0:00

We have 30 million software

0:02

engineers in the world, and

0:04

we have 750 million active

0:06

users of Excel. And now,

0:08

you'll ask, like, how does word

0:10

compare to Excel? And what Excel

0:13

did in the 80s to data analytics

0:15

and numbers is what we're

0:17

trying to do to AI.

0:19

So in the 80s, you

0:21

either had to have a

0:23

team of data analytics people

0:25

or engineers, or you're using

0:27

a calculator calculator. And I would

0:30

say the calculator equivalent here is

0:32

the chart to PT that you

0:34

every time you need to redo

0:37

the conversation and you every time

0:39

you need to instil your

0:41

own needs into it. What World

0:43

War is trying to do is

0:45

saying, hey, a lot of the

0:47

things that you do are repeatable,

0:50

similar to Excel, and you can

0:52

encode your taste into it with

0:54

World War. Today

1:09

we're joined by Philip Kazaera, co-founder

1:12

of Wordware, who's building tools to

1:14

help bridge the gap between human

1:16

creativity and AI. Philip shares his vision for

1:18

why English is becoming the new assembly

1:20

language for LLMs, why he believes that

1:23

the future belongs not just decoders,

1:25

but to people he calls word

1:27

artisans, who can communicate effectively their

1:29

creative vision to an AI system, and what

1:31

that means for the future of programming

1:33

computers. Philip, thank you so much

1:35

for joining us today. I'm excited to

1:38

learn about you and Wordware and get

1:40

your take on how programming computers is

1:42

fundamentally going to change with large language

1:44

models. Thank you for the invite. Let's dig

1:47

right in. I want to start with Andre

1:49

Carpothi. He had a tweet in 2023 that

1:51

went viral. The hottest new programming language is

1:53

English. What do you make of that? And

1:55

what does it mean relative to what you

1:58

are trying to build that Wordware? I think

2:00

syntax will not be as important.

2:02

You know, everyone will be somewhat

2:04

of a coder. People used to

2:06

have to know Python. Now English

2:09

is enough. However, you still need

2:11

to know what you're trying to

2:13

say. And in that way, I

2:15

would say, you know, not everyone

2:17

will be able to use it

2:19

because some people don't have that

2:21

much to say. And I would

2:23

maybe rephrase it a little bit

2:26

in a little bit more of

2:28

an exciting way for us, is

2:30

that the assembly language to LLMs

2:32

is English, but you still have

2:34

to structure it in the right

2:36

way, and you still have to

2:38

use some of the concepts even

2:41

from typical programming in order to

2:43

actually make sure that it does

2:45

what it's supposed to do. You

2:47

heard it here first, okay. English

2:49

is not the hottest new programming

2:51

language. It is the new assembly

2:53

language. And at word where you

2:55

were trying to build that new

2:58

programming language to work with that

3:00

new assembly language. I think bringing

3:02

structure to human AI collaboration is

3:04

something that I've chosen to spend

3:06

the next 10 years of my

3:08

life on. And it's an exciting

3:10

problem, because right now we're trying

3:13

to mix the structure of programming,

3:15

which is very rigid and deterministic,

3:17

with something that's intrinsically fuzzy. And

3:19

that marrying of these concepts and

3:21

choosing the right affordances and the

3:23

right obstruction layers for that communication

3:25

to be incredibly... Easy yet enable

3:28

people to do complex things is

3:30

hard. And this is what we

3:32

are trying to do with word

3:34

where as the engine to enable

3:36

the right mix of the two.

3:38

Let's talk a little bit more

3:40

about no code. I remember back

3:42

when when I first got to

3:45

Sequoia in 2018, I remember no

3:47

code was the hottest thing ever

3:49

and it was like, you know,

3:51

we're not going to we're not

3:53

going to program anymore. And, you

3:55

know, obviously some no code companies

3:57

have done extremely well, like retool

4:00

for example. have you know we

4:02

have software engineers today and so

4:04

so far that no code promise

4:06

hasn't really come to fruition. What's

4:08

different now like what has changed?

4:10

I think coming back to what

4:12

I just said as well on

4:15

the assembly language is English and

4:17

you I kind of have this

4:19

small insecurity when people call Wordware

4:21

a no-code tool. Because we have

4:23

not yet reached a ceiling with

4:25

any of our clients. So you

4:27

can achieve absolutely everything. We've created

4:29

an ability to put in code

4:32

execution blocks, which if you want

4:34

an escape hatch and Wordware is

4:36

not fully capable of everything, you

4:38

still can do it. And in

4:40

that way, you know, the document

4:42

format of how to structure agents.

4:44

and how to write them down

4:47

one by one is still very

4:49

similar to how code works. We

4:51

still have loops, we still have

4:53

conditional statements, and we, you know,

4:55

have flows calling other flows, which

4:57

really is functioned calling, you know.

4:59

So in a way, we don't

5:02

see ourselves as a no-code tool.

5:04

And we kind of believe that

5:06

the word crafters, the word artisans

5:08

of the future are still coding.

