0:00

Hi, no priors listeners. I hope

0:02

it's been an amazing 2024 for

0:04

you all. Looking back on this year, we

0:06

I hope it's been an amazing

0:08

2024 for you all. of Looking back

0:10

on this year, we wanted to bring

0:12

you highlights from some of our

0:14

favorite conversations. with First up, we have

0:16

a clip with the one and only

0:18

Jensen Huang, of CEO of NVIDIA, the

0:20

company powering the AI Revolution. Since our

0:23

2023 chat with chat with Jensen, NVIDIA's tripled

0:25

in stock price, adding almost

0:27

100 billion of value each month

0:29

of 2024 the $3 entering the More recently,

0:31

More recently, Jensen shared his perspective

0:33

again with us. why This time no

0:35

longer a is no longer a

0:37

chip company, but a data center

0:39

ecosystem. our Here's our conversation with

0:42

Jensen. And has moved has moved and

0:44

larger, let's say say, like unit support

0:46

for customers. I think about I

0:48

from about it going from single chip to, you know, server

0:50

to and VL 72. How do How do

0:52

you think about that progression? progression?

0:54

next? what's NVIDIA Like, data center? do your

0:56

fact, we built center? In fact,

0:58

way that we build everything

1:00

unless you're building build everything, unless If

1:03

you're developing software, you need

1:05

the computer in its full manifestation. in

1:07

its don't we don't build

1:09

PowerPoint slides the ship the

1:11

chips. And we build we built a

1:13

whole data center. until we get we get

1:15

the whole data center built up. How do you

1:17

know works until you get the whole data center

1:20

built up? How do you know whole data you

1:22

know built up? How do you fabric

1:24

works in. works? all the

1:26

things that you that efficiencies to be, how

1:28

do you know it's gonna really work

1:30

at the scale? you know it's And that's

1:32

the reason why at the scale? not unusual. the reason

1:34

why it's not unusual to see

1:37

somebody's performance. be

1:39

be dramatically lower than their

1:41

peak performance as shown in

1:44

PowerPoint slides. And it's, computing is just not

1:46

used to, it's not what it is just not

1:48

be. You know, I say that the it used to be. of

1:50

I say that the new unit of computing

1:52

is the data center. that's what you us. So

1:54

that's what you have to deliver. That's what

1:56

we Now we build we build a whole thing like

1:58

that, then and then we, for every single thing about

2:00

combination, air cooled, X86, liquid

2:03

cooled, grace, Ethernet, infinite band, envy link, no envy, like, you

2:05

know what I'm saying? We built every you know

2:07

what I'm saying? We built

2:09

every single configuration. five We have

2:11

in our in our company today. year

2:13

we're going to we're going to build. easily five

2:15

more. So if you're serious about software, you

2:17

build your own computers. to build easily five about

2:19

software, then you're going to build your whole

2:21

computer, and we build it all at

2:23

scale. This is the part that is really

2:26

interesting. We build it at scale, then and

2:28

we build it to build whole We optimize it

2:30

build it. full stack, and then, and then we then, and

2:32

then we disaggregate everything. and we sell

2:34

and we sell it in parts. That's the part the

2:36

part that is completely

2:38

utterly remarkable about what we do. do.

