0:01

Hey everyone, I am excited. We have

0:03

a very very special guest. Logan Capatrick

0:05

from Google Deep Mind is here and

0:07

he's gonna break down the aspects of

0:09

Google Jim and AI that people aren't

0:12

using, that they should be. Gonna talk

0:14

about where the real value is, gonna

0:16

talk about their brand new image generation

0:18

API, their new reasoning models, deep research,

0:20

so much. If you were trying to

0:23

learn about Google AI, this is the

0:25

show for you. Let's get to today's

0:27

episode. We'll

0:31

be right back to today's show, but first

0:33

here's a quick word from Hub Spot.

0:35

Marketing in 2025 is wild. Savvy

0:37

customers spot fake messaging instantly. Privacy

0:40

changes make ad targeting a nightmare.

0:42

And everyone needs more content than

0:44

ever. That's why you have to

0:47

have Hub Spot's new marketing trends

0:49

report. It doesn't just show you

0:51

what's changing. It shows you exactly

0:53

how to deal with it. Everything's

0:56

backed by research, but focused on

0:58

marketing plays you can use tomorrow.

1:00

Ready to turn your

1:02

marketing hurdles into results.

1:05

Go to hubspot.com/marketing to

1:07

download it for free.

1:22

Logan really excited to have you on the podcast

1:24

big follow of yours on Twitter for a long

1:26

time or X for a long time Excited about

1:28

your move to Google having following that as well,

1:31

but let's maybe start not at the start for

1:33

you But maybe at the start for some of

1:35

us who started following along your journey with being

1:37

an early employed over at Open AI like one

1:39

of the most transformative companies there is One of

1:41

the things we want to start with is it

1:43

must have been a pretty wild time to be

1:45

part of that company What was one of those

1:47

exciting launches that you were there for? One of

1:50

the ones where you were like, wow, this is

1:52

just going to be much more impactful than anyone

1:54

really gives a credit for? Yeah, that's a

1:56

good question. I think the one that always

1:58

stands out for me and I'm Obviously in

2:00

hindsight we all know how impactful this launch

2:03

was. It was really simple and sort of

2:05

had all the fanfare as far as like

2:07

actually being a part of it on the

2:10

ground was honestly the GBT4 launch. When we

2:12

launched GBT4 if folks remember the live stream

2:14

that Greg Brockman did where he sort of

2:16

talked about the models and he sort of

2:19

did the infamous example of like drawing a

2:21

little picture of a website and showing the

2:23

model and then it writing the code like

2:25

you think now what. that model was able

2:28

to do two years ago versus today, it's

2:30

like we're in an entirely different world than

2:32

we were two years ago. But I think

2:35

that example, I worked with Greg a bunch

2:37

on that demo and was actually, I have

2:39

a great picture like sitting right in front

2:41

of him as part of that demo and

2:44

we were going through it. And I think

2:46

the reflection for me is just like on

2:48

how far we've gotten from that moment, both

2:50

from like an AI like actual capability standpoint,

2:53

but also like I think a lot of

2:55

the innovation has been the infrastructure to the

2:57

infrastructure to. bring AI to like actually be

3:00

useful. And I think even today I was

3:02

having a conversation last night with someone who

3:04

was talking about the raw capabilities of the

3:06

model versus like this sort of AI building

3:09

harness that has been created in the last

3:11

two years and how much that actually makes

3:13

a difference for like the raw capabilities that

3:15

you can get from these models. And I

3:18

think it still feels like we're so early

3:20

in that sense that like you can get

3:22

those like GBT4 level. And we were talking

3:25

off camera before about native image generation, which

3:27

I'm and like. having this like GPT-4 level

3:29

moment of people seeing, wow, this is like

3:31

an incredible experience like just out of the

3:34

box. So it's still so fun to see,

3:36

you know, you think all the juice has

3:38

been squeezed out and then actually you're right

3:41

around the corner from like something that's going

3:43

to change how people think about the world.

3:45

Yeah, there was a couple of quick pointers

3:47

about that specific video. Kipp and I covered

3:50

that video. I feel like that demo was

3:52

like the nearest we've ever had to like

3:54

the iPhone moment for it was like very

3:56

similar to like that iPhone moment, but for

3:59

AI where people were like, holy smokes, like

4:01

this is, I get it, right, like they

4:03

collect. world who watched that demo that one

4:06

succinct use case really kind of sparked creativity

4:08

across such a wide range of folks internally

4:10

were you always all kind of constantly surprised

4:12

by how big those launches were like just

4:15

how transformative they were or did you kind

4:17

of know this is going to rapidly change

4:19

the way that people think about the world

4:21

like when we launched GD4 everything is different?

4:24

Yeah I think GD4 actually was one of

4:26

those moments and there's sort of some somewhat

4:28

differing perspectives about chat GBT and how much

4:31

of a unknown success story that was going

4:33

to be. I think the GBT four folks

4:35

knew like the real reason that chat GBT

4:37

came out was so that open AI could

4:40

try to experiment with different experiences of bringing

4:42

capable models to the world knowing that GBT

4:44

four was coming. So like chat GBT was

4:46

sort of intended to be the early, you

4:49

know, get feedback from the world. Does this

4:51

chat about thing actually end up being useful

4:53

for models? Because at that point GPT4 had

4:56

already finished training and you know opening I

4:58

was trying to figure out how do we

5:00

productize this really really cool technology but I

5:02

think folks knew GPT4 was going to be

5:05

that useful like they'd been sitting on it

5:07

for a very long time like I think

5:09

the model finished training in summer of 2022

5:11

and sort of made its way out to

5:14

the world in March of 2023 so there

5:16

was a long while where like folks had

5:18

almost fully wrapped their head around this technology

5:21

and what I was capable of, sort of

5:23

experience of experience of getting to sit on

5:25

the technology for that long and actually having

5:27

it be differentiated with what was available externally

5:30

in the world and thinking about now, I

5:32

think about for us, it's like, you know,

5:34

model comes out of the oven from being

5:36

trained and it's like, it has all the

5:39

safety stuff baked in and like, let's get

5:41

it out to the world in 24 hours.

