0:01

Welcome to Preparing for AI , the

0:04

AI podcast for everybody . With

0:07

your hosts , jimmy Rhodes and

0:09

me , matt Cartwright , we

0:11

explore the human and social impacts

0:13

of AI , looking at the impact on

0:15

jobs , ai and sustainability

0:17

and , most importantly , the urgent

0:20

need for safe development of AI governance

0:23

and alignment . urgent need for safe development of AI , governance and

0:25

alignment . Looking for some happiness , but

0:27

there is only loneliness to find . Jump

0:29

to the left , turn to the right , looking upstairs

0:31

and looking behind . Welcome to preparing

0:33

for AI with me , smirvish D Wool

0:35

.

0:36

And me , Trevor Herndon .

0:38

Yeah , that's right . Yeah , yeah , Good . Yeah , I'm really

0:40

tired this week . I was going to tell people that to

0:42

start . I'm really tired this week . I was going to

0:44

tell people that I'm wrecked . Jimmy's

0:46

Jimmy's tired . I've come to return the

0:48

podcast and found him asleep

0:50

. Um , he's literally just woken up

0:52

. I've just got back from South Africa

0:55

and New . York , where you've been doing

0:57

some AI research .

0:58

Uh yeah , In the Outback , not

1:00

Outback , Safari place .

1:02

You were trying to find places in the world

1:04

that were untouched by AI , right ?

1:07

So you went to New York , yeah , yeah

1:09

. And then you went to South Africa

1:11

. Yeah , all the way around have they been touched by AI

1:13

? South Africa

1:15

felt very much like it has not New

1:19

York . Actually , I've never been

1:21

to the States before at all . It

1:24

felt very familiar and not

1:26

very AI-ish .

1:28

Have you been touched ?

1:29

by AI . Yeah , definitely Were you touched

1:31

by AI on your trip . Not

1:33

inappropriately .

1:34

Well , anyway , welcome to Preparing for AI . And

1:36

today's episode is not officially

1:39

a roundup episode . I

1:41

don't know why it's not officially one . I think only because the

1:43

one we released two weeks ago was um

1:45

, it wasn't two weeks ago , it was ages ago actually

1:48

but this is a roundup episode , um

1:50

. So we're just going to do the usual kind of looking

1:52

at the latest ai news . Um , there's

1:55

been quite a lot yeah , we're going to be .

1:57

Yeah , there's loads to round up actually , so

1:59

should we talk about gemini

2:01

2.5 ? Okay , uh

2:04

, just off the top of your head , so I think I was gonna do

2:06

this one . Um , gemini

2:08

2.5 . So yet

2:11

again . And , uh , we say this almost

2:13

every roundup , to be honest , but

2:15

, like , yet again , the there's a new

2:17

llm at the top of the leaderboards .

2:20

Um , so let me just say , because it like happened

2:22

so quickly that when you came back

2:24

the first time I saw you and

2:26

we said , oh , we needed an episode , and

2:28

you told me that gemini 2.5 was the best

2:30

model , and I was like , oh really , oh , I didn't

2:32

know . I mean , it's like , yeah , we used to know every

2:35

time . Now it was just like I don't know , I

2:37

and I sort of don't really not , I

2:39

don't care , but it's like it sort of seems irrelevant , because

2:41

by the time this goes out it might not be no

2:44

, it's true .

2:44

I mean to explain , uh well , what's

2:46

been at the top of the leaderboards recently .

2:48

So grok , grok 3 deep

2:50

seek was the first one to kind of shake it up

2:52

, wasn't it ?

2:53

deep seek the thinking model . So that was all this year

2:55

as well , and we're in march but in sorry

2:58

, 3.7 claude possibly

3:00

or roundabout 3.7 was

3:02

briefly at the very top for specific

3:04

use cases . I think the difference

3:06

with there's been , there's been the

3:09

the thinking model revolution recently

3:11

which , like has happened basically over

3:13

the last couple of months , um , I

3:16

think I think , gemini 2.5 is a

3:18

little bit different in that it it

3:20

was . It's actually

3:23

basically is top of

3:25

all the benchmarks . So Claude

3:27

3.7 was the best model

3:29

for coding , but wasn't necessarily

3:31

better in other areas . Gemini

3:34

2.5 , so it's a model by Google

3:37

. Google have been pretty quiet , but

3:39

we've said for a long time that they're one of these

3:41

companies that potentially are just going to do something like this

3:43

Bizarre to call it like the dark horse , but it kind

3:45

of feels like that , doesn't it ? But

3:47

Gemini 2.5 has been , and

3:49

when we talk about top of the leaderboard , just to sort

3:51

of clarify a little bit , so there's something

3:53

called LLM Arena for

3:56

people who aren't familiar , and this is basically

3:58

where humans use these

4:00

models , but you don't get to to , you don't know which

4:02

model you're using , so it's all like

4:05

a blind test and then

4:07

humans ask the same

4:09

question and get given different answers

4:11

and then choose which one they think is

4:13

the best . And so gemini

4:16

2.5 was this mystery model for

4:18

a while on llm arena and

4:20

they tend to release them a little bit earlier there as well

4:22

. And

4:24

yeah , gemini was like topping all the charts

4:27

there and then they released it and

4:29

then you found out it was like top of all these benchmarks

4:31

, and so it's literally I

4:33

think there's like eight different benchmarks around

4:36

human type reasoning , natural

4:39

language processing , coding , that

4:41

kind of stuff , and Gemini just

4:43

is like the best at all of them and

4:45

it's free and it's completely free

4:47

uh , at the moment , I guess we

4:49

don't know how long that lasts , but the moment is free , isn't it ? it's

4:51

completely free . Yeah , this is the thing about it . So

4:53

I was . I was quite impressed because I

4:55

I mean , I , we're claude fans

4:58

. There's no secret there . Uh

5:00

, I've been using claude's 3.7

5:02

for doing coding and all this stuff , and this

5:04

model came out again , sort

5:06

of just out of nowhere , and when

5:08

I went to try it , it's like it's really

5:10

good , but it's also really fast . It's

5:12

much faster than Claude Probably

5:15

, yeah , at least two or three times faster

5:17

.

5:18

Reasoning or non-reasoning is this this

5:20

is thinking , so we're reasoning .

