03:41Um but, I think that the the core it's

1:03:45never about the ups and downs of the

1:03:47news cycle. It's about progress towards

1:03:50AGI, towards it benefiting everyone. And

1:03:53that is something that I think we've

1:03:53been extremely focused on. The team's

1:03:55executing extremely well. And so, I just

1:03:58want to say that you can't always tell

1:04:00from the outside,

1:04:02but if you think about the things we

1:04:03talked about, that there is this focus,

1:04:05but on many timescales that all add up

1:04:07to where we need to go. Yeah, there's a

1:04:10million questions I want to ask with the

1:04:11time we have left. But, one of them we

1:04:13we touched on this, like making sure a

1:04:15bunch of models are in the hands of a

1:04:17bunch of people and they have that

1:04:18powerful technology. But, we've reached

1:04:20a point where now some of these models

1:04:21are too powerful, we're being told, and

1:04:23they're they're gated for only certain

1:04:25companies.

1:04:26Claude and Mythos have really made a lot

1:04:29of headlines and I think a lot of a lot

1:04:31more fear. And I'm wondering what you

1:04:34guys think of this new moment. Are we

1:04:35getting to a point where like we need to

1:04:37start keeping these powerful models

1:04:39behind the scenes rather than in the

1:04:40hands of everybody?

1:04:42There are people in the world who

1:04:44for a long time have wanted to keep AI

1:04:46in the hands of a smaller group of

1:04:48people. Um

1:04:50you can justify that in a lot of

1:04:51different ways. And some of it's real,

1:04:53like there are going to be legitimate

1:04:54safety concerns.

1:04:56Um but, I expect But, if what you want

1:04:59is like, we need control of AI just us

1:05:01cuz we're the trustworthy people, I

1:05:02think the the fear-based marketing is

1:05:04probably the most effective way to

1:05:05justify that. Um that doesn't mean it's

1:05:08not legitimate in some cases. Uh but, it

1:05:12is

1:05:13you know, clearly incredible marketing

1:05:15to say, "We have built a bomb. We are

1:05:17about to drop it on your head. We will

1:05:18sell you a bomb shelter for $100

1:05:19million. You need to like run across all

1:05:21your stuff, but only if we like pick you

1:05:22as a customer." And

1:05:24it

1:05:25the way that we view balancing these new

1:05:29capabilities that are going to come with

1:05:32uh

1:05:33still our

1:05:35our belief that the world needs to get

1:05:37and use and understand and come up with

1:05:39new ideas for this technology is not

1:05:41always easy. We We have had

1:05:43um cybersecurity in our preparedness

1:05:45framework for a long time, and we have

1:05:47been building mitigations to figure out

1:05:48how we release this,

1:05:50um how we put these models in the hands

1:05:51of a trusted access program, how how we

1:05:55then put more capable models in the

1:05:57hands of everybody. Um

1:06:00but, there will be a lot more rhetoric

1:06:02about models that are too dangerous to

1:06:04release. There will also be very

1:06:05dangerous models that will have to be

1:06:06released in different ways.

1:06:08But, to the point Greg was making about

1:06:12the goal here is to benefit everybody.

1:06:16And also to

1:06:19I don't want to say market this in a

1:06:20way, but but like get the world to come

1:06:22along on this journey with us where

1:06:24where it's like we are going to give you

1:06:25more powerful technology. There's going

1:06:27to be responsibility that goes along

1:06:28with that. We are going to help set up

1:06:29the world for success as much as we can.

1:06:32Um

1:06:33but,

1:06:34we will try to avoid the fear-based

1:06:36marketing as much as we can. So, I'll

1:06:38I'll just ask you directly. Do you think

1:06:39Mythos is just a lot of marketing and

1:06:41not I'm sure it's a great cybersecurity

1:06:43like I We've been talking about this for

1:06:44a long time. Like this has been in our

1:06:46in our

1:06:47model, but there's there's a way of

1:06:48saying like like our version of this is

1:06:50to say

1:06:51these models are going to get much

1:06:52better at cyber. We have this

1:06:53preparedness framework category. Here's

1:06:55our plan for how we deploy this into the

1:06:56world. Here's how this trusted access

1:06:58program looks like. Here's the

1:06:59mitigations we put on the models.