2:40

The The complexity of that is

2:42

just insane. And And the reason for

2:45

that is that is be able to

2:47

to be able to our infrastructure into GCP,

2:49

GCP, AWS, OCI, Azure, their control planes, All

2:51

of their control planes, security planes

2:53

are all different. all of all of

2:55

the way they think about their

2:57

cluster sizing, all different. And, but yet we make

3:00

it possible yet, we make it possible

3:02

for them to all invidious architecture, that

3:04

that could could be everywhere. That's

3:06

really, really in the end. thought,

3:08

you know, that we would like to you know,

3:10

that we would like to have

3:12

a computing platform that developers could

3:14

use that's largely consistent, you know, 10% here and

3:17

you know, 10 % here and there

3:19

because people's infrastructure are slightly optimized

3:21

differently differently, and % here and there, but,

3:23

but everything they, they build run

3:25

everywhere. This is kind of of of

3:27

the principles of software that should

3:29

never be given up should never be given

3:31

up and we it quite it quite dearly. makes

3:33

it possible for our software engineers

3:35

engineers to build run everywhere. everywhere. And that's because because

3:37

we recognize that the investment of

3:39

software is the most expensive investment. It's

3:42

easy to test. Look at the Look at

3:44

the size of the whole hardware

3:46

industry. And And then look at the

3:48

size of the world's industries. trillion on

3:50

top of this of trillion industry. And

3:52

that tells you something. tells The software

3:54

that you build, you have to. you

3:56

build, you have to, you know, you basically as long as

3:58

you shall live. as you shall live. course, have

4:00

to mention our conversation with the lovely

4:02

Carpathie, we dig into the future of

4:04

AI as an exo -cortex, an extension

4:07

of human cognition. Andre,

4:09

who's been a key figure in

4:11

AI development, from open AI to

4:13

Tesla to the education of us

4:15

all, shares a provocative perspective on

4:17

ownership and access to AI models,

4:19

and also makes a case for

4:21

why future models might be much

4:23

smaller than we think. If we're

4:25

talking about a exo -cortex, that feels

4:27

like a pretty fundamentally important thing

4:29

to democratize access to. How do

4:31

you think current market structure of

4:33

what's happening in LLM research, you

4:35

know, there's a - small number

4:37

of large labs that actually have

4:40

a shot at the next generation

4:42

progressing training. Like training. how does that

4:44

translate to what people have access

4:46

to in the future? So you

4:48

kind of alluding to maybe is the state of

4:50

the ecosystem, right? So we have kind of like an

4:52

oligopoly of a few closed platforms and then we

4:54

have an open platform that's kind of like behind so

4:56

like MetaLama, et cetera. And this is kind of

4:58

like mirroring the open source kind of ecosystem. I do

5:00

think that when this stuff starts to, when we

5:02

start to think of it as like an exo -cortex,

5:04

So there's the, there's a saying in crypto, which is

5:06

like not your keys, not your, not your, Not

5:08

your keys, yeah. Like, it the case that if it's

5:10

like not your weights, not your brain? That's

5:12

interesting, because a company is effectively controlling your -cortex

5:14

in their part of Yeah, it starts to feel

5:16

kind of invasive. If this isn't my Exocortex. I

5:18

think people will care much more about ownership, yes. Like, yeah,

5:21

you realize you're renting your brain,

5:23

like it seems strange to rent your

5:25

brain. thought experiment is like, are you willing

5:27

to give up ownership and control to rent

5:30

a better brain? Because I am. yeah. so

5:32

I think that's the trade -off, I think we'll see

5:34

how that works, but maybe it's possible to like

5:36

by default use the closed versions because they're amazing,

5:38

but you have a fallback in various scenarios.

5:40

And I think that's kind of like the way

5:42

things are shaping up today even, right? Like

5:45

when APIs go down to some of the closed

5:47

source providers, people start to implement fallbacks to

5:49

like the open ecosystems, for example, that they fully

5:51

control and they're empowered by that, right? So

5:53

So maybe that's just the extension that will look

5:55

like for the brain, you fall back on

5:57

the open source stuff. Should

5:59

I? I? anything but most of the time you

6:01

of the time it's quite important that the open source

6:03

stuff to. that think so source stuff this is

6:06

not like an obvious point or something that people

6:08

maybe agree on right now But I think

6:10

100 or I guess one thing I've been wondering

6:12

about a little bit is now, but I think 100%...

6:14

I is the smallest about a model that

6:16

you can get to in some sense,

6:18

either in parameter size or everyone to

6:20

think about it? you can little bit curious

6:22

about your sense, either thought a lot about size

6:24

distillation, small models, you know? a I

6:26

think it can be surprisingly small. view.

6:28

And I do I do think that the

6:30

current models are wasting a ton

6:32

of capacity remembering stuff that doesn't matter,

6:34

matter. they remember they remember like the

6:36

ancient ancient... the data set is not curated.

