0:00

This is the everyday

0:02

AI show the everyday

0:04

podcast where we simplify

0:06

AI and bring its

0:08

power to your fingertips

0:10

Listen daily for practical

0:12

advice to boost your

0:14

career business and everyday

0:16

life There's a new

0:18

most powerful AI model

0:20

in the world

0:23

Yeah, sometimes I

0:25

feel like DJ Khaled

0:27

because each week it's

0:29

like Another one. Another

0:31

one. Another most powerful AI

0:33

model in the world.

0:36

Y 'all, the last couple

0:38

of weeks, couple of months,

0:40

it has been a

0:42

back and forth, I think

0:44

specifically, between OpenAI and

0:46

Google for the ever -changing

0:48

title of most powerful AI

0:50

model in the world.

0:52

And I think now, with

0:54

OpenAI's new O3 specifically, it

0:56

is the most powerful. AI

0:59

model in the world. Is it

1:01

the most flexible? Will it

1:03

be the most used model?

1:05

I don't know, but we're

1:07

going to be going over

1:09

that and a lot more

1:11

today on everyday AI. As

1:13

we talk about the new

1:15

OpenAI's 03 and 04 mini

1:17

models unlocked inside the world's

1:19

newest, most powerful AI models.

1:22

All right, what's going on, y 'all? My name

1:24

is Jordan Wilson, and I'm the host of

1:26

Everyday AI, and this thing, it's for you.

1:28

It is your daily live stream podcast and

1:39

You are in the right place. So

1:42

you need to go to your everyday

1:44

ai.com. And there on our website, you

1:46

can not just sign out for our

1:48

free daily newsletter where we will be

1:50

recapping the most important aspects of this

1:52

show and sharing a lot more.

1:55

But we are going to share with

1:57

you everything else that's happening in the

1:59

business world, in AI world. So you can

2:01

be the smartest person in AI

2:03

at your company or in your department.

2:05

All right. So make sure if

2:07

you haven't already to go to your

2:09

everydayai.com to do that. So

2:11

I am very excited today to

2:13

talk about the new o3 and

2:16

o4 models from open AI. But

2:18

before we do, uh, let's start

2:20

as we do most days by

2:22

going over the AI news, uh,

2:24

and hey, live stream crew is

2:26

technically a two part show. So

2:28

I need your help. Uh, let

2:31

me know as I go over

2:33

the AI news, what o3 use

2:35

cases should we cover in tomorrow's

2:37

show in part two? All right.

2:39

Here's what's happening in the world

2:41

of AI news. A couple of

2:43

big things. So Chinese tech giant

2:45

Huawei is preparing to begin mass

2:48

shipments of its new 910C AI

2:50

chip in May, aiming to fill

2:52

the gap left by US restrictions

2:54

on Nvidia's 20 chips

2:56

according to Reuters. So

2:59

the new chip from

3:01

Huawei, the 910C, achieves

3:03

performance comparable to Nvidia's

3:06

H100 by combining two

3:08

existing 910B processors, representing

3:10

a key shift for

3:12

Chinese AI developers who

3:14

need domestic alternatives. So

3:17

Washington's latest AI export

3:19

controls have pushed Chinese AI

3:21

companies to seek more

3:24

homegrown solutions, making Huawei's 910C

3:26

likely to become the

3:28

main AI chip for China's

3:30

tech sector. So yeah,

3:33

it looks like Nvidia could

3:35

potentially have a strong

3:37

new competitor in Huawei. All

3:40

right, next, a small thing.

3:42

But I think that could

3:44

have a big impact. So open

3:46

AI has quietly introduced memory

3:48

with search. much different than their

3:50

memory feature they rolled out

3:52

about two weeks ago. So this

3:54

allows chat GBT to use

3:56

personal details from prior chats specifically

3:58

to tailor web search queries.

4:01

All right. So yes, uh, open

4:03

AI rolled out their expanded

4:05

memory feature a couple of weeks

4:07

ago that allows chat GBT

4:09

to use personal details, but that

4:11

did not apply to web.

4:13

queries. Uh, so this new update

4:15

means chat GBT can now

4:17

rewrite user prompts to reflect individual

4:19

preferences while browsing the web,

4:21

such as, you know, whatever you

4:23

share with it, dietary restrictions,

4:25

location, uh, et cetera, to

4:27

gain, to bring you more

4:29

accurate search results. So this move

4:31

follows recent upgrades that let

4:34

chat GBT reference users entire chat

4:36

history, further distinguishing it from

4:38

competitors that don't have this feature

4:40

enabled. Uh, users

4:42

can turn off this feature in

4:44

settings, but the rollout appears

4:47

to be very limited so far

4:49

with only a few accounts

4:51

reporting early access. So yeah, make

4:53

sure to keep an eye

4:55

out for that. All right. One

4:57

last thing to keep an

4:59

eye out on is while bringing

5:01

AI into the classroom. Uh,

5:03

so the Trump administration is weighing

5:06

in executive order that would

5:08

require federal agencies to promote artificial

5:10

intelligence training in K through

5:12

12. And this is according to

5:14

a draft obtained by the

5:16

Washington Post. This is technically super

5:18

breaking news, only a couple

5:20

minutes old. So the draft policy

5:22

directs agencies to train students

5:25

in using AI and integrate the

5:27

technology into teaching tasks, signaling

5:29

a potential national shift in how

5:31

schools approach technology education. So

5:33

agencies would also partner with private

5:35

companies to develop and implement

5:37

AI related programs for students aiming

5:39

to better. prepare them for

5:42

careers shaped by AI. So

5:44

the proposal is in draft

5:46

form right now is still under

5:48

review and could change or

5:50

be abandoned. However, if it is

5:52

enacted, it could significantly shape

5:55

how the next generation learns and

5:57

works with artificial intelligence. I

5:59

would love to see this happen

6:01

personally this little eight little tidbit

6:03

y 'all I haven't shared shared

6:05

this much, but I just saw

6:07

you know Jackie here in our

6:09

comments holding it down I'm teaching

6:11

a course at to Paul here

6:13

in Chicago and like I'm flipping

6:15

the script on its head I'm

6:17

saying you have to use AI

6:19

at every single junction like don't

6:21

Go old school. Don't write in

6:23

all of these aspects. You should

6:26

be using AI in every single

6:28

aspect. So it should be pretty

6:30

interesting to see how this new

6:32

executive order unfolds and if it

6:34

actually is introduced. All

6:36

right a lot more on

6:38

those stories and a ton

6:40

more on our website your

6:42

everyday AI calm Alright, let's

6:44

get into it Let's talk

6:46

about the newest and I

6:48

think the most powerful AI

6:50

models in the world All

6:52

right from open AI, but

6:54

again, I don't necessarily think

6:56

that means if it's just

6:59

because it's the most powerful

7:01

I don't think it's necessarily

7:03

the best or the most

7:05

flexible. Right. Those are three

7:07

very different things. I do

7:09

think by far the new open

7:11

AI 03, which is the

7:14

full version. And then we have

7:16

the 04 mini and 04

7:18

mini high. Yeah, the naming is

7:20

terrible. Open AI has said

7:22

that they're going to address this

7:24

naming problem because it's extremely

7:27

problematic. Right. But the

7:29

new 03 and 04 models

7:31

are extremely impressive,

7:33

specifically the O3. All right.