1:07:01Um

1:07:02I think I'm talking about how it is not

1:07:03how it is one of the categories in their

1:07:05preparedness framework.

1:07:06Uh so, I'm sure Mythos is a great model

1:07:08uh for cybersecurity.

1:07:10But, I think we have a plan we feel good

1:07:12about for how we put this kind of

1:07:14capability out into the world. When the

1:07:16Anthropic stuff was going down with the

1:07:18Department of War, I mean, from my

1:07:20perspective, it looked like

1:07:22you had sort of David Sacks and people

1:07:25with long ties to Elon. I could see

1:07:28putting pressure to make some of these

1:07:29things happen and have the government

1:07:32focus it on that.

1:07:33You guys kind of came in pretty quick

1:07:35after and and had an announcement of

1:07:37your own that was more favorable.

1:07:39Um

1:07:41I mean, you know, Elon's

1:07:44aggressive against you guys as well.

1:07:47And some of this is funny to me that I

1:07:49mean, there is a world where like you

1:07:50guys and Anthropic are actually on this

1:07:53this other side and and it feels like

1:07:55Elon and even

1:07:57Zach to some degree, my understanding,

1:07:59kind of

1:08:00pressuring things. As a side, yeah, I

1:08:03mean, did you did you feel like

1:08:04Anthropic was treated fairly in that? Do

1:08:07you feel um No.

1:08:09>> [snorts]

1:08:09>> Um I don't

1:08:13Well, I think there was like a lot of

1:08:14bad behavior to go around there, but I

1:08:15don't think Anthropic was treated well.

1:08:17Can you explain more? Like what was

1:08:18standing out to you about what was

1:08:20particularly wrong? Um

1:08:25I don't like look with the

1:08:28things that have happened since with,

1:08:30you know, these models reaching this

1:08:32cybersecurity threshold that is clearly

1:08:34in the national security interest. I I

1:08:36think it all has a little bit of a

1:08:37different flavor to it, but

1:08:40you know, like threats of

1:08:43using the DPA and actually using supply

1:08:46chain risk designation, like this is

1:08:49this is not the relationship that I

1:08:50think our government and our AI efforts

1:08:53need to have. Uh

1:08:56We really care about supporting the US

1:08:58government. I think that's going to

1:08:59become increasingly important as these

1:09:01models get more capable.

1:09:03And I certainly don't think it's like a

1:09:05good stance for

1:09:07uh labs to say, you know, we have the

1:09:08super weapon by the way we're not going

1:09:10to work with you to help help you defend

1:09:11the country.

1:09:12Um but I also don't think it's good for

1:09:14the government to be like

1:09:18fighting this stuff out in the press and

1:09:19using the big hammer that the government

1:09:21has that needs to be used very rarely.

1:09:23So,

1:09:23I

1:09:26A thing about OpenAI is we generally try

1:09:27to be like moderate and centrist and

1:09:29reasonable. And that is what we've tried

1:09:32to do here with the US government, but I

1:09:34see no future no good future where

1:09:38leading AI efforts don't assist the US

1:09:40government. Uh if everything we say is

1:09:42right, we believe the things we're

1:09:44saying about where the models are going

1:09:45to go, which I certainly do, then the

1:09:48government needs our help and we are

1:09:51honored to get to provide it. You know,

1:09:52when I am old as a tech reporter, I

1:09:54mean, when I was first

1:09:56writing about tech, you know, the big

1:09:57war was it was basically Microsoft

1:09:59versus everybody else and it was they

1:10:00were the big evil um proprietary

1:10:03software company. You had the Yeah. open

1:10:05source companies counter. You had Mark

1:10:06Andreessen with Netscape as you know,

1:10:08and and you know, like things would play

1:10:10out in the press and it would be a

1:10:11little vitriol here and there, but you

1:10:14know, and then people had their

1:10:16philosophical and religious camps.