6:38

is not curated, the best. Yeah, and I think this will go

6:40

away and I think we just need to get

6:42

to the cognitive core think we I think the cognitive core

6:45

can be extremely small And I think it's just

6:47

this thing that can be And if it needs to look up

6:49

information, it knows how to use different tools. Is that

6:51

it needs billion up Is that

6:53

how billion? how to use a billion, tools. and surprises. We'll

6:55

probably get to that point and the models can can

6:57

be very, very small. And I And I

6:59

think the reason they can be very small

7:01

very small is think just like distillation works and

7:03

maybe like the only It would say. the only

7:05

works say. well. works like Distillation is where you

7:07

get a really big model you get huge

7:09

amount of compute or something like that, of computers,

7:11

something a very small model. a very Our conversation

7:13

with Brett Taylor, with Brett Taylor, board member

7:16

and founder of Sierra of a really different

7:18

picture of how we interact with businesses

7:20

in the future. in the future. a

7:22

clip of Brett explaining company agents and why

7:24

the website is going to take a back

7:26

seat. going other category seat. The the

7:28

area that my company which is there that

7:30

what I call company agents. and

7:32

is it's really agents. simply about

7:34

automation or autonomy, but in this

7:36

world of conversational AI, AI, how does

7:38

how does your company exist

7:40

digitally? I'll use the metaphor use the

7:42

metaphor of it. We're 1995. you know, if you you

7:44

existed digitally about having a

7:47

website and being in Yahoo and being

7:49

in Yahoo In 2025, existing digitally

7:51

will probably mean having a branded

7:53

AI agent that your customers can

7:55

interact with to do. to that they

7:57

can do on your website, do on your it's

7:59

whether know, asking about your products

8:01

and services. and services, doing commerce, doing

8:03

doing customer service. That domain, I think, is

8:05

that domain I think shovel ready

8:08

right now with current technology. Because again,

8:10

like the persona base agents, it's

8:12

not the proverbial proverbial ocean. You know, you you

8:14

know, you have well -defined processes for

8:16

your customer experience, well -defined systems that

8:18

are your systems of systems of it's

8:20

really about saying in this world

8:22

where world where... We've gone from websites to

8:25

apps to now, conversational experiences, what is the

8:27

conversational experience you want around your brand?

8:29

And it doesn't mean it's perfect or

8:31

it's easy, otherwise we wouldn't have started

8:33

a company around it, but it's least

8:35

well -defined. around it. And I think well defined. And

8:37

I Right now right now in AI, you're

8:39

working on artificial general intelligence, your

8:41

version of your means something different, and

8:43

that's means That's just a different

8:45

problem to be solved. to be solved. But I

8:47

think, you know, particularly areas that Sierra works a

8:49

lot a lot of the companies that

8:51

you all have invested in, is in,

8:53

it saying, you there there shovel ready opportunities

8:55

right now with existing technology? And I

8:57

absolutely think there are. Can you describe

9:00

the, think there a shoveling cycle of building

9:02

a company of Like, what is the gap

9:04

between research and reality? Like, how do you,

9:06

reality? do you invest in as an engineering

9:08

team? as an how do you understand the

9:10

scope of different customer environments? the Just, like,

9:12

what are the sort of vectors of investment

9:14

here? And maybe as a as a starting point

9:17

worth even be worth also defining are what are

9:19

the products that provides today for its customers customers

9:21

and then. do you want that to go and

9:23

then maybe we can feed that back into

9:25

like what are the components of that because

9:27

I think what are the folks are really emerging

9:29

as a leader and you're vertical but it'd

9:31

be great just for a broader audience to

9:33

understand what you focus on. your sure. just give

9:35

a couple of examples to make it concrete.

9:37

So if you buy a new what you focus on.