5:22

Everything it does . Yeah , yeah . So the specific

5:24

use case I've been using lms

5:26

for the most recently and where

5:28

you , where I need the best one is code

5:30

, is coding using cursor , which I talked about

5:32

on a previous episode , and

5:34

for coding up like an entire

5:36

uh website and , and

5:39

actually cursor hasn't even been optimized

5:41

for for gemini yet . Um

5:43

, it's just significantly quicker

5:45

than claude . Like claude , when you're

5:47

doing stuff like that , it can take quite a while . Um

5:50

, and this , this , uh gemini models , like

5:52

incredibly , incredibly fast . So , yeah

5:54

, if you want to check out the

5:56

cutting edge models , do a quick search

5:59

for google gemini , I

6:01

think it's like 2.5 pro

6:03

pro yeah yeah , and you can um , I

6:06

think you go . You go to like a . It's

6:08

like a workshop type thing , it's not like a

6:11

typical chat gpt type interface

6:13

, uh , and you can use it there and it's multimodal

6:16

, like it can . I haven't even had a go

6:18

with it , but there is an option to stream your desktop to it as well , which looks

6:20

pretty interesting . I think you , you basically do like a , almost like a screen share , like you're sharing

6:22

your desktop and I as well , which looks pretty interesting . I think you , you basically do

6:24

like a , almost like a screen share , like you're sharing

6:27

your desktop , and I would imagine

6:29

you can chat with it about what's going on on your desktop

6:31

in real time .

6:32

So what so almost agentic ?

6:35

so I don't think it can interact with it in

6:37

terms of being a genetic , but it can view

6:39

it , so I presume it can like

6:42

solve problems with you in real time based on what

6:44

you're looking at on your screen .

6:45

I presume I haven't tried it actually , so yeah

6:49

, I mean when , when we

6:51

said a minute ago that we'd um , that

6:55

we sort of called or I called it a dark

6:57

horse , and I was thinking back probably

7:00

a month ago , we recorded the episode where we looked

7:02

at like the best models at

7:05

the moment , and I remember specifically

7:07

calling google the kind of forgotten

7:09

model , and we didn't really talk about it much

7:11

, which was , you know , appropriate , because at the

7:13

time , I mean , I guess the thing is it was never

7:15

a , it was never that gemini was a bad model , it

7:18

was just that there was nothing , it had never been

7:20

the first one so it always come out and there

7:22

was already , you know , good models , so it didn't

7:24

do anything different . It never kind of had

7:26

anything that seemed to make it stand

7:28

out . Um , but

7:31

we always thought that google had , because

7:33

they've got the the whole kind of ecosystem

7:35

, that it was likely that at some

7:37

point , if they were not going to be the best model

7:39

like they , I

7:41

think we said at one point , like anthrop , anthropic , we

7:44

saw potential problems for because they were

7:46

not big enough and they didn't have enough of an

7:48

ecosystem there to sort of build

7:50

on . I'm not sure if that's the case , because actually it seems

7:52

like , because they have a lot of their revenue through coding and

7:54

through the API , actually they have got their

7:56

kind of niche to some degree , but with

7:58

Google , well , they've got Amazon backing them yeah

8:00

, true , yeah

8:05

, and backing them , yeah , true , yeah , it's not like they're a little startup , is it ? um , but it does feel

8:07

like with google now , like at this point it's like right , they're banging

8:09

it now and I and I kind of feel like they've

8:12

got their act together , like it was also the issues they had

8:14

. Remember the whole kind of woke thing

8:16

where they had the images of what black nazis

8:19

and all kinds of weird historical

8:22

images where they were . They were , you know , using

8:24

, I mean , as a black nazi . Thing's a bit weird . So

8:26

I don't know that there weren't black nazis like that's

8:28

not necessarily a thing that didn't exist , but I know it was it

8:30

was basically trying to be too diverse

8:33

in the way that it was generating diverse

8:35

images of historical figures .

8:36

So you could ask it for a image

8:39

of a red indian and it would give you a white version

8:41

of one , and then , equally

8:43

, you could have like yeah , black nazis

8:46

, interesting example , but yeah well , that

8:48

was one that came up when I saw it yeah , but

8:50

then the more I think when I , when I said it and thought about

8:52

it , I thought well , actually , like there are black nazis

8:54

, there are nazis of all colors , so that's not

8:56

actually historically incorrect . No

8:58

, it's probably an unusual

9:00

it's an example .

9:02

It is an unusual example but in a sense , like sometimes

9:05

the thing that we talk about these kind of image generation

9:07

models and I remember the episode where amy

9:09

amy aisha brown was on and she was talking

9:11

about how , if you ask for a picture

9:13

of an autistic child , it always

9:15

brings up a sad white boy . So

9:18

there was something built into the kind

9:20

of biases there that I think what they've done is like turn

9:22

the biases the opposite way , flipped

9:24

it . Yeah , anyway , that was . That was the kind

9:26

of I think what everyone remembers is the kind of

9:28

big mistake that google made .

9:30

I think for a while they they were kind

9:32

of tarred by that they made a few faux pas

9:34

, I think , um , and also it feels

9:36

like , I mean , we , we we're obviously

9:38

really into it , so we probably sort of see

9:41

all every single piece of news , but it

9:43

feels like google are just very quiet in that respect

9:45

, like they're not . Claude is all . Claude

9:47

has always been the

9:49

best at coding . Um . Grok

9:53

is famous for being

9:55

, you know , complete , like , like

9:57

having no biases and no filters

9:59

in there , no guardrails , um

10:01

, although most of them don't now anyway . And

10:04

then , uh , chat gpt is just famous

10:06

for being the first and and and

10:09

actually quite often been overall

10:11

the best multimodal and most capable model

10:13

overall .

10:14

Um which I guess we're going to talk about a little

10:16

bit later in the episode yeah , totally

10:18

.

10:19

But whereas , whereas google sort of have a

10:21

usp on it almost in a way , I

10:23

think the other thing don't still don't , though , do they .