5:10

It's just, you know, they are

5:12

structuring the English in a very

5:14

precise way to make sure that

5:16

the prompt is populated in the

5:19

right manner. Word crafters and word

5:21

asins. I like it. Wordware engineers,

5:23

you've heard it here first. How

5:25

are you using the English language

5:27

then to make, you know, since

5:29

you bristle at the no code

5:31

term, how are you making no

5:34

code more code like? And maybe

5:36

this is a good time to

5:38

just give a 30 second overview

5:40

on, you know, the word where

5:42

product and how people use it

5:44

to build things. Sure. So by

5:46

trying to bring that structure to

5:49

the intrinsically fuzzy English language, we've

5:51

created an editor where you can

5:53

use the similar concepts to software

5:55

engineering, functions, and marry it in

5:57

that editor, in that natural language

5:59

IDE, in a way that people

6:01

can construct agents. Something that we're

6:03

not, we're not coaching, we go

6:06

straight to agents. And in my

6:08

definition, agents are almost like, they

6:10

are still software, they are still

6:12

a little bit like software taking

6:14

in inputs and outputting outputs, but

6:16

some of the stages of what

6:18

they do is fuzzy. So marrying

6:21

that structure in the editor, whatever

6:23

you build there, whatever you iterate

6:25

on there, you then have three

6:27

different ways of deploying it. You

6:29

can deploy it as an API

6:31

that will power your product or

6:33

that AI button or that AI

6:36

chatbot that is a little bit

6:38

more complex than just doing a

6:40

vanilla API call to Claude or

6:42

Open AI. That's number one. Number

6:44

two is you can deploy it

6:46

as a workflow where the main

6:48

part of the workflow is not

6:50

like Zapier, it's more AI native.

6:53

and the greedy, the brain of

6:55

the AI is a little bit

6:57

more complex than a couple prompts

6:59

strung together. And the third thing

7:01

that we're building is the GitHub

7:03

for AI, let's say, for people

7:05

to share and what they've done

7:08

and other people are able to

7:10

then fork it and use these

7:12

things as components. Again, marrying some

7:14

concept from software engineering, you need

7:16

other people's components and libraries and,

7:18

you know, you want to be

7:20

building on top of the, on

7:22

the shoulders of the job. of

7:25

these AI thinkers. So again, Wordware

7:27

engine, editor, way to actually create

7:29

these agents and then three different

7:31

ways to deploy. I think one

7:33

of the beauties of code is

7:35

just its expressibility and its precision.

7:37

You know, you know, exactly what

7:40

you were telling the machine to

7:42

do, and you were, you know,

7:44

expressing it in some languages in

7:46

the most precise way possible. English

7:48

is not like that. To your

7:50

point, it's fuzzy. and the steerability

7:52

of that? Or, you know, what

7:55

is the, I guess, abstract thing

7:57

you were doing with English to

7:59

make it more programmable? Yeah, I

8:01

think you're exactly correct. We are

8:03

trying to bring a little bit

8:05

more structure. It's not all the

8:07

way, because if you go all

8:09

the way to a programming language,

8:12

you lose the fuzziness and you

8:14

lose the power of it. But

8:16

it's hard. And right now, most

8:18

of the people, you know, we

8:20

had this way for evaluation software

8:22

companies at some stage. And what

8:24

we've realized with a bunch of

8:27

companies that we work with is

8:29

that they don't know, like they

8:31

don't have these data sets to

8:33

use for evils. And what we

8:35

came up with is that editor

8:37

very quickly gets you to understand

8:39

and develop an intuition of what

8:42

works and what doesn't work. And

8:44

for now, this is the most

8:46

important thing. It's clicking run a

8:48

hundred times quickly and making sure

8:50

that what you've written here. Has

8:52

enough structure in order to output

8:54

things on the right side that

8:56

you want and in that way,

8:59

you know as long as you

9:01

know what you're trying to achieve

9:03

And this is very hard, you

9:05

know, we have a lot of

9:07

companies coming in and just saying

9:09

AI predict weather I literally had

9:11

a big customer say can you

9:14

guys like predict weather for me?

9:16

And you know, that's not the

9:18

case. You need a document where

9:20

you outline what you're trying to

9:22

do and you know, even that

9:24

has enough structure If you have

9:26

done an intro, those are the

9:29

inputs. You'll be playing around with

9:31

images and PDFs, and then you'll

9:33

manipulate it in this way, and

9:35

then you'll get some outputs. And

9:37

developing that intuition is just enough

9:39

structure for today. Soon, maybe we'll

9:41

be able to do better evolves,

9:43

but for now, a lot of

9:46

people really don't know. At the

9:48

moment when they start playing around

9:50

with Wordware, they don't know what

9:52

we are trying to achieve. And

9:54

one of our customers coined this

9:56

term of speed of creativity with

9:58

Wordware is higher. So they learn

10:00

what they are actually trying to

10:03

do as they encounter problems

10:05

with the underlying models and

10:08

they realize Gemini 2.0 pro

10:10

might be better and maybe

10:12

Gemini can take huge PDFs

10:14

and Claude can kind of

10:17

take PDFs but smaller and

10:19

then GPT4 cannot take PDFs

10:21

and you know they develop

10:24

that understanding and that helps

10:26

them to structureize their faults.

10:29

How do you instruct the machine

10:31

to go from intent to outcome?

