0:00

Alright, Freeburg is back. Welcome back to

0:02

the all in podcast episode 160 something

0:04

your favorite podcast in the world yada

0:06

yada yada With me again the

0:08

chairman dictator from off polyhappetea little rain

0:11

man Yeah, definitely David Sachs is

0:13

here and back from his

0:16

time in the metaverse We

0:18

found him somewhere out in space in

0:20

the solar system in his

0:22

Apple goggles your favorite Sultan of

0:25

science David Freeburg is

0:27

back from the metaverse. I

0:29

miss you guys. Welcome home Thanks

0:31

for having me. What did you discover

0:33

when you went to Uranus in Google

0:36

class? Sorry, happy to actually use

0:38

the Apple Vision Pro takeout. I ordered them I

0:40

ordered them and I walked by the Apple Store

0:42

and I was gonna go in and try them

0:45

and there were so many lunatics In there. I was like,

0:47

yeah, I'm not doing it, but I ordered them

0:49

you use you actually use them Ordered one online

0:51

to be delivered and it was like delayed by

0:53

a month. So I went down to the Apple

0:55

Store and picked one up Okay,

0:58

and my kids cannot stop using

1:00

it. Really? I went down to the Apple Store

1:02

But got cleaned out by the thief to stole

1:04

everything That

1:20

was crazy, that's crazy We'll

1:26

put the video in here to the idiots who

1:28

are robbing Apple stores All the devices get pricked

1:30

when you steal them and they all have GPS

1:32

in them. Have you tried it? You know,

1:35

I was too busy working out making love and

1:37

winning. Oh So

1:39

you were you were making sweet love You

1:42

were watching your portfolio go up and you

1:44

were just generally winning. Got it. Got it.

1:46

Yeah Yeah, so freeburg the rest of us

1:48

were being men in the world

1:51

accomplishing stuff But but do tell us about

1:53

your time in the metaverse do those goggles

1:55

come with a lifetime prescription

1:58

of SSR eyes sound

2:00

like one of these like tech journalists that

2:02

are actually anti-tech people? You guys are... Actually,

2:05

tech journalists like it. Talking about the

2:07

next gen computing platform. I remember when the iPad came

2:09

out and everyone poo-pooed the iPad. I thought it was

2:11

stupid. I tried to use it. I couldn't get any

2:13

value out of it. And in 2010

2:15

or 2011, when did it come out? 2010. 2011, we started using it with

2:20

our sales team selling to farmers. And we gave

2:22

every sales guy an iPad and they went out

2:24

in the field with 3G. And they

2:26

were able to close sales in the field meeting with farmers,

2:28

which had never been done before. Usually, I had to get

2:30

a farmer to come into an office. How many iPads did

2:32

you sell? To sell the product. Oh.

2:35

So we had like... No, they were selling

2:37

climbing.com software. We had dozens of

2:39

these sales guys. We gave them out to our

2:41

sales agents as well, the independent agents. They started

2:44

using them. And it was like a real game

2:46

changer in how sales was done in agriculture. And

2:48

I had never even contemplated that when I first

2:50

used the iPad. All right. So

2:52

let's get to brass tacks here. What is the killer app?

2:54

What do you think in the next five years people are

2:56

going to be doing with this thing on a daily basis?

2:58

Is there a daily use case? I'll say

3:00

a couple things. One is like, I feel the same way I

3:02

did about the iPad, which is I don't know what it is

3:04

today, but I can tell that there's something there. And I'll give

3:06

you an example of something I thought about. First

3:08

of all, the AR is game changing. Okay. If

3:11

you've used like the meta... Yeah. The

3:13

Oculus Quest, it like makes

3:15

me super dizzy, makes my head hurt, makes

3:17

my eyes hurt. Like you're super disoriented. What

3:19

Apple solved is that you're like still in

3:22

reality. But then you get to

3:24

interact with these three dimensional kind of objects

3:26

in reality. And it's like really well done.

3:28

It's definitely V1 and there's going to be

3:30

incredible changes in the next couple of generations.

3:33

But it gets rid of all that dizziness,

3:35

disconnected kind of stuff that happens with

3:37

the full VR experience, which I thought

3:39

was really incredible. Then last week,

3:41

and I'm sorry I missed the show, we have a

3:44

facility with my company in North Carolina. We have

3:46

this giant greenhouse facility and I was doing meetings

3:48

with farmers and stuff. I go to the

3:50

greenhouse facility and there's so much work that

3:52

the greenhouse techs and lab

3:55

techs are doing where they're using an

3:57

iPhone and a barcode scanner and a printer

3:59

and they're... holding all these pieces of

4:01

equipment, scanning the QR codes on flowers, taking

4:03

the pollen out, putting it in the next

4:05

flower, training each other how to do it. And

4:08

I was like, I put this Apple Vision Pro on, and

4:11

I was like, man, all the apps and

4:13

all the tools that we had all these

4:15

different pieces for that was taking people tons

4:17

of time, image collection, data collection, could all

4:19

just be done streamlined while

4:21

you're working. You can have

4:23

a tap with a report. Yeah, you have a tap with a

4:25

little into a head thing. Cameras are

4:27

taking images in the middle. QR codes are

4:30

automatically scanned, data is being ingested. The task

4:32

list is kind of giving folks next steps.

4:35

They could listen to music while they're working. And I

4:37

realized for that job, and I met with all the

4:39

team out there and spent time with them, and I

4:41

actually did the work that they do to get a

4:43

better sense for the workflow. And I was

4:45

like, man, literally every aspect of this job will be

4:47

massively improved and productivity will go up by 10x with

4:49

these goggles. Will it happen in the next couple of

4:51

weeks or months? I don't know, but my engineering team

4:53

is looking into it. Can we take it? Can

4:56

we use some software? Can we build some software? Can we put

4:58

this on folks? Give

5:00

them a better work experience, increase our productivity,

5:02

to do automated data capture. So I don't

5:04

know exactly where it goes, but I could

5:06

start to see how this can become a

5:08

more ubiquitous part of a workforce setting, and

5:10

that just be a video game and movie

5:12

tool for consumers. So I'm reasonably optimistic about

5:14

where this goes. It's definitely V1. I feel

5:16

like it's the iPad days where no one's

5:18

really sure where the applications are, but yeah.

5:21

Yeah, enterprise applications. Unbelievable. Makes

5:23

total sense. And also training, training,

5:26

right? Early line workforce, sure house

5:28

workers, where you're getting

5:30

real time kind of task updates, data's being

5:32

ingested all in real time. And

5:34

by the way, the other thing I'll say is training

5:36

is incredible. There's spatial video recording on it. So

5:38

it looks like you're living through the experience that

5:40

someone else had. So you can train

5:42

someone how to do a difficult task, and rather

5:45

than have a human go spend hours training a

5:47

workforce, the workforce can be trained by the goggles

5:49

in a way that you cannot do a two

5:51

dimensional video today. So I don't know. I'm

5:53

pretty optimistic. Very strange news, right? You're

5:56

a fan of Side Five. I remember strange days. Totally. happen

6:00

first here? Are humans

6:02

going to become more like robots by putting these

6:04

on and do this factory work? Or

6:07

is Elon with optimists and some of humane I

6:09

think is the other one? There's a couple of

6:11

other people building a general use robots. Figure is

6:13

the other one. Figure yeah. Which one wins

6:15

the day? Is it going to be humans having eyes

6:18

and data collection like robots

6:21

or robots having appendages like

6:23

humans? Well, let me put

6:25

two ideas together and see

6:28

what you think of this argument. If

6:30

you think about the generation of

6:32

human beings that have as

6:35

close to any

6:37

other generation before it lived in a totally

6:40

immersive world, I would say the best

6:42

representation of that are current teenagers

6:46

and 20-year-old people and

6:48

maybe at the upper edge the early 30s people.

6:52

And why is that? They've lived

6:54

inside of social media their entire lives. They've lived

6:56

inside of immersive video games their entire lives. But

6:58

the question is, are they better

7:01

off and happier as far as we

7:03

know from an evolutionary perspective? And I

7:05

would tell you that the answer is

7:07

a huge gaping no. So

7:11

if you believe that the

7:15

rise in depression, the rise

7:17

in suicide, the dependency on drugs,

7:20

the dependency on SSRIs, the sexual

7:22

promiscuity, the lack of marriage, the

7:25

lack of kids, if

7:27

all of those things are in some ways

7:31

a correlated byproduct, let's not say it's

7:33

causal, right? Let's say it's a correlated

7:35

byproduct of this entire

7:37

immersive almost exclusionary

7:40

detached world that these folks

7:42

have grown up in, taking

7:44

that to the limit, I'm

7:47

just going to put out there it may not be the solution to

7:49

our problems. And so I guess

7:51

the more direct answer to your question is I

7:53

would hope that the latter wins So

7:56

that we take these goggles off and actually learn how to talk

7:58

to each other and look each other in the eyes. Get.

8:01

Married and have children because I think that's actually

8:03

better for the world. And

8:06

I would probably say that it's almost better for the

8:08

world than a connecting a productivity. Interesting.

8:11

And then you see the correlation

8:13

to cancer and disease that is

8:15

disproportionately higher amongst these young people.

8:18

So. I think it's at some point that

8:20

ask ourselves. What? Is structurally happening

8:22

in the lives of these sixteen?

8:24

you know, fifteen to thirty one

8:26

year olds that is just so

8:28

poor in terms of outcomes. And.

8:31

If you look at some of the environmental variables

8:33

that they live in and then take some of

8:36

those and take them to the limit. I think

8:38

that there's a reasonable argument that their lives get

8:40

worse before it gets better. Money

8:43

amount of time you spend on social media

8:45

is correlated with the most of the have

8:47

to sing just this immersive like idea to

8:49

attach real world and. Lived

8:51

through a microphone and glasses taken to

8:54

the limit I'm not sure is the

8:56

solution. To. These kids feeling

8:58

detached, lonely, isolated are Slim

9:01

Jim. Ja hyun. It's the

9:03

it correlates All of these things that we're seeing

9:05

in the summer. Generation correlates with the international how

9:07

to be more fun activity device. Yes, of course

9:09

I hope it's a good product. A device? yes.

9:11

But if we try to make it the panacea

9:13

for anything and everything. Has. The Intergalactic.

9:16

We're. Going to compound. The. Systemic

9:18

issues that these young people. Had

9:21

and I suspect. On

9:23

the margin if you're going to bed. All. Of

9:25

these things that we see in these young people today will

9:27

get worse. As a byproduct of technology,

9:29

not necessarily get better. So if you can.

9:32

Take a different path. Like. Optimists

9:34

or the Sigur Ai robots were

9:36

that work is done. At.

9:38

Least we have a different problem.

9:41

probably maybe even more existential abundance.

9:43

But. A different problem which is now, how do you

9:46

find purpose? But maybe you can find purpose to connection

9:48

and that types of things that humans have been bred

9:50

over. Billions. Of years to

9:52

actually optimized for talking. Sachs.

9:55

I remember when. You.

9:57

Were starting crap. You.

10:00

The fire it up like a group for the are

10:02

you knew that pretty heavy into you made a couple

10:04

of small bats I remember. How. Did I

10:06

don't they get any of it worked out? Really to tell me

10:08

if I'm wrong here but you got in a little bit early.

10:10

There are many you talk about. The business case for this. And

10:13

has that changed? Because he knew? Believe I believe

10:15

the lot of folks thought hey, maybe this is

10:17

the time. When. Zoc really

10:19

start you know had bought hockey listen and they

10:21

started putting out so good product. Seem.

10:24

Like was a false start. Is this the actual

10:26

starting pistol and is is the sort of the

10:28

the are A are. Adoption

10:30

Race. I. Don't think we're quite

10:32

there yet. We've. Been

10:35

talking about Vr being a saying

10:37

for. Over. A Decade

10:39

is no more like Thirty Member the

10:41

on Nintendo Vr. Stuff. With my

10:43

always on the verge of happening or six

10:45

up. the big complain about the Apple devices

10:47

as lotta capability but it's still a pretty

10:50

huge device to wearing your forehead. The Smart:

10:52

really? Can. Be careful, not. To.