5:43

And I think about like, actually, there's a

5:46

real trade-off and I think I'm not sure

5:48

how relevant this is for people who aren't

5:50

releasing models to the world who aren't releasing

5:52

models to like. figure out the story and

5:55

like explore and build it and like we're

5:57

actually figuring out a lot of that stuff

5:59

like with the external world as we're all

6:01

using this model together publicly and maybe that's

6:04

actually the best thing for the world because

6:06

you don't want to be like sitting on

6:08

the alien technology for you know six months

6:11

before it actually makes its way out but

6:13

I think you only get that level of

6:15

like really cool demo that Greg was able

6:17

to do by being able to sit on

6:20

the technology and like really internalize and I

6:22

think Greg to his credit as part of

6:24

that demo like he drove that whole thing

6:26

and like he was able to put it

6:29

together because he had fully internalized like what

6:31

the model was actually capable of and like

6:33

I don't see a lot of that happening

6:36

with like today's era of launches which is

6:38

really interesting. Yeah I think the point you

6:40

bring up is really important is that like

6:42

2020-25 is like a year in which a

6:45

decade is going to happen right like the

6:47

pace is very aggressive and if you're watching

6:49

the show and you may be a distant

6:51

observer of AI What it generally is doing

6:54

for everybody, any company, is that the expectations

6:56

of pace and speed have just gone up.

6:58

It's not just the AI models, it's literally

7:01

every company. And so you have to know

7:03

what your stories are, kind of your core

7:05

principles of what you're building, so that you

7:07

can kind of continue to build the story

7:10

and the product in parallel, because it's rare

7:12

you're going to have the like, hey, I

7:14

know everything, I've got this smooth six months

7:17

launch period, gone are those days right now,

7:19

right, right, right, especially if you are out

7:21

there building stuff, because that six months somebody

7:23

might build something way better and the work

7:26

you have is just completely obsolete and that

7:28

is I think just what 2025 I will

7:30

probably remember the most is that just kind

7:32

of core change in speed and trajectory. Yeah

7:35

I wish for my own sanity that that

7:37

was not the case because it feels like

7:39

this is like the three-year sprint that never

7:42

stops so it's also at a very human

7:44

level feels more and more important than ever.

7:46

And this is what gets me so excited

7:48

about, you know, the human experience. This world

7:51

of AI and all the innovation, the pace

7:53

that's happening, like at the end of the

7:55

day, it just like further exacerbates how. important

7:57

it is for like to do all the

8:00

things that it means to be human and

8:02

to like have those experiences and I so

8:04

fundamentally believe that. Yeah, it gets me happy

8:07

on both ends of the spectrum because I

8:09

love to see all the coolness of stuff

8:11

coming out in AI but like I'm also

8:13

just so bullish on like the human experiences

8:16

that only humans can have and and create

8:18

and all that stuff even in a world

8:20

where AI is intelligent. So you changed, you

8:22

went to Google, Google's pace is also, I

8:25

think, expediated. Like they have some really killer

8:27

launches. I think maybe they're a little more

8:29

understated than Open AI or some of these

8:31

other companies because they have such a whole

8:34

host of other challenges to navigate when they

8:36

release AI. But what are maybe a couple

8:38

of subtle differences that you've seen between how

8:40

Open AI approach and AI in general and

8:43

what Google's approaches? Like what are some of

8:45

the subtle differences you've seen how those

8:47

two companies are trying to expediate AI

8:49

to the world? Yeah, I'll maybe take two

8:51

ends of the coin here. First from

8:53

like the core technology standpoint, but second

8:55

like how we show up in the

8:57

world and talk about our AI stuff,

8:59

which I think will be helpful for

9:01

folks who are thinking about this. On

9:03

the core technology side, one of the

9:06

things that Google really benefits from is

9:08

like the breadth of all the work

9:10

that's happening. So our team's now a

9:12

part of Google Deep Mind. On one

9:14

end of the spectrum, we have, you

9:16

know, alpha fold and protein folding and

9:19

Nobel Prize, you know, image generation models

9:21

and there's weather models and there's Gemini

9:23

and there's there's this huge breadth of

9:25

different stuff that's happening from an AI

9:27

perspective and Deep Mind is really the

9:30

only place in the world where that

9:32

depth is actually happening and you can look

9:34

at all the other labs and you can

9:36

see the models that they're creating and like

9:38

the models that are being created at other

9:40

labs is like one pillar of what Deep

9:42

Mind is doing and I think the really

9:45

exciting thing to me is like if you

9:47

buy into the idea that this like multi-modal

9:49

multi-capability enabled model is going to be the

9:51

thing that enables humans to do all the

9:53

things that we want to do like there's

9:55

only one place in which it's possible that

9:58

that's going to be created and it's inside

10:00

of DeepMind because of this breadth

10:02

of all the work that we're doing.

10:04

And we see this actually happening

10:06

in practice with the cross -pollination of

10:08

research from AlphaFold, again, to weather models,

10:10

to AlphaProof, which is our math

10:13

model, and how all that actually trickles

10:15

back into the mainline Gemini model

10:17

that is available to consumers and is

10:19

available to developers. So I'm super

10:21

excited about that. I'm excited for us

10:23

to lean more into that story

10:26

and lean more into that advantage that

10:28

DeepMind has on the other end

10:30

of the coin after the fundamental research

10:32

and product creation happens. I think

10:34

there's a huge difference in how we

10:36

approach going and telling the world

10:39

about the products and the models that

10:41

we're building. And I think a

10:43

lot of this is just grounded in

10:45

the positions that these different companies

10:47

are in. There's a

10:49

lot of factors at play.