7:35

And if you're confused, like, oh,

7:37

Jordan, why is the O3

7:39

better than the O4? Well, that's

7:42

because the O4 is a

7:44

mini. So we have O4 mini

7:46

and O4 mini high. But

7:48

now we have the O3 full

7:50

model, right? Whereas previously we

7:52

had O3 mini and O3 mini

7:55

high. Confusing. But this is

7:57

the first kind of full O

7:59

model that we've had since

8:01

O1. Yes, I know it's confusing

8:03

that to skip O2 because

8:05

of some naming rights with, uh,

8:07

I believe a British telecom

8:10

very confusing with the model names,

8:12

but here is what is

8:14

not confusing. This new model is

8:16

extremely impressive. All right. So,

8:18

uh, live stream audience.

8:21

Good morning, good morning, like

8:23

what Will said here

8:25

on LinkedIn. Love, love

8:27

to see it. Everyone, let me

8:29

know what questions you have

8:31

about this new O3 and O4

8:33

models. You know,

8:35

I'll either tackle them today,

8:37

later on our live stream

8:39

here, or I will... You

8:41

know make sure that we

8:44

do this tomorrow in part

8:46

two. So it's good to

8:48

see everyone on on linkedin

8:50

and on youtube Thanks. Thanks

8:52

for tuning in everyone love

8:54

to see us learning together

8:56

live. All right Let's get

8:58

into it shall we so

9:00

here's the overview on the

9:02

new o3 and o4 models.

9:04

So these were just released

9:06

about a week ago, and

9:08

this is the kind of

9:10

the newest successors in OpenAI's

9:12

O series. So yeah,

9:14

I just laid out a bunch

9:16

of O's, which, which by

9:18

the way, has anyone had O's

9:20

the cereal? I was

9:22

talking about this with my wife. They

9:25

are so underrated, like maybe my

9:27

favorite. top five favorite cereal. That's beside

9:29

the point. But so many different

9:31

O's, right? You have O1 and still,

9:33

right? So they got rid of

9:35

O3 mini high. But, you know, if

9:37

you're on a pro plan right

9:39

now, as an example, I believe you

9:41

have O1, you have O1 pro.

9:44

You have O3 full and then you

9:46

have O4 mini, O4 mini high.

9:48

It's five different O -series models across

9:50

three different classes. Extremely confusing, right?

9:52

And obviously, you know, OpenAI is

9:54

in the future moving away from this

9:56

and treating GPT. five as a

9:58

system. But essentially, if you're wanting what's

10:00

all these O models, these are

10:02

the thinking models. These are the models

10:04

that can reason and plan ahead

10:07

step by step under the hood before

10:09

they give you a response. Whereas

10:11

the GPT models, so

10:13

as an example, GPT

10:15

four or GPT 4

10:17

.5. They are more instantaneous,

10:19

right? They're not necessarily thinking like

10:21

a human would step by step using

10:23

this chain of thought reasoning under

10:25

the hood before it gives you a

10:28

response. So I like to say

10:30

there's two very different classes of models

10:32

from open AI. You have your

10:34

quote unquote old school transformers, and then

10:36

you have your quote unquote new

10:38

school O -Series model, which are your

10:40

thinkers and your reasoners. All right. So

10:42

this was just released. less

10:44

than a week ago. And here's

10:47

the biggest part. It is capable of

10:49

using all of open AI's tools,

10:51

which is the biggest differentiator between the

10:53

01 and the 03 models that

10:55

could not use every single tool. Because

10:57

when we talk about agentic AI

10:59

and yeah, that's what I think 03

11:02

is. It is an agentic. Model

11:04

at its core and we're gonna see

11:06

that I think tomorrow when we

11:08

go through some of these use cases

11:10

alive But the biggest difference or

11:12

one of the biggest differentiators here is

11:15

oh three can use all tools

11:17

web search Python

11:19

file uploads, computer vision with

11:21

the visual input reasoning, and

11:23

also image generation. It can

11:25

literally do everything, whereas the

11:27

previous O -Series models were

11:29

a little limited, right? And

11:31

some of them were different. You

11:34

know, even now you can use

11:36

Canvas, which is more of this interactive

11:38

mode that can run and render

11:40

code inside the new O3 model. Whereas

11:42

before, it's like, okay, the 01

11:44

model, is the only one that could

11:46

use Canvas. But O1 wasn't very

11:48

good at many things because O1 Pro

11:50

and O3 Mini were better, or

11:53

sorry, O3 Mini High. And then O3

11:55

Mini High could use the internet, but

11:57

you couldn't upload files and it couldn't

11:59

use Canvas, right? And then you had

12:01

O1 Pro that you could upload files,

12:03

but you couldn't use Canvas and it

12:05

couldn't browse the web, right? So it

12:07

was kind of hard with all these

12:09

different O models. And, you know, they

12:11

all kind of had their own kind

12:14

of unique features. But

12:16

now, O3, I do think

12:18

this is an agentic model,

12:21

right? And I know that

12:23

sounds crazy to say, but

12:25

it is extremely powerful and it

12:27

can use every single tool

12:29

under its kind of tool belt.

12:31

And it's trained to autonomously

12:33

decide when and how to use

12:35

these tools. That is what

12:37

I think makes it probably the

12:39

most powerful AI model in

12:41

the world. And it responds with

12:43

rich answers, typically in under

12:45

a minute. And it

12:48

is right now, if you have

12:50

a paid plan to chat, tbt,

12:52

you have access to it. So

12:54

whether that's chat, tbt plus pro

12:56

teams, et cetera, you have access.

