0:00

We said from the very beginning we were

0:02

going to go after AGI at a time

0:04

when in the field you weren't allowed to

0:06

say that because that just seemed impossibly crazy.

0:09

I remember a rash of criticism

0:11

for you guys at that moment. We really wanted

0:13

to push on that and

0:16

we were far less resourced than DeepMind and others

0:18

and so we said okay they're going to try

0:20

a lot of things and we've just got to

0:22

pick one and really concentrate and that's how we

0:24

can win here. Most of the world

0:26

still does not understand the value of like a

0:28

fairly extreme level of conviction on one bed. That's

0:31

why I'm so excited for startups right now is

0:33

because the world is still sleeping on all of this

0:35

to such an astonishing degree. We

0:43

have a real treat for you today. Sam

0:46

Altman, thanks for joining us. Thanks Gary.

0:48

This is actually a reboot of your series, How

0:50

to Build the Future and so welcome back to

0:53

the series that you started. That was like eight

0:55

years ago I was trying to think about that,

0:57

something like that. That's wild. Very good. That's

0:59

right. Let's talk about your newest

1:02

essay on the age of intelligence. Is

1:05

this the best time ever to be

1:07

starting a technology company? Let's

1:10

at least say it's the best time yet. Hopefully there

1:12

will be even better times in the future. I sort

1:14

of think with each successive major technological

1:16

revolution you've been able to do more than

1:18

you could before and

1:20

I would expect the companies to be

1:22

more amazing and impactful and everything else.

1:25

So yeah I think it's the best

1:27

time yet. Many companies have the edge

1:29

when things are moving slowly and

1:31

not that dynamic and then

1:33

when something like this or mobile or

1:35

the internet or semiconductor revolution happens or

1:37

probably like back in the days of

1:40

the industrial revolution that was when

1:42

upstarts have their edge. So

1:44

yeah this is like and it's been a while

1:46

since we've had one of these so this is like

1:48

pretty exciting. In the essay you actually say a

1:50

really big thing which is ASI,

1:52

super intelligence, is

1:55

actually thousands of days

1:57

away. Maybe. I mean that's our

1:59

hope I guess. I

8:00

always felt a little bit weird about that. And

8:02

then I remember one of the things I thought was so

8:04

great about YC, and still that I care so much about

8:06

YC about, is it

8:08

was like a collection of the weird people who were just like,

8:10

I'm just going to do my thing. The

8:13

part of this that does resonate as an accurate

8:16

self-identity thing is I do think you

8:18

can just do stuff or try

8:20

stuff a surprising amount of the time. And

8:24

I think more of that is a good thing. And

8:26

then I think one of the things that both

8:28

of us found at YC was a bunch of

8:30

people who all believed that you could just do

8:33

stuff. For a long time when I

8:35

was trying to figure out what made YC so special,

8:37

I thought that it was like, okay,

8:39

you have this very amazing

8:43

person telling you, you

8:46

can do stuff I believe in you. And

8:48

as a young founder, that felt so special

8:51

and inspiring, and of course it is. But

8:53

the thing that I didn't understand until much later was

8:55

it was the peer group of other people doing that.

8:59

And one of the biggest

9:01

pieces of advice I would give to young

9:03

people now is finding that peer group as

9:05

early as you can was so important to

9:07

me. And

9:10

I didn't realize it was something that mattered. I kind of

9:12

thought, ah, I'll

9:14

figure it out on my own. But man,

9:17

being around like inspiring peers was

9:20

so, so valuable. What's funny is both of us

9:22

did spend time at Stanford. I actually did graduate,

9:24

which is, I probably shouldn't have

9:26

done that, but I did. You

9:29

pursued the path of far greater

9:32

return by dropping out. But

9:34

that was a community that purportedly

9:37

had a lot of these characteristics,

9:39

but I was still beyond surprised

9:41

at how much more potent it

9:43

was with a room full of founders. I

9:45

was just going to say the same thing. I liked Stanford

9:47

a lot, but

9:51

I did not feel surrounded by people

9:54

that made me want to be better

9:56

and more ambitious and whatever else.

9:58

And to the degree I did, The thing you

10:00

were competing with your peers on was like, who

10:02

was going to get the internship at which investment

10:04

bank? Which I'm embarrassed to say,

10:06

I fell on that trap. This is like how powerful peer

10:09

groups are. It's

10:11

a very easy decision to not go

10:13

back to school after seeing what the

10:15

YC-5 was like. There's

10:17

a powerful quote by Carl Jung that

10:19

I really love. The

10:21

world will come and ask you

10:24

who you are and if you don't

10:26

know, it will tell you. It sounds

10:28

like being very intentional about who you want

10:30

to be and who you want to be

10:32

around as early as possible is very important.