10:54

Be something that people want to use all

10:56

the time. Of

10:59

and is also a question of use cases

11:01

but. They're. Getting their with these

11:03

cases. And they

11:05

were like I do think that Apple Vision

11:07

Pro is. It's like I said

11:09

last week, it's a. Useful. Prototype

11:11

are proof of concept. And

11:14

it will get better. So. I'm glad

11:16

they did it. Does. I think you need

11:18

to start somewhere in that skewed been A rating. But.

11:20

Eventually for this to I think

11:23

really take off. you need to.

11:25

Shrink the form factor miniaturized the

11:28

technology. Does every version of it

11:30

make it simpler, a lighter, easier

11:32

to use. Yeah, I mean eventually

11:35

it'll. Feel. Like sunglasses and so

11:37

that is, I guess. If. They

11:39

become a regular glasses and we all agree

11:41

I becomes in. I don't know, I know

11:43

you. I feel like it's pretty damn comfortable

11:45

and as you guys gets, haven't really use

11:47

it by that's what I've heard. That's a

11:49

surprise people. are I online saying some like

11:51

any other had said I've ever worn they

11:53

didn't. Incredible job designing two and I feel

11:55

like ski that black goggles. It. Doesn't

11:57

feel heavy. It doesn't feel pressure, pressure.

12:00

It appeared to ski goggles have you are

12:02

wearing ski goggles it's less constricting and ski

12:04

goggles it's more comfortable of it like floats

12:06

on you a little bit. They did a

12:08

great job with this cushioning device, a belt

12:10

and the ban on it feels very natural.

12:12

Is Apple design right? As like a really

12:14

well designed products that unlike anything else you've

12:16

ever tried. I've always felt like when Apple

12:18

comes into the race that's the starter pistol

12:20

and I think this is that because I've

12:22

I've heard the same thing from everybody. Here

12:25

you have to try it. It feels like different

12:27

than populace in some of those roses that came

12:29

out previously. And. They

12:32

have the app ecosystem and I would not

12:34

discount that when he don't have the ability

12:36

to monetize use the app ecosystem have all

12:39

the people who are already building the com

12:41

after you were braff, Whatever. Notion.

12:43

Will you know all the stuff that people

12:45

use in Love Spot? a fi you tube

12:47

and then ported over here fortnight whenever. I

12:50

think that's gonna be the magic. and

12:52

the statistics are not lying here. I

12:55

mean, this is unbelievable. They've sold already.

12:57

two hundred thousand units. Which. Doesn't

12:59

seem like a lot but for v one that is a

13:01

lot and are going to sell a half million this year.

13:03

Because do like. The sometime I

13:05

was a couple of billion metre sells more.

13:08

They do. Yeah but you this

13:10

is four thousand dollars. This. Isn't

13:13

five hundred so to sell that many have

13:15

a four thousand dollar devices in quite a

13:17

prefer concept. It's not like a. Regular.

13:20

Apple product that. Is

13:22

a mass market device up. Tens.

13:24

Or hundreds. Most people get by, but it puts them

13:27

on a path. Where they can

13:29

the iterative you here butter. See. I

13:31

think the end this as I get from an

13:33

s free bird. Would you compare this to buying

13:35

a Macbook Pro. Buying. And I

13:37

phone. Or buying the archaeologists. you

13:39

know, whatever they you know five hundred our

13:41

unit because. Everybody I see talking

13:43

about online is comparing it to the purchase

13:46

of a laptop. Because. Of

13:48

the desktop and you can can add to your coat

13:50

hanger, surf the web and do all that where it

13:52

Where do put. This is buying a Tv As a

13:54

buying a laptop is a buying a smartphone. What would

13:56

you have more to be really productive on it? Ah.

13:59

if you're going to use it for writing purposes or coding purposes.

14:01

So it doesn't really work with

14:03

just the headset, but you could do that.

14:06

Yeah, it's definitely like buying a new computing device,

14:08

but people felt the same way about the iPad.

14:10

Go back to 2010 when the iPad came out

14:13

and everyone was like, who is it for? It's

14:16

a whole new computer, who's it for? You already have a

14:18

phone, you already have a computer, why do you need an

14:20

iPad? And then they sell tens of millions a

14:23

quarter now. So I really, as I do

14:25

the math on this, I was just kind of doing some back of

14:27

the envelope stuff. I think they're gonna sell $100 billion of

14:30

Apple Vision Pros, not this version, but

14:32

this version plus the next version probably

14:35

over the next. I

14:37

would guess for them to get to $100 billion in sales, it'll

14:39

take them less than five years.

14:42

I think they're gonna run the table on everybody. I

14:44

think they're gonna own the entire space. I think everyone's

14:46

underestimating this as a new computing

14:48

platform. And once these applications, particularly in the

14:50

enterprise settings, start to kick in. And

14:52

I will say that the movie watching experience is way

14:54

better than watching on a TV in your living room.

14:57

My kids cannot stop asking me to use the goggles

14:59

to watch instead of an iPad or TV. Because

15:02

you see 3D, like all Pixar movies are

15:05

natively 3D. And so you've

15:07

got the Disney Plus app on there, you watch a Pixar

15:09

movie and you're watching in 3D, the kids are blown away.

15:12

So I think we're all gonna be surprised by how this

15:14

goes. Disney's all in on it. I remember when our parents

15:16

told us not to sit too close to the TV. Now

15:20

we're just strapping the thing to our face. Yeah,

15:23

I had the most Silicon Valley moment ever.

15:25

I go to buy a cup of coffee. I was going for

15:27

a little walk. I see blue bottle. I'm like, oh, you know

15:29

what? I give myself a mocha. I lost a little bit of

15:32

weight. I'm gonna treat myself. $9 for a mocha.

15:35

Number one, that tilted me. $9

15:37

for a mocha? Well, it was $8 and then I

15:39

gave a dollar tip. And

15:42

then I felt cheap giving a dollar tip. You know,

15:44

it's $8.99 for a carton of clover of milk, all

15:46

organic. You

15:49

can make infinite lattes at home. Anyway. Where

15:51

did you go for your $9 mocha? I'm

15:54

in Palo Alto right now because we lost power.

15:56

Blue bottle. Yeah. I posted

15:58

this. I'm like $9. What am I doing? You

16:01

know, I just I felt like buying a chocolate bar and putting

16:03

in a couple of clips. Look at the stain. Your dirty lips

16:05

left on the cup. Oh my God. Look at the look of

16:07

that. You know what? You're a little obsessed with my

16:09

lips. Take it easy, dude. Yeah. So

16:12

anyway, then there's a kid in the place

16:14

wearing the goggles with the keyboard.

16:16

No, stop. He's pounding. He's

16:19

getting work done. This kid was doing work.

16:22

And I tell you, he was putting in the

16:24

hours. He was putting in the hours. No one looked at your laptop.

16:26

No one looked at your screen. That's what I wrote

16:28

about it. All your work without anyone seeing what you're

16:30

doing. This kid had four desktops up. This guy was

16:32

probably on Pornhub, Spotify,

16:35

writing code. How many words did

16:37

this person say to another human being while you were

16:39

there? No, zero. And you know what? When

16:42

they're on a laptop, they're the same. What's the difference? He's coding.

16:44

Nobody bad it. And I think this is going to run

16:47

the table on this. I think it's 100 billion sales.

16:50

100 billion sales under five years. I take the over.

16:52

I take the over. What do you

16:54

got? The over, the under. And if they keep it

16:56

at three grand, they got to sell 30 million units

16:59

to get to 100 billion. They're going to

17:01

make up a lot of money in this app store, too, by the

17:03

way. I think that you guys are right that it's going to be

17:05

successful in terms of revenue. What I'm

17:07

asking is a more societal question. Do you guys actually

17:09

think it's better? No, I don't want my kids in

17:11

this all day now. And I could see this becoming

17:13

super-dicking. Hey, Freiburg, can I buy three for your kids?

17:15

Just have them walk around with them? No, I have

17:17

a no-win in the house rule as well. But wait

17:19

a minute. Hold on. What about

17:21

productivity, Freiburg? My kids aren't trying to be

17:23

productive. They're using a burn part. I call childhoods

17:25

not people. You don't even have a productive childhood.

17:28

It's supposed to be not productive. You guys understand

17:30

that at some point, you guys will be the

17:32

only six kids whose

17:34

parents haven't given them the stupid thing to

17:36

put on their face? No, this

17:38

is going to be time-restricted. I have a no

17:40

iPad, no phone, no... I

17:42

let them use the headset because we got it for

17:45

them. No, no, it burns their brain away. Burns

17:47

their brain away. No way. Man.

17:50

Absolutely. I totally agree with you. I think that's

17:52

a great question. The loss of our ability to

17:54

communicate as human is critical and it's a fail

17:56

point. I do think that there are applications where

17:58

these things create great unlocks. I think this

18:00

is an enterprise device. Can you imagine giving the

18:03

sales team on the farm to go there? They

18:05

can take off their sweaty headset when the Sun

18:07

is shining and then give it to the farmer

18:09

to put on and then he can put it

18:11

on and feel the Sweat and the headband will

18:13

be wet. No, that's not the use case. It

18:15

does it by the way It's a very personal device

18:18

in order to log in, you know, it does like a eye scan

18:21

Or you have to have like a lock-in like login like you

18:23

do with your phone But then you got to reset the eye

18:26

because it automatically sets the eye position So

18:28

when you put on someone else's headset, you got to

18:30

reset the IP. It's a whole thing. So it's not

18:32

a transferable device It's a very personal computing, you know

18:34

kind of thing So I don't think

18:36

it's gonna be the same as like an iPad or a phone. It's

18:39

a very different kind of thing I don't know what it's gonna look

18:41

like yet. I know I say next week We do the show inside

18:43

of these or at least me and you free bird will be will

18:45

be There's

18:48

a There's an app. There's an

18:50

avatar thing So what it does it scans your

18:52

face while you're talking and then all four of

18:54

us can see each other as the avatar Yeah,

18:56

let's do it. It'd be hilarious. I

18:58

had a moment this week in parenting I had

19:00

a moment this week where I told

19:02

one of my children that when I send

19:04

a text message I expect an

19:07

immediate response. Hmm Otherwise,

19:09

I am going to cancel that

19:11

child's phone and take it away and

19:13

then separately when they respond It has

19:15

to be in structured well thought out perfectly

19:18

formatted English and then

19:20

then third I said every single email I

19:23

see from you interacting with your teachers or

19:26

Anybody else that's there to help you needs

19:28

to be incredibly well written and formatted and

19:30

if I see garbage English I'm

19:32

gonna take your phone away Okay,

19:35

so you don't want them on their phones, but

19:37

they have to respond right away Well, they have very

19:40

strict rules about they can use they're there for

19:42

literally that all they can do is communicate like

19:44

they can use I message But

19:46

it is shocking to me that despite the lack of

19:49

games that they have or whatever how

19:51

poor they are in being

19:54

able to communicate and what

19:56

little access to devices they

19:59

have have already made them orders

20:02

of magnitude less able to communicate than frankly I was

20:04

able to when I was their age. And so I

20:08

can just imagine what happens when you become

20:10

even more ensconced in something that you can

20:12

cocoon yourself with I

20:14

don't agree with you. I don't disagree with you. Not to

20:16

say that it's not gonna be a

20:18

revenue generator, but I think that you could just

20:21

as easily, frankly, instead

20:23

of impacting Apple as revenues, you can probably

20:25

go along the makers of SSRIs, here

20:30

comes the spread trade, Bumble

20:33

and Tinder, and you'll get to the same place

20:35

economically. All right. All right, here

20:37

we go. We've got a lot on the, what a

20:39

great leap forward for humanity. I can't wait. I

20:43

just see this as the laptop replacement. Okay.

20:46

I wanted to talk a little bit about what apparently

20:48

is gonna be the spread trade of the last

20:50

year. Meta is continued

20:53

their unbelievable run and

20:56

Snap dropped like 30%.