10:33

Like let's say I'm a brilliant

10:35

filmmaker and I want, you know,

10:38

I want to use Wordware to

10:40

create the next hit. What do

10:42

I do in Wordware in order

10:44

to make that happen? Yeah, so

10:47

for now we focus on knowledge

10:49

workers where that is a little

10:51

bit easier. I think using your

10:54

example for figuring out how the

10:56

future will be, you know, if we

10:58

have... let's say, you know, George

11:00

Lucas, playing around, hypothetically

11:02

playing around with GPT-7

11:04

and trying to create

11:06

Star Wars, he might

11:08

type in just a prompt

11:10

being like the two sentences,

11:12

and he would just say,

11:14

hey, create a movie about

11:16

words between Star Systems. And

11:18

that's just enough to give

11:20

a model, an ability to

11:22

run on its own. and

11:24

this is actually not what

11:26

you want. You want to

11:29

convey your creative vision, but

11:31

for now in order it's

11:33

just knowledge work, but soon

11:35

it will be all work where it

11:37

needs that human sprinkle, that human

11:39

paste. And these are the things

11:41

that I really value is like,

11:43

people say, oh, nobody will have a

11:46

job whatsoever. I don't agree with it

11:48

at all. I think the human taste and

11:50

how you do things will matter even

11:53

more. And I use this George Lucas

11:55

example because it's a little bit easier

11:57

to understand how taste is influencing that.

11:59

But... everywhere, writing a good

12:01

email is dependent on good taste.

12:03

Figuring out I was just hiring

12:06

for an executive assistant and like

12:08

everyone in our company needs to

12:10

show that taste and needs to

12:12

show a little bit more conviction

12:14

in the way that they do

12:17

things. So for executive assistant, like

12:19

she needed to choose a right

12:21

restaurant for our off-site, you know?

12:23

And that also has taste. I

12:25

don't want to trust in AI

12:28

with this. So yeah, that's kind

12:30

of that sprinkle of human touch

12:32

is very important. So taste as

12:34

the last bastion of humanity? I

12:36

think so. I think creativity and

12:39

taste. Do you think machines can

12:41

learn human taste? I think they

12:43

can. But that's not the point.

12:45

There is an interesting analogy here.

12:47

They put humans into an MRI

12:50

machine and they've shown them two

12:52

different pieces of art. Both of

12:54

them were done by AI. And

12:56

they told the people, hey, one

12:58

is created by a human artist

13:01

and another one is created by

13:03

AI. Our brains work completely differently

13:05

and our different parts of the

13:07

brain fire when we assume human

13:09

intent behind something. So you know,

13:12

I can create a song and

13:14

just like it will be a

13:16

good song with Suno or whatever

13:18

and send it to my friend

13:20

and he'll be like, yeah, it's

13:23

a cool song. But if I...

13:25

ingrained my intent and I'll create

13:27

a song about our, you know,

13:29

skiing trip to Shamboni and I'll

13:31

make it funny and I will,

13:34

that intent will be there. I'm

13:36

pretty sure his brain will be

13:38

firing in a completely different way,

13:40

giving him like, giving him a

13:42

completely different experience in that way.

13:45

Does that make sense? It makes

13:47

a lot of sense. Yeah. I

13:49

love it. I'm going to transition

13:51

to talking about the next billion

13:53

developers, and you've alluded to this

13:56

a few times in the conversation,

13:58

and I really want to just

14:00

pull on that thread. So you

14:02

started this, you know, this. conversation

14:04

by saying, you know, there's a

14:07

certain set of people in the

14:09

world that know how to code,

14:11

but there's a difference of people

14:13

in this world that have creativity

14:15

and have ideas. Do you think

14:17

that set of people is larger?

14:20

Is that how we get to

14:22

the next billion developers? There's an

14:24

interesting analogy here. We have 30

14:26

million software engineers in the world,

14:28

and we have 750 million active

14:31

users of Excel. And now... you'll

14:33

ask, like, how does word work

14:35

compared to Excel? And what Excel

14:37

did in the 80s to data

14:39

analytics and numbers is what we're

14:42

trying to do to AI. So

14:44

in the 80s, you either had

14:46

to have a team of data

14:48

analytics, people or engineers, or you're

14:50

using a calculator. And I would

14:53

say the calculator equivalent here is

14:55

the chart to PT that you

14:57

every time you need to redo

14:59

the conversation and you every time

15:01

you need to instil your own

15:04

needs into it. What World War

15:06

is trying to do is saying,

15:08

hey, a lot of the things

15:10

that you do are repeatable, similar

15:12

to Excel, and you can encode

15:15

your taste into it with Word

15:17

War. And I believe that... that

15:19

taste will be important as mentioned

15:21

before and I think the next

15:23

500 million or a billion users

15:26

of AI might be calling them

15:28

I don't know it will be

15:30

the term hopefully it's word we're

15:32

engineered but you know word artisan

15:34

or whoever and the really important

15:37

part here is that they need

15:39

to know what the AI is

15:41

supposed to do many many times

15:43

you know we are a horizontal

15:45

tool and people come to us

15:48

and they say, hey, what can

15:50

I do with AI? And I

15:52

tend to explain it as if

15:54

for now it's an intern, but

15:56

intern after university, and you need

15:59

to write out on a piece

16:01

of paper, a couple of things.

16:03

that you want to do with

16:05

the intern. So you need to

16:07

say, hey, this is your job.