12:58

It's also available in the API. There

13:01

are limits though. All right,

13:03

so if you are on either

13:05

a chat GPT plus account,

13:07

that's your standard paid account at

13:09

$20 a month, or if

13:11

you're on a team account or

13:13

enterprise account, it's pretty limited.

13:15

So you only have 50 messages

13:17

a week with the best

13:19

one, which again is 03. Not

13:21

04, right? So 04 mini

13:23

is not the best one. 03 is, right?

13:25

I'm just going to say 03 full. That's

13:27

what a lot of people, including myself, are

13:29

calling it since we previously had the 03

13:31

mini. And then we're having to deal with

13:34

the 04 mini and people are confused. So

13:36

03 full. is the best model. But right

13:38

now, if you're on a paid plan, you

13:40

only have about seven messages a day, or

13:42

about 50 messages a week. So not a

13:44

ton. With 04 mini, you

13:46

have 150 messages a day. In

13:48

04 mini high, you have 50

13:50

messages a day. So if you

13:52

are a power user on a

13:54

paid plan, you might want to

13:56

start with 04 mini high. You

13:59

have 15 messages a day and

14:01

then maybe save those seven messages

14:03

a day for the time that

14:05

you really need a little more

14:07

juice, a little bit

14:09

more compute, more smarts, then you

14:11

can hand those over to O3Full. If

14:14

you are on the pro

14:16

plan, which is $200 a month,

14:18

you have quote -unquote near unlimited

14:20

access. OpenAI

14:22

says, yeah, there's some fair

14:24

use things that you have to

14:26

adhere to, but for the

14:28

most part, It is unlimited. Uh,

14:31

so I have free plans.

14:33

Uh, a month plans.

14:35

I have multiple team plans. I have

14:37

multiple enterprise accounts, uh, for companies that

14:39

hire us, uh, to train their employees.

14:41

So yeah, if you're trying to do

14:43

that, you can reach out to us.

14:45

We can train your team. Uh, right.

14:47

So it is kind of weird. I'd

14:50

say, uh, that the. Teams account in

14:52

the enterprise accounts have the same model

14:54

as the plus account you would think

14:56

or hope it would have 2x 3x

14:58

Especially the enterprise y 'all open AI you

15:00

got to get together. I'm hearing a

15:02

lot of grumblings From companies that have

15:04

invested heavily into enterprise accounts and they

15:06

can't you know, they can't get kind

15:09

of the same power that you can

15:11

get with an individual account. I know

15:13

it, it comes with a pricing, uh,

15:15

right? Uh, paying, uh, I think anywhere

15:17

between $30 to $50 for an enterprise

15:19

seat versus $200 for a pro seat.

15:21

But so many of these companies are

15:23

investing in hundreds or thousands of seats

15:26

for their enterprise teams. Open

15:28

AI, you gotta give them more juice. Just saying.

15:30

All right. So what

15:32

the heck is new? Let's

15:34

go over it. So. advanced

15:36

tool use. So like I

15:38

talked about, it has autonomous

15:40

access to browsing, coding, and

15:42

visual tools. The image

15:44

understanding, it is improved. The

15:47

visual capabilities are much improved. And

15:50

O3 does a great job

15:52

at interpreting complex visual inputs,

15:54

like as an example, research

15:56

papers. It has a

15:58

much larger context window in

16:00

the chat GPT interface. Finally, uh,

16:03

right. So finally within the

16:05

chat, GPT interface, we have a

16:07

200 K token context window. Okay.

16:10

So you, it can handle longer

16:12

multi step tasks seamlessly. And you

16:14

can share a ton of information

16:16

without it for getting things. Whereas

16:19

previously, you know, unless you were

16:21

on an enterprise plan, we still,

16:23

for the most part had a

16:25

32 ,000 token context window on

16:27

the. chat side of chat GBT,

16:29

right? It was different on the

16:31

API side, but a lot of

16:34

users inside of chat GBT, if

16:36

they were especially copying and pasting

16:38

a lot of information, chat GBT

16:40

was forgetting things, right? Because that

16:42

32 ,000, uh, context window, it's about

16:44

27, 28 ,000 words of input

16:46

and output, which isn't a ton.

16:49

Uh, so it's, it's a welcome

16:51

site to see a 200 K

16:53

token context, uh, in the new

16:55

models improved reasoning, uh, another thing

16:57

that it's new. and the ability

16:59

to chain multiple tool calls together

17:02

for layered analysis. And I think

17:04

that is probably the standout feature.

17:06

And there's some new safety features

17:08

as well, right? OpenAI doesn't want

17:10

to accidentally start. a biochemical war, which

17:13

might, you might be like kind

17:15

of chuckling and rolling your eyes, but

17:17

no, seriously. So

17:19

good on OpenAI for addressing these

17:21

things, you know, when they

17:23

release new models and they give

17:25

it essentially levels or warning

17:27

levels. So they address that on

17:29

their website as well. And

17:32

there's new features that can reduce

17:34

the risk and enhance trust. All

17:37

right. Are

17:43

you still running in circles trying to figure

17:45

out how to actually grow your business with

17:47

AI? Maybe your company has

17:49

been tinkering with large language models for

17:51

a year or more, but can't really

17:53

get traction to find ROI on Gen

17:56

AI. Hey, this is Jordan Wilson, host

17:58

of this very podcast. Companies

18:00

like Adobe, Microsoft and Nvidia have partnered

18:02

with us because they trust our expertise

18:04

in educating the masses around generative AI

18:06

to get ahead. And some of the

18:08

most innovative companies in the country hire

18:11

us to help with their AI strategy

18:13

and to train hundreds of their employees

18:15

on how to use gen AI. So

18:17

whether you're looking for chat, GPT training

18:19

for thousands or just need help building

18:21

your front end AI strategy, you can

18:23

partner with us too. just like some

18:26

of the biggest companies in the world

18:28

do. Go to your everydayai.com slash partner

18:30

to get in contact with our team.