1:10:18Obviously, like this stuff where we are

1:10:20in AI land is insane. Um

1:10:23that so much of this seems tied to the

1:10:25personalities of

1:10:27you, Elon, um

1:10:30Zuck, Dario, and and Demis and the

1:10:32various animosities and world views.

1:10:34Like

1:10:35you wrote about this the Shakespearean

1:10:38drama of it. Like how do we possibly get

1:10:39past this? I mean, you guys are all so

1:10:42um I feel

1:10:43I don't see a path out and and obviously

1:10:46like we mentioned there's a lawsuit

1:10:47coming that will only Wouldn't all the

1:10:49hands on stage? I'm so sorry to say.

1:10:51>> Some some of the people involved only

1:10:56trust themselves to get it right. And

1:10:59because they think the stakes are

1:11:01infinite and because they don't, you

1:11:02know, for different reasons, uh they

1:11:04don't think anybody else can

1:11:07do it right or they don't want anybody

1:11:09else to do it. Um that leads to

1:11:13I think some very toxic behavior. We

1:11:15can't control the behavior of other

1:11:16people. Uh but we will continue to

1:11:19advocate for doing this as a

1:11:22collective project that humanity has to

1:11:25get right together and not something

1:11:26where it should be about

1:11:28one person or one ideology winning or

1:11:30losing or you know, having a

1:11:34sole victory or something like that. I

1:11:36mean, we'd be remiss and I know it's

1:11:38super sensitive, but I mean, it

1:11:40you know, if you look at what happened

1:11:41to you personally over the last week,

1:11:44there's tons of science fiction books

1:11:46that have been written about what

1:11:47happens when

1:11:49um

1:11:50the factions that really don't want to

1:11:51see technological progress happen get um

1:11:54you know, things get more extreme. Um

1:11:58I don't know. Like there's an argument

1:11:59to be made that this this switch is now

1:12:02flipping.

1:12:03Along those lines. With people that want

1:12:05to like stop. Yeah, I mean, you know,

1:12:06I've I've read Richard Clark's book

1:12:09years ago where a Bill Joy-like figure

1:12:11decides AI can't happen and he's running

1:12:13around destroying data center. You know,

1:12:15this stuff has been out there forever.

1:12:16It just feels like, God, I mean, I don't

1:12:18see how things could get like more

1:12:19heightened and so

1:12:21um it doesn't seem like it's getting

1:12:22better.

1:12:23I I assume it will go up and down, but

1:12:25directionally more heightened.

1:12:27This must be like you wrote about. I

1:12:29mean, clearly this must be horrifying.

1:12:30Horrifying. Yeah, yeah, I mean, I don't

1:12:32really have anything like deep to say

1:12:33here.

1:12:36That was a crazy way to wake up. Um

1:12:40The the first day I was sort of in this

1:12:42light kind of adrenaline shock about it

1:12:44and just trying to like figure out

1:12:45logistics.

1:12:47And then the day after I was just like,

1:12:49you know,

1:12:50there's going to be more stuff like this

1:12:51and it's incredibly disheartening and I

1:12:53went through a real depressive cycle

1:12:54about it. Um, but it's very scary and I

1:13:04Yeah, I don't think I have anything

1:13:05super deep to say. I think the doomerism

1:13:07talk hasn't helped.

1:13:09Uh,

1:13:10I think the way certain other labs talk

1:13:12about us hasn't helped. I actually don't

1:13:14want to make like a a side statement. I

1:13:15think the way Anthropic talks about

1:13:16OpenAI doesn't help. Um,

1:13:19and

1:13:22you know, I I I like

1:13:24I hope that cooler times will prevail.

1:13:26One One thing I do want to say, just to

1:13:30you know, the point of we can't control

1:13:31other people, we can control ourselves.