16:10

This is the title of what

16:12

you're trying to do. Here are

16:14

some of the documents or input

16:16

that you'll be working with. Here

16:18

are the data sources, and here's

16:21

the output that I'm expecting of

16:23

you. And the important caveats here

16:25

that people don't often understand is

16:27

that the data sources has to

16:29

be something that you trust. You

16:32

can't just say hey go and

16:34

search the internet because often you

16:36

end up with things that you

16:38

don't Agree with and if the

16:40

intern works on top of that

16:43

that's a problem And another one

16:45

is and this is very important

16:47

is that you're gonna trust the

16:49

intern with this So you know

16:51

if you want to send a

16:53

thousand emails to every person that

16:56

needs a response in your inbox,

16:58

with some people you just want

17:00

trust and intern to do this.

17:02

And this is how AI works

17:04

right now. So as long as

17:07

you have a job right now,

17:09

that you say, hey, if I

17:11

haven't one intern, I could easily

17:13

explain it to them. And a

17:15

lot of our work is like

17:18

this right now. We often read

17:20

an email, go search and drop

17:22

box, go search notion, and then

17:24

we create a response that is

17:26

essentially... based on this database that

17:29

we curate, then you can be

17:31

using AI. And I think more

17:33

and more, this knowledge work is

17:35

gonna be automated. And I think,

17:37

no, in that way, the next

17:40

billion people are gonna be, they're

17:42

gonna need two things, intent, what

17:44

actions to happen, and taste, how

17:46

do you want to do this?

17:48

and all of the rest will

17:51

feel like CEOs of the biggest

17:53

enterprise because we will have a

17:55

thousand knowledge workers working beneath us

17:57

and trying to actually execute on

17:59

these two things. What I heard

18:02

just now was a lot of

18:04

automation about knowledge work. I mean,

18:06

the thing that I'm most intrigued

18:08

about within AI and within generative

18:10

AI is its generative capacity, including

18:13

the ability to create, you know,

18:15

you mentioned the... George Lucas example,

18:17

but also to create, you know,

18:19

new applications, new marketplaces, you know,

18:21

new products. And so do you

18:24

imagine, do you see Wordware primarily

18:26

serving, you know, making the knowledge

18:28

worker more productive? Or do you

18:30

see it also assisting in kind

18:32

of the creation of new products,

18:35

services, you know, pieces of art?

18:37

For now, it's mostly about the

18:39

productive work, I would say. It's,

18:41

you know, the AI engine is

18:43

the AI heart of your product

18:46

is wordwork. Currently, we have not

18:48

dipped into the generative UI part

18:50

of things. We're not lovable. We've

18:52

actually used lovable to, you know,

18:54

wrap our AI heart for some

18:57

of our customers. And that has

18:59

worked great. I'm just so impressed

19:01

with their product. Talk more about

19:03

this. Loveable, I think, also sees

19:05

themselves as, you know, enabling the

19:08

next billion developers. You see, you

19:10

have a similar vision. You know,

19:12

how do you think your view

19:14

of how the world will go

19:16

foots with their view? And like,

19:18

why didn't you choose to make

19:21

a, you know, that style of

19:23

no code tool? Yeah, because the

19:25

way that I see the word

19:27

developer is a little bit different.

19:29

they see the word developer as

19:32

what developers do today, which is,

19:34

you know, a lot of SAS

19:36

is a wrap around a database

19:38

with some dashboard and ability to

19:40

manipulate that data. They are creating

19:43

a much more personalized dashboard, you

19:45

know, and a lot of people

19:47

are going to create incredible vertical

19:49

SAS based on loveable. And I

19:51

think that's incredible. And the one

19:54

thing that was missing through all

19:56

of this is that they not

19:58

only grab the UI part, they

20:00

also grab the database part, which

20:02

many people do not know how

20:05

to manipulate and hence they unlocked

20:07

a lot more use cases. Whatever.

20:09

But what we are trying to

20:11

say is that this part of

20:13

creating a generative, creating a UI

20:16

on top of a database is

20:18

not the future. The future is

20:20

to actually utilize this reasoning engine

20:22

that an alum is in a

20:24

productive manner, and we focus on

20:27

that substance of AI at the

20:29

beginning. In the future, we might

20:31

want to expose that engine. as

20:33

a in a UI, maybe it's

20:35

a chatbot, maybe it's, you know,

20:38

digesting some images, etc. But the

20:40

real important part is the AI

20:42

engine. And yeah. So if you

20:44

think about an app as, you

20:46

know, there's the UI, there's the

20:49

application logic, and there's a database,

20:51

what you're saying is, you know,

20:53

you really want to just, you

20:55

know, knock it out of the

20:57

park on the application logic, so

21:00

to speak. And right now we're

21:02

able to work on a lot

21:04

more data, which is not structurized.