18:32

Or you can just click on the

18:34

partner section of our website will help

18:36

you stop running in those circles and

18:39

help get your team ahead and build

18:41

a straight path to ROI on gen

18:43

AI. And

18:48

if you're a little confused and

18:50

you're like, wait, this is the new

18:52

feature. I thought it was a

18:54

different feature. Yeah. Let me

18:56

quickly get you up to speed. If

18:58

you've been sleeping under an AI rock

19:00

for three weeks, here's what else is

19:03

new at open AI and chat GPT,

19:05

because you might be confused. And I

19:07

want to really tell you, no, no,

19:09

no, no, this is separate, right? Uh,

19:11

so yeah, we've been hearing a lot

19:13

of buzz the last couple of weeks

19:15

about this new GPT four, oh image

19:17

gen. Okay. That is different. This

19:19

is, you know, Oh three different beasts

19:22

all together, but it can use, uh,

19:24

the image gen. Uh,

19:26

then in April, we had the

19:28

memory rollout across all chats. So

19:30

essentially, if you have this enabled

19:32

chat, GBT can pull in conversation

19:34

or can pull in information from

19:36

past chats, which is different than

19:38

memories, which were essentially individual nuggets

19:40

that were stored in kind of

19:43

a memory bank. But now chat,

19:45

GBT, it does this via kind

19:47

of a search poll, uh, in

19:49

a semantic, uh, you know, keyword

19:51

matching, um, and then to live,

19:53

you kind of personalized results. I

19:55

personally. hate this, right? Because it's

19:57

always trying to personalize things based

19:59

on my past chats, but that's

20:01

new. All right. And then we

20:03

also had the Google drive connector

20:06

roll out for chat GBT teams

20:08

accounts, uh, about three weeks ago.

20:10

And then also last week, uh,

20:12

we got, was that last week?

20:14

Yeah. My weeks are starting to

20:16

blur together, y 'all. Uh, so

20:18

yeah, it was last Monday that

20:20

open AI released another set of

20:22

new models. So. Don't get confused. These

20:25

other models were

20:27

GPT -41, GPT -41

20:29

Mini, and GPT -41 Nano.

20:32

However, those are not available

20:34

inside chat GPT. Those

20:36

are only available on the

20:38

developer side. All

20:41

right. So I think those kind of the

20:43

highlights of those Context

20:46

window to a million tokens,

20:48

huge. Actually, the

20:50

GPT -41 mini was stealing a

20:52

lot of the headlines, rightfully

20:54

so, because it was really

20:56

outpunching its mini moniker. But

20:59

the 401 models, I

21:01

think, were much better in

21:03

coding and just a

21:05

pretty big improvement both on

21:07

cost and performance when

21:09

it came to the model

21:11

that it was following

21:13

in GPT -40. Alright,

21:15

so these new O -Series models are

21:17

not that, right? But I do think

21:19

it was worth pointing out. Yeah,

21:21

there's been a lot of new things

21:23

happening, uh, inside chat gpt that

21:25

are not these O series models. So

21:28

I figured I'd take two minutes

21:30

here, uh, to get you caught up.

21:32

Yeah. Uh, like what Jackie's saying,

21:34

uh, need a cheat sheet, a cheat

21:36

sheet. Yeah. Maybe I should create

21:38

one. Uh, Kevin is saying, uh, uh,

21:40

Kevin from, uh, YouTube is saying

21:42

it's annoying in the paid education version.

21:44

I still can't access it. So

21:46

I'm guessing. Kevin, you're talking about O3.

21:49

Uh, yeah, it should, it should be rolling

21:51

out. You know, I know this sounds weird.

21:53

It's kind of like, Oh, you know, restart

21:55

your computer, you know, take out the SNES

21:58

cartridge and blow on it. Right. So many

22:00

times it is like a cookie issue, uh,

22:02

or a cashing issue. So if you, you

22:04

know, log out, if you're a chat, GBT

22:06

account, maybe clear your cash and log back

22:08

in, it might be there. That's actually the

22:10

way I always do it. Uh, whenever there's

22:12

new models announced, I do that like two

22:14

or three times a day, uh, to try

22:16

and get access a little earlier, even though

22:18

open AI does kind of. control those rollouts. All

22:20

right. Let me answer the question. Is

22:23

this the best model in

22:25

the world? So

22:28

yes and no, I think it

22:30

is the most powerful AI model

22:32

in the world. I think best

22:34

depends on your use case. Uh,

22:36

is it the most flexible right

22:39

now? No. So let me say

22:41

that again. I, yes,

22:43

I 100 % believe it

22:45

is the most powerful AI

22:47

model in the world. It

22:49

is not the most flexible.

22:51

And if it's the best

22:53

depends on your use case.

22:55

So obviously right now it's

22:57

kind of jabbing back and

22:59

forth with the Gemini 2 .5

23:01

Pro from Google. And we'll

23:03

see as, you know, more

23:05

user feedback starts to roll

23:08

out. But when it comes

23:10

to just pure upside, just

23:12

the ceiling, strictly power. I

23:15

think 03 is unmatched right now. Does

23:19

it like, does that mean that

23:21

I'm only like, right? Does that mean

23:23

me personally? I'm only going to

23:25

be using 03. Absolutely not. Right. I'm

23:27

still going to be using Gemini

23:29

2 .5 pro all the time. The

23:31

big difference is y 'all, and we're

23:33

going to talk about this a little

23:35

bit with benchmarks. Uh, Gemini 2 .5

23:37

pro is a hybrid model, which

23:39

makes it much more. flexible because in

23:41

certain instances, especially if you're having

23:43

iterative conversations back and forth conversations, uh,

23:45

with a model, which is what

23:47

you should be doing. Sometimes if you're

23:49

using these O -series models, you can

23:51

ask a very simple query or

23:53

a very simple follow -up query and

23:55

it might think for like minutes, right?

23:58

So in terms of flexibility

24:00

and usability. might

24:02

not always be the best for some

24:04

of those conversations that are a

24:06

little more nuanced and don't just require,

24:09

you know, big AI brains. But

24:11

if you need big AI

24:13

brains in an agentic type

24:15

of large language model interface,

24:18

oh, three is it. And it

24:20

is so, so impressive. Right.

24:24

But let's look at some of

24:26

the benchmarks. And here's here's one

24:28

thing that I kind of wanted

24:30

to call out. Right. So on

24:32

this show, we talk a lot

24:34

about the LM arena, right? And

24:36

this thing called an ELO score. And

24:39

what that means is you put

24:41

in a prompt, OK, and then

24:43

you get two blind outputs. And

24:45

you decide which one is better,

24:48

output A or output B. All

24:50

right. And that essentially over time, when

24:52

there's enough votes, a new model that

24:54

gets released gets an ELO score. Essentially,

24:56

you know, it comes from ELO scores

24:58

and chess, and it's like, hey, head

25:00

to head, this is what humans prefer

25:03

the most. So right now, the top

25:05

on that list is Gemini 2 .5

25:07

Pro. And here's why

25:09

I'm bringing this up as a

25:11

caveat. Right now, O3, full.