10:26

Even with this model . It's the best

10:28

at everything , which we've said

10:30

to people a lot of time but it doesn't really matter to you if it's

10:32

the best , because unless you're doing certain things you've given

10:34

the example of coding . There will be

10:36

other examples where people

10:39

who have particular things

10:41

they need to do , so I saw that programming , creative

10:43

writing were two things that it was particularly

10:45

good at . There'll

10:47

be people who have specific use cases where they will definitely

10:49

value it , but for most people they

10:52

wouldn't necessarily notice a difference

10:54

. Maybe they notice it's quicker . And also , there

10:56

still isn't , in a way , a thing to bring you to the

10:58

model , apart from the fact it's free . I

11:00

do question , because you've used it a lot more than

11:02

me . I do question , like because you've used it a lot more

11:04

than me , like when we say it's free . You know ChatGPT is free

11:06

, but if I use ChatGPT once I've done something for half an

11:08

hour , it

11:12

says you need to use the other model until 7 o'clock tomorrow morning . So is it free , like

11:14

not unlimited , but for a reasonable amount of use

11:16

, or is it free just for you to use it for 10

11:18

minutes and then it you know ?

11:29

it puts , you get to use it . It might not be the

11:32

other amazing thing . Which is genuinely amazing

11:34

with uh uh , gemini

11:36

2.5 and specifically for coding

11:38

, but for other stuff as well is it's got a 1 million

11:41

token context window , um

11:43

, which is significantly that than almost

11:45

all of them .

11:46

That was the thing that Gemini did have before was

11:48

that it actually was already a factor

11:50

, but no one really talks about it . It had a mass even

11:52

on , I think , 1.5, . It had a bigger context

11:55

window . It was a million . Yeah , can you explain

11:57

what that is for people actually ?

11:59

Yeah , so token , you can roughly equate

12:01

a token to a word . It's

12:05

you can roughly equate a token to a word . It's not quite one for one . So , like

12:07

some words are made up of multiple tokens . But if you , to keep it simple , if you think of as

12:09

a token has been a word um , then

12:12

then it's basically how

12:14

many words it

12:16

can retain in its memory before

12:19

it almost like forgets the first one , so

12:21

to speak . So with a

12:23

lot of models when they first came out so went like

12:25

GPT-3 , I mean , I know we're going back a

12:27

little way there I think it had something like a

12:29

4,000 token context window

12:31

, which is quite

12:33

short . If you think about 4,000 words , and actually

12:36

it's probably more like 2,000 words

12:38

, then you

12:40

know after 2,000 words it

12:42

would start to if you carried on the conversation . It would start to if you carried on the conversation

12:45

it would start to forget the things you were talking about first

12:47

. Um . Most models now

12:49

have something like 120 000 token

12:52

context windows as a maximum . Um

12:55

and google

12:57

has a million tokens

12:59

. And once you start get to a million tokens

13:02

, then you're talking about being

13:04

able to have a whole novel in its memory

13:06

effectively .

13:07

Do you know what Precursor 7 Sonnet is

13:09

? I think

13:12

I remember that it was a big leap forward

13:14

.

13:14

I'd have to look it up , but I still think it's probably

13:17

maybe a quarter Wow , so

13:21

maybe 200,000 compared

13:23

to Google's a million , and I think there's a version

13:26

. I'm not sure this is the version that you can access

13:28

uh for free , but

13:30

I think they do have a two million context

13:32

window version as well , so double it so anyway

13:35

. I mean , it's something that google's models are

13:37

already famous for it's

13:39

. It also lends itself very much

13:42

to coding , so it's not only the

13:44

best model at coding , but like . A long

13:46

context window for coding is quite important

13:48

because when you're using

13:50

things like cursor , you can end up having to like

13:53

, especially as

13:55

the code gets more complicated , as your code

13:57

base gets more complicated , you can end

13:59

up having to upload a lot of information

14:01

into the context window every time you're talking

14:04

to it .

14:11

The last point on this before we move on . Um , I just remembered one other thing that was supposed to be

14:13

, you know what it kind of excels at , and that was it's the reduced hallucinations

14:16

. Down now every model that's released . Now

14:18

, one of the things that they talk about how

14:20

is how it's reduced hallucinations , and it seems like

14:22

there is you a methodology that's been

14:24

sort of generally applied . Yeah

14:29

, I mean , it hasn't ruled , it

14:31

hasn't stopped hallucinating and I

14:33

guess you know , in the near future

14:35

they probably never will . And

14:39

I think it's the reason

14:41

for me it's kind of interesting is there is an acknowledgement

14:43

, the fact that every model that comes out now is talking

14:45

about reduced hallucination . Yeah , there's

14:47

an acknowledgement . It's a key thing to

14:49

try and reduce down hallucinations

14:51

. I don't know if you've noticed anything from from

14:53

using it , like I've noticed that models

14:56

, the recent models that have come out , they

14:58

tend to be better at telling you when they're hallucinating

15:00

or warning you about it , um , which kind

15:03

of helps . But I don't know if you've noticed with gemini 2.5

15:05

that there's a noticeable difference in hallucinations

15:07

I haven't noticed that .

15:09

The . The main thing I've noticed with with

15:11

um gemini 2.5 is

15:13

it started telling me what to do , uh

15:16

, which is a bit mad , I think I was talking to you

15:18

about .

15:18

We thought we're a few years away from that , but we've already started

15:20

, haven't we ?

15:21

yeah , yeah . So when I've been coding with it , it's

15:24

like cause , when you're coding with a model , it's

15:26

sometimes you have to do stuff to help it

15:28

do what it needs to do , because you cause

15:30

it can't necessarily carry out as a jet , it's kind

15:32

of an agentic thing . It can't necessarily

15:34

carry out some actions for you . And so the

15:36

other day when I was coding with it , it

15:38

was . I

15:48

basically hadn't realized that I

15:50

needed to take these actions , so I

15:52

asked it , I asked it , I told it that it wasn't

15:54

working and it said it reminded me that

15:57

I needed to go and take some actions . And

15:59

then later on , in the same conversation , it started

16:01

like adding reminders for

16:03

user needs to do this , user needs to do that

16:06

at the end of the , at

16:08

the end of each conversation , just as a

16:10

because it basically picked up on the fact that I

16:12

wasn't really paying attention were

16:15

you jet lagged ?

16:17

uh , well , yeah , was that pre-jet lag ?

16:18

and then it was just the other day , so okay , well , that

16:20

makes sense . It probably knew , yeah nice

16:23

.

16:27

So should we talk about the new ? Well

16:30

is it . It is new , isn't it ? The gpt

16:32

open ai is chat gpt

16:34

image generation model

16:36

. I don't know what it's actually called , but it's now integrated

16:38

with chat gpt rather than being a

16:40

kind of standalone yeah , so

16:42

um , the

16:45

big news here is everyone's decided studio

16:47

ghibli is the ghibli

16:50

ghibli , ghibli .