10:51

I think people perceive Google as

10:53

a company very differently than

10:55

people perceive OpenAI as a company.

10:57

OpenAI also has a very

10:59

different product offering. Google has many,

11:02

many, many products across many,

11:04

many different domains. And there's just

11:06

a lot of downstream impacts.

11:08

Beautiful thing for OpenAI is they

11:10

have a clean slate in

11:12

many ways. So people either haven't

11:14

formed a prior about some

11:16

specific angle of what they're doing

11:18

or even something very tactically.

11:20

They have an open namespace. They

11:22

can call their products anything.

11:24

They can use whatever URL they

11:26

want, et cetera, et cetera,

11:28

because there's nothing conflicting with that.

11:30

They don't need to worry

11:32

about the crossover between these different

11:34

products. And I think about

11:36

this a lot because we get

11:38

a lot of feedback from

11:40

the external world that, hey, we

11:42

wish this thing was simpler.

11:44

We wish this naming schema was

11:46

a little bit easier to

11:48

follow. And a lot of this

11:50

is just the artifact of

11:52

the complexity that Google has because

11:54

of how large of a

11:56

company it is. And then in

11:58

turn, we have to find

12:00

ways to try to more authentically

12:03

lean into the things that

12:05

matter to us. And I think

12:07

this is one of the

12:09

biggest challenges for Google. Like, I'm

12:11

sure there's folks in your

12:13

audience who... have thought about this or experienced this at

12:15

different companies, like it's just hard to tell a really authentic

12:17

story as Google, not because there's not interesting authentic stories happening

12:19

inside of Google, it's just because it's the artifact of the

12:21

size of a company that you're a part of the size

12:23

of a company that you're a part of, and I think

12:25

as the size of a company that you're a part of,

12:27

and I think as the size of a company increases, like,

12:29

about how we can tell this authentic story because

12:31

I feel like we miss telling the magic

12:33

of why this technology is so important when

12:36

you sort of don't go the authentic route.

12:38

And I like feel these so deeply inside

12:40

of me these like really authentic interesting AI

12:42

stories that again that only Google can tell

12:45

and like but the only way to land

12:47

that message is in a really authentic way

12:49

and there's just like so many different angles

12:51

of this tension to reconcile and I just

12:54

don't think that open AI is an example

12:56

like they don't have to deal with this

12:58

because they're just not the size of a

13:00

company where like they have all those different

13:03

dimensions of tension right now. The one thing

13:05

I would probably add Logan as a consumer

13:07

of both products is that like one of

13:09

the advantages Google has is that like the

13:12

breadth of all of their tools and being

13:14

able to just like seamlessly integrate Gemini and

13:16

even if the story isn't there. When I

13:19

opened up Google Maps and I looked at

13:21

a place and Jim and I was now

13:23

on the place listings and I could just

13:25

ask Jim and I anything about that place,

13:28

I was like, holy cow, this is incredible.

13:30

I could just ask it, oh, I'm

13:32

going to this place in New York.

13:34

If I get there at 10 a.m. How

13:36

long am I going to have to wait

13:38

in line? Basic things like that that would

13:41

have been impossible or I

13:43

would have read like 30 credit threads

13:45

to find out the answer is now

13:47

like a 10. into those wow moments.

13:49

It's almost the story is getting told by

13:51

the user just discovering those features too. Yeah,

13:53

I think that's a great point. And I

13:55

think this speaks to like, again, one of

13:57

those advantages from the breadth perspective is like.

13:59

when Deep Mind builds Gemini. We're not

14:02

building it for a chat app. Like

14:04

if you think about like where Gemini

14:06

is actually being integrated, it's like across

14:08

some of the largest product suites that

14:10

touch the most users in the entire

14:12

world and like that has a very

14:14

different set of constraints that would potentially

14:16

be built and like Gemini is powering

14:18

search, Gemini is in YouTube and like

14:20

these are like billion billion billion user

14:22

products that like all of these very

14:24

nuanced characteristics matter a lot. And actually

14:26

we've seen like lots of really great

14:28

examples of you know the requirements for

14:30

a model to be really good for

14:32

search, actually leading to like something that's

14:34

really great from a developer perspective. We've

14:36

seen this with a few of the

14:38

last 1.5 series of models where, you

14:40

know, the search team needed something and

14:42

it ended up being like a really

14:44

great tradeoff of capabilities that developers also

14:46

ended up wanting those things and it's

14:48

cool to have those levels of sort

14:50

of internal engagement from these teams. One

14:52

of the things you mentioned Logan was

14:54

there's just like a plethora of different

14:56

AI tools available kept mentioning that we're

14:58

going to see a decade of progress

15:00

in a single year. And so one

15:02

of the things we wanted to do

15:04

in this show is to try to

15:06

distill it down into like what do

15:08

you think are great AI use cases

15:10

for our audience to take away from

15:12

the show and start to implement and

15:14

maybe we can specifically think about Gemini

15:16

and Google, but what are some of

15:18

the AI use cases today? Like that

15:20

you can actually just go and start

15:22

using Google Gemini for today that you

15:24

think are like widely underestimated or underused

15:26

by the average consumer of AI, the

15:28

person who maybe isn't in the details

15:30

and in the weeds day in day

15:32

out. This is a good question. I

15:34

think one of the challenging things about

15:36

this question is that it is a

15:39

ever evolving answer because literally the capability

15:41

flywheel is spinning as we speak, which

15:43

is awesome. I think today some of

15:45

the things that are getting me most

15:47

excited is in December, we launched deep

15:49

research. So we launched sort of the

15:51

world's first iteration of deep research, which

15:53

if folks haven't used it, is essentially

15:55

a research assistant. You can put in

15:57

whenever your query is and the model

15:59

will go off and search. in the

16:01

context of our deep research, visit, you

16:03

know. thousands potentially of different websites to

16:05

answer the question. I think that simple

16:07

product artifact of like showing you the

16:09

number of websites that the model browsed

16:11

through in order to get you the

16:13

answer is the thing that just makes

16:15

that product experience work for me. I'm

16:17

like in zero percent of any of

16:19

the research I've ever done in my

16:21

life, have I looked at more than

16:23

10 websites? So the fact that the

16:25

model went out and looked at a

16:27

thousand websites like just gives me like

16:29

a lot of confidence in the model.