9:39

Yeah, sure. you're having technical issues with your speaker,

9:41

you get the dreaded So light, new you'll now

9:43

chat with having is powered by speaker, help you

9:45

dreaded flashing help you debug whether it's

9:47

a hardware Sonos AI, -Fi issue, by Sarah. things

9:49

like that. you're a If a XM

9:51

subscriber, their AI agent agent is named which

9:53

which I think is a delightful

9:55

name. and it's everything from upgrading and

9:58

downgrading your subscription level to if you... get

10:00

a trial when you purchase a new vehicle, speaking

10:02

to you about that. Broadly

10:04

speaking, I would say we help companies

10:06

build branded customer facing agents. And branded

10:08

is an important part of it. It's

10:10

part of your brand, it's part of

10:12

your brand experience. And I think that's

10:14

really interesting and compelling because I think

10:16

just like you know, I go back

10:19

to the proverbial 1995, you your website

10:21

was on your business card. It was

10:23

the first time you had this digital

10:25

presence. And I think the same novelty,

10:27

and probably we'll look back at the

10:29

agents today with the same sense of,

10:31

oh, that was quaint. You know, I

10:33

remember if you go back to the Wayback

10:35

Machine, you look at early websites it

10:37

either someone's phone number and that's it, or

10:39

it looked like a DVD intro screen

10:42

with like of graphics, You a lot of

10:44

the agents that customers start with are

10:46

often around areas of customer service, which is

10:48

a really great use case. But I

10:50

do truly believe if you fast forward three

10:52

or four years, your agent will compass

10:54

all that your company does. I've used this

10:56

example before, but I like it. But

10:58

just imagine an insurance company, all that you

11:00

can do when you engage with them.

11:02

Maybe you're filing a claim. Maybe you're comparing

11:04

plans, We were talking about our kids earlier,

11:06

maybe you're adding your child to your insurance

11:08

premium when they get old enough to have

11:10

a driver's license, all of the above you

11:12

know, all of the above be done by

11:15

your agent. So that's what we're helping companies

11:17

build. Next, we talked to the Sora

11:19

team at OpenAI, which is building an

11:21

incredibly realistic video AI generation model. In

11:23

this clip, we talk about their research

11:25

and how models that understand the world

11:27

fit into the road to AGI. Is

11:30

there anything you can say about how

11:32

the work you've done with Sora sort

11:34

of affects the broader research roadmap? Yeah,

11:37

so I think. something here is about. the

11:41

knowledge that Sora ends up learning about

11:43

the world just from seeing all this

11:45

visual data. It understands 3D, which is

11:47

one. cool thing because we

11:49

haven't trained it to. We didn't

11:51

explicitly bake 3D information into it whatsoever.

11:54

We just. trained it on video

11:56

data, and it learned about 3D because 3D

11:58

exists in those videos. And it learned. learned

12:00

that when you take a bite out of a hamburger

12:02

that... you leave a bite mark. So

12:04

it's learning so much about our world. And.

12:07

And when we interact with the

12:09

world, so much of it is visual.

12:11

So much of what we see and

12:13

learn throughout our lives is visual information.

12:15

So we really think that. just

12:17

in terms of intelligence, in terms

12:19

of leading toward AI models

12:21

that are more intelligent, that better understand

12:23

the world like we do. This will actually

12:25

be really important for them to have

12:27

this grounding of like, hey, this is the

12:29

world that we live in. There's so

12:32

much complexity in it. There's so much about

12:34

how people interact, how how. things happen,

12:36

how events in the past end up

12:38

impacting events in the future, that this

12:40

will actually need to just much more

12:42

intelligent AI models more broadly than even

12:44

generating videos. It's almost like you invented

12:46

like the future of visual cortex plus

12:49

some part of the uh,

12:51

reasoning parts of the brain or something

12:53

of simultaneously. Yeah. And, and that's

12:55

a cool comparison because a lot of

12:57

the intelligence that humans have is

12:59

actually about world modeling, right? All the

13:01

time when we're thinking about

13:03

how we're going to do things. We're playing

13:05

out scenarios in our head. We have dreams

13:07

where we're playing out scenarios in the head.