16:52

Sorry , definitely not ghibli is the uh trending

16:54

.

16:54

what's the word Viral ? Yeah

17:03

, it's the new viral .

17:04

Yeah , exactly , it's kind of a new meme . It's

17:07

to do everything in Studio Ghibli

17:10

style using ChatG

17:12

, gpt . So , um , this has

17:15

actually been in the news quite a lot , so if

17:17

you haven't been under a rock , you've probably seen

17:19

something about this . Um , but , yeah , gpt

17:21

have released a . Um , it's actually

17:23

a phenomenal image open ai open

17:26

.

17:26

Ai have released . Sorry open .

17:27

Ai have released a new

17:30

method of image generation which takes

17:32

much longer but is actually pretty

17:34

phenomenal . It's like much like I mean . Previous

17:37

image generation stuff couldn't get text

17:39

right . It couldn't spell things right .

17:42

You can edit these now like , completely

17:44

, perfectly , the text . You can write like a 22

17:46

line text prompt and it will take everything

17:48

into account Like it's . It's another level , isn't

17:51

it ? It's a completely different level . Yeah , take everything into account , like it's

17:53

it's another level , isn't it ?

17:53

It's a completely different level , yeah , totally Totally . If you haven't had a chance to

17:55

go with it , I think I think it's only

17:57

on paid accounts , um , but it's

17:59

, um , it's pretty awesome , uh

18:02

, and some of the images that can produce like proper

18:05

, like just like next level . It

18:07

doesn't actually a diffusion

18:09

model , as far as I can tell , as

18:11

far as anyone can tell , because it's not . It's

18:13

actually closed um source

18:15

as usual from open ai . Uh

18:18

, so , so , like you , but

18:20

apparently it doesn't use the standard diffusion

18:22

model that's been used in almost almost

18:25

all image generation , um , for the last

18:27

couple of years since this stuff came about . So

18:29

, um , yeah , I'd say

18:31

, just check it out , like , really , really , really

18:33

impressive . If you haven't seen all the

18:35

studio Ghibli stuff online , um , I'm

18:38

not saying that , right , you are , you are

18:40

now . I am now Okay . Studio

18:42

Ghibli . Anyway , apparently the

18:44

, the creator of the studio Ghibli stuff , is

18:46

absolutely horrified .

18:48

But I was going to say this is the irony

18:50

of it , the fact that that's what's gone viral , is that

18:52

he is literally the person who hates

18:54

AI more than anyone in the world

18:56

, yeah , and thinks it's destroying

18:58

the entire sort of industry

19:01

and art that he loves . And

19:03

it makes sense because he draws everything from hand and

19:11

that's why you know this Jujo Ghibli stuff is . I mean

19:14

, if you I don't know if most

19:16

people listening will know what it is Things

19:19

that spirited away , how's moving castle

19:21

, totoro , I guess , the sort of famous ones

19:23

, but it's everything is

19:25

kind of hand drawn . The amount of frames that

19:27

they have to draw to do it , it's like so

19:29

, so labor intensive , like

19:32

kind of have have just

19:34

pushed back against any kind of technology to do

19:36

these really really kind of um

19:38

, you know , authentic vintage

19:40

style of of cartoons , and now ai

19:43

is basically just recreating it all . So it

19:45

does seem kind of cruel that the viral

19:47

thing that is . I mean it is really cool , like people

19:50

are doing it . The main thing is that family pictures , so

19:52

people get family pictures and do them , and it's not

19:54

just studio ghibli style . You

19:56

can actually like , you

19:58

can actually choose specific films , and it will

20:00

do that and it's not just that like you can do it for anything

20:02

, can't you can do like 1980s baseball

20:05

.

20:05

um yeah , people have been doing

20:08

baseball game . You can do any kind of theme that

20:10

you want and it's amazing , south park

20:12

, like yeah , whatever style you want . And

20:14

the difference is like

20:16

previous image generation models would sort of

20:18

have a go at this , but they would . They would like

20:20

mess around with the picture , like so

20:22

, if you know , like you say , if you had a family

20:25

photo , it would like

20:27

they would . Previous models would sort of do

20:30

some kind of really rough approximation , but

20:32

generally like wouldn't really have any

20:34

past , any resemblance . These ones

20:36

look like it's genuinely

20:39

like . It's genuinely

20:41

like an artist has taken that photo and produced

20:43

it in a different style .