16:31

I'm happy to up and we can

16:33

talk more about this, but I think

16:35

this is one of the biggest challenge

16:37

for AI products, which is like in

16:39

many cases AI products are like asking

16:41

the user to basically do all this

16:43

up front work in order to provide

16:45

them value. And I think deep research

16:47

is this great example of like the

16:49

model of the technology, just sort of

16:51

doing the heavy lifting for you and

16:53

you as a user get to just

16:55

like ask your silly question or your

16:57

serious question, and then the model goes

16:59

and finds it. So I love deep

17:01

research. line of reasoning models, which is

17:03

awesome. I agree with you, Logan, that

17:05

like far and away the number one

17:07

thing, if nobody's really using AI for

17:09

much than just some random questions, they

17:11

should use deep research. Deep research is

17:14

incredibly powerful. And the amazing thing about

17:16

Google Gemini, and you correct me if

17:18

I'm wrong, is that deep research is

17:20

a free feature. I don't think there's

17:22

any hard limits on deep research yet.

17:24

And really, the biggest thing there is

17:26

you then rolled out these reasoning models,

17:28

which makes it think through the ability

17:30

of what sources and what follow-up questions

17:32

to ask around those sources far, far

17:34

better than it was just a couple

17:36

of months ago. So you could enter

17:38

this drop-down and you would pick whatever

17:40

model you want to use. I don't

17:42

think people fully always grock. When do

17:44

I use this deep research? Like what's

17:46

the perfect way to delineate between flash,

17:48

which is also incredibly great an AI

17:50

model. Now I need the deep research,

17:52

plus you mentioned something really important, which

17:54

is now you've integrated reason and into

17:56

that. Maybe just explain what we... mean

17:58

to users that you've integrated reasoning to

18:00

this deep research? Yeah, I think that

18:02

basically the land is like your everyday

18:04

questions, you know, if it's a simple

18:06

question. just use 2.0 flash, it's going

18:08

to be very quick. It'll get you

18:10

an answer like almost instantly. If you

18:12

really do need something that is not

18:14

surface level, like if you're looking for

18:16

like, you know, who won the Cubs

18:18

game yesterday, you know, if you don't

18:20

need deep research for that, if you're

18:22

trying to understand like why the Cubs,

18:24

and I'm based in Chicago, and I'm

18:26

not even a Cubs fan either, but

18:28

I'm using the Cubs example, but if

18:30

you're trying to. understand why the Cubs,

18:32

you know, builds whatever the ivory wall

18:34

is around the back of the field

18:36

and like what the technique that they

18:38

used to build that and who the

18:40

people were who worked on it and

18:42

like what type of permitting they needed

18:44

in order to put that together, like

18:46

that level of depth in the question

18:48

that you have. There's no product that

18:51

can do that besides deep research. And

18:53

I think the reasoning models is the

18:55

sort of key enabler this and we

18:57

initially launched deep research back in December

18:59

with Jim and I 1.1.1.5 Pro. It

19:01

was really powerful, but a lot of

19:03

the techniques being used by 1.5 pro

19:05

were actually like trying to get it

19:07

to do what a reasoning model actually

19:09

does, which is be able to sort

19:11

of have essentially this like inner monologue

19:13

of, you know, thinking through different pieces

19:15

of a question, like actually reflecting back

19:17

on the initial answer that it's given

19:19

in like trying different versions of this.

19:21

And you can sort of think about

19:23

how we as humans think through this

19:25

process, by default AI models just kind

19:27

of. spit out an answer as quickly

19:29

as they can is basically the way

19:31

that models are trained today. And the

19:33

thinking models are trained to like don't

19:35

actually spit out the answer as quickly

19:37

as you can. iteratively go through this

19:39

process, try a bunch of different things,

19:41

make sure that you're sort of covering

19:43

the breadth and depth of what a

19:45

user might be asking for. And it

19:47

actually leads to some pretty substantively different

19:49

outcomes. So if folks have tried stuff

19:51

before with AI and you're like, ah,

19:53

the models are just dumb and don't

19:55

really have the ability to do these

19:57

things, try the new sort of generation

19:59

of reasoning models. I think there's a

20:01

lot of use cases that just were

20:03

not possible before that like all of

20:05

a sudden just work today. I am

20:07

very convinced that most humans do not

20:09

realize what like deep research especially is

20:11

capable of. For example, like I was

20:13

wanting to get rid of like a

20:15

tree in my yard and I was

20:17

like, I don't know what the permitting

20:19

process is. I don't know what you

20:21

would need to do. It did all

20:23

of that told me exactly what I

20:26

could and couldn't do, what the rules

20:28

are. And I literally just said. I

20:30

wanted to remove a tree and gave

20:32

it my address and it did everything

20:34

else, right? And it's like people just

20:36

wouldn't think that it could do things

20:38

like that. Like I had to estimate

20:40

an entire construction project. It's like, well,

20:42

all right, somebody's giving me this estimate.

20:44

What does deep research say? And it's

20:46

incredible how detailed it is and its

20:48

ability to go through complex documents and

20:50

frame things in simple ways is really,

20:52

really good now. Yeah, and I think

20:54

it's wild to just reflect on that.

