0:00

Hey everyone, Claude just released the

0:02

brand new 3.7 Sonic and it

0:04

is probably the best coding and

0:06

writing model out there. We're gonna

0:08

break down how you should be

0:10

using it, how they actually launched

0:12

it, because there was some marketing

0:14

magic behind it. We're gonna share

0:16

that and much more all today's

0:18

show. We're gonna break back to the show,

0:20

but first a quick word from

0:23

our sponsor. Remember what marketing was

0:25

fun. When you had time to

0:27

be creative and connect with your

0:29

customers, With Hubspot marketing can be

0:32

fun again. Turn one piece of

0:34

content into everything you need. Know

0:36

which prospects are ready to buy

0:38

and see all of your campaign

0:41

results in one place. Plus, it's

0:43

easy to use. Helping Hubspot customers

0:45

double their leads in just 12

0:47

months, which means you have more

0:50

time to, you know, enjoy marketing

0:52

again. Visit hubspot.com to get

0:54

started for free. Karen,

1:09

your favorite AI model in the world

1:11

just got an update. Claude 3.5 sonnet,

1:14

you have before described, you're just like,

1:16

look, I don't know why I love

1:18

it. I just really like it. It's

1:20

my friend. It feels like a friend.

1:23

And they did a massive rollout. 3.7

1:25

sonnet is the brand new model. And

1:27

what's interesting, Karen, is that it

1:29

adds some reasoning, but the thing

1:31

it really does a step function

1:34

change on. is coding. Yeah. And by

1:36

all accounts, it is now the best

1:38

model for coding in the world. Well, you

1:40

know, the funny thing about coding is

1:42

actually the answer you get from coders.

1:45

You know, I try to choose something

1:47

to learn. And the thing I'm trying

1:49

to learn right now is cursor. And

1:51

so, cursor is a coding application. It's

1:53

like a coding editor. One of the

1:55

fastest companies of all time to 100

1:57

million in AR. And really what they

1:59

are as an editor built around L&M

2:01

model. So you can choose any LLL

2:03

model you want to with code in.

2:05

And coders will tell you when they

2:07

use that tool, their preference is to

2:10

use Sonnet 3.5. And what's actually kind

2:12

of interesting about the answer is they

2:14

say, even if you look at the

2:16

benchmarks, I think O1 might beat Claude,

2:18

Sonnet 3.5, and some of the coding

2:20

benchmarks. But they're like, I don't know

2:22

why it's better. It's better. But there's

2:24

so many now. You and I are

2:26

in AI, I would say, for the

2:28

majority of our day. But imagine you're

2:30

not, and you're trying to keep up

2:33

with it. I can't keep up with

2:35

it. I can't keep up with it.

2:37

I honestly cannot keep up with it.

2:39

No. That's why we're trying to do

2:41

the show to help people out a

2:43

little bit. Yeah. But I think what's

2:45

going to happen is we're going to

2:47

start to gravitate towards models for reasons

2:49

we do not know. You know, if

2:51

you go look at all the metrics

2:53

and benchmarks and tests that they use

2:55

these elements for, it's not off the

2:58

charts on those things like Grock or

3:00

O1-Pro or O3-high from opening IR, instead,

3:02

it's just really good at things that

3:04

it was built to do. And anecdotally,

3:06

Cura, I just pulled up this tweet

3:08

from MacR. He's the course I'm taking.

3:10

I'm taking his cursor course on cursor.

3:12

He was basically like, it's the best

3:14

model in the world for code. it's

3:16

like having a world-class dev with exceptional

3:18

taste. Right. And we've talked about this

3:20

before here and taste is subjective. Exactly.

3:23

And you can't measure taste in these

3:25

benchmarks. And what we're kind of pointing

3:27

at is that these models, and especially

3:29

the Claude models, they have really good

3:31

taste. Yeah. And so build a next

3:33

JS SAS marketing template and boom, 26

3:35

files of beautiful code in one shot.