10:35

Yeah, this was definitely one of my

10:37

takeaways, at least for myself, is no

10:40

one is immune to peer pressure and so all you

10:42

can do is pick good peers. Yeah. Obviously,

10:44

you went on to create

10:46

looped, sell that, go to

10:48

Green Dot and then we ended up getting to

10:50

work together at YC. Talk to me about the

10:53

early days of YC research. One of the really

10:55

cool things that you brought

10:57

to YC was this

10:59

experimentation. I remember

11:01

you coming back to partner rooms and talking about

11:04

some of the rooms that you were getting to

11:06

sit in with the Larian surrogates of the world

11:08

and that AI was at

11:10

the tip of

11:12

everyone's tongue because it felt so

11:14

close and yet that was 10 years ago. The thing I always thought

11:23

would be the coolest retirement job was to get to

11:25

run a research lab. It

11:28

was not specific to AI at

11:30

that time. When we started talking about YC

11:33

research, well, not only was it going to it,

11:35

it did end up funding a bunch of different

11:37

efforts. And I wish I

11:39

could tell the story of what was obvious that AI

11:41

was going to work and be the thing, but we

11:43

tried a lot of bad things too around

11:45

that time. I read

11:48

a few books on

11:50

the history of Xerox, Park and

11:52

Bell Labs and stuff and I think there were a lot of people, it

11:54

was in the air of Silicon Valley at the time, that

11:56

we need to have good research labs again. And I

11:59

just... I thought it would be so cool to do.

12:01

And it was sort of similar to what YC

12:04

does in that you're gonna allocate capital to smart

12:06

people and sometimes it's gonna work and sometimes

12:08

it's not going to. And

12:10

I just

12:12

wanted to try it. AI for sure

12:14

was having a mini moment. This

12:17

was like kind of late 2014, 2015, early 2016, was

12:21

like the super intelligence

12:23

discussion, like the book super intelligence

12:25

was happening. Yeah,

12:27

the DeepMind that had a few

12:29

like impressive results but a little bit

12:31

of a different direction. You know, I

12:33

had been an AI nerd forever. So I was like, oh,

12:36

it'd be so cool to try to do something but it

12:38

was very hard to say what to do. Was ImageNet out

12:40

yet? ImageNet was out. Yeah. Yeah. For

12:42

a while at that point. So you could tell if it was

12:44

a hot dog or not. You could sometimes.

12:46

Yeah, that was getting there, yeah. You

12:49

know, how did you identify the initial people

12:51

you wanted involved in, you know, YC

12:54

research and OpenAI? Greg

12:57

Brockman was early. In retrospect, it feels like this

12:59

movie montage and there were like all of these,

13:01

like, you know, at the beginning of like the

13:03

Bankai's movie when you're like driving around to find

13:05

the people and whatever. And

13:07

they're like, you son of a bitch, I'm in. Right,

13:10

like Ilya, I like heard

13:13

he was really smart. And then I watched

13:15

some video of his and he's also, now

13:17

he's extremely smart, like true, true, genuine, genius

13:19

and visionary but also he has this incredible

13:22

presence. And so I watched this video of

13:24

his on YouTube or something. I was like,

13:26

I gotta meet that guy. And I emailed him, he didn't respond. So

13:28

I just like went to some conference he was

13:30

speaking at and we met up and then after that we

13:32

started talking a bunch. And

13:34

then like Greg, I had known a little

13:36

bit from the early Stripe days. What was

13:38

that conversation like though? It's like, I really

13:40

like what your ideas about AI and

13:42

I wanna start a lab.

13:44

Yes, and one of the things that

13:47

worked really well in retrospect was

13:50

we said from the very beginning we were gonna go

13:52

after AGI at a time when

13:54

in the field you weren't allowed to say that

13:57

because that just seemed impossibly.

14:00

crazy and borderline

14:02

irresponsible to talk about. So that got

14:04

his attention immediately. It got all of

14:06

the good young people's attention and the

14:09

derision, whatever that word is, of the mediocre old people.