20:58

Here's a chart for y'all of Snap versus

21:00

Meta. You can take a quick look at

21:03

it here. And just for context, both companies

21:05

did great during COVID and ZURP, hit all time highs

21:07

in 2021, but they both got

21:10

crushed due to the ad spend pullback,

21:12

obviously. But then Meta started to get less

21:14

focused on their headsets and more focused on

21:16

AI, started doing their reduction

21:18

in head count, 22% year over year from 86,000

21:21

to 67,000 in

21:25

the last quarter for Meta. And

21:27

their quarterly profits have increased to an all

21:29

time high of $14 billion. That's

21:32

profits folks in Q4 for Meta. All

21:35

time high for the stock price $470 a

21:37

share, 1.2

21:40

trillion market cap, Snap down

21:42

60% from its closing price on

21:44

its IPO day in 2017. Let

21:47

me just jump to Chamath before I get into

21:49

more charts and everything. You pointed out Chamath and

21:51

maybe you could explain to the audience just how

21:53

ridiculous the

21:55

voting rights were and the

21:58

massive dependence that... the

22:01

SNAP team and the executives had

22:03

on stock-based comp two issues for

22:05

you, Chamath? Well,

22:07

I mean, I think I said it before. I think that case

22:10

studies have been written about how tilted

22:14

the governance is in SNAP. I think the

22:16

point is that they basically have infinite

22:18

to zero voting power over common

22:21

shareholders. So there's no real feedback

22:24

loop. And I think that that has

22:26

probably adversely affected the types

22:29

of people that traffic

22:31

in their stock. Now, look,

22:33

activists and

22:36

short sellers sometimes have a very

22:39

bad reputation. But

22:41

if you steel man their side of it, what

22:44

they are there to do is to shine a light on

22:47

inefficiency and in the short seller case,

22:49

sometimes in propriety, but it

22:51

should all lead to companies being better

22:53

run. I think

22:55

Meta had this example where they

22:58

had a really big hiccup and

23:00

everybody, including us, sort of pointed out

23:03

the levels of spend that

23:05

they were making really didn't make any sense. I

23:08

think we had a chart that compared the level of spend

23:11

of Meta second only to like

23:13

the spaceship program, right? Just like

23:15

an enormous amount of money. And

23:18

look, Mark got the message. He

23:20

heard it loud and clear. I

23:22

think he got fed up with whatever was going on there

23:24

and he fixed it. And it's

23:27

in the numbers. Now, I don't know,

23:29

snap, because to be honest with you, I've

23:31

never taken more than one second to look

23:33

at that company. And the

23:35

reason is, there is just zero ability

23:38

for me to have any useful say. So

23:40

I've never honestly looked at its performance. I've

23:43

never studied a single characteristic. I've never

23:45

trended it. And

23:47

I think the point is that I

23:49

am probably where a lot of other reasonably

23:51

smart folks who could give a reasoned opinion

23:53

on how to make it better land. And

23:58

part of the reason is because there is No

24:00

feedback loop that matters. Yeah, and when

24:02

you know that, Why? Would you waste

24:04

your time at least and are other options? Rider

24:06

other options. And then metal was another one. You

24:08

know you can write a letter. It

24:11

gets picked up on C N B C

24:13

and Bloomberg and whatever. And all of

24:15

a sudden they kind of pay attention and

24:17

I think and you look at Disney Nelson

24:19

Pelts goes and gets I'd Perlmutter shares by

24:21

some more takes. oh my goodness and health's

24:23

yeah. we'll see whether that six it's sub.

24:25

The point is that. When. All

24:28

of these other cases. People. Are

24:30

investing the time because. They think

24:32

that there's even a small shred of a

24:34

chance that the company listens. But.

24:36

If you literally have no say. You.

24:39

Couldn't even do a proxy. You couldn't vote

24:41

the shares. Why would you bother? And

24:43

I said that as more of an example works.

24:46

Maybe there is a I so I don't even

24:48

know why. Stop did poorly and again on lock

24:50

and really take the time because it's like why

24:52

bother to consider. A section:

24:54

should they unwind and slight know

24:57

voting Common shares Super voting shares

24:59

nonsense And and should this. Go.

25:01

Away as a concept in the stock

25:03

market wells I mean Facebook. Or.

25:05

Miller has the pretty similar concept. I

25:08

mean I guess Zoc Bird has to

25:10

see percent voting control where as of

25:12

it's because I the nine percent so

25:14

snappers. More. Egregious. The.

25:17

Difference is that Zuckerberg is

25:19

listening and speak your mind.

25:22

But. The reason why. Snap.

25:25

Is doing poorly as not because

25:27

it's revenue as deteriorated so. I

25:29

looked up. For this fast, Such

25:31

a pity. For. Their team after so.

25:34

Assuming Gb to mother was leading. If

25:37

you compare Twenty Twenty One to Twenty

25:39

Twenty Three, their total revenue went up

25:41

from four point one of four Point

25:43

five. Billion. And

25:45

gross profit. Went. From

25:47

call it two point four to two

25:49

point five billion. So. Not. A

25:51

huge increase, but. Revenue. Gross profit

25:54

were slightly up. But. if you're

25:56

operating expenses they went from three

25:58

billion to four billion a year.

26:01

And that is why their operating income or

26:03

operating loss went from a $700 million

26:07

loss to a $1.4 billion loss in two

26:09

years. So that's the source of

26:11

the problem is that they

26:14

increased their operating expense by a billion

26:16

dollars a year from 2021 to

26:19

2023. Yeah. They

26:22

seem like they're the last ones to get the memo. Yeah,

26:24

they were the last ones to get the memo and just

26:26

finish the point. So you saw that a

26:29

few days ahead of this quarterly announcement where

26:32

their stock got crushed, they put out

26:34

a press release saying they're going to cut their headcount 10%. Ahead

26:37

of a little. Yeah, too little too

26:39

late. Yeah, they knew, right? They didn't have a

26:41

problem. So they released the press release

26:43

saying, oh, we're going to cut. Well, you

26:46

should have done what Zuckerberg did. You know, Zuckerberg

26:48

did a 20% cut last

26:51

year. He got serious. He got lean

26:53

and fit. And instead, these

26:55

guys held out, did nothing. Then when they

26:58

know that the market's going to crush them,

27:00

they put out this lame announcement 10%. No, not 10%.

27:04

Really, if you just want to get back to

27:07

where you were two years ago in terms

27:09

of operating expense, you need a 25%

27:11

reduction. Yeah, yeah. But

27:13

it's more than that. If you look at the

27:15

numbers, let's use operating cash flow with $165 million

27:18

for SNAP for the quarter. So their

27:20

operations generated $165 million of

27:22

profit. But for the entire year,

27:25

because they lost money in the quarters prior, they generated

27:28

free cash flow of only $35 million.

27:31

So the business net produced

27:33

$35 million of incremental cash.

27:35

You know how stock based comp accounting

27:37

works, the charge happens when it vests.

27:40

So this is what employees are vesting. During

27:43

the year of 2023, employees vested

27:45

$1.3 billion of stock based comp.

27:47

So that means new shares or

27:49

options were issued that on an

27:51

accounting basis, the options are valued using black shoals,

27:53

and the shares are valued based on the share

27:55

price. So they issued $1.3 billion

27:57

of stock based comp. So they generated $35 million. Free

28:00

cash. And the youth point

28:02

three billion dollars to compensate employees the

28:04

under our bags. So that means that

28:06

they paid employees. Forty. Times the

28:09

free cash flow that was generated for

28:11

shareholders during the year. Which. Is

28:13

also equivalent to ten percent.

28:16

Of. The enterprise market value of this company. So.

28:18

The Enterprise value the company is fifteen billion

28:21

dollars. Ten percent of that it was issued

28:23

two employees to compensate them. Now let me

28:25

give you that that the story of another

28:27

city. Mehta and by the way, snaps

28:29

your account because they issued all the stock.

28:31

The. Number of shares outstanding increased by

28:33

four percent during the year. During.

28:36

The your met his number of shares

28:38

outstanding decreased. By. Half a percent

28:40

because they use cash to go and buy

28:42

back stocks are they were able to reduce

28:44

the shares outstanding. Now as you guys talk

28:46

about medical employee town for twenty percent. And.

28:49

Snap Cut employees had cows by

28:51

three percent during the year. But

28:53

here's the crazy difference in performance

28:55

the stockade com expense. For.

28:57

Met A during that year. Was. About

29:00

fourteen billion dollars the that your.

29:02

That. Company generated seventy one billion

29:05

or property cashflow. So. I'm

29:07

while Snap keep employees forty

29:09

times. The free cash flow met

29:11

a employees you know about our twenty percent

29:13

of the of the free cash flow and

29:15

then and then matter when around and i

29:17

use of have that extra cash to buy

29:19

back twenty billion dollars the stock so they

29:21

bought back more shares than what the employees

29:23

were his you that that your work. So.

29:26

It shows such a difference. And

29:28

looking out for shareholders. So if I'm an investor

29:30

and by the way that is treating him like

29:32

twenty five times free cash flow which is not

29:34

a crazy multiple given all the new businesses are

29:37

they have and lama to and the progress into

29:39

cloud and other things have a my damned if

29:41

I'm looking at those two businesses have a shareholder.

29:43

you got this. Data controls the whole start. He's.

29:45

Giving employees and billion three of share the

29:48

year. When. He's only making thirty million

29:50

dollars a freak outs for your and then the

29:52

other guy. Is. Issuing. Fourteen.

29:54

Billion dollars have shared buying them all back. and

29:57

he's making seventy billion a free cash flow your i don't

29:59

know it's very hard to decide which one to go

30:01

after. Well, Spiegel brought it up in an

30:03

interview I saw. And a lot

30:05

of the layoffs were top heavy. So he got

30:07

rid of a lot of the top people who

30:09

had these huge comp packages. And

30:11

then what I'm hearing from a

30:14

lot of executives is cutting

30:16

these highly stock comped executives

30:18

who also have big cash comp, cutting

30:21

them, putting lieutenants in charge, and

30:23

then moving more jobs

30:25

to other locations where people don't

30:27

expect stock-based comp. If you're in

30:29

India, or you're in South

30:31

America, whatever, stock-based comp is

30:33

not like the obsession it is here. So

30:36

as everybody optimizes these businesses, I

30:38

mean, Facebook even didn't give a damn. Why

30:40

are these 5,000 employees? So

30:42

they announced roughly 500 job

30:44

cuts out of, what, 5,500 employees. I

30:49

mean, should that company

30:51

be operating with 2,000 employees? It's

30:54

a good question, yeah. How long have the number of Twitter employees from 8,000 to

30:57

1,500? When

30:59

you look at the number of apps that they're running

31:01

and the number of products that they're running compared to

31:03

Meta, right? Meta has far more apps, far more infrastructure.

31:06

Meta is serving 3.2 billion daily

31:09

active users. Snap is about 400 million. So

31:12

Meta is 8X the users

31:14

with many more applications and much

31:17

more infrastructure. So

31:19

I think it's another great kind of ratio to

31:21

look at the performance of these two. I

31:24

think you're exactly right, Zach, yeah. The

31:26

other advantage that Meta has is because

31:28

they're so profitable, they have the resources

31:30

to go big in AI. Big

31:33

time. Which is very expensive. So,

31:35

yeah, so they are the leader. You get

31:37

all this option value at Meta, which you

31:39

don't get at Snap. There's all this infrastructure

31:41

that they can leverage, much like Amazon did

31:43

with AWS, into things like cloud, AI

31:46

tools for third-party developers,

31:48

third-party applications. Meta

31:52

is the biggest advertising platform next to

31:54

Google in the world now, and

31:56

there's much more that they can start to do to

31:58

extend further into the future. to the platform.

32:00

They did get an awesome save. Remember

32:03

Apple screwed them and was like, you

32:05

can track devices now. And that just

32:07

took a massive hit in

32:09

the ad network. And it was all those headwinds. They

32:11

were like, okay, we're just gonna use AI to

32:13

optimize ads. And supposedly the AI optimization of ads,

32:15

I was talking to somebody on the inside. They

32:18

said like, yeah, we got it

32:20

all back. We gained it back. We've

32:22

got massive AI advertising optimization going on.

32:24

So yeah, that's great that Tim Cook

32:27

kicked us in the nuts, but we don't

32:30

care. By the way, that's a great point,

32:32

Jay Cal. It really says a lot about

32:34

how Meta was able to respond to that

32:36

change, which a lot of people speculated would

32:38

destroy the advertising business. And the fact that

32:40

they were able to engineer solutions to drive

32:43

advertising revenue up to $40 billion, it's

32:45

just mind blowing. It's a really

32:47

kind of impressive outcome for the team. And I think it

32:50

speaks a lot to the quality of the engineers there. I

32:52

think it's a great point. Sacks, you tweeted

32:54

that you're seeing a little SaaS bounce back

32:57

all of a sudden. That's interesting. I'm

32:59

seeing something similar. Last year, last two years, you

33:01

had a ton of people cutting their

33:04

SaaS spend, maybe removing the number

33:06

of SaaS vendors they had, consolidating

33:08

vendors. You tweeted, many public and

33:10

private software companies are experiencing accelerating

33:12

growth after six to seven

33:15

quarters of deceleration. SaaS recession

33:17

appears to be over, according to the

33:20

SaaS master, David Sacks. You

33:23

wanna pack this for us? What are you saying?