1:13:32Like one thing I I've been just really

1:13:35impressed by and astounded by with Sam

1:13:37is just how throughout this time, like

1:13:39that very day, he was there doing things

1:13:43that absolutely required him. Like he

1:13:45just like keeps pushing on the mission.

1:13:48And

1:13:49I think that is something that I don't

1:13:51take for granted at all. Like I just

1:13:52think that the degree of resilience that

1:13:54is represented by Sam is is extreme and

1:13:56I think very underappreciated.

1:13:58The I mean

1:14:02I've seen the same thing. I don't know

1:14:03what people will be like, oh, you're

1:14:04taking it easy or something like that. I

1:14:05don't know. I I am

1:14:07It's hard to think of a like another

1:14:09individual over the last 3 years who's

1:14:11been through more dramatic business and

1:14:14and personal good

1:14:15>> The narrative has been quite dramatic.

1:14:17>> be sympathetic cuz they're like you're

1:14:18rich.

1:14:19>> be sympathetic cuz they're like And

1:14:20that's fine. Um,

1:14:22I I I said this the night before

1:14:25that this thing happened at my house.

1:14:27Um, I had some people over for dinner

1:14:30and we were talking like uh some work

1:14:31colleagues and we were talking about the

1:14:32next phase and they're like, uh, it's

1:14:34been like a you know, a brutal time for

1:14:35you in the press, but I I guess it's

1:14:36like good in some sense and I was like,

1:14:38uh, you know, at least no one tried to

1:14:39kill me.

1:14:40Um, You said that the night before?

1:14:42Uh, and

1:14:46and it but it really did put into

1:14:48perspective that like, you know, people

1:14:50can say all the mean things they want

1:14:51and as long as they only say the mean

1:14:53things, it's like it's really not that

1:14:55big of a deal. You keep going.

1:14:56>> You always told me your dream is to like

1:14:59retire out in that someday. So like why

1:15:01not We got a lot of work left. But like

1:15:03somebody's going to get It's clear

1:15:05somebody's going to get to AGI now. We

1:15:06got like five companies that could

1:15:08reasonably do this.

1:15:10>> Which company would you most like to get

1:15:11there first? Is that Is that what it

1:15:13comes down to?

1:15:13>> cuz look, I think that the point cuz the

1:15:16the missing thing is this point of how

1:15:19do we help people really understand what

1:15:21it is that this technology can do for

1:15:23them.

1:15:24And I think that's something that every

1:15:25company can contribute to, right? That I

1:15:27think is something we have a perspective

1:15:29on. We've talked a lot about it in this

1:15:30podcast. And fundamentally, I think that

1:15:32that is

1:15:34something that is kind of the

1:15:36responsibility and on all the people who

1:15:38are trying to create this technology is

1:15:40also to show its benefits and why people

1:15:42should want it. Why should people be

1:15:44protecting and defending the ability to

1:15:46create it? Like why does America need to

1:15:48have leadership here? Like why is all

1:15:49this good for not just a country, but

1:15:52you personally, your future, the future

1:15:54of your kids? And that's something that

1:15:55we wake up thinking about like like I

1:15:58would say every day. Like it I don't

1:16:00think it's an overstatement. Like this

1:16:01is something we talk about constantly.

1:16:03We think about

1:16:04whether it's weird creative, you know,

1:16:06legal structures for a company. We've

1:16:07done that many times

1:16:09because we're trying to solve for this

1:16:11mission, for this thing that we think is

1:16:13so important. And that's something where

1:16:14if other people want to contribute to

1:16:16that, then more power to them. Like

1:16:17that's something we should all be doing,

1:16:18but it's something that drives us

1:16:20uniquely. We so believe that if we can

1:16:23deliver technology that enables

1:16:24prosperity for everybody, if we can give

1:16:26people more agency over the future,

1:16:28uh that will

1:16:30through some fits and starts lead to a

1:16:32better world. I don't think all

1:16:35people working in the field believe

1:16:36that, but anyone who does, we are

1:16:38delighted to work with and that is the

1:16:40the that, you know, we we want to move

1:16:41the world towards. Go Going back Well, I

1:16:44think we have time for like a couple

1:16:45questions.