13:09

We're thinking in advance of doing things. If

13:11

I did this, this thing would happen. If

13:14

I did this other thing, what would happen,

13:16

right? So we have a world model and

13:18

building Sora as a world model is very

13:20

similar to a big part of the

13:22

intelligence that humans have. How do

13:24

you guys think about the sort of

13:26

analogy to humans as having a very

13:28

approximate world model versus something that is

13:31

as accurate as, let's say, a physics

13:33

engine in the traditional sense? right? Because

13:35

if I hold an apple and I

13:37

drop it, I expect it to fall

13:39

at a certain rate. most humans do

13:41

not think of that as articulating a

13:43

path with a speed as a calculation.

13:46

Do you think that sort of

13:49

learning is like parallel in large

13:51

models? I think it's

13:53

a really interesting observation. I

13:56

think how we think about things is that it's

13:58

almost like a deficiency, you know, in humans that it's

14:00

not so high fidelity. So know the

14:02

fact fact that we We actually can't do very

14:04

accurate long -term prediction when you get down to

14:06

you get down to a narrow set of

14:08

physics is something that we something that

14:10

we can improve upon with some

14:12

of these systems. And so we're

14:14

optimistic that Sora will supersede that

14:16

that kind of capability and will, long

14:18

enable it to more more intelligent

14:20

one day than humans as world models. models.

14:22

But it is, it is an an

14:24

existence proof that it's not

14:26

necessary for other types of intelligence.

14:29

Regardless of that, it's still

14:31

something that that SORA and models in the the future

14:33

will be able to improve upon. improve Okay, so

14:35

it's very clear that the trajectory that the for

14:37

like throwing a football is gonna be

14:39

better. a football is the next, next

14:41

versions of these models versions of these

14:43

let's say. than I could add something

14:45

to that, this relates to the

14:47

paradigm of scale. of scale and the

14:49

better lesson a bit lesson a bit about

14:52

how we want methods you as you increase

14:54

compute, get better and better. and and

14:56

something that works really well in this

14:58

paradigm. paradigm. is doing the simple. but

15:01

challenging just of And

15:03

you can try data. with more And

15:05

you can try coming up with

15:07

more complicated tasks, for example, something

15:09

that. but use video explicitly,

15:11

but is maybe in some like

15:14

space that simulates approximate things or something.

15:16

But all this complexity actually isn't

15:18

beneficial when it comes to the scaling

15:20

laws of how methods improve as

15:22

you increase scale. And what works really

15:24

well as you increase scale is

15:26

just just predict data, and that's

15:28

what we do with do with we just

15:30

predict text. And And that's exactly what

15:32

we're doing with we're with which is

15:34

we're not making some complicated trying

15:36

to figure out some new thing to

15:38

optimize. We're saying, hey, the best

15:41

way to learn intelligence in a scalable

15:43

matter matter is to just predict data. That

15:45

makes sense in relating to what

15:47

you said, Bill, what you said, Bill, just get

15:49

much better with no necessary limit

15:51

that approximates limit that We also

15:53

sat down with Dmitry We also sat down

15:55

with of Dolgov. Today, the company

15:57

is scaling its self -driving fleet,

15:59

completing over ,000 fully autonomous rides

16:02

per week in cities like San Francisco

16:04

and Phoenix. It's my

16:06

favorite way to travel. way to In

16:08

this trip, Dmitri explains why

16:10

achieving full autonomy, removing the driver

16:12

entirely the achieving entirely, accuracy 100% than

16:14

99 .99 % accuracy in self -driving

16:16

is much harder than it might

16:18

appear. appear. Why is it breaking from like,

16:20

you know, know, say let's say advanced

16:22

driver assistance to work in more seems

16:25

to work in more and more

16:27

scenarios versus let's say full

16:29

autonomy? autonomy? What's the the what's the Yeah.

16:31

Yeah. the number of the number of Right?