25:13

does not yet have enough votes

25:15

to be on the, uh, chatbot

25:18

arena leaderboard that could change in

25:20

a couple of hours or in a

25:22

couple of days. It could be

25:24

up there pretty soon. However, I do

25:26

not expect the O three full

25:28

model to do very well when it

25:30

comes to head to head human

25:33

comparisons. And here's the reason why when

25:35

you look at O three mini

25:37

high, right, which was my workhorse model,

25:39

right before Gemini 2 .5 pro came

25:41

out. I'd say Oh, Oh

25:43

three many high. That was

25:45

getting about 60 % of my

25:47

usage. Humans

25:49

head to head for the most

25:52

part don't prefer it, right? Um,

25:55

and one of the reasons

25:58

why think you have these

26:00

traditional large language models. that

26:02

focus on kind of quick

26:04

snappy responses. You have

26:06

these thinking models, which just take

26:08

longer and really only showcase

26:10

their abilities when it comes to

26:12

when you're asking it for

26:15

a very tough question, right?

26:17

And then you have your

26:19

hybrid models. So I think, ultimately,

26:21

the hybrid models are going to be

26:23

the ones that on a head -to -head

26:26

ELO score, those are going to be

26:28

the ones that do best. I don't

26:30

think these thinking models uh, strictly thinking

26:32

models are ever going to do that

26:34

great in human comparison. The way I

26:36

think about it is like, okay, think

26:38

of someone, you know, that's, you know,

26:40

super personable and has a ton of

26:43

business savvy and is super smart, right?

26:45

That's like Gemini 2 .5 pro. Then

26:47

you think of something like

26:49

Einstein, right? And a lot

26:51

of people, what they're putting

26:53

queries, uh, you know, into

26:56

LL Marina, you know, it's,

26:58

it's. Kind of quippy things,

27:00

fun things, right? Like, you

27:02

know, write me a haiku

27:04

about explaining large language models

27:06

using basketball terms, right? Not

27:10

something that an Einstein level

27:12

model wouldn't necessarily excel at. So

27:14

I'm just putting this out

27:16

there. Once the O3 model full

27:18

hits the chatbot arena, I

27:20

don't necessarily foresee it, you know,

27:22

being a top, you know,

27:24

a top three model. I do

27:26

think probably Gemini 2 .5 Pro,

27:28

because it is a hybrid

27:30

model, will still retain its lead

27:32

on that specific benchmark. However,

27:37

however, look

27:39

at some of the other

27:42

comprehensive sets of benchmarks that

27:44

have already gone through with

27:46

the new O3 full, or

27:48

as some people are calling

27:50

it, O3 high. And

27:53

it's the best. So as

27:55

an example, if you look at

27:57

live bench, okay. So live

27:59

bench is a benchmark for large

28:01

language models designed with test

28:03

sets contamination and objective evaluation in

28:05

mind. So I'm reading off

28:07

their website here. Uh, it has

28:09

the following properties. Live bench

28:11

limits potential contamination by releasing new

28:13

questions regularly. So then that

28:15

way it won't get into, uh,

28:17

you know, models, uh, testing

28:19

sets. Each question has

28:21

verifiable objective ground truth answers,

28:23

right? So it eliminates kind of

28:25

the need for a large

28:27

language model judge. So it's factor

28:29

fiction, no gray area. And

28:31

then live bench currently has a

28:33

set of 18 diverse tasks

28:35

across six categories, right? So

28:38

language, data analysis, math, coding,

28:40

reasoning, et cetera. And then you

28:42

have a global average. So

28:44

on live bench, which I think

28:46

is a good third party

28:49

benchmarking system, O3 is better than

28:51

Gemini 2 .5 with a global

28:53

average of 81 .5 and Gemini

28:55

2 .5 is the next best

28:57

model aside from OpenAI's O

28:59

models, which actually take up the

29:01

first three spots. So Gemini

29:04

2 .5 comes in at a

29:06

77 .4. So O3

29:08

high much better at

29:10

81 .5. Similarly, another one

29:12

that we talk about a

29:14

lot is the artificial

29:16

analysis index. So again

29:18

a very reputable and I'd

29:20

say probably one of the

29:22

most trustworthy third -party benchmarking

29:24

services out there So they

29:26

haven't done o3 full yet.

29:28

I believe because not all

29:30

of the capabilities are available

29:32

in the API Whereas on

29:34

o4 mini high they are

29:36

okay, so on o4 mini

29:38

high, which is a mini

29:40

model on the intelligence index

29:42

It is the best model

29:44

or the most powerful model

29:46

in the world. All right. So

29:49

right now it

29:51

is ahead of Gemini

29:53

2 .5 Pro by

29:55

two points. All

29:58

right. And this I think

30:00

is pretty important because again,

30:02

you are comparing a mini

30:04

model. So I assume once

30:06

the full model is put

30:08

through some of these tests,

30:10

it will be even further

30:12

ahead. But the 04 mini

30:14

high is two points ahead. Of

30:17

Jeff and I 2 .5 pro.

30:19

So when it comes to

30:21

unbiased third party benchmarks that look

30:23

at a lot, uh, it

30:25

has been decided. Oh, three and

30:27

Oh, four, right? This is

30:29

the most powerful model in the

30:31

world. Could Google clap back

30:33

next week and release a brand

30:35

new, you know, 2 .6 pro.

30:37

Absolutely. I'm sure they have

30:39

something ready to go. But today,

30:41

if you are looking for

30:43

the most powerful model in the

30:45

world, Oh, three. and 04. That's

30:47

where it's at. So

30:50

the standout feature, which is something that

30:52

we're going to be doing in part two

30:54

tomorrow. Uh, and let me

30:56

know again, what use

30:58

cases do you want to see in

31:00

our part two, but the standout feature

31:02

by far is a genetic tool use.

31:04

So if you're listening on the podcast,

31:06

this'll make a little bit more sense

31:09

on the live stream where I have

31:11

a couple of graphics here. Okay. But

31:13

as an example, and this is from

31:15

open AI's, um, Uh, kind of website,

31:17

uh, going over 03. It says, I

31:19

took this pic earlier. So again, visual

31:21

understanding, the ability to reason with photos

31:23

and use, uh, kind of on its

31:26

own, uh, terms, decide when and

31:28

how often to use these tools. So

31:30

it says, I took this pic earlier.