1:16:46>> Actually, I have wish list questions.

1:16:47>> I've got Can I'm going to I'm going to

1:16:49give you two answers real fast.

1:16:50>> I actually want to end on the one that

1:16:51I'm about to ask. So, I'll do this one

1:16:53and then [laughter]

1:16:55All right, we're ready. We're ready. I

1:16:56mean, how you

1:16:59There is like a

1:17:00How existential do you view this trial?

1:17:04I I actually think it's a real

1:17:06opportunity for us to tell our story.

1:17:08Because if you look at the course of

1:17:10OpenAI, we have really let

1:17:13the other sides, when there's splits,

1:17:15tell the story. And we really have done

1:17:17our best not to sort of say, well, well,

1:17:19that's not really what happened. Like,

1:17:21let's talk about the truth. And the In

1:17:24this case,

1:17:26we finally have no choice, right?

1:17:27Because we have to defend ourselves. We

1:17:29have to tell the truth. We have to tell

1:17:30what happened. And I'm extremely proud.

1:17:32Like, I've spent a lot of time looking

1:17:34back at the history at a bunch of

1:17:35different messages. And of course,

1:17:37there's always things that you can like

1:17:38try to like cherry-pick, be like, "Aha,

1:17:40you said this thing."

1:17:40>> Your diaries are are famous. I know,

1:17:43right? But the thing is, that they're

1:17:44not because And first of all, incredibly

1:17:46personal documents. Extremely painful to

1:17:48have something so personal be sort of

1:17:50taken from you and then, you know,

1:17:52attempted to be weaponized. But those

1:17:54particular, you know, sentences are kind

1:17:56of the worst thing that the opposition

1:17:58could find. You're like, "Really?" And

1:18:00the thing is, they're all taken out of

1:18:01context, right? The question of, you

1:18:03know, "Hey, like, we're in the middle of

1:18:05this negotiation, right? We've all

1:18:07agreed the only path forward for OpenAI

1:18:09is a for-profit." Sam, Ilya, Greg, Elon,

1:18:12we all agreed on this. We all said this

1:18:13is what we got to do. This is literally

1:18:15the thing for the mission. And now

1:18:17you're in this crazy negotiation, right?

1:18:18Elon's like, "You need majority equity.

1:18:20You need to

1:18:21to be CEO. You need full control." And

1:18:24we got so close. It's like, "Okay, fine.

1:18:26Okay, we're not going to be equal

1:18:27partners. You need this this massive

1:18:29amount of equity. If that's what you say

1:18:31you need, like,

1:18:33we would like you to be involved. We can

1:18:34get there. Okay, fine. You know, Sam

1:18:36will be CEO. Elon will be CEO. He needs

1:18:39it so that everyone knows he's in

1:18:40charge. Fine.

1:18:41But absolute control.

1:18:44Absolute control over OpenAI. Even if

1:18:46you say, "Well, I'll dilute down. I'll

1:18:47give it up in the future." And you're

1:18:49like, "What is our mission?

1:18:51Like, do we really believe in our

1:18:52mission? Do we really care about this

1:18:53picture of

1:18:55we want this technology to benefit

1:18:56everyone, and there shouldn't be one

1:18:58person person in charge of the whole

1:18:59future.

1:19:00Doesn't matter who that person is. Like,

1:19:02that was the breaking point. That was

1:19:04the thing that caused us to say no.

1:19:06And so, we've never told that story for

1:19:08years. But now we will. And so, I I

1:19:11think that it is a real opportunity for

1:19:12people to understand what truly

1:19:14motivates us, what we truly stand for. I

1:19:16think it's insane that he's doing this,

1:19:18but I'm kind of

1:19:19my fear at this point is he decides to

1:19:21like drop the case right before the

1:19:22trial, and we don't get to do all this.