20:56

This is like the V0 of the

20:58

product. Like truly, like, this is the

21:00

bare bones version of what deep research

21:02

can possibly be. And like today, it's

21:04

essentially just using search. And there's more,

21:06

and I actually think at the time

21:08

of this going out. because it's rolling

21:10

out today, the ability for deep research

21:12

to be combined with audio overviews from

21:14

Notebook LM is also rolling out. So

21:16

now you can sort of take that

21:18

deep research experience that you had and

21:20

then actually just click a button, turn

21:22

the entire thing and do an audio

21:24

overview and then you have a podcast.

21:26

Let's do that. And you're learning about

21:28

the permitting process in whatever city you

21:30

are. And you can ask it questions

21:32

and interrupt that podcast and be like,

21:34

well, I don't understand what you mean

21:36

here. Yeah, so I think the whole

21:38

experience and the way in which humans

21:40

sort of get this type of information

21:42

is changing, which is I think a

21:44

good thing. I'm curious if you had

21:46

this perspective on the permanent, I just

21:48

wouldn't do it to be honest or

21:50

like, or I would pay like, you

21:52

end up paying like this like market

21:54

inefficient price because you're like, well, I'm

21:56

not willing to. interesting to me so

21:58

like therefore I'm gonna get gouged by

22:00

somebody and they're gonna charge me twenty

22:03

thousand dollars to take a tree out

22:05

of my backyard or something crazy like

22:07

that. But it's also not even the

22:09

price right it's the time to get

22:11

to that same outcome would have taken

22:13

me weeks yeah and it took like

22:15

five minutes yeah right it's like I

22:17

can just accomplish so much more in

22:19

my life and I don't think any

22:21

human realizes like the rate of progress

22:23

you can now make. So

22:30

if you're listening to this podcast,

22:33

you are more than likely thinking

22:35

about business ideas and how to

22:37

market your business. Every entrepreneur has

22:40

a million ideas vying for attention

22:42

in their brain. And there's probably

22:44

that one idea you can't stop

22:46

thinking about. The one that came

22:49

to you at 2am and you

22:51

got up and you wrote it

22:53

down somewhere. Well, I'm here to

22:55

tell you that that idea could

22:58

be worth $5,000 right now. The

23:00

hustle and indie hackers are teeming

23:02

to find the next great founder

23:04

and get him or her started.

23:07

with $5,000. All you have to

23:09

do is record a 60-second video

23:11

pitch telling them your business idea

23:13

and how it makes money and

23:16

why it's awesome. One winner gets

23:18

$5,000 cash plus promotion to millions

23:20

of hustle readers. Make sure you

23:23

enter by April 4th at the

23:25

Hustle.co/Big Break. The winner will be

23:27

announced on April 11th on the

23:29

Hustle Daily Show podcast, the Hustle

23:32

newsletter, and on the Hustle's social

23:34

channels. Now get to pitch it.

23:37

It's a combination of there's things you would

23:39

never have done because you didn't have

23:41

the time or money and now you can

23:44

do and just how quickly you can do

23:46

them. And one of the ones I just

23:48

don't want to gloss over because you mentioned

23:50

something that is actually one of my favorite

23:53

workflows for deep research. We did an entire

23:55

video of it, which is deep research into

23:57

notebook and then turned it into audio, create

24:00

podcast and ask podcast questions. That's actually

24:02

how I've been learning about call options. which

24:04

this is the worst time to start doing

24:06

call options, but that's a whole other thing

24:09

that can I complain around. No financial advice,

24:11

but that's one of the things we've been

24:13

playing around with. But when you say that's

24:15

being integrated, do you mean I can just

24:18

go to Google Gemini, like you're integrated in

24:20

LM notebook, into Gemini, or is the

24:22

research still going to exist separately, or are

24:24

they going to be combined? How is that

24:27

integrated together? Yeah, I think for today's launch,

24:29

when you do a deep research query inside

24:31

of the Gemini app, you'll see an option

24:33

when the deep research process is done to

24:36

like convert the deep research into an audio

24:38

overview. And I actually think maybe, don't quote

24:40

me on this, I think maybe, don't

24:42

quote me on this, I think there's maybe

24:45

like a direct way on this. I think

24:47

there's maybe like a direct way on this,

24:49

I think maybe, don't quote me on this,

24:51

like, would go back in a notebook alum.

24:54

But it is cool to just be able

24:56

to have that single shot version of the

24:58

overview created for you. Yeah, maybe since

25:00

we're on notebook alum, why don't you give

25:03

the users a little bit of your ways

25:05

of using notebook alum? Because that's actually a

25:07

favorite tool of this show. Again, I don't

25:10

think enough people even know that tool or

25:12

are using it enough. Yeah, I think for

25:14

folks who haven't tried notebook LM, you can

25:16

think about notebook LM sort of as the

25:19

knowledge assistant that you might have, like

25:21

imagine or a learning assistant or a tutor

25:23

that you might have, or honestly, just like

25:25

a way to bring content to life is

25:28

another version of this. Josh Woodward, who leads

25:30

the Google Labs team, has this idea of

25:32

sort of like infinite repurposing of content and.

25:34

I think notebook alum is like a really

25:37

great example of this. You can take, imagine

25:39

you have, you know, something really boring

25:41

like a onboarding manual to set up, you

25:43

know, a vacuum cleaner. And for whatever reason,

25:46

you're one of the 10 people who actually

25:48

reads the onboarding manual and you have to

25:50

read it, but you're really bored and you

25:52

don't. actually want to go through the whole

25:55

thing. You could take a PDF version of

25:57

that, you know, a 150-page onboarding manual that

25:59

says all the different gizmos and gadgets

26:01

and things that your product does, stick it

26:04

in a notebook album, and really quickly generate

26:06

like a summary, a learning guide, a principally,

26:08

like one of the things that folks are

26:11

most excited about, a podcast level conversation with

26:13

like really sort of witty and smart sort

26:15

of nuance in the conversation about whatever the

26:17

content is that you're uploading and you're

26:19

uploading. has features like you can actually interject

26:22

mid-conversation and say things like, you know, hey,

26:24

this is actually really boring, spice things up

26:26

a little bit, or, you know, I don't

26:29

understand this point that you were just making,

26:31

can you sort of re-articulate what this is?