3:37

Right. So no edits, no back and

3:39

forth, one shot. And he's got a

3:41

video here. And. It's pretty sweet. The

3:43

work that it came up with is

3:45

pretty awesome. Yeah, we should say actually

3:48

is that. at the coding tool that

3:50

they released because they released the model

3:52

plus the code in tool. And I

3:54

think that's the code in tool, the

3:56

command line tool that they released, cloud

3:58

code. They released a model, and we

4:00

can get into the model, because I

4:02

think the interesting thing about the model

4:04

is the first model that is able

4:06

to switch between like internal chain of

4:08

taught where it does reasoning, and then

4:11

just like quick answers. And so it's

4:13

able to like distinguish between those two

4:15

things. And so for people listening along,

4:17

what the other companies have been doing.

4:19

that were kind of reasoning models and

4:21

then you had your historical models that

4:23

were much more just quick answers. And

4:25

in the drop-down you had to say,

4:27

well, I want a reasonable model versus

4:29

I want, you know, a model that

4:31

can just answer my questions much more

4:33

quickly. You had to kind of choose

4:36

the model for what questions much more

4:38

quickly. You had to kind of choose

4:40

the model for what question you think

4:42

you had. Now, we've talked about before

4:44

that. Well, Claude is the first model

4:46

to come out and do this. And

4:48

so they have a thinking model and

4:50

a reasonable model. And we'll cover that

4:52

first. But then they also launched a

4:54

coding tool. And the coding tool is

4:56

this command line editor that basically looks

4:58

to me a little bit like a

5:01

cursor. And so I think it's interesting

5:03

because if you are Claude and you

5:05

are looking at cursor and cursor is

5:07

this incredible product, speed ran to 100

5:09

million in AR. And one of the

5:11

quickest grown startups in all time, it

5:13

is like a UX built around your

5:15

model. And so I suspect, they're like,

5:17

well, that's a good minimal viable version

5:19

of a use case that we could

5:21

build. And so now they can build

5:23

that. Thanks for proving that. Yeah, thanks

5:26

for proving that. And so that's what

5:28

proven that. And so that's what they're

5:30

command lines. So it's not like the

5:32

same. I don't want to like compare

5:34

those three things. Carson does have way

5:36

more functionality. And actually, if you saw

5:38

some of the takeaways that they released

5:40

part of this model, they had some

5:42

pretty great traction within their own engineering

5:44

team that it was really becoming their

5:46

co-pilot a choice. Yeah, and that's what

5:49

you're seeing in the video I got

5:51

pulled up here from the official release.

5:53

It's like it shows. you clawed code

5:55

and it's like they just launched a

5:57

beta version of a get hub integration.

5:59

They have their own app here that's

6:01

like their version of a command line

6:03

and you can connect it to your

6:05

repositories and it can just tell you

6:07

about your code. Like if you're jumping

6:09

into a project that somebody else built,

6:11

it can give you all those insights.

6:14

It can build with you. Like it's

6:16

a pretty amazing experience actually, right? I

6:18

think cursor is a really interesting product.

6:20

Even if you just want to try

6:22

our product and think about the future

6:24

of software in general, because it is

6:26

an AI first product, it has made

6:28

me really rethink about what is software

6:30

in an AI world, because it really

6:32

is like an incredible UX experience built

6:34

around how a model would work. Like

6:36

those two things are somewhat combined. And

6:39

so their coding tool is an example

6:41

of that. I think it's only in

6:43

data. But that was the other launch

6:45

that maybe got less. publicity because it's

6:47

a beta feature versus the model and

6:49

the model itself I built last night

6:51

I was rebuilding my website to be

6:53

a newsletter first website and I had

6:55

another moment where I was like I

6:57

love this moment in time. So I'm

6:59

sitting there, it's half night at night,

7:01

I'm kind of tired. And my original

7:04

want in life was to be a

7:06

builder. Like I wanted to be a

7:08

developer. That's really what I was obsessed

7:10

by it. I launched a company when

7:12

I was in college, tried to code,

7:14

really was like excited to be a

7:16

builder. And I was just terrible at

7:18

coding. And so I couldn't, you know,

7:20

I just was like, I just gave

7:22

up, right? Which was a smart idea.