21:06

And this is the big, big

21:08

difference right now. A lot of

21:11

Suss's right now will still work

21:13

in a similar manner, just the

21:15

database is fuzzy. And the database

21:17

might be what you see every

21:19

day. And how the hell would

21:22

you put that in a typical

21:24

normal, you know, database? And I

21:26

think working on top of that

21:28

context is the really exciting part

21:30

for me. Let's go back to

21:33

this concept of word artisans or

21:35

word wear engineers if everything goes

21:37

right. You know, what's your vision

21:39

of what a word wear engineer

21:41

looks like in 10 years? Oh,

21:44

that's a tough one. I've been

21:46

thinking a lot about what does

21:48

work look like in 10 years

21:50

for human beings? And I was

21:52

struggling with this at the beginning

21:54

because it's really hard. to understand

21:57

people's jobs even today. And often

21:59

I boil it down to the

22:01

software that they use. They can

22:03

talk a big game about strategy

22:05

and you know I set the

22:08

mission and in the end of

22:10

the day I ask them do

22:12

you do meetings, do you do

22:14

email, do you do PowerPoint presentations,

22:16

do you work in Excel, or

22:19

do you work in code? And

22:21

I just want to understand what

22:23

does work look like in 10

22:25

years and like what are you

22:27

really working with you know is

22:30

it a hair like interface when

22:32

you just talk to the AI

22:34

and it does a lot of

22:36

work for you I think to

22:38

be honest voice is kind of

22:41

not the best modality to express

22:43

that hence I kind of think

22:45

that in its simplest form, word

22:47

where it's a document where you

22:49

jot down your thoughts and you

22:52

do it in a more structured

22:54

way, word where a co-pilot AI

22:56

is helping you throughout structuring it

22:58

and in the end of a

23:00

day you behave like a CEO

23:03

which sets the strategy intent and

23:05

all of us on that piece

23:07

of paper essentially on these blank

23:09

canvas and you draw that vision.

23:11

Maybe it's even more than words,

23:14

it's just, you know, you generate

23:16

this vision of how your own

23:18

enterprise works. And, you know, I

23:20

look at different things around us,

23:22

and I see furniture or shoes

23:25

or whatever, and I think there

23:27

is taste ingrained into what kind

23:29

of shoes you like to make.

23:31

So in 10 years, if somebody

23:33

wants to become a creator of

23:36

the best brand of shoes. It

23:38

becomes about, that true becomes a

23:40

luxury object, which has ingrained taste

23:42

and intent in it. And then

23:44

a bunch of things in the

23:47

end will be, will happen on

23:49

its own. The really tough parts,

23:51

even manufacturing it and so on,

23:53

will happen on its own. But

23:55

what's your job in the future

23:58

is talking to other CEOs. will

24:00

not, but humans don't want

24:02

to lose that control. So you

24:04

will talk with other CEOs about

24:06

maybe doing a partnership with your

24:08

shoe brand and somebody else. You

24:10

will have to be still critical

24:12

about the intent of the other

24:14

person and you have to instil taste

24:17

and your own creative vision into that

24:19

shoe. Do you think that, you know, do

24:21

you think a billion people globally will

24:23

be? capable of programming a machine in

24:25

English language the way you describe or

24:27

in Wordware documents the way you describe

24:29

because it does require you know it's

24:31

it's almost like pseudo coding and you

24:33

know there is logic there's there's loops

24:35

and things like that I guess maybe

24:38

talk about today like do you need

24:40

to be technical in order to use

24:42

Wordware like who is the ICP today

24:44

and what is needed to move that

24:46

ICP so that you can reach a

24:48

billion developers yeah I think

24:50

Right now, what we've enabled

24:52

is people who are somewhat

24:54

technical, CEOs, technical PMs, high

24:57

up in the org chart

24:59

to engrain their own, like,

25:01

kind of think that they know

25:03

what needs to happen and get

25:05

there quicker. You know, so Max

25:08

from InstaGuard, for example, he

25:10

is a founder and he

25:12

spent four days just, you

25:15

know, refining his idea. in

25:17

Wordware, instead of hiring a

25:19

whole team, but he is

25:21

somewhat analytically minded. And for now

25:23

that's the case. We did not want

25:26

to make too much magic because the

25:28

models were not there. Right now what

25:30

we're doing is we're moving more

25:32

into that blank canvas when you

25:34

just describe the idea and we

25:36

take care of guessing the right

25:38

structure. And you still will be

25:41

able to, you know, in a

25:43

very fine-grained way, edit it. but

25:45

you will start playing a lot

25:47

more, we'll probably use O3 to

25:49

get you to the first draft

25:51

of how that flow works. And

25:53

when we kind of loop back

25:56

to the future of how it

25:58

will all look like and... really

26:00

whether we'll have one billion developers,

26:02

you know, working in word where

26:04

it becomes a much bigger question

26:06

here. It becomes a question of

26:08

like, will a billion people want

26:10

to do productive work? Like, you

26:12

know, we just talked about the

26:14

shoe. How many people will have

26:16

the drive to put out something

26:18

to the world and they will

26:20

want to express that creative vision?

26:23

Maybe in, you know, post resource

26:25

scarcity world, most of it will

26:27

want work. but I think we'll

26:29

still have the equivalent of billionaires

26:31

and it will be about influence,

26:33

it will be about taste, and

26:35

it will be about how you

26:37

utilize your own resources and how

26:39

do you multiply it to have

26:41

the equivalent of future money. And

26:43

I went a little bit deep

26:45

here, but I for what it's

26:47

worth, I think that I think

26:49

the innate drive to create is

26:51

like a deeply human drive and

26:53

I think that exists in a

26:56

post-cabilistic world. I also have that

26:58

opinion and I really believe in

27:00

humans. Like, I want them to

27:02

succeed. Like, somebody asked me, one

27:04

of our prospective employees asked me,

27:06

what, like, Philip, in 10 years,

27:08

what do you want there to

27:10

be, like, what have you done?