26:33

And one of the most common flows that

26:35

I actually see folks who I work with

26:38

and I work remotely so I don't

26:40

commute to the office, even though I probably

26:42

should go to the office every once in

26:44

a while, folks take a bunch of like

26:47

work documents, they put them in notebook alum,

26:49

create audio reviews, and they like listen to

26:51

them on the drive of the work. Yes,

26:53

exactly. I know Kip has a ton of

26:56

use cases here, but just two really quick

26:58

ones. So this one here is exactly

27:00

what you said. It's like a strategic knowledge

27:02

of this, and that's specific to one of

27:05

the cross-functional pods that I run here. And

27:07

one of the cool things I can do

27:09

is like add all the documentation. So I

27:11

add all like meat and transcripts, add all

27:14

the docs that have been created that month

27:16

that week. And then when I'm out walking

27:18

to dogs, I can actually either listen

27:20

to it or listen to it. way of

27:23

being able to talk to the podcast directly

27:25

and ask questions. And so it's a pretty

27:27

cool way where you can actually take knowledge

27:30

with you with work and actually conversate with

27:32

it. Like if you're out of the office,

27:34

if you're driving, if you're walking, I have

27:36

knowledge assistance trained on every single project

27:38

and I can always ask them questions. And

27:41

so it's like you have a project assistant

27:43

slash executive assistant for every project. And as

27:45

I said, I wasn't lying. Like all the,

27:48

this is like one of the example where

27:50

I had deep research. Teach me about call

27:52

options and then try to pick out like

27:54

five call options that are underpriced and why

27:57

believe so things are underpriced. priced, using

27:59

third party sources, and what I say is

28:01

I use the most trusted third party sources

28:03

to come up with that hypothesis. And then

28:06

I had a whole conversation with it in

28:08

this interactive mode when I loaded the podcast

28:10

and had a whole conversation that lasted like

28:12

15, 20 minutes, all around its prediction on

28:15

why meta was a good call option. And

28:17

I kept telling you that I was

28:19

actually, that was done four weeks ago, and

28:21

actually it was pretty right, because I think

28:24

Kipp has already done that trade. So, maybe,

28:26

pretty powerful. I don't keep you love this

28:28

tool, but that's pretty powerful stuff. Well,

28:30

hold on. One of the threads we're

28:32

going here, and I do think this

28:34

is true, is that the Google suite

28:36

of AI products that you all

28:38

have wrote out, I think are the best

28:41

to help humans learn. Because

28:43

I'm about to pull a deep

28:45

cut out of one of my favorite

28:47

Google tools that nobody talks about,

28:49

because we've talked about deep research, we've

28:51

talked about, about, notebook, you can basically

28:53

just decide you want to learn about

28:56

something like how do I make pasta

28:58

from scratch for example and what I

29:00

love about learn about it's like where

29:02

deep research is like a deep dive

29:04

on a topic to gain perspective and

29:07

then you can kind of ask follow-up

29:09

questions on notebook alum. This is

29:11

much more like almost like a structured

29:13

learning process and course and it creates

29:15

this amazing composable web experience over here

29:17

and then you can break down and

29:19

dive into different components and basically It

29:22

solves a lot of the empty box

29:24

problem of like, hey, I don't know

29:26

anything about this thing. I don't even

29:28

know what to ask. And it's prompting

29:30

me a bunch of different aspects about

29:32

making pasta. I happen to know how to make

29:34

pasta, but it's like, if you don't know how

29:36

to add pasta to boiling water because you've never

29:38

done it before, like this is very helpful, right?

29:40

And so I think if you look at Deep

29:42

Research, Notebook LM and Google Learn, like there

29:45

are three tools that are really master classes

29:47

that are really master classes in helping people

29:49

learn. Yeah, I love that. And I feel like

29:51

the magic, the real magic and the bow of

29:53

this is like, how can you bring all that

29:56

together, understand a user's intent, put

29:58

the right product. in front of

30:00

them. Yes. Yes. This is going to be

30:02

my point. This is the challenge. The sentence

30:05

that I just said sounds, you know, simple

30:07

to say, when you actually look at it,

30:09

like these are the engineering and product problems

30:11

of the decade. Like really, these are like

30:14

not a trivial thing to bring together that

30:16

level of technology, especially in a world where

30:18

you're sort of balancing this user context

30:20

problem, which is like if folks. talk

30:23

to AI models all the time, like

30:25

every time you talk to a new

30:27

AI model, it has no context of

30:29

who you are, it doesn't know what

30:31

you've done before, which is another thing

30:33

that sort of just landed in the

30:35

Gemini app, which is the ability to

30:37

personalize model responses with your Google search

30:40

history. And the model can really intelligently

30:42

say, like, here's what this user has

30:44

done before, here's what they're interested in,

30:46

how can we actually use that to

30:48

sort of prime the model to give

30:50

you the model to give you the right?

30:52

pulling in the right like for you capital

30:54

know like here's this learn about like experience

30:57

because Kip's been doing a bunch of stuff

30:59

over here and this other product versus you

31:01

know search for call options and bringing that

31:03

stuff so there's a whole like spectrum of

31:05

different use cases and I think having the

31:08

personalized context means that you can get the

31:10

right product surface or you can get the

31:12

right product experience in front of the right

31:14

user persona which I think is not something that

31:17

happens in today's software 2.0 product suite.