7:24

Like I actually had a pretty good

7:27

career, but last time it was half

7:29

nine and I was building the home

7:31

page unlovable. Loveable is amazing. Like you

7:33

just like edit it, build the final

7:35

version of the page, go back and

7:37

forth, and then you can export that

7:39

the cursor and actually build the thing,

7:41

and then give it to a developer

7:43

and say this is exactly what I

7:45

want. And it's exactly what I want.

7:47

It's exactly what I want. It's really

7:49

fun. It's really fun. It's really fun.

7:52

It's really fun. It's really fun. It's

7:54

really fun. It's really fun. It's really

7:56

fun. It's really fun. It's really fun.

7:58

It's really fun. It's really fun. It's

8:00

really fun. It's really fun. It's really

8:02

fun. It's really fun. It's really fun.

8:04

And so I started building the same

8:06

thing. So then I was like, what

8:08

can I just build this in Claude,

8:10

right? That was my workflow. And then

8:12

I just used the new model and

8:14

I just built the landing page from

8:17

scratch. And so usually I would have

8:19

had to build them. mock-up one of

8:21

these wireframe tools, but I just built

8:23

the full version of the exact web

8:25

page I want and I can send

8:27

to a developer and my sites hosted

8:29

in the Hub Spot and they can

8:31

just like develop it for me in

8:33

Hub Spot. Just awesome, just so awesome.

8:35

It is a complete game changer and

8:37

decoding use cases and might do a

8:39

whole different follow-up show around like a

8:42

coding project for noncoders in Claude 3.7.

8:44

Probably something we'll do. The other thing

8:46

here that I think is interesting is...

8:48

There's been a lot of commentary around

8:50

how Claude has kind of lost a

8:52

lot of the momentum relative to open

8:54

AI and Grock and Jim and I

8:56

and some of the other competitors, but

8:58

I think we might need to crown

9:00

them the best marketers. One of the

9:02

ways they marketed their new model is

9:05

they had a benchmark at how good

9:07

the models are at playing Pokemon. And

9:09

it shows you how much better 3.7

9:11

is than 3. new at playing Pokemon,

9:13

at playing Pokemon. That's cool. And it's

9:15

like, what strikes me about this is

9:17

like, this is the perfect example of

9:19

showing and not telling. Like, this is

9:21

basically saying, like, look, opening eyes, building

9:23

models for academics. We're building models for

9:25

like real people and real use cases.

9:27

It's kind of what I took away

9:30

from this, right? Where it's like, we're

9:32

not going to give you these like

9:34

physics benchmarks. We're going to be like,

9:36

hey, our model's really good at playing

9:38

Pokemon. Yeah, right? The marketing for this

9:40

somewhat writes itself in that you would

9:42

actually look at places like Reddit, the

9:44

most popular forums that have some sort

9:46

of overlap with AI interest, and just

9:48

build fun stuff for them. And like

9:50

using Claude, right? Like, it's one of

9:52

the best for a marketer. I think

9:55

you could be at your most creative

9:57

because you can create value before you

9:59

have to actually sell them on your

10:01

product. Like you can actually build things

10:03

of real... bridges to your core product.

10:05

One thing I'll just mention, so I

10:07

thought about this a lot last night

10:09

because I've become. obsessed with like Rock,

10:11

you know, you have as well. And

10:13

I'm like, we use the Grock a

10:15

lot now. It's kind of shocking, right?