1:19:23But I am happy to like explain all this

1:19:25to the world and have this chapter

1:19:26behind us. Okay. So, my my question is

1:19:29rounding back to the beginning of this

1:19:30podcast. Like, I was talking about this

1:19:32narrative I feel like I as an AI

1:19:34reporter grew up with hearing of, you

1:19:36know,

1:19:37"We're all going to die, and we're all

1:19:38going to lose our jobs." I feel like I

1:19:40heard that time and time again, and now

1:19:43portions of your personal notebook are

1:19:44public. Um

1:19:46I'm curious, again, how would you have

1:19:48talked about this differently with

1:19:50hindsight being 20/20?

1:19:53Well, I think that the way that we think

1:19:55about it today

1:19:57is the way that I wish we had talked

1:19:59about it in some sense. But I don't know

1:20:01we could have with the knowledge we had

1:20:02at the time. For example, even some of

1:20:04it is related to the technology itself.

1:20:06We had a picture in 2017, 2018, the way

1:20:10we would build AGI was through a

1:20:12competitive multi-agent simulation.

1:20:13Like, imagine an island of like a

1:20:15thousand agents that all are in a battle

1:20:17to survive and replicate. And you could

1:20:19see that, if you put a ton of compute

1:20:21into it, maybe it would build something

1:20:22very smart. But that smart thing

1:20:25that would not be connected to human

1:20:26values at all. You'd have to have a

1:20:28separate step of figuring out, how do

1:20:29you even talk to this thing, right?

1:20:31Didn't grow up at all in the real world,

1:20:33doesn't have any concept of language,

1:20:34doesn't have any connection to our

1:20:36reality. It's just smart and powerful.

1:20:38That's a very scary system, right? You

1:20:40start from a place of

1:20:42trying to think about how could you

1:20:44possibly align it. But instead we have

1:20:46this language model route, which is

1:20:47rooted in our values, which is rooted in

1:20:50understanding humans, and we have chain

1:20:52of thought that actually you can

1:20:54monitor, that actually if you approach

1:20:56things right in a technical sense, you

1:20:57have a path to make that be faithful. So

1:20:59it really represents what is truly

1:21:01motivating the AI. And you just realize

1:21:03we have a totally different

1:21:04technological path to get to the outcome

1:21:07we're talking about, and it's a much

1:21:09more optimistic one. And so I think that

1:21:11there was some technical learnings we

1:21:13had to have for what is this technology

1:21:15truly going to be, how will it be

1:21:17created, what are the ways in which you

1:21:18make it useful. And that is something I

1:21:20think we could not have appreciated, but

1:21:22we do today.

1:21:23Um it's been a fascinating discussion.

1:21:26It's um really kind of you guys to both

1:21:28show up and do this together. Um

1:21:31we we have been a humble podcast. We're

1:21:33honored um for for you guys to spend

1:21:35this time. And I guess I mean the

1:21:38obviously we covered a lot of ground,

1:21:39but the headline is in relatively short

1:21:42order, new new models Yeah. are really

1:21:44good new models. Yeah. Um well, thank

1:21:46you guys. Thank you guys so much.

1:21:48>> Thank you for having new models useful

1:21:49for everyone.

1:21:50>> [laughter]

1:21:51>> Thank you guys. Thank you. Thank you.

1:21:56The Core Memory podcast is hosted by me,

1:21:58Ashley Vance, and or Kylie Robinson, or

1:22:01both of us together. It is produced by

1:22:04me and David Nicholson. [music]

1:22:06Our theme song is by James Mercer and

1:22:09John Sortland, and the show is edited,

1:22:12always, by the John Sortland. Thank you

1:22:16so much to Brex and Combinator Ventures

1:22:19for all [music] your support, and thank

1:22:21you most of all to everybody for

1:22:23listening or watching. We love you.

1:22:26Please leave us a like, a review, a

1:22:28subscribe, all those tremendous things.

1:22:31Thank you and we'll see you again.

1:22:33[music]