31:19

software is static, it's like predefined for

31:21

you in a lot of ways. Just

31:23

to touch on this, because I think

31:25

this is a really important point you're

31:28

making, and actually something even internally we

31:30

are dealing with a little bit, which

31:32

is what you're basically saying is instead

31:34

of at some point, go into this

31:36

kind of drop-down experience, you will have

31:38

one assistant that you talk to, and

31:40

then the background the assistant is basically

31:42

deciphering intent, and the assistant can pass

31:44

you to whatever model that thinks fulfills

31:46

that. then you all have to do.

31:48

We have a sales assistant that can

31:50

actually help sell through chat. We have

31:52

a support assistant that can help do

31:54

a bunch of support tickets. We need

31:57

to actually create an upsell assistant so

31:59

we can actually upgrade. customers to different

32:01

tiers and sell them on those different

32:03

tiers. And what we have is like

32:05

these individual assistants. And the way we

32:07

decide for intent is somewhat like wherever

32:09

you are in the go to market,

32:11

we say, well, this is probably the

32:13

right assistant. But even us, we need

32:15

multi-bought orchestration at some point where you

32:17

have one assistant, the assistant can say

32:19

you're trying to buy product for the

32:21

first time. You have a support question.

32:23

You are our existing customer and you

32:25

need to upgrade. And that. is really

32:27

complicated. And so I can't even imagine

32:29

how complicated it is for a Google

32:31

where the intent is like any single

32:33

possible thing in the world. Like it's

32:35

like, how do you ever decipher that

32:37

intent? But I think what you're saying

32:39

is even when I think about my

32:41

AI experience with Google, you all have

32:43

a ton of great products and I

32:45

would love to end with one of

32:47

the ones that I think is. maybe

32:49

transformational for this year, the image generation,

32:51

but they're disparate. We jumped into notebook

32:53

algorithm, kept jumped into the learning product,

32:55

we jumped into all these different AI

32:57

models you can choose through to the

32:59

drop-down. Is the plan for Google to

33:01

like pull them all together at some

33:03

point? So you just have like one

33:05

AI interface as a consumer. Yeah, this

33:07

goes back to one of the challenges

33:09

that Google has. It's a company that

33:11

has lots of product services. I do

33:13

think more and more of the Gemini

33:15

app is becoming this sort of unified

33:17

place to get a lot of these

33:19

experiences and sort of a externalization path

33:21

for Kipa showing learn about these experiences

33:23

and sort of a externalization path for

33:25

Kipa showing Learnabout, which is one of

33:27

the Google Labs experiments. I think more

33:29

and more of those experiences are finding

33:31

their way into the Gemini app. user

33:33

journeys. Like I think about the four

33:35

or five different products that I use

33:37

like all day every day and like

33:39

there's different journeys apart of those products

33:41

for different users and I think they'll

33:43

continue to be like different products. We

33:45

didn't talk about AI Studio but like

33:47

A.S. Studio is another one of these

33:49

where like that's actually how folks access

33:51

the native image generation capabilities today and

33:53

like I think in the future we're

33:55

still going to have A. I. Studio

33:57

because like the user persona that we

33:59

care about A. is a developer who's

34:01

sort of exploring these models and wants

34:03

to build something, which is very different

34:05

in a lot of cases than sort

34:07

of the main experience for users going

34:09

to the Gemini app who like are

34:11

trying to use the Gemini app really

34:13

as like an assistant on a daily

34:15

basis. It's like a daily active user

34:17

product versus AI studios really intended as

34:19

sort of the portal to the developer

34:21

world for folks interested in Gemini. Maybe

34:24

talk about. the latest image release. When

34:26

I think about AI and where we

34:28

might have been a little wrong, I

34:30

thought video would be further along than

34:32

it is today. And I mean, AI's

34:34

ability to go from text to video,

34:36

right? It's still pretty clunky in a

34:38

lot of cases. It's not like production

34:40

ready. And then the other one was

34:42

images, text to image. And I felt

34:44

like the first iteration of that were

34:46

all these great tools where you could

34:48

go to image, but for image to

34:50

be really useful for people, there had

34:52

to be great editing tools as part.

34:54

Google's latest release, Kip and I were

34:56

playing around with it last week, is

34:58

really awesome. And maybe you could just

35:00

give the context on what that model

35:02

is and why it's so good. Yeah,

35:04

no, 100%. So for folks who haven't

35:06

been following closely, we launched Gemini 2.0

35:08

back in December, and we showcased these

35:10

capabilities actually, and we rolled them out

35:12

to a small group of trusted testers

35:14

to get some initial feedback. And then

35:16

last week, we rolled out to every

35:18

developer, the ability to use Gemini's native

35:20

image generation. And I think the thing

35:22

that's actually capturing a lot of interest

35:24

is the native image editing capability. Because

35:26

the model is natively multimodal, you can

35:28

pass in an image. And you can

35:30

say. hey, update this image to, in

35:32

the example we're looking at on screen,

35:34

add a little chocolate drizzle to these

35:36

crescents, or add a strawberry drizzle to

35:38

these crescents, or I've seen a bunch

35:40

of really cool examples, which I did

35:42

not think about, of taking black and

35:44

white images and actually asking the model

35:46

to colorize those images. And you can

35:48

bring them back, you can take in

35:50

two images and you can sort of

35:52

fuse them together. You know, I saw

35:54

a funny example of like. like hot

35:56

dogs and stuff like that and merging.

35:58

together into this comedic image and lots

36:00

of different random things like this that

36:02

is the full spectrum of really useful

36:04

to really silly. But the thing that

36:06

I think is capturing folks' attention is

36:08

if you think about how you would

36:10

have had to do this workflow, pre-native

36:12

image generation and image editing, it's just

36:14

hard to do. Like the number of

36:16

people who can do that in whatever

36:18

the professional tool is, is pretty limited.