27:12

And I want to save, like,

27:14

the human creative vision. I don't

27:16

want everything to be AI. I

27:18

really have the pleasure when I

27:20

go to an artisan shop on

27:22

my holiday, and I know that

27:24

somebody put in the intent and

27:26

put in the work, and I

27:29

want to interact with the story

27:31

of it. Okay, so today your

27:33

ICP is the analytical creative, which

27:35

is a little bit of a

27:37

unicorn, and over time, as you

27:39

can lower, as the models get

27:41

better, as you iterate on your

27:43

interface, you'll lower the bar, so

27:45

they'll really be just more the

27:47

creative, is your ideal user, you're

27:49

going to lower the bar of

27:51

how analytical you need to be

27:53

in order to use word. Yes,

27:55

but at the same time, you

27:57

know, my use of the word

28:00

creative is not. to what most

28:02

people associated with right now. I

28:04

think a good creative is also

28:06

using growth channels in the right

28:08

manner. They are creative about everything

28:10

that they do in this new...

28:12

uncertain word of AI where everything

28:14

is changing. And, you know, I'm

28:16

not thinking only about an artist

28:18

that's painting on the canvas. It's,

28:20

I think creativity is, can basically

28:22

show itself in so many different

28:24

aspects of work. Let's talk about

28:26

user interfaces and, you know, the

28:28

future gooies, the gooies of the

28:30

future. right before we film this

28:33

podcast you made the analogy that

28:35

you know transformers of the new

28:37

transistor maybe say a little bit

28:39

more about that and what you

28:41

think the new gooey is going

28:43

to be. So I think the

28:45

analogy here is that if the

28:47

LLM is the well if Transformers

28:49

the the new transistor and it's

28:51

being packaged as the model the

28:53

model is kind of the mainframe

28:55

let's call it you know and

28:57

then we took our sweet time

28:59

to utilize the power of that

29:01

mainframe in a GUI that's accessible

29:03

by billions of people. You know,

29:06

there has been really two big

29:08

spikes there. The first was the

29:10

desktop, you know, and Apple came

29:12

in coming up with their GUI.

29:14

And the second one's mobile. And,

29:16

you know, right now, we are

29:18

almost like exposing the numbers and

29:20

the logic in a chat style

29:22

thing, and nobody has had a

29:24

better idea. We think that the

29:26

document style is better for doing

29:28

kind of more complex work. Because

29:30

often when you try to achieve

29:32

something you just give it two

29:34

sentences and the model just runs

29:36

on its own and it's just

29:39

enough the two sentences our lead

29:41

investor recently said that the two

29:43

sentences is just about enough for

29:45

a model to hang itself on

29:47

and You know you will get

29:49

something completely different than what you

29:51

actually want it and this is

29:53

a problem of like lovable and

29:55

devins of this world as well

29:57

But I basically think that there

29:59

are better is coming and you

30:01

know whether they will be based

30:03

on AR or you know there

30:05

will be an assistant that's listening

30:07

to everything that we do I

30:09

that was actually my first company

30:12

augmenting human memory of always on

30:14

listening devices using GPT to and

30:16

Bert I've been in this since

30:18

the GPT two days I mean

30:20

I my research wasn't into LSTM

30:22

which are the precursor to the

30:24

transformer architecture and I've been in

30:26

this for a while I think

30:28

nobody has yet delivered on this

30:30

I want everything that I hear

30:32

to be somewhere in a searchable

30:34

database that also has the perfect

30:36

context about me, you know, the

30:38

way that I want to do

30:40

things. And I think those affordances

30:42

and those like, we called it

30:45

gooey, but it's really the underlying

30:47

way of interacting with intelligence, is

30:49

not going to be mainly chat.

30:51

I just don't believe it. Programmable

30:53

documents. Do you think that is

30:55

fundamentally what word word looks like,

30:57

u.y. wise, and call it five

30:59

years? I think there is more

31:01

and more magic in it, and,

31:03

uh... I would believe that I

31:05

want people still to be able

31:07

to do that fine grain work.

31:09

You know, we've linked it with

31:11

George Lucas doing the movie, you

31:13

know, in a way you almost

31:15

want to firstly start with the

31:18

high level thing, the two centers

31:20

description, and then zoom in, and

31:22

zoom in, and zoom in, and

31:24

create, you know, modules which make

31:26

the best scene that is five

31:28

seconds, and then combine them together

31:30

in that way. So what I

31:32

would like Wordware to be is

31:34

to transcend. abstraction layers and you

31:36

know be able to zoom it

31:38

all out start with a sentence

31:40

and have it run maybe see

31:42

whether it's working in the right

31:44

manner and then as you see

31:46

that some things are not doing

31:49

the thing that you want them

31:51

to do is to be able

31:53

to zoom in and so and

31:55

see maybe you know four sentences

31:57

of exactly what is doing and

31:59

what are they inputs to this,

32:01

what it's trying to do in

32:03

the middle and what are the

32:05

outputs, you know, that's kind of

32:07

the most simple one level in,

32:09

and then you want to zoom

32:11

in more and more and more

32:13

as you redefine and reiterate on

32:15

your idea of how this should

32:17

be done. How did you arrive

32:19

at the current user interface? I

32:22

think it does feel really novel

32:24

compared to how others are, you

32:26

know, enabling AI builders today. How

32:28

did you arrive at the current

32:30

user interface? Was it more experimentation,

32:32

listening to users? Was it you

32:34

philosophizing about, you know, what it

32:36

should be? I think currently the...