20:44

Basically , I don't know if you're aware of this , but because

20:46

I was doing a bit of research for this earlier

20:48

and um , because so

20:51

previously when you used chat

20:53

gbt and you tried to create an image and

20:55

use what's called dali at the time , basically

20:57

dali wasn't integrated . So you'd give a prompt and

21:00

what it would essentially do is go and send that prompt into

21:02

dali and then pull the image back in like it wasn't

21:04

integrated . Now it's integrated

21:06

4o , I think , is the model that it uses

21:08

. So it's it's not using the most up-to-date model

21:10

, but it now has access to all of 4.0's

21:13

training data to help it to create the image

21:15

. And that's one of the reasons why

21:17

this is better . And there was a couple of examples I'd given

21:19

. Someone said create an image that

21:21

shows why san francisco

21:24

is foggy . And it created an image

21:26

. You know it was able to reference 4.0

21:28

to find out why san francisco is foggy and

21:30

then to put that into then a kind of educational

21:33

image , whereas before you would have had to say

21:35

you could say make me an image of san francisco

21:37

that is foggy , but if you wanted to explain

21:39

why it couldn't because the image couldn't reference

21:41

training data the image could just follow your text

21:43

prompt right . There was another example which I

21:46

I wasn't that blown away by this until

21:48

they kind of explained it that they asked

21:50

it to create this comic book , um

21:53

, based on a unicorn

21:55

with an upside down horn , and

21:57

they said every other image creation

21:59

tool out there will turn

22:02

the horn the right way up , just can't do it

22:04

. We couldn't do it , but this one with this whole

22:06

thing kept the horn the wrong way . Now

22:08

that doesn't sound that phenomenal , but the point here is

22:10

that you know when you're prompting it to

22:12

do things . It's not

22:14

kind of limited . In the same way , you

22:16

gave the example of text . I think for me that has been

22:18

the biggest thing with image generation is even the

22:20

best ones that promise that the text will be perfect

22:23

. It wasn't perfect , you know it was . It

22:25

had got better , but it would still make mistakes . With text

22:27

you still saw the kind

22:30

of not so much the kind of six fingers

22:32

on girls , but you know there was still a bit of that

22:34

to some degree , the making mistakes . Now

22:36

these images are like you can

22:38

kind of create what you want to

22:40

do . It sort of feels sad in a way

22:42

that you know I think

22:45

this is it diminishes art even more . It diminishes art even

22:47

more and it's better , but you know

22:49

, it is incredibly fun and it's good

22:52

enough that you can use it , I guess , for

22:54

actual practical uses . So it's not now

22:56

just a pure novelty tool . The quality

22:58

of it is good enough you can use it for , I think

23:00

, things which are , you know , genuinely kind of worthwhile

23:03

and will will make time savings

23:05

or , you know , make things better than they were previously

23:07

. So , although I have a kind of sadness

23:09

for what this is doing to

23:12

art , because I think it is , you know , taking

23:14

away yeah , I don't think it's destroying you , but I think it's

23:16

taking away . Um , it does also

23:18

have like a lot of merits , yeah for sure

23:20

.

23:20

And like , like I mean so you could create

23:22

logos previously , but not if they

23:24

had text in them . Really like now you

23:27

can just create whatever logo you want and

23:29

it'll , it'll work we

23:31

said we need to update our logo .

23:32

This is our 50th episode , so maybe we'll

23:35

use the image generation to

23:37

you know , sex up our current

23:39

logo rather than creating a completely

23:41

new one yeah , yeah , just pimp it out and pimp

23:43

it up a little bit with , yeah , a bit of bit of gpt4o

23:46

love yeah so

23:52

deep seek .

23:53

I know literally nothing about this , so

23:55

I'm quite intrigued .

23:56

We did . We did an episode on deep seek . Then

23:59

we did an episode which that episode on

24:01

deep seek was our most popular episode

24:03

since the first episode

24:05

, which is , which is pretty awful I wish people would stop

24:07

listening to the first episode and then

24:09

we said we'd stop talking about deep seek , so we need a . We

24:11

did a second episode about deep seek

24:14

and then I think we did another episode just after

24:16

where we also talked a

24:18

lot about deep seek . So you know

24:20

, I don't want to talk about deep seek anymore , but

24:22

we need to talk about deep seek again

24:24

. Um , because their

24:26

impact I think is still kind of resonating

24:29

and is still creating kind of effects

24:32

that this is about deep sea but it's also a

24:34

kind of china and chips and kind

24:36

of huawei piece . So the first bit , deep

24:39

seek have brought out um v3.1

24:42

, um , which is

24:44

um , a non-reasoning

24:47

model . So this is not the kind of

24:49

there is going to be , I think r3

24:53

or r2 , r2 is

24:55

it r1 or r2

24:57

, so the first one was r1 . Yeah , well , anyway

25:00

, there's gonna be a new reasoning model which apparently is

25:02

going to kind of potentially be at the top

25:04

again , and there's all this kind of expectation

25:06

. V3.1 is not that . V3.1

25:09

is an upgrade on their original model

25:11

which was released in December when no one actually heard

25:13

about it , which was version three For

25:17

the technical stuff . So it's got transformer-based

25:19

architecture , 560 billion parameters

25:22

. It uses that mixture

25:24

of experts model . It has support

25:27

for a context window of a million tokens , so

25:29

it matches Gemini

25:31

2.5 . Like I said , this is not a

25:33

reasoning model , um , but it shows

25:35

, you know , really really high performance

25:37

um support for 100 languages

25:39

with near native proficiency

25:42

. It has apparently demonstrated

25:44

38 reduction in hallucinations

25:46

. So again talking about reduction hallucinations , but

25:49

those of you who listen to the model episode

25:51

will remember me saying that um , deep

25:53

seek was the sort of worst hallucinating

25:55

so that is , I like I like

25:57

the statistic a 38 percent reduction

25:59

in hallucinations .

26:00

That sounds like a hallucination

26:03

in itself . It does , doesn't it ? Yeah ?