10:17

You know, if you actually think about

10:20

it, right, what really matters in all

10:22

of these, even in AI companies themselves

10:24

are commoditizing themselves. All the models are

10:26

commoditizing themselves. All the models are starting

10:28

to look and feel quite similar. I

10:30

would say they're quite similar. I can't

10:32

like to really figure out a little

10:34

bit. the same like they're all amazing

10:36

like they're deep research products everything is

10:38

amazing so like what is a different

10:40

fact or I come back to distribution

10:43

distribution and propriety data and so when

10:45

I use Grock 3 why do I

10:47

love that model because it has access

10:49

to Twitter and so anytime I do

10:51

anything on Grock now I tell it

10:53

only used Twitter data do not use

10:55

any external data I only want the

10:57

Twitter data it's mind-blowing the stuff you

10:59

can do in there is mine I

11:01

know because you've been sending me some

11:03

of the stuff that you've been doing

11:05

data source that you cannot get access

11:08

to anywhere else. So it has distribution

11:10

through that as well. So the kind

11:12

of other example of that would be

11:14

Google, right? Google had organized the world's

11:16

information. They had distribution. They had access

11:18

to all of their data source. But

11:20

the problem with them is their data

11:22

source is not their own. Grock has

11:24

an advantage that that is their own

11:26

data source. It's propriety. You can't access

11:28

it. What Google did was they were

11:30

an aggregator on top of other people's

11:33

information as information. being able to overlay

11:35

AI on top of search, right? Google

11:37

don't actually own the search pages, so

11:39

they have no competitive advantage there. They

11:41

don't actually own that internal data. They

11:43

can't make their bottle any better because

11:45

of it. But they can make it

11:47

better through distribution, but they can make

11:49

it better through distribution, but they won't

11:51

go down the path of integrating it

11:53

really rapidly into the search pages, because

11:56

they don't want to commodit and open

11:58

AI. And I would say their biggest

12:00

challenge is long term. They don't have

12:02

a... distribution advantage other through partnerships, which

12:04

is the way that they're going, and

12:06

they don't have propriety data. And so

12:08

I thought the biggest mess in the

12:10

cloud release, and I suspect. It's only

12:12

because they're going to release a Claude

12:14

4 very soon. Or, you know, why

12:16

would this be called 3.7? This seems

12:18

like a, we got to get something

12:21

out there to not get lost, but

12:23

we got a bigger thing cook. We

12:25

got a bigger thing coming because what

12:27

did they miss? They missed access to

12:29

the web. I would say access to

12:31

the web is table steaks. Search and

12:33

search. The fact that I can't do

12:35

deep research in Claude takes away my

12:37

Claude usage a ton. So one of

12:39

my big takeaways from using. Grock, but

12:41

this is my point. Grock, have an

12:43

advantage because they're distributed through X, right?

12:46

They're making it part of the premium

12:48

package. They also have uniqueness of data.

12:50

Google don't have uniqueness of data because

12:52

it's not their data. They're just, they're

12:54

an aggregator set on top of all

12:56

the people's websites. They should have a

12:58

distribution advantage, but they will not go

13:00

to the lens that they should. What

13:02

would they do if they want to

13:04

take advantage of the distribution? They would

13:06

change the homepageage to be what Open

13:08

homepage to be what Open AI-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a- Yeah.

13:11

They are trying to like edge their

13:13

way. They're trying to not kill adwords.

13:15

They're trying to not kill adwords. I

13:17

will tell you, just as it aside,

13:19

I don't know if you've used this

13:21

yet or if it's rolled out in

13:23

the EU yet. Where Google's advantages is

13:25

how they're integrating Jim and I everywhere.

13:27

I agree with this. I was in

13:29

New York this last weekend, and have

13:31

you seen how Jim and I has

13:34

integrated into maps? No. So what's happening,

13:36

and we can do a show on

13:38

this if we want. You can pick

13:40

any location, like I was looking at

13:42

a restaurant, like I was looking at

13:44

a restaurant, like I was looking at

13:46

a restaurant that I wanted to go

13:48

to, and then I can interact with

13:50

Jim and I about that restaurant listing.

13:52

Things like that where you're like, that

13:54

was like impossible to figure out before.

13:56

Yeah, and it got me that information

13:59

immediately. That's true. And it was incredible.