32:38

the approach to creating these agents

32:40

was a block based on the

32:42

2D canvas. And once, you know,

32:44

I've been building agents for a

32:46

long time, you know, I think,

32:48

you know, March 2023, I put

32:50

out the first article about how

32:52

to build agents and me and

32:55

Robert, my co-founder, we've been in

32:57

this for a long time. And

32:59

the more, the better the models

33:01

got, the prompting became more difficult.

33:03

because you can do more complex

33:05

things with it. So at some

33:07

stage there was this movement of

33:09

like the prompt is going away,

33:11

so on. We actually really disagreed

33:13

with it. And that idea is

33:15

gone a little bit. It's like,

33:17

you know, we came back, did

33:19

a loop again, and be like

33:21

actually communicating your vision is really

33:23

important. And when we tried to

33:25

communicate our vision, which was a

33:28

little bit ahead of what the

33:30

models could have done at the

33:32

time, we started to notice that

33:34

the 2D canvas. It's just not

33:36

enough. Like if you do a

33:38

reflection loop inside of a reflection

33:40

loop, you run out of dimensions.

33:42

And we basically really like the

33:44

way that code is structured. Code

33:46

has an ability to express very

33:48

very complex concepts in a way

33:50

that is still like you can

33:52

still much. manipulate it and understand

33:54

it. Think about trying to structure

33:56

the, you know, the whole Uber

33:58

app with all of like everything

34:01

in it on the 2D canvas.

34:03

It would become so cluttered and

34:05

so messy. You know, you can

34:07

do the big picture thing, but

34:09

not really the, you know, you

34:11

don't want engineers to be interacting

34:13

in that way. You want the

34:15

engineers on the future World War

34:17

engineers to be interacting with something

34:19

that's easy to grasp the structure

34:21

of very complex systems. Whereas the

34:23

Uber app actually could probably be

34:25

described in pseudocode. And it seems

34:27

like you're, you know, you're getting

34:29

people closer to that vision versus

34:31

the 2D campus. Yes. And I

34:34

think, you know, the most important

34:36

part here is that Uber has

34:38

an agent equivalent and this is

34:40

what we're trying to build, you

34:42

know, if you... want an agent

34:44

to decide where is that person

34:46

going and where they starting their

34:48

journey and where they will accept

34:50

that charge or you know you

34:52

want to maybe make sure that

34:54

the charge is right for that

34:56

particular person. There is an agent

34:58

equivalent there and you know people

35:00

are going to like people can

35:02

build that agent on word where

35:04

it's not like you're going to

35:07

create that whole UI with with

35:09

with for Uber and I think

35:11

you know probably Uber. is the

35:13

right obstruction layer. You don't want

35:15

to be ordering an Uber through

35:17

a chatbot or through like a

35:19

voice-based thing or, you know, but

35:21

you might want an Uber to

35:23

be ordered for you if you

35:25

have a counter-invite. So, you know,

35:27

in a way that like for

35:29

your personal use Uber is nice

35:31

because you can click around and

35:33

the agent will not always know.

35:35

But I was coming here and

35:38

I wanted a way more, actually

35:40

way more kind of get that

35:42

far yet. But I wanted a

35:44

way more to be ordered and

35:46

to be ordered perfectly when I

35:48

need this. And it's almost like

35:50

a assistant, personal assistant would do

35:52

this for me. And now that

35:54

capability is open to everyone. So

35:56

we'll soon have these kind of

35:58

affordances and these kind of obstruction

36:00

layers there. I think that's a

36:02

great note to end on. Should

36:04

we end on a lightning round?

36:06

Let's go. Okay. One or two

36:08

sentence answers only. Okay, first question.

36:11

What is your most hot take

36:13

or contrarian taken AI not related

36:15

to word where or everything we

36:17

just discussed? Pre-training will still gonna

36:19

matter. And Deep Seek is a

36:21

little blimp that people liked to,

36:23

people jumped on because people love

36:25

a good drama and it was

36:27

connected to China and actually it

36:29

doesn't matter that much. Okay, I

36:31

know I said lightning round, but

36:33

you have to say more. What

36:35

do you mean it doesn't matter

36:37

that much? I mean, they utilized

36:39

some cool techniques and the rest

36:41

of their community is going to

36:44

learn from that. However, you know...

36:46

Like the fact that they like

36:48

trained it for a little bit

36:50

cheaper for like a lot cheaper

36:52

Does not involve all the experimentation

36:54

that they did before that and

36:56

You know I'm I don't know

36:58

if I'm supposed to say it

37:00

But I'm pretty sure they had

37:02

access to the best invidious as

37:04

well for that experimentation and It's

37:06

not that novel like people jumped

37:08

on it because they were like,

37:10

oh my God, China is taking

37:12

over the race and so on

37:14

and invidious stock price like plummeted.