26:05

but also like . So 62 percent of hallucinations

26:07

are still there . That also sounds not that

26:09

impressive , to be honest

26:11

yeah , I suppose . So it depends how what

26:14

percentage it had before but then I researched

26:16

this on perplexity so it could be a perplexity

26:18

hallucination yeah , exactly , um

26:20

, anyway , I'll believe it , but

26:23

yeah , um , like yeah , enterprise

26:25

customers have api

26:27

use of it . It has apparently

26:30

a chrome extension that's coming out . And the one thing

26:32

that I kind of wanted to talk about here , because it's something I've

26:34

noticed in the in the last few weeks . So

26:36

I talked a few times on the podcast

26:38

about how you're being in china

26:41

, about the difference in the way that ai

26:43

is integrated with stuff and that ai

26:45

is integrated with a lot of things , um

26:47

, that is not integrated within

26:49

other countries , but that we didn't really have

26:51

people using chatbots on their

26:53

phones or iPads

26:55

or whatever in the same way as we did in the

26:57

West with sort of ChatGPT and

27:00

Claude and things like that , and how DeepSeek

27:02

was what had kind of really pushed that into the mainstream

27:04

. The other thing that I've noticed recently

27:07

is like DeepSeek in China

27:09

is integrated in everything

27:11

, like everything . Like

27:13

you go into Baidu Maps Baidu Maps is this

27:15

sort of equivalent of Google Maps here and

27:18

you've got a DeepSeek logo

27:20

and you click on it and there's an AI feature . I

27:23

mean it is like it is crazy

27:25

and talking to friends who

27:27

are , you know , relatively

27:30

senior in companies or who run their own businesses or

27:33

people who are working in sort of accounting and stuff

27:35

like that and they're all talking about yeah

27:37

, we've got , you know , we've got ai integrated

27:39

, they've got deep seek integrated in this and we're like it only

27:41

came out two months ago and it's just

27:43

integrated in everything , like I

27:45

. It more and more makes me question . Like

27:48

this idea that deep seek is this you know

27:50

bunch of guys in a cupboard with a few

27:52

gpus I mean that was sort

27:54

of proven to be nonsense . They had a lot of money

27:56

, but that this was a side project , like I'm

27:58

pretty sure the you

28:00

know chinese communist party is . You

28:03

know , if they weren't backing them now this , then they're certainly

28:05

backing them now . But they've got a lot more behind them

28:07

than we thought , because the way in which

28:09

they bought that model out was kind of shocking

28:12

. It shook the foundations of the american

28:14

companies . It's kind of bought open source

28:17

forward . It's changed

28:19

, which we're going to talk about in a minute , but it's changed

28:21

the way open ai are potentially going to do things . They're

28:23

potentially going to go back to being a bit more open . You

28:25

know it's been phenomenal , but we kind of said that with

28:27

the models themselves . Well , actually

28:30

, like it's not that the model is that much better

28:32

. That's the key point . It's the economics . It's

28:34

how , um , how cheap it is . It's

28:36

how um , efficient it is

28:38

. I think the way you're seeing this integration

28:41

kind of shows that the examples

28:43

that I was giving are maybe the kind of business

28:46

like some of these , like I say , are big businesses

28:48

. Some of them are not that big and they're talking about

28:50

we've got ai integrated with deep

28:52

seek is like it's so much cheaper

28:54

that maybe this is the thing that businesses

28:56

that previously would be like you

28:58

know they might have thought about it , but like they're

29:01

not sure , they're not quite sure if it's the right time now

29:03

. It's so cheap with deep seek and

29:05

it seems that deep seat must have just gone on this

29:07

massive charm offensive again to

29:09

do all this stuff . I think they've got a lot more

29:11

. You know people working with them and sales

29:14

and etc . To get that integrated . But it's

29:16

phenomenal like have a look . You maybe

29:18

don't use many chinese apps , as me , but

29:20

if you look on chinese apps now you'll

29:22

just see deep seek integrated all over the place . We

29:25

chat deep seek integration . I'm not

29:27

sure if it's there now , but there's supposed to be a

29:29

way you can use it through the , through the

29:31

mini app . I'm pretty sure they're using it on the background

29:33

of , you know , delivery apps , everything like

29:35

that . It's it's . It's crazy

29:37

and they've got this new model coming , which you

29:40

know , even if it's not the best model

29:42

, I think it will be the like up there and

29:44

it will be cheap and it will have some that there's

29:46

going to be something about it oh yeah , and I think

29:48

the way these things work is they they obviously

29:50

develop the standard model first

29:52

and then they add the reasoning later , which is

29:55

the same for the other companies .

29:56

It's weird to talk about that , because reasoning

29:59

in itself is something that's only a few

30:01

months old . On that point

30:03

, I do wonder . It's something

30:06

I don't think we'll ever know , I wonder

30:08

. I think

30:10

there's three possible scenarios . I think one

30:12

of them is that the Western

30:15

companies , like OpenAI , were already looking

30:17

at reasoning but were basically

30:19

for want of a better word sandbagging .

30:23

And like they were just they were just keeping it behind

30:26

.

30:26

They were keeping it back so they could , like

30:28

, release things more slowly . And

30:31

then the release of R1 has kind of made all

30:33

these other companies release these thinking models straight

30:36

away . The other option , which I think

30:38

is less likely , is that it just

30:40

completely caught them by surprise and they've scrambled

30:42

to catch up and hadn't even thought of thinking

30:44

model , reasoning models , which I think

30:46

is less likely . I think they probably but they got it

30:48

.

30:48

They got their act together afterwards pretty quick .

30:50

I think they were sad . It was like within a week or something .

30:52

I think they were sad .

30:54

And I don't know what the third option

30:56

was . I'm tired and jittery

30:58

.

30:59

You've mentioned that . Yeah , yeah

31:04

, the thing with DeepSeek that

31:06

is kind of for me is so

31:08

I

31:10

don't want to say amazing , but it

31:13

is still amazing that they did it without

31:16

access to the best hardware , which is

31:18

what all like they don't have ?

31:21

I mean , maybe they do have the best chips

31:23

and you know they've just got them through back

31:25

channels .

31:25

Yeah , that's the thing They've

31:28

innovated , in a way

31:30

that , even if it involves some degree of copying

31:32

from chat , gpt and you know there are various

31:34

kind of accusations of how they did it even

31:37

if they did all that , even if they

31:39

still had to innovate to find the way to do

31:42

it cheaper and to do this inference

31:44

cheaper and , to you know , find

31:46

a way to make it more efficient . I mean

31:48

, you know they're also like in just

31:50

in terms of how they're advancing . So v3

31:52

that was released in December that's the one just

31:54

before that had 128,000

31:57

tokens and now it's a million

31:59

Like , just like the

32:01

leap forward in terms of how

32:04

quickly they've kind of made that

32:07

leap is just pretty staggering

32:09

.

32:10

Yeah , they seem to have basically levelled

32:12

the playing field almost immediately

32:15

. And who

32:17

knows , I mean it's possible that

32:20

R2 , the next reasoning

32:22

model they release , which probably won't be that

32:24

far away , might leapfrog everything

32:26

again . I don't know . It's

32:29

really interesting like it like for

32:32

a while it looked like open ai

32:34

were going to have the very best models , like almost

32:36

all the time , and they don't seem to anymore

32:38

. Like it's gemini is the best model now

32:41

. Claude was for a while like it

32:43

just seems to be switching hands all

32:45

the time right now and I think open ai

32:47

well , maybe they didn't have the most

32:49

resources available to them , like clearly xai

32:53

and and google obviously do have

32:55

enormous resources available

32:57

to them . I haven't heard from llama for a while . It's

32:59

possible we'll get a new open source model from

33:01

llama tomorrow . That is better

33:03

than all these models .