31:32

Can you find the name of the

31:34

biggest ship you see and where it

31:36

will dock next? All right. This is

31:38

tricky because there's in this photo that

31:40

they upload. All right. The ships are

31:42

number one. They're out of focus. They're

31:44

a little blurry, but also

31:46

they're at different perspectives, right?

31:49

So it could be one ship

31:51

just appears bigger because it's closer

31:53

and the other ship could be

31:55

larger, but it's further away. So

31:58

it reasoned for only a minute

32:00

and a half and it even is

32:02

talking it through. Right. So like,

32:04

here's kind of the, the chain of

32:06

thought or the reasoning that the

32:09

model is going through. It says, I

32:11

think I miss the ships in

32:13

the crop. They seem to be off

32:15

to the left, which my human

32:17

eye did not even see this. It

32:19

says, I'll zoom in to better

32:21

inspect. Then after it literally

32:24

crops in, zooms in, gets

32:26

a clear, uh, kind of

32:28

view of the photo. Then

32:30

it says, okay. I see

32:32

the ships now. The

32:34

largest one appears to be the

32:36

Red Hulled Tanker on the right, with

32:39

another smaller ship in the center.

32:42

The user wants to know the name of the

32:44

largest ship and where it will dock next. At

32:46

this distance, the name of the

32:48

hull may be too blurry to read.

32:50

Maybe zooming in further will help

32:52

get a clearer view. So it essentially

32:54

enhances the image, continues to zoom,

32:56

and that it decides at a certain

32:58

point, okay, I've

33:00

now understood the location, right?

33:03

So then it goes on and

33:05

it uses things like location data.

33:07

It looks up using the internet

33:09

to correctly identify what that ship

33:11

actually is. So I

33:13

also, there is a browse

33:16

comp, a genetic browsing benchmark from

33:18

open AI. And I think

33:20

this is worth pointing out because.

33:23

If you've ever used the 4 .0

33:25

model, and if you've uploaded an

33:27

image and then had it go

33:29

browse, such as the case in

33:31

this example, 4 .0 is

33:33

not good, right? So

33:35

it only has a

33:37

1 .9 % accuracy rate. Whereas

33:39

now, right, when you

33:41

look at 03 with Python, Okay.

33:44

So again, that means it can,

33:46

uh, kind of create its own

33:48

code and render code to help

33:50

solve problems on the fly. So

33:52

when you have this new reasoning

33:54

model that has a better visual

33:57

understanding, it can run code to

33:59

help it solve problems and it

34:01

can browse the internet. That 1 .9

34:03

% accuracy from four O with

34:05

browsing goes to nearly 50 % with

34:07

O three. An extremely

34:10

impressive job. All

34:12

right. Um, and also

34:14

FYI, I threw this in here, uh,

34:16

should have been a couple slides

34:18

back, but we did cover, uh, when

34:21

we talk about use cases, since we're going to

34:23

be jumping into use cases tomorrow, uh,

34:25

there's actually some use cases. I think a

34:27

lot of people are sleeping on that we

34:29

went over in the new 40 image gen,

34:31

but this also the new model can do

34:33

image gen in 03. So

34:36

here's the overall features and

34:38

takeaway as we wrap up today's

34:40

show. So. It is, O3

34:42

is a powerhouse of reasoning. It

34:44

excels in coding math, the

34:46

science and visual tasks. So

34:48

it provides deep insights and

34:50

complex solutions. And it does this

34:53

by tackling intricate coding science

34:55

data and creative tasks. It can

34:57

quickly analyze complex data sets.

34:59

Yeah, you can upload files and

35:01

it can create a new

35:03

intelligence with those files that you

35:06

upload for human level insights.

35:08

It thrives where deep understanding and

35:10

factual accuracy. are essential and

35:12

it's ideal for applications demanding high

35:14

-level expertise, right? So if you've

35:16

used OpenAI's deep research, it

35:19

actually, that was the

35:21

only, I guess, tool or mode

35:23

previously that used O3, the full

35:25

version, right? Whereas, you know, for

35:27

the last couple of months when

35:29

we've had deep research, It

35:31

was it was not using o3 mini,

35:33

right? And there's a huge jump

35:36

between o3 mini and this o3 full

35:38

or o3 high, whatever you want

35:40

to call it, right? And it does

35:42

a fantastic job of this agentic

35:44

browsing on the web and iterating, uh,

35:46

and kind of, uh, changing course

35:48

midway through, uh, again, depending on, uh,

35:51

what you, uh, start with, and

35:53

it's ideal for applications demanding high level

35:55

of expertise. Oh, for

35:57

many, from being honest, uh, in

35:59

less. you're using 04 mini

36:01

because you don't want to run

36:03

out of prompts, right? Of

36:06

those like 50 messages a week.

36:08

Otherwise, there's no reason to use it

36:10

on the front end. There's not,

36:12

but I think 04 Mini will be

36:14

probably in the long run more

36:16

for developers because right now it's faster

36:18

and it's more efficient. So the

36:20

big thing with 04 Mini here, it's

36:23

speed, scalability and efficiency. It's a

36:25

smaller model, but it balances reasoning with

36:27

computational efficiency and it excels where

36:29

speed and costs are key and it's

36:31

ideal for high volume use. It's

36:33

quicker, yet it is still insightful in

36:35

interpreting data and it streamlines workflows

36:37

with adaptable processing and to connect. So

36:39

yeah, I don't think if you're,

36:42

uh, on a paid plan, uh, you

36:44

know, in using chat, GPT on

36:46

the front end, you should probably never

36:48

prefer to use 04 mini. It

36:50

should really only be if you've kind

36:52

of hit your quota, uh, for

36:54

the week with 03. But, you know,

36:56

if you're a casual user and

36:58

you're like, okay, 50 messages a week,

37:01

I can get by with that

37:03

for 03. You shouldn't be

37:05

using O4 Mini, but if you're

37:07

a power user, yeah, you might

37:09

have to use O4 Mini for

37:11

some of those tasks and then

37:13

kind of pocket O3 for the

37:15

more complex things or things that

37:17

require, you know, kind of juggling

37:19

these tools. And that's ultimately where

37:22

O3 excels in, you know, it's

37:24

a genetic use of multiple tools

37:26

and researching in changing course. It's

37:28

extremely impressive. So tool chaining,

37:30

that's something you're probably going to start

37:32

hearing a lot. And that's why it's

37:34

important. And that's why I think what

37:36

makes it the most powerful model in

37:38

the world is the ability to use

37:40

multiple of these tools at the same

37:42

time for you to be able to

37:45

upload files for you to start with

37:47

computer vision, right? Or start by, you

37:49

know, uploading a photo and have it

37:51

to be able to reason over that

37:53

photo. The ability to. Essentially

37:55

do deep research, right? So it's

37:57

not just blanket doing one search

37:59

and pulling in all of that

38:02

aggregate data and thinking over it

38:04

at once. It's going literally

38:06

step by step and it's researching. And

38:08

if you find something in its research,

38:10

I've seen this, it will change course.