16:33

And it's the nature of the nature of

16:35

this right? If If you think about where

16:37

we started in 2009, in of

16:39

our first first, you know, mile stops. One goal that

16:41

we that we set for ourselves

16:43

was to drive to drive, you know, 10 routes. Each

16:45

Each one was 100 miles long

16:47

all over the the Bay Area. You know, freeways, downtown

16:49

San downtown San Francisco, around Lake

16:51

Tahoe, everything. And you had to do

16:53

100 miles miles no intervention. So the

16:55

car had to drive autonomous from

16:58

beginning to end. That's the goal

17:00

that we created for ourselves. the

17:02

goal that we have seen about a dozen of us. You

17:04

know, about a months, of us took that. 18

17:06

months. We achieved that. 2009. no image net,

17:09

no confidence, no no no

17:11

big models, tiny computers, you know,

17:13

you know, how this, right? to get started. get

17:15

It's always been the property. with

17:17

every wave of technology, wave of very

17:19

easy to get started. to get started.

17:21

that the hard problem. And it's of the curve has

17:23

been getting that the early and steeper, but that's

17:25

not where the complexity is. The complexity is

17:28

in the long tail of the many, many,

17:30

many nights, and you don't see that if

17:32

you go the a prototype, if

17:34

you go for a is in the long and

17:36

this is where many, spending all of

17:38

our, that's the only hard part of

17:40

the problem. if you go for a I guess

17:42

nowadays it's always been getting easy

17:44

with every technical kind of cycle.

17:46

is where, you know, can take with all

17:48

of the advances the advances of an AI. and

17:51

especially in the general area I will and the LLLMs and

17:53

you can take kind of of an

17:55

almost off -the -shelf, know, know, transformers

17:57

are amazing. are amazing. The are

17:59

amazing. amazing. you can take a kind of

18:01

a VLLM that can accept images

18:04

or accept images where

18:06

you can give it text decoder

18:09

where you can text and you of book fine

18:11

tune it you can just it with just

18:13

a little bit of data from, go say,

18:15

say, camera data on a car to instead of words

18:17

to trajectories or or you know you decisions you might

18:19

the thing because a know box, because a black

18:21

box you know you been trained for a little

18:23

bit, and you fine tune it a

18:25

little bit. And you let me fine tune it a

18:27

I little think, if you and like that without you know

18:29

science to build ask AV good this is

18:31

what they would do. science to And out

18:33

of the box, today this something do yeah and out

18:35

it's amazing, right? something that of transport. right yeah

18:37

the power of is mind with powers violism is just a

18:39

little bit of effort, right you get something

18:41

a the road, and of a you bit of effort

18:43

don't get effort of miles something you get your mind. something

18:45

you get But then is is that enough? that

18:48

Is that enough to remove the

18:50

driver and drive and of miles and have

18:52

a safety record? You know, that is just really better than humans?

18:54

No, right? I guess this guess know, with

18:56

every, you know, evolution, technology, and a breakthrough

18:58

in AI, they've seen, like, I'm about, appreciate

19:00

it. Up next, we we have my dear

19:02

friend, friend, Dylanfield, CEO of Figma. Dylan shares

19:04

his prediction for how user interfaces

19:07

will evolve in an AI -driven world.

19:09

While While many predict a shift

19:11

toward conversational or agent -based interfaces, Dylan

19:13

suggests that new interface paradigms will

19:16

complement existing ones. He also

19:18

highlights the exciting potential of potential

19:20

and intelligent cameras as the

19:22

next frontier as input methods. in How

19:24

do you think about How you

19:26

UI the general? in general? that's is gonna come

19:28

with AI. AI. A lot A lot of things are kind

19:30

of collapsing in the short in chat interfaces. There's a

19:32

lot of people talking about a future a lot of world.

19:34

talking which a away with most UI world,

19:37

which and away with stuff happening in the

19:39

background. it's just all How do you think

19:41

about where UI is going in general right

19:43

now? where UI is I kind of in kind of

19:45

comes back to the rabbit point I was making earlier. comes back

19:47

to the rabbit there's a lot earlier.