33:04

I don't know so I was listening to

33:06

nate whitmore's podcast

33:08

the other day and he was talking about a

33:11

kind of theory that what china

33:13

is potentially doing here is , you

33:16

know where the

33:18

us had the lead in terms of those closed

33:20

source models and the kind of software side

33:22

of it that they could sell to people is by

33:25

open sourcing and kind of just throwing that out

33:27

, it kind of gets rid of the

33:30

ability for china , for us to kind

33:32

of lead and generate all the revenue through

33:34

that , because it's all kind of open source , and then china

33:36

at some point later can pick up what

33:38

it does well , which is creating the kind of hardware

33:40

and not necessarily like the top end chips , but , you

33:43

know , selling the actual physical stuff and

33:45

creating it . And it makes a lot of sense . Like

33:47

it absolutely makes a lot of sense , like you know how

33:49

they did things previously is follow the same kind

33:52

of model , because it does feel like the closed

33:54

source models , like maybe

33:56

the maybe it was never going to . You

33:59

know it was never in the long run going to work

34:01

, because you've always said , like you know they're

34:03

not that far behind . But deep

34:05

seek was the thing that has kind of shaken

34:07

that up . And there was also like and

34:09

this is not just in one place like there's a lot of this

34:12

. Um , even people like mark andreessen

34:14

, who you know were

34:17

saying previously that the us was like two years

34:19

ahead of china , and now they're saying , oh

34:22

, it's three to six months in some areas

34:24

, and in some areas it's probably level

34:26

and in some very specific areas

34:28

, china may even be ahead like it's

34:30

like how quickly it's made that up

34:33

is absolutely crazy and

34:35

a lot of this is driven and you know , credit to

34:37

kapila gray shaw , you know christy loke

34:39

, who've had on this podcast for for kind of talking

34:41

about this stuff . You know well , before we thought about

34:44

it but a lot of this stuff has happened because

34:46

of chip controls and because you know the us

34:48

has kind of driven this innovation

34:51

yeah , they've pushed .

34:52

They've pushed them into a corner where they've had to innovate and

34:55

there are um , there are

34:57

a lot of smart people , like you have to . It's

35:00

been talked about a lot recently . Like you people

35:03

have still got . I think there's

35:05

a misconception that china is

35:07

still in this place where it's copying the

35:10

West , where that was the

35:12

case and it still is in some areas

35:14

, like it definitely still is in some areas but they're also innovating

35:16

and starting to really really

35:18

catch up and even lead in some

35:21

areas . I mean , the classic example

35:23

right now is electric cars

35:25

. It's got nothing to do with AI

35:27

, but , like electric cars , china went

35:29

from catching up to now leading

35:31

. Basically , they have the some of the cheapest

35:34

and best electric cars on the planet but

35:36

I can give you ai examples .

35:37

I mean , I said to you I was in wuhan a few months

35:39

ago and there are , you know , plenty

35:41

of self-driving taxis on the road

35:43

there . I know there are some in the us , but you

35:46

know that there's there's a lot of them in

35:48

Wuhan . It's not the only city that has that

35:51

. There are examples in Shenzhen

35:53

, which is the sort of tech hub

35:55

just over the border from Hong Kong

35:57

. Drone delivery services are

36:00

apparently just kind of the norm

36:02

. There they're happening . I

36:04

mean , when I say the norm , I'm not saying like every delivery

36:06

is taking place via a drone

36:09

, but it's there . That kind of

36:11

the low altitude economy

36:13

is a big part of China's five

36:15

year plan . It's a massive

36:17

thing . I mean , they're already talking about autonomous

36:21

helicopter transport

36:23

in the next five helicopter taxis , autonomous

36:25

helicopter taxis in the next five

36:27

years or so in parts of China . Like there's some

36:29

phenomenal stuff . I want to talk about

36:31

Huawei , actually , because I

36:34

think this is kind of important in terms of , like

36:36

it's Huawei's chip technology

36:39

that is contributing a lot towards

36:41

this . So , because of those sanctions , huawei are

36:43

basically you know , I mean they're not

36:45

an SOE , but they're basically the national

36:47

tech company of china . Now let's be honest . Um

36:50

, yeah , they've got these chips the ascend 910c

36:53

, which is designed to rival

36:55

nvidia h100s . They're

36:57

not as good as , obviously , the very top chip , but

37:00

I think the point is where we thought

37:02

previously is like you have to have the absolute top chip

37:04

. With the improvements in the architecture that

37:06

that we've seen from the likes of deep seek , it's like

37:08

they need chips , but they don't necessarily need

37:10

the very top chip . There's also some

37:13

talk 40 of nvidia's revenues

37:15

apparently may come from china

37:18

, not necessarily directly , but because a lot

37:20

of chips are sold to vietnam and singapore

37:22

and they're then illegally kind of exported

37:24

into china that way . So I

37:26

don't know , I mean , maybe that means like there are more nvidia

37:28

top end chips in china than there's

37:30

supposed to be . Yeah , um , but it definitely

37:32

feels like this sort of huawei have really

37:35

really been pushed to

37:37

innovate and and create these kind of better

37:39

chips . And I , I like , I'm wondering for the

37:41

first time it's like it's not like it's china gonna win

37:43

, because I don't think there is like a winner . It's not

37:45

like going to the moon , it's like one gets there , no

37:48

, okay , if we got to the moon , but whatever , no , but are

37:50

they can , then that's it . But it's like can they be

37:52

ahead at some point and can they be neck and neck and

37:54

pushing forward and going backwards ?

37:56

you know , I think probably they can

37:58

now they seem to , yeah , in general in

38:00

ai they seem to be , they seem to be

38:02

roughly in that kind of space . I

38:04

don't know , like being three months behind doesn't

38:06

seem that .

38:07

They're collaborating with DeepSea

38:09

and it feels like it's the DeepSea-Huawei collaboration

38:11

.

38:16

That is the kind of key thing in terms of what China is potentially going to do . But also

38:18

, if you're three months behind , that's basically effectively not

38:21

really behind at all . Two years behind

38:23

is significant . Two years behind

38:25

could mean , by the time you've caught up

38:27

, like the game's over okay but is

38:29

it because two or three ?

38:30

months behind . But is it because

38:33

six months ago no , not

38:35

six months . Three months ago china was two years

38:37

behind . Apparently now it's three months behind or

38:39

not ? So like they were the . The

38:41

point is like , when people say they're two years behind

38:43

or they're six months is like well , if they were two years behind

38:45

, but then two months later they were three months

38:47

behind . They were obviously never two years behind

38:50

no , no , it's all .

38:50

It's all kind of like they . Clearly it's just a figure thrown

38:53

out there isn't it ?

38:53

they clearly were they clearly weren't , but

38:55

I guess they were they're two years behind in what

38:57

people knew about . yeah , exactly exactly

39:00

like at that point deep seek hadn't come out

39:02

and no one has seen anything like this . So I

39:04

think I think it was just more that they showed

39:06

their hands , so to speak . But

39:10

what I mean is like

39:13

pundits thought they were two years behind

39:15

, and the difference between being two

39:17

years behind and three months behind

39:19

is like basically

39:22

you're out of the game versus like

39:24

you're neck and neck your

39:33

neck and neck right .

39:34

So , in terms of like the choreography of this episode , this is a bit of a balls up , because I'm

39:36

going to talk about open ai again . Um , which I probably should have talked about

39:38

. We talked about the open ai model , but anyway

39:40

, it's a different .