36:20

You can just do it with like

36:22

a very, very simple text prompt. So

36:24

like now the opportunity space of people

36:26

like creating dynamically edited images is now

36:28

essentially every human on earth is able

36:30

to do that, which is just. This

36:32

is the thing that continues to blow

36:34

my mind is you get a capability.

36:36

And this is what AI is enabling

36:38

across so many different domains. It's this

36:40

thing that only a few people could

36:42

do. And then overnight it ends up

36:44

being this thing that everyone can do.

36:46

And then overnight it ends up being

36:48

this thing that everyone can do. And

36:50

I think actually coding is like before

36:52

image generation is also having this like

36:54

parallel moment with vibe coding and everything

36:56

they want. pretty much get what they

36:58

were looking for. It's the same thing

37:01

with image editing today. Before you had

37:03

to be very good at using one

37:05

of these tools in order to do

37:07

this stuff, and now it's like every

37:09

human on earth can do this, and

37:11

it just changed overnight, which is just

37:13

such a weird experience to think about.

37:15

Yeah, vibe design. I like that. This

37:17

is the kind of dream, I think,

37:19

for AI in general, or the transformative

37:21

mission of AI in general, as it

37:23

unlocks. creativity somewhat overnight because it allows

37:25

people now to unleash their creativity and

37:27

they're no longer hindered by having to

37:29

learn the tools. And I know that

37:31

sounds like a lazy way to think

37:33

about creation, but I don't think it

37:35

is because I think the creation part

37:37

is the important part and the learning

37:39

the tools shouldn't hinder that part of

37:41

you, like the ability to create things.

37:43

Is this visual story? Could you just

37:45

end on what this product is? Because

37:47

we discovered this today. I'd love to

37:49

know like, what is the way we

37:51

should think about visual story, which again,

37:53

for people who were following along on

37:55

YouTube or even on RS. These are

37:57

all available in AI Studio and you

37:59

should really go in and play with

38:01

that product because there's a bunch of

38:03

great tools in there Yeah, AI Studio

38:05

is again. It's our surface for developers

38:07

intended to bring the models to life

38:09

in a way that ultimately wants to

38:11

get you to build with them. So

38:13

here we're trying to, this example that

38:15

we're looking at, trying to capture a

38:17

developer, sort of imagination of what products

38:19

they could go and build themselves, but

38:21

principally AI studio is not intended to

38:23

be like the daily assistant product. Like

38:25

it's a very thin surface on top

38:27

of the models. We make a bunch

38:29

of like very opinionated decisions to keep

38:31

the core AI studio experience the same

38:33

as the experience. you would get in

38:35

the API. So we don't have a

38:37

bunch of like fancy bells and whistles

38:39

that you yourself couldn't do as a

38:41

developer in AI Studio, which makes it

38:43

limiting as a product. Like if you

38:45

want, like people are always like, why

38:47

don't we have deep research in AI

38:49

Studio? I want that. It's like because

38:51

deep research is something you could build

38:53

sort of a similar deep research experience

38:55

using the API, but it's not available

38:57

to developers today. So we don't want

38:59

that experience in AI Studio. Yeah, so

39:01

if you're someone who wants to build

39:03

stuff, getting your Gemini API key, all

39:05

that stuff, testing out the latest capabilities

39:07

of the models, happens in AI Studio,

39:09

and our team sits in Google Deep

39:11

Mind right next to the model team

39:13

physically, so that oftentimes our product is

39:15

sort of the fast path to externalize

39:17

the latest Gemini models, which is a

39:19

lot of fun, and it's cool to

39:21

be on the frontier, and it's cool

39:23

to see the excitement with native image

39:25

generation. Very cool. kind of like the

39:27

way ChatGBT was for text. I feel

39:29

like this is similar for images. It's

39:31

really the first time I've seen the

39:33

ability to kind of get the image

39:35

crisp and concise to the way you

39:38

want it. This was awesome. Logan, we

39:40

really appreciate you coming on and giving

39:42

us a deep dive into what it's

39:44

been like to be part of this

39:46

kind of AI journey. You're working with

39:48

two of the most transformational companies there

39:50

are you or you have and also

39:52

just going deep into like how people

39:54

can start using the Google tools straight

39:56

after this episode. Yeah, this is the

39:58

time for having me. Hopefully the call

40:00

options go well and we'll... I'll be

40:02

starting somewhere together in a few months.

40:04

That's our infamous test to whether AI

40:06

is truly good or not. Is it

40:08

making us money with call options? I

40:10

love it. Thank you so much Logan.

40:12

This was awesome. Appreciate the time in.

40:14

Yeah, of course. Right

40:34

back to today's show, but first,

40:36

I want to tell you about

40:38

a podcast I love. It's called

40:40

Nudge. It's hosted by Phil Agnew,

40:43

brought to you by the Hubspot

40:45

podcast network. Ever notice that the

40:47

smallest changes can have a big

40:49

impact? On Nudge, you learn simple

40:51

evidence back tips to help you

40:54

kick bad habits, get a raise,

40:56

grow a business. Every bite-sized 20-minute

40:58

show comes packed with practical evidence

41:00

from admired entrepreneurs and behavioral and

41:02

you can apply. If you're a

41:04

marketer, you're going to love this

41:07

show because it discusses the psychology

41:09

behind stellar marketing. They just did

41:11

a great episode titled 99.9% of

41:13

ads are genuinely awful with Tom

41:15

Goodwin. You can learn why so

41:18

many ads today are ineffective and

41:20

what marketers are getting wrong. Listen

41:22

to the nudge wherever you get

41:24

your podcast.