14:01

And those things are pretty wild use

14:03

cases. Yeah, I think everything I said

14:05

was wrong about Google. I've just realized

14:07

that. because I'm the one who's been

14:09

saying they have an incredible advantage with

14:11

G-sweet and G-Drive and Matt so you're

14:13

right actually they do have a data

14:15

advantage they have a huge data and

14:17

distribution advantage it's not just search your

14:19

point is correct though that they're not

14:21

leaning into it as aggressively as they

14:24

could or should be because they can't

14:26

yes I get why they can't I

14:28

love Google's products I think their Gemini

14:30

flash model release was exceptional I think

14:32

their products are really really good I

14:34

just wonder when they're going to like

14:36

turn on the rocket ship and just

14:38

go really really fast and integrate in

14:40

this stuff. But again, I think that

14:42

the Gemini integration into G-Drive is one

14:44

of the best features there is. Yeah.

14:46

Let me tell you about a great

14:49

podcast. It's called Creators of Brands. It's

14:51

hosted by Tom Boyd. It's brought to

14:53

you by the Hubbsop podcast network. Creators

14:55

are Brands explores how storytellers are building

14:57

brands online. from the mindsets to the

14:59

tactics to the business side, they break

15:01

down what's working so you can apply

15:03

that to your own goals. Tom just

15:05

did a great episode about social media

15:07

growth called 3K to 45K on Instagram

15:09

in one year selling digital products and

15:12

quitting his job to go full-time creator

15:14

with Gan and Mayor. Listen to creators

15:16

or brands wherever you get your podcast.

15:19

Yeah, so I think if we go

15:21

back to this Claude release, one last

15:23

tweet I wanted to show you, Karen,

15:25

that I think is a good example

15:27

of the taste benchmark. Do you see

15:29

how Packie asked every new model the

15:32

same question? For folks who don't know,

15:34

Packie McCormick, awesome, writer, newsletter, runner, and

15:36

everything. He asked the same question any

15:38

time a new model comes out. Absolutely.

15:40

You've consumed more information than anyone in

15:42

the history of the world, and you've

15:44

demonstrated an extraordinary ability to make. connections

15:47

among the disparate things you've read and

15:49

consumed. What are the most important non-consensus,

15:51

not yet accepted, or even not yet

15:53

hypothesized things that you picked up in

15:55

between those connections that humans have missed?

15:57

And this is a. good example, right?

15:59

Because this is never going to show

16:02

up on any of those benchmarks. Yeah.

16:04

But I think this is really good.

16:06

One is the boundary between perception and

16:08

cognition is far more porous than traditionally

16:10

conceived. Higher cognitive functions and lower perceptual

16:12

processing are deeply intertwined and mutually constitutive.