33:25

Well, it's still pretty early because not everyone's reported.

33:27

But if you looked at the big tech cloud

33:29

performance in Q4, you

33:32

could see that there's a bounce back in

33:34

here. This is net new ARR added

33:37

for AWS, Azure, and Google

33:39

Cloud. So you see here

33:41

in Q4, that there's

33:43

a huge increase in net new ARR

33:46

for the big cloud computing platforms. And

33:49

then I think another bellwether is

33:51

Atlassian. So we're still waiting to

33:53

hear from HubSpot, and the details

33:55

for Zoom, Adobe, companies like that. They haven't reported yet, but

33:58

if you look at it last year. Jira amongst other

34:01

products. They're based in Australia. Yeah, the

34:03

major. Yeah, exactly. Collection of SaaS companies,

34:05

right? It's a collection of SaaS products.

34:08

Yeah. So, net new ARR would be

34:10

the amount of growth in that quarter.

34:12

And this is on a

34:14

year-over-year basis. So, you can kind of see Q4 of 21 was

34:17

the absolute peak and then plummeted. And

34:23

then it actually went negative for about a year.

34:26

That's tough to be in a company with net new

34:29

ARR going negative. Yeah. This

34:31

doesn't mean, by the way, the company is shrinking.

34:33

It just means that the amount of

34:35

net new ARR, which is the amount

34:38

of growth, is actually smaller than that

34:40

same quarter of a year before.

34:43

Yeah. And then in

34:45

Q4, you could see there's some acceleration here.

34:47

That they're starting to add more. They added more

34:50

net new ARR, I guess 33% more in Q4 than they

34:54

did over the previous year. And part of that,

34:56

SaaS, is because the comps are lower and they

34:58

kind of bottomed out. Yeah. They

35:00

bottomed out. Now they're re-accelerating. Yeah, that's great. You

35:02

know, we're starting to see this in some of

35:04

my board meetings as well, where in

35:06

2022, everybody

35:09

was missing their numbers and re-forecasting down,

35:11

and then they would miss the re-forecast.

35:14

Yeah. So, by 2023, the

35:16

forecasts were very, very conservative. And I would

35:19

say, now I'm seeing companies

35:21

beat

35:23

the lower forecasts in

35:25

Q4. This wasn't happening

35:27

earlier in the year, but finally, I think

35:29

people are starting to beat their

35:32

lower forecasts for Q4. That's the question

35:34

that I was curious about. What do you actually

35:36

think is happening? Is that we've

35:38

re-baseline these businesses. So now, what would

35:41

have looked like just a massive miss

35:43

over the last two years now looked

35:45

like a beat because we've just completely

35:48

reset expectations? Is it that? Or

35:50

is it that the economy is

35:52

actually expanding and we can count

35:55

on some reasonable growth

35:57

rates? Is it a combo of the two? What do you think

35:59

is actually is. Yeah,

36:01

I mean it's definitely a new baseline

36:04

in the sense that if you go back to

36:06

2020 or 2021,

36:09

we considered good growth to be, you know,

36:11

two to three X year over year. And

36:14

now if it's going from 60

36:16

to 80% growth year over year, you're

36:18

happy. So there's definitely been a lowering

36:20

of expectations. That being said,

36:22

you still see in these numbers, there

36:24

has been a bottoming out and we're

36:26

starting to now grow from this new

36:28

baseline. So

36:30

for example, I think

36:33

with Atlassian here, we are seeing

36:36

an increase in spend basically in growth,

36:38

right? So the way our recession is

36:40

typically defined is two quarters of negative

36:42

growth, right? We had six to seven

36:45

quarters of decelerating or

36:47

negative growth in SAS, in

36:49

tech, which is why I called it the

36:52

SAS depression or B that, yeah, it was actually

36:54

kind of a depression, you're right. But

36:56

now we're seeing quarter over quarter growth.

36:59

So growth is reaccelerating. Growth

37:02

is higher than it was. So is he going to

37:04

get to where it was? That probably will take some

37:06

time, but it feels like the

37:08

problems in the ecosystem worked themselves out and now

37:10

we're back to growth again. Yeah, I can

37:12

add psychologically, because I'm on a couple of SAS

37:14

boards as well. And psychologically, it felt like you

37:17

tell me if I'm right, SAS, if you saw

37:19

the same thing. There were two years

37:21

of calling up customers. And they were like, we're,

37:23

we're consolidating vendors. And by the way, we did

37:25

a riff. And so we need

37:27

20% less seats. So we're going to have

37:29

20% less SAS

37:31

companies that we're buying from, and we're going

37:33

to have 20% less seats. So you start

37:36

putting that all together, man, everybody

37:38

was just in psychological free age mode, we cannot

37:40

spend money, I don't want to lose my job. So

37:42

you're if you're a procurement person, you're the CTO, you

37:44

don't want to lose your job, you don't want

37:46

to have more cuts. So you're like, well, I can

37:48

cut some software costs. Do I get

37:51

points for that? And the point you would score

37:53

for the last two years was cutting costs. But

37:56

the market ripping, and

37:58

you now got a really, you know, efficient

38:00

company, you're like, hey, can we spend a little

38:02

bit on SaaS to make the remaining employees even

38:04

more, you know, productive?

38:06

Okay, maybe that's a reasonable discussion. And then

38:09

people are playing ball in

38:11

terms of negotiating prices. So that's the

38:13

other thing I see is like, people are like, well, we'll take

38:15

your software, but here's what we want to pay. And then they're

38:17

coming to the board and saying, can we do this deal? Would

38:19

have been a million dollar deal, but it's a $200,000. Again, take

38:21

the money. Take the

38:23

money. Let's, let's bear hug that customer. The

38:26

market is generally an escalator on the way

38:28

up and elevator on the way down. So

38:30

the recovery is going to take a long

38:32

time, but at least we bottomed out and

38:34

we're in recovery as opposed to continuing declines.

38:37

Yeah, by the same token, if you're

38:40

a startup and you're not seeing improvement

38:42

in your Q4 sales, then you no

38:44

longer have a macro excuse for why

38:46

you're not doing well. Interesting. And

38:48

then Freeberg, you added, you know, you're

38:51

like, I'll make my own software. You said,

38:54

you know, some software is too expensive. I'll

38:57

put a developer on it. And so how's that working out

38:59

for you? Are you still in that mindset of like, yeah,

39:01

maybe we just build our own software? Yeah, I

39:04

mean, I, it's not just us. I think

39:06

we're seeing a lot of companies pursuing

39:08

this path. A couple engineers

39:10

can rebuild the functionality of

39:12

core applications, particularly because I

39:14

think if you think about

39:16

the business model that makes SAS so great, is

39:19

they could value share rather than

39:21

charge the cost of an engineer

39:24

plus some margin, the

39:26

great business model, the equity value that comes

39:28

in software, you can build

39:30

something once that creates $100 of

39:32

value, you could probably charge your customer

39:34

$30, $40 for that product, because

39:37

it's saving them 60 bucks, 70 bucks, and they'll

39:39

make that switch to software. So,

39:41

you know, the ROI driven value

39:44

share model in SAS

39:46

has made it incredibly valuable. The

39:48

problem now is that

39:50

an engineer can be hired to

39:52

build the replacement. And so it

39:54

creates price compression. So the SAS

39:56

company can no longer capture that much value,

39:59

because the savings is actually less than that.

40:01

Because the enterprise might say, hey, I'm going

40:03

to hire someone. And instead of spending 60

40:05

grand a year on your software, I'm going to

40:07

allocate a quarter of an engineer's time to build that

40:10

software. And it's going to replace that

40:12

cost. So I think that that's still the case.

40:14

So while there might be bookings, there's still, which

40:17

are driven largely by a search for efficiency

40:20

gains, a search for more profitability,

40:22

for more productivity within an enterprise. There

40:24

are other options for that enterprise to

40:26

realize that productivity gain today. And

40:29

that's what's going to cause, perhaps,

40:31

price compression and more competition

40:33

than has been the case. But I don't

40:35

think that the adoption of software is going

40:37

to slow down. It certainly seems to be

40:39

re-accelerating, which is great. More competitive,

40:41

right? We're moving into a hyper-competitive market, especially

40:44

with AI. It's a mix of internal software.

40:46

It's a mix of internal software and a

40:48

cool. As you guys know, there

40:50

are very few traditional non-tech enterprises now that

40:52

don't have a software team that can write

40:54

code. Now that so many companies

40:57

have software teams that write code, they're all going to be

40:59

asking the question, should we be buying

41:01

this software, or should we be building something

41:03

internal? Yep. It's a classic buyer build situation.

41:05

All right. Let's talk a little bit about VCs

41:08

and how they're investing in AI. There seems to be

41:10

three camps shaping up here, Tama. One

41:12

group is like, the incumbents are going to

41:14

win. Microsoft, Google, Amazon, everybody,

41:17

they're going to win the day. So

41:19

they're going to wait and see. Then

41:22

there's another group who's sitting it out because they're

41:24

like, hey, open source is going to win. Meta

41:26

is committed to open source and

41:29

collaborative platforms. I've been playing with

41:31

Hugging Face with Sandeep

41:34

as well, as Uchima, and it's pretty amazing

41:36

what's happening over there. And then a

41:38

bunch are obviously placing bets right now. The

41:40

valuations are absurd. Founders

41:42

Fund and Andreessen Horowitz, two notable firms,

41:45

are approaching it differently. Founders Fund

41:47

bought into open AI at a $29 billion valuation. But

41:51

aside from that investment, they're generally

41:53

avoiding the AI deals. My other hand,

41:55

Andreessen, is betting heavily.

41:58

Character AI, Replit, 11L. You're

42:01

also in Riplet, Saxton. So

42:04

what do you think? Is open source going to

42:06

win the day? You've been picks and shovels the whole way.

42:08

You've been talking about compression. Maybe this isn't actually a

42:11

good market. What

42:13

you're thinking as a capital allocator, Chamath? I

42:15

think foundational models will have

42:17

no economic value. I think that they

42:19

will be an incredibly powerful part of

42:22

the substrate and

42:24

they will be broadly available and

42:26

entirely free. Wow. If

42:28

you think about that, any closed

42:30

model, especially a closed model that

42:33

operates on the open internet

42:35

is not very valuable. Any

42:39

open source model that trains

42:41

on the open internet will

42:44

make that so. In

42:46

that world, things like Mistral and

42:49

LAMA will

42:51

essentially decay the market to zero. If

42:55

you're looking at any economic value that has been captured

42:57

up until today, if it has

42:59

been captured by having a

43:01

proprietary closed model trained

43:04

on open data, that

43:07

economic value will go away. I

43:09

think Google and Microsoft and Facebook

43:11

and Amazon and

43:14

all these startups have a deep

43:16

economic incentive actually to make that

43:18

so. So now you

43:20

can evaluate what that means. If

43:23

you get an open model from Hugging Face that's

43:25

just kick ass, where do you

43:27

spend money? Well, you're going

43:29

to have to spend money to actually train it,

43:33

to fine tune it, maybe to have

43:35

some pretty zippy inference. All

43:39

of that means that there's a new kind of substrate

43:41

that has to be built, which is all around the

43:44

way that the tokens per second are provisioned

43:47

to the apps that sit on top of the model. What

43:49

that means is you need to go back to 2006 and 2007 and say,

43:51

okay, when we first created

43:54

the cloud, who

43:56

made money? And fast forward

43:58

18 years later. It's the

44:00

same people that are still making money. So

44:03

the people that made money in 2006 and 2007 were Amazon,

44:05

principally, because of EC2 and

44:09

S3. The perfect

44:11

analogy of EC2 and S3 in 2024 is the token per

44:13

second provider. Now

44:18

there you have to double click and say, okay, well,

44:20

what does a token per second provider need to do to

44:22

make a lot of money? And

44:25

I think the ultimate answer is you need your own

44:27

proprietary hardware. So who is in a position to do

44:29

that? Amazon has announced that

44:31

they have an inference and training

44:33

solution. For training,

44:35

Cerebras has announced a pretty compelling solution.