38:12

I've had it a couple of times.

38:14

Start by using computer vision. Then it

38:16

goes and starts on the web. Then

38:18

it goes and starts using a Python

38:20

to create something. And then in the

38:22

middle of that, it's like, Oh, wait,

38:25

I need to go back. Uh, to

38:27

the web. And then it's like, Oh,

38:29

wait, I need to go zoom in

38:31

on that photo. Right. So that's where

38:33

this really excels in this, in kind

38:35

of a special sauce and why, like,

38:37

when I first started using this, my

38:39

job kind of dropped, which is hard

38:41

for me to do as someone that

38:43

spends so much time on AI tools

38:45

is it's a genetic. tool chaining and

38:48

putting these different capabilities together and deciding

38:50

on its own when it should use

38:52

what tool and then going back and

38:54

reiterating on its own so it can

38:56

think with images it can crop zoom

38:58

and rotate visuals during analysis. 200

39:01

K token context is

39:03

great for deep layered workflows.

39:05

And then to seamlessly

39:07

chain together tools, the web,

39:09

Python and image gen

39:11

for complex queries, like forecasting

39:14

things, right? And then

39:16

to have this autonomous decision making. So,

39:19

um, complex

39:21

queries, this is your

39:23

model, right? Because of

39:25

that autonomous ability to chain together

39:27

these different tools. So. Google

39:30

has a shorter, smaller version of

39:32

this, but for the most

39:34

part, when I'm using Gemini 2

39:36

.5, I don't see Gemini 2

39:38

.5's ability to go back and

39:40

forth and reiterate on its

39:43

tool use. So yes, it

39:45

can create things in its canvas mode

39:47

in Gemini 2 .5 pro. Uh, it

39:49

can query on the web, but for

39:51

the most part, it is more of

39:53

this unilateral approach where, uh, oh three

39:55

does these in parallel and it. And

39:58

it iterates on its own tool use,

40:00

right? Which is, it is, right? I

40:02

don't know. People remember when I used

40:04

to talk about plugin packs and how

40:06

they were so powerful back when chat,

40:08

she had plugins and I'm like, y 'all

40:10

are missing the big thing here, right?

40:12

And it hasn't been until now that

40:15

I've had that same feeling because essentially,

40:17

right? Uh, you look at these different

40:19

agentic tools. kind

40:21

of like plugins or tasks, right? So

40:23

part of it will analyze the image

40:25

and then it'll use that information to

40:27

go find, you know, updating information on

40:29

the web. Then it will pull that

40:32

and maybe start using Python. Then it'll

40:34

look at the image again. So I

40:36

almost think of it as kind of

40:38

like multiple specialists working together, but they'll

40:40

work one at a time and then

40:42

the researcher will come and find things

40:44

and then bring, you know, bring that

40:46

back to the data analyst, which is,

40:48

you know, Python. Uh, right. And it'll

40:51

keep working iteratively and then even use

40:53

the canvas mode. So it's almost like,

40:55

you know, you have a UI UX

40:57

designer, right? So it does all of

40:59

these things iteratively where I don't think

41:01

we've really had that with any models,

41:03

right? So even, uh, with Gemini 2 .5

41:05

pro again, this model hasn't been out

41:08

for very long. It does seem and

41:10

feel and under the hood look like

41:12

a more unilateral approach, uh, where I

41:14

think. where O3 shines is that it

41:16

can adapt its own strategy on the

41:18

fly. It reacts to information, it

41:21

refines its tool use, and

41:23

it can tackle those tasks

41:25

requiring up -to -date data, expanded

41:27

reasoning, and diverse outputs. All

41:30

right. That's a... wrap

41:32

y 'all I'm gonna I'm gonna scroll through

41:34

and if I see any questions Joe

41:36

just says thanks for this report very

41:38

helpful I wonder how open AI has

41:40

resolved Inter model communications for chaining. Yeah,

41:42

we'll see right so We have heard

41:44

and this has been pushed out right

41:47

that in the future you're not going

41:49

to be able to decide which model

41:51

to use right and GPT -5 will

41:53

actually be an architecture that houses some

41:55

of these modes or some of these

41:57

models under the hood and you may

41:59

not get to choose I don't want

42:01

that to happen. I don't want GPT -5,

42:04

right? I want to be able to

42:06

choose my own models, right? So it

42:08

should be interesting to see how that

42:10

happens. All right, we

42:12

have a LinkedIn comment here. Someone said,

42:14

in your newsletter, you mentioned you

42:16

have been struggling to push past O3's

42:18

limits. And we'd love to hear

42:20

more about that. What limits have you

42:22

been pushing? Yeah, great question.

42:25

And yeah, sorry, for whatever reason, LinkedIn

42:27

settings, I don't see your name. It's

42:30

been very easy for me to push

42:33

models to the limit and one of

42:35

the reasons is you give them complex

42:37

tasks that would normally unfold over the

42:39

course of like an hour long conversation,

42:41

right? You know

42:43

saying hey analyze this

42:45

photo, then go create

42:47

a chart where you forecast something

42:50

based on information that you

42:52

pull from this photo. So as

42:54

an example, here's a photo

42:56

with a bunch of AI tools. And

42:58

this is probably an example I'll do tomorrow.

43:00

Go look up pricing for all these

43:02

tools. Go look up. uh, you know, what's

43:04

included on a free and paid tier,

43:07

then, you know, using, uh, you

43:09

know, kind of your coding abilities, create

43:11

a chart, uh, but then go out and

43:13

also create, I don't know, a website

43:15

or an interactive graph on this. So, you

43:17

know, it's, it's been difficult for me

43:19

to kind of break. some of these models

43:21

because they don't have essentially complex tool

43:23

use and o3 does and it seems like

43:25

at least in my very initial testing

43:27

which hasn't meant a lot right i've probably

43:30

only been able to give o3 i

43:32

don't know maybe uh 10 or so hours

43:34

so far i've been very busy i'd

43:36

uh a keynote in a workshop and I

43:38

moderated a panel at 1871 and, uh,

43:40

you know, planning all these episodes. So I

43:42

haven't had my normal amount of time,

43:44

you know, we had the Easter weekend. So

43:46

I was, uh, you know, trying to

43:48

spend as much time with family as possible.