19:49

Yes, there's a lot of happening in terms

19:51

of agents, but I think like in terms

19:53

of the way that we way that we... use UI

19:55

UI to interact with agents, we're just

19:57

the beginning. just the beginning and

20:00

I think that the interfaces will get

20:02

more sophisticated, but also even if they

20:05

don't. I suspect

20:07

that it's just like any new media

20:09

type. When it's introduced, it's not like

20:11

the old types go away. right? just

20:13

because you have TikTok doesn't mean

20:15

that you, you know, hunger watch

20:17

YouTube. Even if it's true

20:20

that a new... form

20:22

of interaction is via chat interfaces, which I'm not

20:24

even sure I believe. But if, if we take that

20:26

as a prior. on the

20:28

No Briars then I think

20:30

that you still have UI and

20:32

actually I think you have

20:34

more UI and more software. than

20:36

before. Do you have any predictions

20:38

in terms of multimodality? Like, do

20:40

you think there's more need for boys? Like, Like,

20:42

so, you know, a lot of the debates have

20:44

is like, When are you going

20:47

use voice versus text versus other

20:49

types of interfaces? And,

20:51

you know, you imagine arguments in all sorts

20:53

of directions in terms of you know, do you

20:55

use what and things like that. And a

20:57

lot of people are not a lot, some people

21:00

are suggesting because of the rise of multimodal

21:02

models, you'll have like voice input or more things

21:04

like that because you'll be able to do

21:06

real -time sort of smart contextual semantic understanding

21:08

of like conversation. And so you

21:10

have more of a verbal

21:12

conversational UI versus a -based UI. And

21:14

so it kind of changes how

21:16

you think about design. So, I

21:18

was just curious if you have any thoughts on

21:21

that, that sort of future stuff. There's

21:23

all sorts of contexts where a voice

21:25

UI is really important. And

21:27

I think that,

21:29

it might be that we find that

21:31

voice UIs start to

21:33

map to more traditional UIs. because

21:37

it's something that like you obviously do. in

21:40

a more generalized way. Bye.

21:43

But yeah, I mean, personally,

21:45

I don't want to navigate.