44:38

Google obviously has TPU. Then

44:40

there's a handful of startups, including one that I helped get off

44:42

the ground in 2016 that

44:44

I funded called Grok. All

44:47

of those companies are in a position to

44:49

build a tokens per second service. Then

44:52

you have companies like Together AI, which basically just

44:54

go and take venture money and

44:56

wrap NVIDIA

44:58

GPUs. And

45:01

you can debate what the advantage will be there.

45:03

One could say, well, it's not

45:06

really a huge advantage over time. So

45:09

my refined thoughts today are

45:11

sort of what my initial guess was when

45:14

we started talking about AI a year ago, which

45:17

is the picks and shovels providers can make a

45:19

ton of money. And

45:21

the people that own proprietary data can

45:23

make a ton of money. But I think

45:25

open source models will basically crush

45:28

the value of models to zero economically, even

45:30

though the utility will go to infinity, the

45:32

economic value will go to zero. Did any of

45:35

you guys see Chamath's interview with Jonathan Ross? No,

45:38

not yet. You put it out, right, Chamath? You

45:40

made it public? You know, I did it just

45:42

for my subscribers, but Jonathan is the founder and

45:45

CEO of Grok, the company that I just mentioned.

45:47

And the quick version

45:49

of that story is I would pour

45:52

over the Google earnings results in

45:54

the mid teens of 2000 because I was pretty

45:56

actively investing in a bunch of different Public

45:59

Equities. And Sundar said in a

46:01

press release he mentioned that they had ruled

46:03

they're on silicon for machine learning. I'll

46:06

keep you. And. I

46:08

was like what is going on that Google

46:10

thinks. That they. Can.

46:12

Actually roll their own silicon? What must they

46:14

know? That the rest of us don't know. And.

46:17

So it took me about six or nine months. but

46:19

from sunny. I. Got introduced to

46:21

Jonathan and then. We.

46:23

Were able to get Jonathan to leave Google and he

46:25

started and you Jonathan was a founder of Tp You

46:28

at Google. And then he started

46:30

rock which I was able to. Lead.

46:32

That funding round and. Since. Any sixteen

46:34

so years ago. I

46:37

was. I get a. A space. as with

46:39

Jonathan talking about the entire ai landscape

46:42

and ai acceleration to my subscribers. but

46:44

it was so good I gotta say

46:46

he is. He. Was so

46:48

impressive. That

46:50

we kind of like. Figured. Out

46:52

a way to just. Play. The

46:54

space. And tape it and

46:56

then we published it to everybody. So it's it's

46:59

on, it's on my twitter for anybody that wants

47:01

to was no Will get is awesome Amazing He

47:03

is. Really? Impressive. I.

47:05

Was sitting on a seventeen going

47:07

to Santa Cruz Rock Moving. For.

47:10

Of hour and a half and I listen to

47:12

it says I've kept me alive. Gotta

47:14

I gotta wonder what are you saying. He's.

47:16

Great know. He's. Got some great.

47:19

Insights and I think he's very

47:21

compelling. And arguing

47:23

why? Some. Of the big. Cloud

47:26

providers today. That. Are

47:28

offering infrastructure for a I've. Modeled.

47:31

Training and incidents. Are.

47:33

Going to be challenge if someone can build

47:36

full stack. And be six and

47:38

do it successfully. So. It was

47:40

a really good interview. I actually think it's really worth listening.

47:42

Thera. But. I

47:44

enjoyed it. Yeah, thanks for printed out there. I was

47:46

like literally just city are sitting in a car browsing,

47:48

Put her and I saw your thing and I clicked

47:51

on it. much of sets it up when I was

47:53

a little. It's a little hard actually want to do

47:55

a space for your subs? You. Can't actually

47:57

just. Flip. A switch and and release it

47:59

to. All of your followers. So

48:01

we actually had to like. Literally.

48:04

Play it. And. Then just

48:06

capture the audio out and then republish

48:08

it. But anyways, despite that inconvenience, If

48:11

anybody's interested in learning about Ai hardware,

48:13

he is very compelling and he's very

48:16

educational. Two sacks your thoughts on just

48:18

how you're approaching investing in a I

48:20

if you're specifically investing in. The.

48:22

Underpinnings of ai picks and shovels, yada, yada

48:24

or of your to suck on the application

48:26

level and it's. Finances. You

48:28

know, that kind of the front. While. We

48:31

we devise it the space into three categories.

48:33

That one is with them, the Miles himself

48:35

a foundation models which can be either. Open.

48:38

Source: Close Source. There's.

48:41

Infrastructure So much more. Saying to

48:43

be like Model training. They

48:46

can be vector databases. Tools.

48:48

That developers use. To.

48:51

Create the I started. Really, it's their

48:53

enterprise. And then the third would be

48:55

applications which can be things like whole pilots or

48:57

to be a free. A. I

48:59

that's using a I tell

49:02

a turbocharged. His capabilities.

49:04

Yeah most ass would be in

49:07

the. Application Bucket and says principally

49:09

we're we're focused. although we do look

49:11

at. Infrastructure. Plays and Models.

49:13

however I do think there is an

49:15

argument for and we really were with

49:18

a question of. Come. Out as

49:20

a symbol like all the model companies to

49:22

get to like monetized. Really?

49:24

Were time I open a i write because of the

49:26

leader. So the question is can they maintain their lead?

49:29

I. Do think there is an

49:31

argument? Is that open A

49:33

I will. Say.

49:35

And lead and actually do quite well. And.

49:39

I think as a few points their one is.

49:41

That. If you're a consumer you just when he

49:44

is the best gp t you want to use google.

49:46

He got his as it's just like search right if

49:48

google is a little better. Are. The factors

49:51

as a little better than being or up

49:53

the other search engines. You. don't

49:55

win a plurality of search traffic you actually

49:57

and of winning at all because consumers just

49:59

want the best one. So most

50:02

of the tests show that OpenAI is still

50:04

ahead of the open source models, and

50:06

I think even people in the open source movement

50:08

will tell you that OpenAI is, call it

50:10

six months ahead. They have no doubt

50:13

that open source will get to where OpenAI

50:15

is now in six months. Nonetheless,

50:17

if OpenAI just maintains a little bit

50:19

of a lead over

50:22

open source, then it could

50:24

compound. It can

50:27

basically win the vast, vast majority

50:29

of the call it consumer search

50:31

or consumer GPT market. So

50:33

that's point number one. Point number two is,

50:36

now that OpenAI has these

50:38

hundreds of millions of consumers

50:40

using it, that's a pretty

50:42

attractive audience for developers to want

50:45

to reach. And OpenAI

50:47

has done a really good job creating

50:49

a platform for developers to

50:51

create what are called custom GPTs.

50:54

So most developers don't

50:56

want to go through the hassle of

50:59

training a model, fine tuning a model, doing all

51:01

of that work that you have to do in

51:03

the open source ecosystem. They just want

51:05

to point chat GPT

51:08

at a repository of

51:11

data or documents, information, have it

51:13

learn what it needs to learn,

51:15

fine tune it in that way,

51:18

maybe add some lightweight functionality using

51:20

OpenAI's platform to create a

51:22

custom GPT. That's what I think most developers

51:24

want is they just want a

51:26

simple stack to work with. And

51:28

they're going to prize, again, simplicity

51:31

and the power of the

51:33

developer tools over the theoretical

51:35

control they get by rolling

51:37

their own models, training and fine tuning their own models

51:39

in open source. And so I

51:41

think what you're seeing now is, I mean, how many

51:43

custom GPTs have already been created on the platform? It

51:46

might be tens of thousands. I mean, there's

51:49

so many millions. Yeah, so easy to create

51:51

them in. So you have a classic developer

51:53

network effect where you've got open AI aggregating

51:55

hundreds of millions of consumers, because they perceive

51:58

that chat GPT is the best. then you've

52:00

got developers wanting to reach that audience.

52:02

So they build custom GPTs on

52:05

the OpenAI platform that actually gives

52:07

chat GPT more capability. And

52:10

that's something that open source can't easily catch up with.

52:12

Well, actually, actually. Now, let

52:14

me just finish the point. So it

52:17

is a flywheel where, you know, classic

52:19

operating system developer

52:21

network effect where you want to use the operating

52:23

system as the most programs

52:26

written for it. And

52:28

interestingly, Hugging Face has realized this, and

52:30

Hugging Face released this week their own

52:33

version of GPTs, which is really interesting.

52:36

And you can pick SACS, which

52:38

open source project you want to use to make

52:40

it. So unlike GPTs on chat GPT, we have

52:42

to pick theirs. On

52:44

the Hugging Face one, you could pick,

52:46

you know, LAMA or whichever one you

52:49

want. There's an account called artificial analysis

52:51

that you can follow. The thing

52:53

to keep in mind, SACS, is that for any of

52:55

this to be true, these APIs need to be usable,

52:57

right? I mean, I don't know if

53:00

you remember, but when we were building apps,

53:02

even as back as the late

53:04

2000s and early 2010s,

53:07

one of the things was there was a pretty important

53:09

paper that was published by Google about attention span. And

53:12

it would look at page load times in

53:14

a cold cache environment, right? And it

53:16

basically said you have to be at like 150 milliseconds, right?

53:20

That's like best in class performance

53:22

or faster. And I remember when we

53:24

read that at Facebook, we went crazy. So

53:27

much so that at one point, a small team and

53:29

I kind of actually launched a

53:31

stripped down version of Facebook to compete with

53:33

Facebook. If there's a, Nick, you can

53:35

probably find this article on TechCrunch. And we did

53:37

it without telling everybody it was called like Facebook

53:39

Zero. Anyways, the point is speed matters. Because

53:42

in the absence of having very snappy response,

53:44

you could have the best model in the

53:46

world. But if it takes 10,

53:48

20, 30 seconds to basically initiate

53:50

and get back data from a fetch request,

53:53

it's an impossible thing to do. So

53:56

I think one of the things that you have to keep in mind

53:58

is that there are these two things that need to move

54:00

at the same time. One is the quality

54:02

of how the model is, but two is

54:05

the speed and its responsiveness, which is a

54:07

function of, again, hardware and your ability to

54:09

basically tokenize tokens per second

54:11

very, very quickly. So that developers are

54:13

incentivized to not just play around in

54:15

a sandbox, but to actually build

54:17

production code. And I don't think we've

54:19

seen that second thing happen, because nobody is delivering it.

54:21

And that's the big thing that nobody talks about. For

54:24

example, like AWS, if you look inside of

54:26

how expensive it is to build an app

54:28

there, I've tried, even when

54:30

they give you credits, the credits they give you aren't

54:33

sufficient enough to even pay for half the power.

54:36

And then the way that they schedule, and the way

54:38

that they try to orchestrate you to use hardware, makes

54:41

building production apps unless you are willing to

54:43

spend millions and millions of dollars for a

54:46

very slow app unfeasible.

54:48

And so if you go back to

54:50

a startup economy raising money here, the

54:53

venture investor should start asking

54:55

the question, well, what

54:58

is the speed and usability of these services

55:00

that I'm funding? And the

55:02

reason is because you could build the best experience in

55:04

the world that runs on localhost. But

55:06

if all of a sudden you actually try to launch it as an

55:09

app, and the thing takes 35 and 40 seconds

55:11

to generate something, it's DOA. And

55:14

I don't think enough people ask those questions or

55:16

understand that that's true. So this is why I

55:18

think you have to sort of be looking at

55:20

both of these two things at the same time.

55:23

But this account is interesting

55:25

because it kind of just strips

55:27

things down to the bare facts,

55:30

and they start to allow you as a

55:33

third party to understand what you

55:35

can do. Speed is

55:37

just such a critical component of this. And

55:39

what Google found was, as you

55:41

know, free brokers, you were there. Every time

55:43

they lowered a certain number of milliseconds, usage went

55:45

up, right? People did more searches, which makes sense

55:47

if you get your results back faster. Yeah,

55:50

it was a key metric from day one

55:52

at Google. Marissa Mayer, she ran all the

55:55

consumer-facing product, she was like, I'm gonna take

55:57

Google during this earlier era. beat

56:01

it into the team. I mean, if you guys remember, one of the

56:03

first, the first kind of

56:05

early feature of the Google results page was the

56:07

amount of time it took to load the results.