37:17

And I just think it's another

37:19

place where some models were trained

37:21

that were open sourced and it's

37:23

not gonna, you know, we're not

37:25

gonna remember it in like a

37:27

year or like even six months

37:29

or maybe they will take over,

37:31

but the model doesn't really matter

37:33

that much. How you kind of

37:35

work with that best model out

37:37

there, that's what matters. That is

37:39

a hot take indeed. Okay. Next

37:41

question. Who's going to have the

37:43

best frontier model next year? Oof.

37:45

I think Open AI is always

37:47

super bullish and they always promise

37:50

a lot. And then I was

37:52

just going to talk with Sam

37:54

Altman on the YCAI retreat and

37:56

the O3 that the way that

37:58

he pictured it sounded great. But

38:00

I think we both know that

38:03

they over promise a little bit

38:05

a lot. And I love Entropic.

38:07

I think their kind of vision

38:09

and their kind of the way

38:11

that they've created this is great.

38:14

But recently Gemini 2.0 Pro with

38:16

their abilities to ingest 6,000

38:18

pages of PDF is really blowing

38:20

my mind. So end of the

38:23

story is. I have no clue.

38:25

This is a place where it's

38:27

a place where it's super fragmented

38:29

and people have zero loyalty. Pre-training

38:32

is hitting a wall. I think,

38:34

you know, famous people including Ilea

38:36

have been have included saying something

38:39

to that extent recently.

38:41

Agree or disagree? Disagree. Right now

38:43

I think, you know, it's the

38:45

intelligence of a model is

38:48

linked logarithmically to the resources

38:50

that is needed to train

38:52

it. But doing a 2X of

38:54

intelligence is on its own exponential.

38:58

Like if I'm smarter 2X

39:00

than somebody else, it doesn't

39:02

mean I'll do 2X of

39:04

the work. It means that

39:06

I'll find ways that probably

39:08

mean I'm a 10X or

39:10

even more. Favorite new AI

39:12

app, not word where? I would say I

39:14

started to edit content because we need

39:17

to explain and educate. people a little

39:19

bit more about both word were in

39:21

AI so the script is something that

39:23

I've been I've been loving and I

39:26

use granola every day and the newest

39:28

model that I'm really impressed is the

39:30

German I 2.0 pro I really like

39:32

it. That is that's a hot take

39:35

as well. I haven't heard much of that

39:37

from people. I think they came out like

39:39

four days ago so people have not

39:41

been playing around with it. Their PDF

39:43

capabilities are awesome. What application

39:45

or application category do you think

39:47

will really go mainstream and hit

39:50

this year? I would love

39:52

to see I'm personally very

39:54

very involved with that whole

39:56

AI having the context of

39:59

your life. and being able

40:01

to, you know, basically make better

40:03

decisions based on the context. And,

40:05

you know, I've rewind, which, you

40:07

know, I think they are called

40:10

limitless right now. I've ordered their

40:12

pendant, by the way, it's been

40:14

like a year and a half

40:16

and I still don't have it.

40:19

I don't know. Send it to

40:21

me or... If you're listening, please

40:23

send it. And I had to

40:26

change a color, because I know

40:28

they didn't have the color, but

40:30

I would love... for there to

40:32

be a provider which has a

40:35

lot more context and can do

40:37

the personal stuff for me. Don't

40:39

you think that's Apple over time?

40:41

I was just about to say,

40:44

I think, ideally that N421 model

40:46

or whatever it's called of the

40:48

AR glasses that they are trying

40:50

to push out there, which I

40:53

think Facebook has taken over a

40:55

little bit. Maybe we'll see early

40:57

stages of that and I think

40:59

they're the only ones. where the

41:02

privacy really like they have a

41:04

good brand around privacy and two

41:06

even if your new AR glasses

41:09

run out of battery it's still

41:11

cool to be wearing a $5,000

41:13

you know a piece of hardware

41:15

and maybe that's the UX but

41:18

I don't know what's that UX

41:20

and like a microphone so far

41:22

failed. Yeah. single piece of content that an

41:24

AI official should read or watch? I would

41:26

say all of the deep learning dot AI

41:28

resources. Everyone, like we have a bunch of

41:30

candidates apply for jobs. By the way, we're

41:33

hiring whatever I should be looking very, very

41:35

aggressively. So come join Wordware. But the deep

41:37

learning that AI resources are awesome and they

41:39

explain everything from from the bottom layer all

41:41

the way to the practical layer of how

41:43

to actually get it done. I also think

41:45

if you don't understand the the

41:47

underlying technology go Go see Free

41:50

Blue One Brown, incredible channel

41:52

on channel on YouTube, everything super

41:54

well everything super well. think think.

41:56

Wonderful. your lightning was full

41:58

of full of I didn't

42:00

even have to ask

42:02

you for a specific to

42:04

ask you for a specific thank

42:07

you so much for

42:09

coming on. I really

42:11

enjoyed chatting about on. know

42:13

How you see the

42:15

world evolving from how you

42:17

see the to from developers to word

42:19

you know you know, word where

42:21

engineers, if goes right, and appreciate you

42:24

you sitting down to

42:26

show your vision and

42:28

your hot and your hot you

42:30

for having me. Thank

42:32

you me. Thank you. You