21:47

the information spaces that I

21:49

interact with every day,

21:51

all day via voice. I

21:54

also don't want to do it in

21:56

minority report style on the vision pro

21:58

exactly either. Maybe with with

22:00

a keyboard and mouse and like

22:02

an amazing vision pro, monitor setup or Oculus,

22:04

like that could be cool, but I be cool, but

22:06

I don't want to do the board

22:08

thing. thing. it's And so it's so I

22:11

It's interesting, so I think that we get

22:13

these new glimpses that interaction patterns that

22:15

are really cool cool and the natural natural inclination is

22:17

to extrapolate and say they're gonna be

22:19

useful for everything. be useful I think that they

22:21

have I think of their role. sort of

22:23

their role and it doesn't mean that It

22:26

doesn't mean that they're gonna be ubiquitous

22:28

across every interaction we have. we have. But that's

22:30

but that's a natural cycle to

22:32

be in. it's good. It's I think

22:34

it's good, of that it's healthy to have

22:36

sort of that. couldn't do, because

22:38

around. have that, then you do,

22:40

because if you don't have that, then

22:42

you don't get to find out. of And

22:44

so I'm supportive of people exploring as

22:46

much as possible you kind of that's how

22:48

you kind of HCI on HCI how I use out

22:51

how to use computers and to the fullest

22:53

potential that could be possible. One of

22:55

of the things I am really bullish on mean, you

22:57

I mean, you just think of it as

23:00

an input mode or a peripheral, but

23:02

it's really hard for people to

23:04

describe things visually. And so

23:06

the idea of intelligent cameras,

23:08

even in the in the basic

23:10

sense. Oh, it worked. It Oh, it I

23:12

think that's actually a I think that's actually a

23:14

really fun space to be, as you

23:16

said, because I because I

23:18

actually think that will be useful. And

23:21

it's And it's something that every

23:23

user is capable of, right, pictures, capturing

23:25

video. And so I think that'll

23:27

be that'll be, I'm that. bullish wrap

23:29

up our favorite moments from

23:31

2024, we have scale CEO, scale CEO

23:34

In this clip, he shares his bold

23:36

take on the road to take on the Alex

23:38

also dives into why generalization into why harder

23:40

than many think, and why solving these

23:42

niche problems and more data in evals

23:44

is key to advancing the technology. and evals

23:46

you believe about AI that other people don't. My

23:48

My biggest belief here is

23:50

that the the the to AGI

23:53

is a lot more that looks a

23:55

lot more like curing cancer

23:57

developing a a vaccine. what I mean

23:59

I mean by that. is I think think that the build

24:01

AGI is to build in, you know, is

24:03

going to be to solve a bunch of small

24:05

problems that where you to have to

24:07

solve a bunch of small problems where

24:10

you don't get that much positive

24:12

leverage between the next one problem to solving

24:14

the next problem. of, you know, it's like curing

24:16

curing cancer, which is then have to

24:18

then each in to each individual cancer

24:20

and solve them independently. And eventually over

24:22

a multi over a time frame, we're

24:24

going to look back and realize that

24:26

we've built AGI, we've cured cancer. But

24:28

the path to get there will

24:30

be be this like, know, quite road of solving

24:32

individual capabilities and building and

24:34

of. individual to support this

24:37

end mission. to support this end

24:39

I think a lot of people in

24:41

the industry paint of people in like, you know,

24:43

the path to just like, you know, we'll get there,

24:45

we'll like, you know, we'll get there, we'll like,

24:47

we'll solve it in it in one fell swoop.

24:50

this is a lot of think there's a lot

24:52

of implications for how you actually think

24:54

about you know, technology arc arc and... and how society is

24:56

going to have to have to deal with it.

24:58

I I think it's actually a pretty

25:00

bullish case for society adapting the technology

25:02

because I think it's gonna be consistent,

25:05

slow progress for quite some time and

25:07

society will have time to fully sort

25:09

of have time to fully to the technology that

25:11

develops. the When you that a problem

25:13

at a time, right, if we just at

25:15

a away from the analogy a little

25:17

bit, should I think of that as think

25:19

of that as... generality of multi -step reasoning is

25:21

really hard, as you as you know, Monte research

25:24

is not the not the answer that people

25:26

think it might be. run into scaling walls like We're

25:28

just gonna run into scaling walls, like

25:30

gonna of like of what are the dimensions

25:32

of I multiple problems? the main thing

25:34

there's I think there's very limited generality

25:36

that we get from these models for

25:38

and even for for for example, saying there's

25:40

no understanding there's no positive transfer from

25:42

learning in one modality to other training

25:44

off of a off of a bunch of

25:46

video doesn't really help you that

25:48

much with your text problems and vice

25:50

versa. and vice And so. so I I

25:52

think what this means is like, each like each

25:55

sort of each niche of each niche of

25:57

capabilities or each area of is going to

25:59

to require. separate flywheels, data flywheels, to

26:01

be able to push through

26:03

and drive performance. You don't You

26:05

don't yet believe in video as basis

26:07

for world model that helps. I think it's great narrative. I

26:09

narrative. I don't think there's strong scientific

26:11

evidence of that yet. Maybe there will

26:13

be eventually. but I But I think that

26:15

this is the, I think the base case, think the

26:18

say, case, let's say, is one where,

26:20

you know... There's not that much

26:22

generalization coming out of the the models. And

26:24

so just actually just need to slowly solve

26:26

lots and lots of little problems to ultimately

26:28

results in AGI. Thank you so much for

26:30

listening in 2024. 2024. We've really enjoyed talking

26:32

to the people reshaping the world for

26:34

for AI. If you want If you want to more

26:36

deeply dive into any of the conversations you've

26:38

heard today, today, we've linked the full episodes

26:40

in our description. Please let us

26:42

know who you want to hear from on what your

26:44

questions are for next year. Happy holidays. are for next

26:46

year. Happy holidays. us on Twitter at

26:48

No Prior's Pod. Subscribe to our YouTube channel

26:51

if you want to see our faces. Follow

26:53

the show on Apple podcast, Spotify, or

26:55

or wherever you listen. That way you

26:57

get a new episode every week. every

26:59

And sign up for emails find

27:02

transcripts for every episode at no every.com.