56:09

They'd show you how many milliseconds it took. Yeah,

56:11

they'd show you that. Yeah, they literally put your

56:13

North Star metric exposed to the consumer, which that

56:15

must have lit a fire under the asses of

56:17

all the developers and server people, yeah? Well, I

56:19

mean, they were kind of showing off the quality

56:21

of the infrastructure and the way they did indexing

56:23

and everything, but the result really

56:26

played out in usage. The faster the results,

56:28

the more frequently you would use the search

56:30

engine and the more likely you were to

56:32

come back. And it's amazing how much consumer

56:34

behavior drifts based

56:36

on milliseconds. Like, you have a few milliseconds of

56:39

the way. McDonald's learned this, right? I mean, if

56:41

you look at the, if you ever see the

56:43

movie The Founder where they explain the McDonald's process,

56:45

they learned it too. Guys, look at this. This

56:48

is really interesting on this analysis. I mean, Chamath,

56:50

are you saying that you don't think OpenAI can

56:52

achieve the necessary levels of performance? No,

56:55

I'm saying two things. OpenAI is three different

56:57

businesses. OpenAI has a closed model

56:59

that's trained on the open Internet. I

57:02

think economically it's going to be very hard to

57:04

sustain that unless they start buying all

57:06

number of apps so that they

57:08

can get some fine tunes that they control that

57:10

are proprietary to them. So, for example, if OpenAI

57:12

were to buy all of Reddit, that

57:15

would be a really interesting development that

57:17

would improve the quality of OpenAI

57:20

in a unique and differentiated way

57:23

relative to where things like Wama and Mistral will

57:25

get to at the same time, as well as

57:27

X's Grok. I think they're

57:29

all going to converge to the

57:31

same quality in the next probably

57:33

12 to 18 months. That's point

57:35

number one. Your belief there is there's enough

57:37

data in those pools that everybody reaches parity.

57:40

No, did you guys? Okay, Nick, did you?

57:42

So, I published this primer on AI.

57:44

Yeah, we saw the primer. Yeah. There

57:47

is a slide in there, Nick, that you can pull out, but

57:49

it just shows you that there is

57:51

a converging in the

57:53

quality of the results as

57:55

the number of the parameters of the model

57:57

gets higher and higher. And what it effectively

57:59

shows you is that we are already in the land

58:02

of diminishing returns when

58:04

models are trained on the same underlying

58:06

data. So if you are using the

58:08

open internet, llama, Mistral,

58:11

OpenAI, they're all getting to the same quality

58:13

code point and they will be there within

58:15

the next six to nine months. So

58:18

that's business number one on OpenAI. Business

58:20

number two is a consumer facing app

58:22

called ChatGPT. That has a

58:24

lot of legs because I think people are,

58:26

you know, develop habits, it'll be very sticky,

58:29

and I think it'll get better and better. And

58:31

then the third business that they're in

58:33

is selling enterprise services to

58:36

large Fortune 500s. In fact, if you look at

58:38

their OpenAI days, what they talk about is they

58:40

sell, they've sold already to like 94% of the

58:42

Fortune 500. What

58:45

does that mean? I think what that actually means

58:47

is they've sold a lot of test environments and

58:49

sandboxing. But again, in order

58:51

to translate that into functional production

58:54

code that's used by Bank of

58:56

America, right, or Boeing in production,

58:59

you have to have zippy, zippy

59:01

fast SLAs and a level of

59:04

performance that no cloud

59:07

provider yet has delivered, none,

59:09

nobody. So Nick, if you just go

59:11

to that, please, the thing, I just wanted

59:13

to show you this, because it's really interesting. Sure, this is

59:15

not mine, this is theirs. If you look at quality versus

59:17

price acts, it starts to show you

59:19

like, where do you want to be? You

59:22

want to be in the upper left

59:24

quadrant in their analysis, right?

59:28

And so the point is, what you can see

59:30

is that a ton of different models are

59:32

getting to this same place. And

59:35

so obviously, you'd want to use the model that's

59:37

the cheapest, or most convenient. Who's

59:40

going to pay for that? If you and your LPs want

59:43

to pay for that, the

59:45

person that figures out the way that it's the cheapest

59:47

to give you the same answer will actually end up

59:50

winning because you will run out of money and they

59:52

will not. I don't know.

59:54

I mean, I think that there's a lot

59:56

of business problems inside companies where people just

59:58

want to very quickly set. up their

1:00:01

own, again, custom GPT without having

1:00:03

to go through the time,

1:00:07

the cost, the hassle of trying to do

1:00:09

model training or fine tuning. So

1:00:11

let's just back up. Here's the path that

1:00:13

OpenAI is on. So step

1:00:15

one, get hundreds of millions of

1:00:17

consumers using it and getting them

1:00:20

to view OpenAI or chat

1:00:22

GPT as the Google in this area, right?

1:00:24

Long presumption, this is just the one you

1:00:27

go to when you have a question.

1:00:31

Step two, these same

1:00:33

people, these same consumers now want to

1:00:35

use chat GPT at work because there's

1:00:37

some research they want to do. So

1:00:40

OpenAI has just rolled out both

1:00:43

enterprise licenses and team workspaces.

1:00:45

So you can work collaboratively on the

1:00:47

same queries in a team workspace. Step

1:00:50

three is rolling out a very easy

1:00:52

to use dev platform that allows developers

1:00:54

to again create custom GPTs

1:00:56

by just pointing OpenAI

1:00:59

at repositories. And

1:01:02

so let's say that you're the customer

1:01:05

support team and you

1:01:07

want to create a GPT

1:01:10

to help customer support answer cases. You

1:01:14

could basically then train

1:01:18

chat GPT on let's say

1:01:20

every customer support ticket

1:01:23

and email that

1:01:26

the company has ever produced, right? Now

1:01:28

you could wait for the company's IT

1:01:31

department to get us to act together

1:01:33

and figure out how to train an

1:01:35

open source model on the same thing. But

1:01:38

do you really want to wait for that or do you just want to get going? And

1:01:41

now OpenAI has given you the enterprise

1:01:43

license that you need to

1:01:47

pacify the concerns about security and privacy

1:01:49

and all that kind of things to

1:01:51

some degree. There's always going to be

1:01:53

those super paranoid Fortune 500 companies that

1:01:55

will insist on owning

1:01:58

everything and doing it. doing it

1:02:00

open source. Let me build on your example.

1:02:02

So I run a small software company during

1:02:05

the day called Hustle. And we

1:02:08

saw a lot of tickets related

1:02:11

to this specific legislation

1:02:13

that exists whenever you're texting

1:02:16

or you're doing auto dialing stuff called

1:02:18

10 DLC. And

1:02:20

so we wanted to

1:02:22

eliminate those tickets, right? So I actually

1:02:25

went and I built a GPT, which

1:02:28

is called the Privacy Policy Generator, because a

1:02:30

lot of these trouble tickets were because the

1:02:32

privacy policies were bad. And

1:02:34

we trained them using a

1:02:37

handful of ones that were good and a handful of ones that

1:02:39

are bad with a bunch of rules. And I

1:02:41

trained them all. And it's wonderful,

1:02:44

except I can't run it in production, because

1:02:47

it's not the kind of thing that is

1:02:50

usable in that way right now. It's still

1:02:52

very difficult. And so all I'm saying

1:02:54

is, I'm happy to keep spending a few hundred dollars

1:02:56

a month, a few thousand bucks a month, whatever it

1:02:58

is that I'm spending, I don't quite exactly know. And

1:03:02

I agree with you, it was very easy.

1:03:04

I think OpenAI does an excellent job of

1:03:07

getting off the ground. But what I'm also

1:03:09

saying is that when you

1:03:11

actually translate that into a mainline

1:03:14

use case, right, where

1:03:16

I want to now give it to my

1:03:18

support team and say, this is now a

1:03:20

tool you can rely on, it's integrated into

1:03:22

your workflow, into your other tools, it's integrated

1:03:24

into how you pipe out data into Salesforce

1:03:26

or what have you. It's

1:03:29

just very hard. And I'm not saying

1:03:31

it's not going to get fixed. I'm saying we're just

1:03:33

not there yet. And one of the ways in

1:03:35

which it's not there is that there

1:03:38

is no place I can go, including

1:03:40

OpenAI, that actually makes it

1:03:42

fast enough to be usable in production. You

1:03:44

wrote this on OpenAI Stack?

1:03:46

You wrote a custom GPT? Yeah, built

1:03:48

myself. Yeah, I think you

1:03:50

could do the monohugging face now. It's going to be

1:03:52

a lot of options. In terms of integrating into your

1:03:54

workflows, I think it's a really interesting point because I

1:03:56

saw a demo somewhere

1:03:59

where now now, actually

1:04:01

I think OpenAI announced this, that you can

1:04:03

at mention a custom GPT.

1:04:05

Yeah, yeah, Sunny showed me that this

1:04:07

week on the pondoon. Yeah, in chat

1:04:09

GPT, you can now at mention a

1:04:11

custom GPT to kind of invoke it. Yeah,

1:04:14

so how it works is you'd say, hey, I'm

1:04:16

heading to New York, what flights can

1:04:19

I get at Expedia, at Kayak, whatever, and

1:04:21

then it gives you, you know,

1:04:24

the results here and you're kind of pulling

1:04:26

that up. Just to the point about

1:04:28

where data advantages lie and that's ultimately going

1:04:30

to drive value. I

1:04:33

cannot, I've tried to think a lot about this,

1:04:35

I cannot think about a

1:04:37

better data advantage that

1:04:41

is orders of magnitude better than

1:04:43

anything else. Say YouTube. Say YouTube.

1:04:46

Yeah. YouTube. It is.

1:04:49

So here's the numbers. I pulled this up. You

1:04:52

guys know like GPT-3 and three and a half were

1:04:54

trained with a heavy weighting on common

1:04:56

crawl, which is this open source. Yeah, we

1:04:58

talked about this before Gil Elbaz runs it

1:05:00

open source crawling of

1:05:02

the web. The total amount of data in

1:05:04

common crawl, which I think accounted and I could be off

1:05:06

on this something like 40 to 60% of the weighting in

1:05:09

GPT-3 or 3.5. I'm

1:05:12

off on this probably. So the total amount of data in

1:05:14

that common crawl data set is about 10 petabytes.

1:05:18

Okay. Based on

1:05:20

YouTube's public statement recently,

1:05:23

they're seeing about 500 hours a minute of the

1:05:26

video uploaded or 720,000 hours a day. And

1:05:30

if you assume somewhere between, you know,

1:05:32

just under 1080p on that video, we're

1:05:34

talking about probably one

1:05:37

to two petabytes of data being

1:05:39

uploaded to YouTube per day.

1:05:42

So if you assume like over time, the

1:05:44

definition of the video has gotten better and

1:05:46

the amount of uploads gotten up, you

1:05:49

could probably assume that there's roughly

1:05:51

I'm guessing there's probably somewhere between

1:05:53

2000 and 3000 petabytes

1:05:55

of data in YouTube

1:05:58

growing by. one to two petabytes

1:06:01

per day, which makes YouTube

1:06:03

data repository 300 times larger than

1:06:06

common crawl, which makes it bigger than

1:06:09

anything else that anyone else has. And here's the amazing

1:06:11

thing about it. It has

1:06:13

video, it has image, it

1:06:15

has audio, it has text, it has

1:06:17

everything. It's multi-minute. And it is growing.

1:06:19

So if you were to take a

1:06:21

bet or build a thesis around this

1:06:23

point that the data advantage is going

1:06:25

to drive value creation, if Google

1:06:28

gets its act together and leverages the

1:06:30

data repository at YouTube, it is an

1:06:32

insurmountable moat that will only continue to

1:06:34

extend because the quality of the YouTube

1:06:36

experience and the network effects continue to

1:06:39

accumulate for them. So I think it's

1:06:41

the most valuable asset in the world

1:06:43

today, based on this thesis, that

1:06:45

AI value is going to accrue to the data

1:06:47

owner. I think you're making such an important point.