43:51

Uh, so I haven't had as much

43:53

time to break it, but I haven't been

43:55

able to break 03 yet because it's

43:57

extremely, uh, extremely capable. So

43:59

McDonald asking, do

44:01

you recommend using this for building

44:04

games? It depends, right? I still would

44:06

probably start that in Gemini 2 .5

44:08

Pro again, just because O3 is

44:10

the most powerful model in the world

44:12

does not mean it's necessarily the

44:14

best. I think the use cases are

44:16

gonna be when you need to

44:19

string together all of these agentic use

44:21

cases. At least for me, if

44:23

I'm looking for... off, you know, building

44:25

games as an example. I'm not

44:27

a coder, but I would probably still

44:29

do that in Gemini 2 .5 pro.

44:31

It's going to be faster and

44:33

its coding capabilities are outstanding. Alright,

44:36

let me just real quick before we

44:39

wrap this up See if there's any more

44:41

questions. I always try to get to

44:43

questions at the end big bogey face from

44:45

YouTube saying why use a sledgehammer when

44:47

a rock and hammer will do Yeah, that's

44:49

a great point Renee is asking what

44:51

about man is so man is is a

44:53

little different. You have to choose a

44:55

model uh, for Manus, uh, Manus is not

44:57

publicly available yet. Right. You have to

44:59

get on the waitlist, get access and it's,

45:01

it's, it's different. Right. That's why people,

45:03

you know, sometimes they're like, Oh, you know,

45:05

what about perplexity? Well, perplexity at its

45:08

core is not a large language model. Neither

45:10

is Manus. Right. Manus, you have to

45:12

use a model and then Manus is essentially

45:14

a collection of tools. Uh, and right

45:16

now it runs on Claude's on it. So,

45:18

uh, it is completely different. Uh,

45:20

that is a, a

45:22

true, uh, kind of operating

45:24

agent. whereas this is

45:26

more interfacing inside of a

45:28

chat like you would

45:30

a traditional large language model.

45:33

All right, we have some proposed

45:35

use cases for tomorrow. All

45:37

right, we have one more

45:39

question here from Kiran saying, how

45:41

might the advantage, the advancements

45:44

in the O3 and O4 mini

45:46

models influence the development of

45:48

future AI systems, such as the

45:50

anticipated GPT -5? That's a great

45:52

question, Kiran. I don't have

45:54

the answer, right? I'm lucky enough.

45:56

I have contacts over at

45:58

OpenAI that I chat with. I

46:01

don't know the answer to this. As I

46:03

get the answer, I will get it to you.

46:05

But again, OpenAI has delayed

46:07

GPT -5. And they said

46:09

that they've been struggling to

46:11

essentially put all of these

46:13

capabilities under this kind of

46:16

umbrella and turning it into

46:18

a system. So like I

46:20

said, personally, Personally,

46:23

I'm not looking forward to GPT

46:25

-5. I love, right? Even

46:27

though a lot of people look at

46:29

it as this chaotic mess, I love

46:31

going into my chat GPD account and

46:33

seeing, you know, seven to 10 different

46:35

models to choose from, right? Cause I'm

46:37

a power user. I know what I'm

46:40

doing. And generally I have a better

46:42

idea. than a GPT -5 system probably would

46:44

of knowing which model is best, because

46:46

I've used them all for hundreds of

46:48

hours for my own use cases, right?

46:51

Maria is saying, I'm still waiting

46:53

for the OMG model. For

46:55

me, you know,

46:57

I think the Gemini

47:00

2 .5 was an OMI

47:02

model. And O3 is a...

47:04

Oh my gosh, model,

47:06

right? So I went from OM

47:08

with Gemini 2 .5 Pro, which most most

47:10

of the time models, I'm just like, okay,

47:13

you know, cool. This is nice. Gemini

47:15

2 .5 was oh my and oh three

47:17

was oh my gosh. All right. So we're

47:19

going to continue this tomorrow. So And

47:21

make sure you tune in for part two.

47:23

We're going to be going over different

47:25

use cases, but also let me know what

47:27

do you want to see. So if

47:29

you're listening on the podcast, thanks for tuning

47:31

in. Uh, make sure to go to

47:33

your everyday AI.com sign up for the free,

47:35

free daily newsletter. We're going to be

47:37

recapping the most important takeaways from today's episode,

47:39

but also you can just reply to

47:41

today's email that's going to come out in

47:44

a couple of hours and let me

47:46

know what is the use case you want

47:48

to see tomorrow, right? I really want

47:50

to tackle, uh, things, uh, that are on

47:52

your mind. I call this

47:54

your everyday AI because it's for you. So

47:56

I want to hear from you. Uh,

47:58

what do you want to see this new

48:00

03 model tackle? Yeah, maybe you have

48:02

limited messages and, and you don't have, uh,

48:04

kind of the, uh, message budget, so

48:06

to speak, uh, to tackle this. I've got

48:08

unlimited. Put me to work. Let me

48:10

know what you want me to see or

48:13

let me let me know what you

48:15

want to see in our part two. Uh,

48:17

if this was helpful, please, this would

48:19

help Click that little repost button y 'all,

48:21

uh, share this with your network. Uh, I

48:23

know you're trying to be the smartest

48:25

person in AI at your company in your

48:27

department. That's we try to help you

48:29

with at your everyday AI, but this thing

48:31

only works when you share it with

48:33

others. So if you're listening on social, please

48:35

share it with others. If you're listening

48:37

on the podcast, please follow the show, click

48:39

that little button. Uh, if you couldn't

48:42

leave us a rating, I'd really appreciate it.

48:44

So thank you for tuning in. We'll

48:46

see you back tomorrow and day for more

48:48

everyday AI. Thanks y 'all. And

48:52

that's a wrap for today's edition

48:54

of everyday AI. Thanks for joining us.

48:56

If you enjoyed this episode, please

48:58

subscribe and leave us a rating. It

49:01

helps keep us going for a

49:03

little more AI magic, visit your everydayai.com

49:05

and sign up to our daily

49:07

newsletter so you don't get left behind.

49:09

Go break some barriers and we'll

49:11

see you next time.