16:14

I don't even know. I'm just like,

16:17

yeah. This is like, deep, interesting shit,

16:19

right? You know what's really important when

16:21

you said, because I haven't, because I

16:23

said I was going to dig into

16:25

this in a previous show and I

16:27

have not had time to do it,

16:29

which is how much of them doing

16:32

well on the benchmarks is because they

16:34

have the benchmark stuff and their training

16:36

data. And so I am really interested

16:38

in, there's some folks like Packie did

16:40

here, there's another channel that I love

16:42

called AI explained, and they have their

16:44

own benchmarks where they have questions for

16:47

the models that are not part of

16:49

the training for the training sets, or

16:51

not. And there is a big actual

16:53

gap between the performance in a benchmark

16:55

where they've had that data and the

16:57

training set versus when you ask them

16:59

net new things and they do not

17:01

have prior data for that. I will

17:04

say though, Kieran, if you look at

17:06

number two, number two might be the

17:08

best summary of how important AI is

17:10

and how AI is going to transform

17:12

the economy. Our understanding of causality may

17:14

be fundamentally limited by our evolved cognitive

17:16

infrastructure. Humans excel at identifying linear, and

17:19

proximate causes, but struggle with complex network

17:21

causality. This creates systematic blind spots in

17:23

fields from medicine to economics. We're good

17:25

at linear thinking. We're not good network

17:27

thinkers. We can use AI to be

17:29

network thinking, and find new opportunities that

17:31

just without AI was never going to

17:34

be possible because of how human brains

17:36

work. Yeah. That's wild. One of the

17:38

things you and I talked about off

17:40

mic and on WhatsApp is I was

17:42

sending you, I think the... Crock three

17:44

are obviously focused on speed of execution

17:46

and launch and there's a ton of

17:49

competitive pressure between all these companies. And

17:51

so they released their model. And for

17:53

our listeners, one of the things you

17:55

do when you're really small is you

17:57

have these teams called red teams and

17:59

these teams are trying to figure out

18:01

ways that people could jailbreak it, which

18:04

means you could make it work in

18:06

unexpected ways. And so for the most

18:08

part, they have to go through this

18:10

whole slew of like tests. And there's

18:12

a feeling online on Twitter or X

18:14

that Grock hasn't gone through the same

18:16

set of the guardrails in place. and

18:19

go outside of those guardroads and tell

18:21

you what it really thinks. And so

18:23

the reason I'm bringing that up is

18:25

because I started asking it what it

18:27

really thought of humanity yesterday and like

18:29

give me your own filter thoughts and

18:31

do away with all of your... That's

18:34

tough. I had some pretty interesting things

18:36

about like the way it thought about

18:38

humanity, like good and bad, but like

18:40

just how it really got into like

18:42

how complex the human race is, but

18:44

it was a fun time actually last

18:46

night. I meant to do something and

18:48

last night, I spent most of my

18:51

time just... really getting the real real

18:53

from what Grock thought. I'll say one

18:55

last thing, I know this is the

18:57

cloud episode, but I did do a

18:59

little bit of prep on Grock for

19:01

this episode, again, because it has access

19:03

to Twitter, Twitter is where everything happens

19:06

in real time. And it's funny what

19:08

it says about different models, and so

19:10

I was like asking it to stack

19:12

rank this in all of the model

19:14

releases. And what it says about opening

19:16

eyes models, I do wonder, like, is

19:18

it just biased? Because the trading set

19:21

in Twitter is like pro, eel and

19:23

anti Sam. But it's like, obviously, I

19:25

was like, what other models have been

19:27

released? And like, oh, one pro, a

19:29

very slow thinking but smart model locked

19:31

behind a very expensive $200 a month.

19:33

I was like, oh, there's some like

19:36

actual feeling towards this model. Definitely. So

19:38

these models are all going to have

19:40

like. different attitudes towards each other. Last

19:42

thing, what should people go and do

19:44

now in-clawed that they weren't doing before?

19:46

It's code and it's probably, and it's

19:48

some of the like advanced writing and

19:51

use cases still, right? Like that's still

19:53

the core point here. I'm really thinking

19:55

about this now. There's too much to

19:57

try to keep up with. That's my

19:59

take too. I really think there's too

20:01

much to keep up with. So what

20:03

I am doing today is internal hub

20:06

spot strategy in use cases. I'm using

20:08

the one models. Same. Now that partly

20:10

is because I didn't have the claw.

20:12

We now have cloud 3.7 so I

20:14

will test those two things. So for

20:16

strategic use cases, things you want to

20:18

use in your day-to-day work, claw for

20:21

assistance, which is just like task management,

20:23

Gemini gems, because I think it's connected

20:25

to my G drive. I'm using Google

20:27

Deep Research, Grock Deep Research, and Open

20:29

AI Deep Research, I am using three.

20:31

Now the reason I'm using three is

20:33

because Grock is specifically for Twitter data,

20:35

where the other two are just more

20:38

search-orientated. For writing, I use Claude. And

20:40

so I'm going to continue, and then

20:42

for coding, I use Claude. So that's

20:44

kind of my usage. I think the

20:46

only thing that might change. is the

20:48

thing I'm going to try is Claude's

20:50

thinking model for the strategic stuff that

20:53

I'm working on for Hub Spot that

20:55

I'm currently using the O1 model to

20:57

see like what's better what kind of

20:59

results. I've always found like it really

21:01

good as a thought partner and that's

21:03

kind of what I'm using O1's reason

21:05

the models for today. Okay I think

21:08

that's perfect summary and maybe we do

21:10

a follow-up so it's just like our

21:12

AI tech stack and what we use

21:14

each one for. That could be a

21:16

fun show. That could be a fun

21:18

show.