1:06:50

This is why the counterfactual is

1:06:52

true, and it's actually showing up in the

1:06:54

data. And Nick will show you this

1:06:56

slide again from the AI primer. But that

1:06:59

is why we're seeing these diminishing returns for

1:07:01

you, Bergen, all of these third-party benchmarks of

1:07:03

these models. Using the same data sets. It's

1:07:05

all using the same data sets. So what

1:07:07

we are proving is not that the underlying

1:07:09

hardware can't scale, nor that transformers are only

1:07:11

efficient to a point. That's not what all

1:07:13

of this convergence is showing. It's

1:07:15

that in the absence of proprietary data, you're just going

1:07:18

to get to the same model quality. And

1:07:20

we're seeing a bunch of different models get

1:07:23

to a very early finish line, which, again,

1:07:25

if people like Facebook are doing for

1:07:27

free, that's much easier

1:07:30

to underwrite because you don't have

1:07:32

to underwrite it being a differentiator

1:07:35

in five years. But if you have a startup

1:07:37

with equity value tied to a model, I

1:07:40

think it's very... It's

1:07:42

much more of a tenuous place to be in the

1:07:44

absence of proprietary data. And everyone

1:07:46

in the world has a camera

1:07:48

and a microphone in their pocket and

1:07:51

high-speed internet now from the phone

1:07:53

in their pocket. And more and more

1:07:55

people are uploading that content, that data that's

1:07:57

being generated. free

1:08:00

data vacuum and it's just out in the

1:08:02

world and most of it's getting up well

1:08:04

it is public facing though so it's not

1:08:06

just true for text it's also true for

1:08:08

you know all of the image generation so

1:08:10

like if you look they can train more

1:08:12

than just an LLM on it right they

1:08:15

can build all sorts of

1:08:17

yeah go ahead no no I was

1:08:19

just gonna say like the version of

1:08:21

common crawl for training these image models

1:08:23

also exists and so to your point

1:08:25

it's like we are all operating from

1:08:27

the same brittle very fixed small quantum

1:08:30

of training information and

1:08:33

so that is why I think like Facebook

1:08:36

and Google are doing a really

1:08:38

important job by deciding that these models

1:08:40

should be free right

1:08:43

and then being able to so then

1:08:45

the question that just accentuates their data

1:08:47

advantage it does and

1:08:49

and I think that it allows them to

1:08:52

decide how much to leak

1:08:54

out so for example whenever like if

1:08:56

you were using a lot of Google

1:08:58

services like GFS big table BigQuery you

1:09:01

know TensorFlow the

1:09:03

versions that you had access to via GCP

1:09:07

was always one or two generations behind

1:09:09

what the Google employees got to use

1:09:11

right but it

1:09:13

was still so much better than anything else that we

1:09:15

could get anywhere else that you would still build to

1:09:17

those endpoints and I think there's a

1:09:20

similar version of this where Facebook and Google probably

1:09:22

realized like look we'll have version

1:09:25

five running internally to optimize ads and all

1:09:27

of this other stuff that makes our business

1:09:29

that much better and we'll expose

1:09:31

version three to the public but version

1:09:33

three is still trained on so much proprietary

1:09:35

data that it's so much better than version

1:09:37

10 and anything else that's just operating on

1:09:39

the open internet and

1:09:42

you know to your point freeberg that's

1:09:44

the outward-facing stuff YouTube is a collection

1:09:46

of things people want to share what

1:09:49

Google also has is Google Docs

1:09:51

and Gmail things that people say

1:09:53

privately so they have a another

1:09:55

data resource there that they can

1:09:57

tap you know and there'll be

1:09:59

regulation and privacy around that, but maybe there's

1:10:01

a difference there. But I honestly can't think

1:10:04

of the quantum coming close to YouTube. Not

1:10:06

even close. Well, the thing to Jason's point,

1:10:08

which is really interesting, is like, you know,

1:10:10

there's a modality in AI called rag, where

1:10:13

you can actually just augment

1:10:15

with very specific training on a very specific

1:10:17

subset of documents to improve. It's like a,

1:10:19

it's like a hacked version of a fine

1:10:22

tune. But the beautiful thing about that is

1:10:24

like, if you have a Google workspace, my

1:10:26

entire company runs on Google workspace. In fact,

1:10:29

most of my companies do at this point, to

1:10:31

click a button, where all of

1:10:33

a sudden now, all of that stuff and

1:10:36

all of my G drives, all of a sudden, is

1:10:38

trainable. So that the n

1:10:40

plus first employee comes in, and has

1:10:43

an agent that's tuned on every deck,

1:10:45

every model, spreadsheet, every

1:10:47

document, that's a huge edge.

1:10:49

Huge edge, by

1:10:51

the way, and as a CEO, if

1:10:53

you gave me that choice, I don't

1:10:56

think anybody underneath that reports to me has

1:10:58

any right to make that decision. But as a CEO,

1:11:00

I would click that button instantly, and I have that

1:11:02

right as a CEO. And so like, that's the CEO

1:11:04

pitch is like, look, I can just give you these

1:11:07

agents that are that are like the

1:11:09

next version of a knowledge base that we've always wanted

1:11:11

inside of a company. Right? notion

1:11:14

has this, you know, they basically

1:11:16

you can start asking your entire

1:11:18

notion instance questions about notion, which

1:11:20

is incredible. And yeah, you

1:11:23

can just add as a CEO, you can

1:11:25

see across everything from off because as you

1:11:27

know, with Google Docs, if you're

1:11:29

in a compliance based industry like finance, you

1:11:32

can see everything every message, every

1:11:35

email, every document, and you can search the

1:11:37

security model and the data model becomes very

1:11:40

complicated in all of that stuff. Like for example, like,

1:11:43

how do you know that this spreadsheet is

1:11:45

actually you should learn on

1:11:47

it. But who gets to actually then

1:11:49

have that added to the

1:11:52

subset of answers, right? All of

1:11:54

a sudden, like salaries, yeah, HR

1:11:56

information, information gets put into the

1:11:58

training model very dangerous. Or

1:12:00

subset A of a company's working on a proprietary

1:12:02

chip design that they actually like the way that

1:12:04

Apple runs highly highly segregated

1:12:07

teams where nobody else can know so

1:12:09

there's all kinds of complicated security and

1:12:11

data model and usage questions there, but

1:12:14

Yeah, brave new world. So there's been a lot

1:12:16

of discussion real estate you shared a video with us Why don't

1:12:18

you kick it off for us here in Freiburg? What's going on

1:12:21

in commercial real estate and SACS? You've got holdings and a lot

1:12:23

of that as well. So let's kick up the

1:12:25

commercial real estate Challenges

1:12:27

of the moment. Well, I mean I

1:12:29

think we're teeing off of Barry's Comments

1:12:32

at this event last week. He and I

1:12:35

met backstage because I spoke right before him

1:12:38

and then he gave this talk Which is

1:12:40

available on YouTube where he talked about the state of

1:12:42

the commercial real estate market and particularly he talked about

1:12:44

the office market just

1:12:47

to take a step back to talk about the scale of

1:12:50

Commercial real estate as an asset class in the

1:12:52

US Nick if you'll pull up this chart The

1:12:55

total estimated market value of commercial real estate

1:12:57

in the US Across

1:13:00

different categories is about 20 trillion

1:13:02

dollars with about three trillion dollars

1:13:04

being in the office market Which is

1:13:06

specifically what he was talking about and he was saying that in

1:13:09

the US We're seeing people not coming

1:13:11

back to work and all these offices are empty and

1:13:13

we've talked a lot about these offices being written down

1:13:15

So how significant of a problem is this so 20

1:13:18

trillion dollar asset class? Obviously the multifamily market

1:13:20

is probably not as bad as office and

1:13:22

retail Which are the most heavily affected each

1:13:25

of which are about three trillion dollars a

1:13:28

piece the rest of

1:13:30

these categories seem relatively unscathed in

1:13:33

comparison industrial hospitality Healthcare,

1:13:36

you know those those real estate sectors

1:13:38

are probably pretty strong data centers Obviously

1:13:40

growing like crazy self-storage at a great

1:13:42

market if you pull up the next image So

1:13:45

it turns out that of the 20

1:13:47

trillion dollars of market value There's

1:13:49

about six trillion dollars of debt. So

1:13:52

you can kind of think about that 20 trillion being

1:13:54

six trillion owned by

1:13:57

The debt holders and 14 trillion by the equity

1:13:59

holders And the debt is owned

1:14:02

roughly 50% by

1:14:04

banks and thrifts. And this was this concern

1:14:07

that we've been talking about with higher rates. Is the debt on

1:14:09

office actually going to be able to pay? The debt on retail

1:14:11

going to be able to pay? When half

1:14:13

of that debt is held by banks and thrifts

1:14:15

that as we've talked about have such a close

1:14:18

ratio to deposits

1:14:21

that you can actually see many banks become

1:14:23

technically insolvent if the debt starts

1:14:25

to default. So the 30th point that

1:14:28

he made was if you look at the office market, which

1:14:31

is marked on everyone's

1:14:33

books as $3 trillion of market

1:14:35

value, he thinks it's probably worth closer to

1:14:37

$1.8 trillion. So

1:14:40

there's $1.2 trillion of

1:14:42

loss in the

1:14:44

office category. And

1:14:46

if you assume 40% of that $3 trillion

1:14:48

is held as debt, you're talking about $1.2

1:14:51

trillion of office debt. A

1:14:53

reduction from $3 trillion to $1.8

1:14:56

trillion means that the

1:14:58

equity value has

1:15:00

gone down from $1.8 trillion to $600 billion. So

1:15:04

they've lost equity holders in

1:15:06

office real estate have probably lost two thirds of

1:15:08

their value, two thirds of their

1:15:11

investment. And who owns all of that? Most

1:15:14

of that, 60 plus percent, call

1:15:16

it two thirds of that, is

1:15:18

likely owned by private equity funds and

1:15:21

other institutions where the end beneficiary

1:15:23

is actually pension funds and retirement funds.

1:15:26

And so if two thirds of the value has to

1:15:28

be written off in these books and it hasn't happened

1:15:30

yet, what's going to happen to all these retirement funds?

1:15:32

And this is where going back to my speculation

1:15:35

a couple months ago, kind of gets revisited. If

1:15:37

you're actually talking about a two third write down

1:15:39

on the value in these funds, most of that

1:15:41

being pension funds, you're not going

1:15:43

to see governments let that happen. You're

1:15:46

going to see the federal government. There's going

1:15:48

to be some action at some point. And

1:15:51

it's unlikely the office market is going to

1:15:53

suddenly rebound overnight. If this stays the

1:15:55

way it is, who's going to

1:15:57

fill that hole for retirees and pensioners? because

1:16:00

we're not going to let that all get written down. Someone

1:16:02

is going to step in and say, we've got to do

1:16:04

something about this. And there's going to need to be some

1:16:06

sort of structured solution to support

1:16:08

retirees and pensioners, because that's ultimately who

1:16:10

ends up holding the bag in

1:16:13

this massive write down. He didn't go all the

1:16:15

way there in his statements. He was talking more

1:16:17

about his estimate of $3 trillion to $1.8 trillion.

1:16:19

And then I tried to connect the dots and

1:16:21

what that actually means. And ultimately, there's going to

1:16:23

be some pain felt by retirement funds that's going

1:16:25

to need to be dealt with somehow. So in

1:16:27

fact, I don't know if that sits right with

1:16:29

you. I mean, I think the big picture is

1:16:31

right. I think you're applying a lot of averages.

1:16:34

I think in the office market in particular, the typical

1:16:37

office deal is more like one third equity

1:16:39

and two thirds debt. There's just a lot

1:16:41

more leverage. Right. So that'd

1:16:43

be point number one, which makes the situation

1:16:46

worse. Even worse. Yeah. So I would say

1:16:48

that there's a huge amount of

1:16:50

equity that's been written off. But in

1:16:52

addition to that, there's a lot of

1:16:54

debt holders who are

1:16:57

in trouble too. And that

1:16:59

debt is held by regional banks. So these

1:17:02

commercial loan portfolios are significantly impaired. That's

1:17:04

what we saw with community bank of

1:17:06

New York is that their stock cratered

1:17:09

when they reported higher

1:17:11

than expected losses in their commercial real estate

1:17:14

portfolio. So Freiburg,

1:17:16

I think the point is just the pain

1:17:18

from this is not just going to be

1:17:21

on the equity holders,

1:17:23

but also on these banks, which can't

1:17:26

afford to lose very much. It's not

1:17:28

evenly distributed. Yeah. Right. Yeah. Right. And

1:17:31

we saw this in San Francisco where some of these buildings have

1:17:33

70% debt to equity ratios. And the value puts them in the

1:17:38

hole and equity is wiped out completely and the debt holders

1:17:40

have to take ahead. And normally that

1:17:42

debt is not really written off very often.