14:12

And I felt like somehow that was a really good

14:14

sign and really powerful. And we were like this

14:17

ragtag group of people. I

14:19

mean, I was the oldest by a decent amount. I was like, I

14:21

guess I was 30 then. And

14:24

so you had these people who were like,

14:26

those are these irresponsible young kids who don't

14:28

know anything about anything. And they're like saying

14:30

these ridiculous things. And

14:33

the people who that was really appealing to, I guess,

14:35

are the same kind of people who would have said like, it's

14:37

a, you know, I'm a sophomore and I'm coming or whatever. And

14:39

they were like, let's just do this thing. Let's take a run

14:41

at it. And

14:44

so we kind of went around and met people one

14:46

by one and then in different configurations of groups. And

14:49

it kind of came together over the course of,

14:52

in fits and starts, but over the course of like nine

14:54

months. And then it started

14:57

happening. And then it started happening. And

14:59

one of my favorite like memories

15:01

of all of OpenAI was Ilya

15:05

had some reason that Google or something

15:07

that we couldn't start in. We announced in December of

15:10

2015, but we couldn't start until January of 2016. So

15:13

like January 3rd, something like that of 2016 for

15:16

like very early in the month, people come back from

15:18

the holidays and we go to Greg's

15:20

apartment. Maybe there's 10

15:22

of us, something like that. And

15:25

we sit around and it felt like

15:27

we had done this monumental thing to get it started. And

15:30

everyone's like, so what do we do now? What

15:34

a great moment. It reminded me of when

15:36

startup founders work really hard to like raise

15:38

a round and they think like, oh, I

15:40

accomplished this. We did it. We did

15:42

it. And then you sit down and say

15:44

like, fuck, we gotta like figure out what we're going to do. It's

15:47

not time for popping champagne. That was actually the starting

15:49

gun. And now we got to run. And

15:52

you have no idea how hard the race is going to be. It

15:54

took us a long time to figure out what we're going to do.

15:58

But one of the things that I'm... really

16:00

amazingly impressed by Hylia in

16:02

particular, but really all of the early people about it, is

16:05

although it took a lot of twists and turns

16:07

to get here, the

16:10

big picture of the original ideas was

16:13

just so incredibly right. And

16:15

so they were like up on like one of

16:17

those flip charts or white boys on I Remember

16:19

Witch in Greg's apartment. And

16:22

then we went off and, you know,

16:24

did some other things that worked or didn't work or

16:26

whatever. Some of them did and eventually now we have

16:28

this like system. And

16:32

it feels very crazy

16:34

and very improbable looking backwards, that

16:36

we went from there to here with so

16:38

many detours on the way, but got where

16:40

we were pointing. Was deep learning even on

16:43

that flip chart initially? Yeah, I

16:45

mean more specifically than that, like do

16:47

a big unsupervised model and then solve RL was on

16:49

that flip chart. One of the flip charts

16:52

from a very, this is before Greg's apartment, but from

16:54

a very early off-site. I think

16:56

this is right. I believe there were three goals

16:58

for the effort at the time. It

17:01

was like, figure out how to do unsupervised learning,

17:03

solve RL and never get more than 120 people.

17:07

Missed on the third one. That's right. The like

17:10

the predictive direction of the first two is pretty good.

17:13

So deep learning, then

17:16

the second big one sounded like

17:18

scaling, like the idea that you

17:21

could scale. That was another heretical

17:23

idea that people actually found even

17:25

offensive. I remember a

17:28

rash of criticism for you guys at that

17:30

moment. When we

17:32

started, yeah, the core beliefs were deep

17:34

learning works and it gets better with scale.

17:38

And I think those were both somewhat

17:40

heretical beliefs. At the time, we didn't know how

17:42

predictably better it got with scale. That didn't come

17:44

for a few years later. It was a hunch

17:46

first and then you got the data to show

17:49

how predictable it was. But people already knew that

17:51

if you made these neural networks bigger, they got

17:53

better. We were sure of that before

17:57

we started. And... You

42:00

mentioned earlier that thing about discover all

42:02

of physics. I was wanting to

42:04

be a physicist, wasn't smart enough to be a good

42:07

one, had to contribute in this other way. But the

42:09

fact that somebody else, I really believe is now going

42:11

to go solve all the physics with this stuff. I'm

42:14

so excited to be alive for that. Let's

42:16

get to level four. I'm so happy for whoever that person

42:19

is. Yeah. Do you want

42:21

to talk about level three, four, and five briefly? Yeah.

42:24

So we realized that AGI had become

42:26

this badly overloaded word and people

42:29

meant all kinds of different things. We tried to

42:31

just say, okay, here's our best guess roughly of

42:33

the order of things. You have these level one

42:35

systems which are these chatbots. There'd

42:37

be level two that would come which would

42:39

be these reasoners we think we got there

42:41

earlier this year with

42:43

the 01 release. Three

42:46

is agents ability to

42:48

go off and do these longer-term tasks. Maybe

42:51

like multiple interactions with an

42:53

environment, asking people for help when they

42:55

need it, working together, all of that.

42:59

I think we're going to get there faster than

43:01

people expect. Four is

43:03

innovators, that's like a scientist and

43:05

that's ability to go explore like

43:07

a not

43:09

well-understood phenomena over

43:13

a long period of time and understand what's just,

43:15

kind of go just figure it out. Then

43:19

level five, this is the

43:21

slightly amorphous like, do

43:23

that but at the scale of the whole company or a whole

43:26

organization or whatever. That's

43:29

going to be a pretty powerful thing. Yeah. It

43:31

feels kind of fractal, right? Like even the

43:33

things you had to do to get to

43:35

two, sort of rhyme with level five and

43:38

that you have multiple agents that then self-correct, that

43:40

work together. I mean, that kind of sounds like

43:42

an organization to me just at like a very

43:44

micro level. Do you think that we'll have, I

43:46

mean, you famously talked about it. I think Jake

43:49

talks about it. It's like you

43:51

will have companies that make billions of

43:53

dollars per year and

43:55

have like less than a hundred

43:57

employees, maybe 50, maybe 20 employees.