1:17:45

Well, this is why the debt holders don't want to foreclose.

1:17:48

They don't want to get these buildings back because when

1:17:50

they do, they're going to have to write down the

1:17:53

loan. As long as the loan

1:17:55

is still outstanding and they haven't foreclosed,

1:17:57

they can pretend that the value of the building is

1:17:59

not imprecise. Paired. Kick the can

1:18:01

down the road is the best strategy for them. So it's

1:18:03

called pretend and extend. So what I

1:18:05

do is they'll work out a deal with

1:18:08

the landlord, the equity holder

1:18:10

that the equity holder will say,

1:18:12

listen, I can't pay the interest. So they'll just tack

1:18:14

on the interest basically as principal at

1:18:16

the end of the loan and they'll

1:18:19

extend out the term of the loan. Which would wipe

1:18:21

out the equity at a certain point, yeah. And

1:18:24

all that. Well, what it does is it

1:18:26

allows the equity holder to stay in control over

1:18:28

the building, right? Because yeah,

1:18:30

the equity holder can't pay make their debt payments today,

1:18:32

but they're going to

1:18:34

postpone those debt payments till the end of

1:18:37

the loan. And

1:18:39

again, in the meantime, just kind of hope that the

1:18:41

market. Couldn't match yet at some point since they have

1:18:43

so little equity in these buildings typically just exceed

1:18:46

the value of the property. And it's like,

1:18:48

I'm just working for the bank now. And

1:18:51

why am I even putting this work in? Because

1:18:53

everyone kind of hopes that the market will recover

1:18:55

the value their equity will go up and

1:18:58

they'll be able to make their debt payments again. So

1:19:01

if you're the equity holder, you'd

1:19:03

rather hold on and have

1:19:05

a chance to your equity being worth something in

1:19:07

recovery, then definitely lose the building. And

1:19:09

if you're a regional bank, you'd rather blend

1:19:12

and extend or pretend and extend as opposed to

1:19:15

having to realize the loss right now

1:19:19

and showing the market that your solvency may

1:19:21

not be as good as you thought. The

1:19:23

same thing happened with government

1:19:25

bonds. For that with SCB and these other

1:19:27

banks, they had these huge held

1:19:30

to maturity bond portfolios. These

1:19:33

are mostly just T-bills that

1:19:36

were worth, I don't know, 60 cents on the dollar when

1:19:38

interest rates spiked from 0 to 5%. But

1:19:42

they didn't have to recognize that loss

1:19:44

as long as they weren't planning

1:19:46

to sell them. Right. And

1:19:48

then when they had the bank run, they had

1:19:50

to sell. Well, yeah, that's right. So when depositors

1:19:52

left because they needed their money or

1:19:54

because there was a run or because they could get

1:19:57

higher rates in a money market fund, all of a

1:19:59

sudden, they were going sudden, these banks have to

1:20:01

sell their health and maturity

1:20:03

portfolios and they have to recognize that loss.

1:20:05

And that's when everyone realizes, wait a second,

1:20:08

they're not actually solving. Okay, so, Chumab, supply

1:20:10

demand matters in real estate. We have a

1:20:12

tale of two cities here on one side

1:20:14

in real estate, for commercial real estate, no

1:20:17

demand for office space, which

1:20:19

is in way too much supply. Paradoxically,

1:20:22

on the other side, we

1:20:24

have this incredible market for developers, which

1:20:26

is, gosh, there's not enough

1:20:28

homes, I think we need 7 million more homes, and

1:20:31

the demand is off the charts for homes, yeah?

1:20:33

Yeah, I mean, I think you're basically right. I

1:20:36

keep trying to explain residential is not a

1:20:38

great market either because interest rates have spiked up.

1:20:40

So, there's not a vacancy problem. Multifamily

1:20:43

developers are still able to lease the

1:20:45

units, they're still able to rent. The problem

1:20:48

is their financing costs have shot through the

1:20:50

roof. So, again,

1:20:52

let's say you were a developer who built multifamily in

1:20:54

the last few years. You took out

1:20:56

a construction loan. That construction loan might

1:20:58

have been at 3%, 4%. Now,

1:21:02

you want to put long-term financing on it. But

1:21:05

if you can even find debt right now, because there's a credit crunch

1:21:07

going on, you might have to pay 8%, 9%, 10%. Yeah,

1:21:10

but at least you can find a renter. You

1:21:13

can find a renter, that's true, but only

1:21:15

at a certain price. And let's say you

1:21:17

unwrote that property to, I don't know, like a

1:21:19

five cap, like a certain yield. But

1:21:22

now, your financing costs are much higher than you thought.

1:21:24

You might be underwater. That

1:21:27

situation isn't as bad as what's happening

1:21:29

in the space. Why? I

1:21:32

think it's worse in some ways. If

1:21:35

you're fully rented and

1:21:37

your building is underwater because now your debt

1:21:39

payments are much higher than you expected, then

1:21:41

there's no business model. Are we seeing that?

1:21:44

Are we seeing tons of multifamily go under?

1:21:46

Can I make two points? One, I think

1:21:48

David is right, which is that I

1:21:51

don't know this market very well, but just as

1:21:54

a bystander, here's what I observed. It

1:21:56

seems that the residential market has a

1:21:59

feature. And I

1:22:01

don't know whether it's good or bad, but that feature is

1:22:03

that you reap price

1:22:05

to market demand every

1:22:07

year. So to the extent that supply demand

1:22:10

is changing and default rates

1:22:12

are up or whatever, that's reflected

1:22:14

in rents. And you see that because

1:22:16

rents change very quickly and most human

1:22:18

beings are signing six months to

1:22:21

one year leases. So that reset happens very quickly

1:22:23

so it can more dynamically adapt. So to the

1:22:25

extent that a market segment

1:22:27

is impaired, you see the impairment quickly.

1:22:31

On the office side, what I see is

1:22:33

that there's been a structural behavior change in

1:22:36

COVID that has reset in every other

1:22:38

part of the world except for the

1:22:41

United States where there are

1:22:43

these, frankly, typically younger, typically

1:22:46

more junior employees that

1:22:48

have held many of these companies hostage in

1:22:50

a bid to return back to office space.

1:22:53

And so we know that there is this

1:22:55

vacancy cliff that's going to hit commercial real

1:22:57

estate. We just don't know

1:22:59

when because they're in long-term leases. They're

1:23:01

canceling these leases over long periods of

1:23:03

time. So the reset cycle is longer.

1:23:06

That's just my observation as an outsider. I don't

1:23:08

know what that means for prices or

1:23:10

anything else, but it just seems that at

1:23:12

least the residential market can find a bottoming

1:23:14

sooner because you can reset prices every

1:23:16

year. But commercial just seems

1:23:18

like a melting ice. direction I correct

1:23:20

you, Sacks? That assessment?

1:23:23

Commercial has both a

1:23:25

demand problem and a financing problem. Multifamily

1:23:27

just has a financing problem, but it's

1:23:29

important to understand. We're talking about office.

1:23:32

There's retail and then there's office

1:23:34

and then there's other industrial. Do you guys see in

1:23:36

China? China has 50 million

1:23:38

homes ahead of

1:23:41

schedule. 50 million additional supply

1:23:43

that can house 150 million people. So

1:23:46

as acute as our issues are, the China issue

1:23:48

might be much, much seismic.

1:23:51

Can I just give you an example on the multifamily

1:23:53

side? Let's say that you bought a building in 2021,

1:23:55

the absolute peak of the market. And

1:24:00

you could get debt at say 4%, okay? And

1:24:03

you penciled out, let's call it

1:24:05

a 6% yield that with

1:24:07

the debt you're getting, so let's say

1:24:09

you did two-thirds debt at

1:24:12

4%, you could now level up that

1:24:14

6% yield to 10%, okay? That's

1:24:17

like sort of the math, right? Now

1:24:20

all of a sudden and to get there, you'd have to do

1:24:22

some value-added work on the property. You have to spruce it up,

1:24:24

okay? Now it's a few years

1:24:26

later and your

1:24:29

short-term financing is running

1:24:31

out and you need to refi. And you've done

1:24:33

your value-added work, but here's the problem. The

1:24:35

overall valuations in the market have come way

1:24:38

down. So before

1:24:40

the bank was willing to give you two-thirds loan

1:24:43

to value, now the values come way down. You

1:24:46

may not even be able to get two-thirds loan to value, so you're going to have

1:24:48

to do what's called an equity in

1:24:50

refinancing. You're going to have to

1:24:52

produce more equity. You're going to have to

1:24:54

pony up more money. So instead of taking equity out, like

1:24:56

when the deal goes well, you're going to have to put

1:24:58

equity in. You may not have that equity if you're the

1:25:00

developer. The other thing is that your

1:25:02

financing costs now might be 10%. So

1:25:05

now you've got negative leverage. You're generating

1:25:08

a 6% yield, but you're borrowing at

1:25:10

10% to generate that 6% yield. So

1:25:13

the debt no longer makes sense. Again,

1:25:16

you're not positively leveraged. You're negatively leveraged. So

1:25:19

you're not going to want to take out that debt. And

1:25:21

if you do take out that debt, the

1:25:23

buildings would be underwater. It's not going to be

1:25:25

generating net operating income. It's going to

1:25:27

be generating losses. So

1:25:30

that's why even categories

1:25:32

like multifamily where

1:25:35

you don't have a vacancy problem, there's strong

1:25:37

demand, those properties still

1:25:39

don't make sense. If you

1:25:41

had long-term debt on your multifamily, if you were

1:25:43

able to lock in that 4% loan for 10

1:25:46

years, you're fine. But

1:25:48

for all the people who are refinancing now, who are

1:25:51

coming up this year, last

1:25:53

year, they're in deep

1:25:55

trouble. And that's why there's a

1:25:57

rolling crisis in real estate is because the debt.

1:26:00

rolls over time. It's not like everybody hits

1:26:02

the wall. It has to refinance at the

1:26:04

same time. Well, thank God, right?

1:26:06

I mean, this would be cataclysmic if it was.

1:26:08

Can you imagine if Silicon Valley

1:26:10

and San Francisco had to say, here's actually

1:26:12

the reality? Anybody want to actually pay for

1:26:14

this office? Call them the same year? That

1:26:18

would be insane. But

1:26:20

the crisis is growing. It's as the

1:26:22

leases roll and those old rents that

1:26:24

were higher the market roll off and now

1:26:26

you have to take on new

1:26:29

leases if you can even get them at a

1:26:31

much lower rate. And as the old loans roll

1:26:34

that were at a much lower interest rate, you have to get

1:26:37

financing even if you get it at a much higher interest rate.

1:26:39

That's when all of a sudden these

1:26:42

buildings go from being basically solvent to

1:26:44

insolvent. Yeah. I mean, Janet

1:26:46

Yellen's just going to bail these folks out. That means

1:26:48

you won't bail out the banks themselves, but you'll bail

1:26:51

out the creditors, obviously, the people holding the

1:26:53

bag. They'll get bailed. Yeah. That's

1:26:55

everybody agrees. Janet Yellen. Yellen.

1:26:58

Our treasury secretary. I don't know

1:27:00

if she's going to be the one to do it. I think there's

1:27:02

going to be congressional action on this stuff. Yeah.

1:27:05

I mean, they tend to lead it to...

1:27:08

All right. For the Sultan of

1:27:10

Science, David Freiburg and David Sachs

1:27:13

and Chamath Pali Haftiya, the Chairman Dictator, I am

1:27:15

the world's greatest moderator. We'll see you next time

1:27:18

on the Roland Pied. Bye-bye. We'll

1:27:52

see you next time. I'm

1:28:01

John B. Beers. What is your favorite

1:28:03

movie? We need to get one for

1:28:05

the group. Why don't you get Murphys?

1:28:08

I'm so annoyed!