39:41

It's a different point , but you could just move this

39:43

bit before the other bit and then the

39:45

bit you've just said won't make any sense yeah

39:47

, so if I've done that and this makes

39:49

no sense , that's why we've done it .

39:51

If not , then then well

39:53

, either way it won't make complete sense , but

39:55

good idea , correct ? So

39:57

yeah , open ai . Um . Apparently

40:00

I haven't actually looked at the um

40:02

tweet or x . What

40:04

is a tweet called ? If it's on x , is it called an x

40:07

?

40:08

yeah , I think it's an x .

40:09

Is it you say I , I did an x , you

40:12

say I x'd it , or do you still say

40:14

I tweeted it , but on x ?

40:16

I don't know .

40:17

Actually , I think it's an x okay , well

40:20

, on x , someone either

40:22

tweeted or x . No , sorry , someone didn't

40:24

. The devil himself , sam outman , did

40:26

um to say

40:28

that open ai plans to

40:30

release a new open weight model , um

40:33

. Open weight is basically

40:36

an open source model um we've

40:38

discussed . It's like the difference in semantics between the

40:40

two . But , for for clarity

40:42

, an open source model . It will be the first

40:44

open source model since gpt2 in

40:47

2019 . I then looked this up

40:49

because I thought let me check that that actually was

40:51

an open source model . It wasn't an open

40:53

source model when they first released it , so it seems

40:55

like OpenAI have never actually been that open

40:57

, but they did make it open afterwards

41:00

. So anyway , they are going

41:02

to release a open source

41:04

model . We don't know if it'll be the best model . We don't

41:06

know why're releasing a a model and

41:08

not saying what it is . But this is in line with their usual

41:10

bullshit let's release 20 models at the

41:12

same time , thing that they were supposed to not

41:14

do . But this is a massive thing

41:16

because it is and I guess

41:18

, is why maybe the deep seat coming before this

41:21

kind of makes sense is open . Ai , the

41:23

company who were all about being open , that

41:25

were the least open company , have now

41:27

been pushed to the point where they are going to release open

41:29

source models . Is this the end of closed

41:32

source models ?

41:33

um , oh , I , I think so

41:35

. Yeah , I think so , like I . I I

41:38

mean , I think google's model

41:40

is still proprietary in that sense , and claude

41:42

is as well , but it feels like

41:44

it's moving more in the direction

41:46

of open source . I don't know why open

41:48

AI are doing this , but I think

41:50

it feels like they're just confused at the moment

41:52

, like everything they do is just

41:55

a bit odd and a bit

41:57

sort of disjointed . Like GPT

42:00

4.5 was clearly released

42:02

hastily in

42:05

my opinion , it wasn't the best at anything

42:07

and the argument for releasing

42:09

it was , it's like a bit more personable

42:11

or something like that um , yeah

42:13

, it was bullshit , wasn't it ? uh , they obviously

42:16

like released all their . You know , deep

42:18

seat came out and then they released . Well , they

42:20

, they already had a thinking model , but they released

42:22

the . They opened it up so you

42:24

could see what it was saying , see what it was thinking um

42:28

, but then not as much as Deep Seek did

42:30

. So they just seem sort of a

42:32

bit baffled at the moment , a bit lost

42:35

. The most exciting thing for

42:37

me is the new image generation

42:39

model , which actually genuinely seems cool but

42:42

then hasn't had a massive hype

42:44

around it .

42:44

Yeah , it hasn't had a big fanfare around its release

42:47

in a way , has it .

42:48

No , and

42:52

they still have this weird thing where , which they said they were going

42:54

to tidy up , but they still have this weird thing where you , when you try and use it , you have to choose

42:56

between eight , six different models , or eight different models

42:59

, and depending on which one you choose , you

43:01

can use different features

43:03

. So you , you can use deep research and

43:05

search and image generation

43:07

. They only work with certain versions and it's not . It's

43:09

really not clear . Ironically , the deep research and search and image generation , they only work with certain versions and it's not . It's really not clear .

43:11

I run it . The deep research is probably the most phenomenal

43:13

thing and that seems to have like . Obviously

43:15

it has been announced but that seems to have kind

43:17

of flown under the radar bit . We talked about

43:19

it a little bit but that's pretty amazing

43:22

. Um , I just I just had a quick look

43:24

at a bit more detail . So it's

43:26

the open weight model is a reasoning

43:28

model , so it's going to be similar to open

43:30

ai's 03 mini model , apparently so

43:33

well , also similar to 01 . So

43:35

a reasoning model , um , multiple

43:37

multilingual problem solving

43:39

, so it will be in multiple languages . Developers

43:42

can access and modify the model's trained parameters

43:45

, the weights , without needing the original training data

43:47

, which facilitates customization . And

43:49

just to clarify , because I actually wasn't 100 sure

43:51

, but open weights apparently is halfway between

43:53

open source and closed , so it gives transparency

43:56

in how the models make

43:58

connections , but it doesn't reveal all

44:00

of the code or the training data so that's

44:03

the sort of main difference , I guess .

44:04

Yeah , that's the whole thing about being able to make

44:06

so basically open weight , change

44:09

the underlying . You can't change the sort

44:11

of base you can't change the base

44:13

model in

44:15

that sense , but you you well , you can't see

44:17

what it's been trained on , but you can . You

44:19

can fine tune it right , because the

44:21

open weights means you can fine tune it .

44:24

I was just thinking we keep slagging off um open

44:27

. Well , I just slag off open ai because of Sam

44:29

Altman . I don't actually dislike open

44:31

AI itself , but we keep slagging

44:33

them off for being closed source . But

44:35

then Anthropic Claude , which we

44:38

love , is also closed source

44:40

. So Anthropic , if you're listening

44:42

, do the right thing .

44:45

Actually , yeah , that was my question . Is

44:48

there any indication as to why they've done this

44:50

?

44:50

from what you saw , I

44:52

mean , there is an indication in the tweet from

44:54

what I can see , but when deep seek came out

44:57

um sorry , when deep

44:59

seek r1 came out , sam outman , literally

45:02

within a day or two , said yeah , we

45:04

don't want to be on the wrong side of history , we think

45:06

open source is the way forward . Which was like

45:08

absolutely like quite obviously just a purely

45:10

reactionary thing . So I think it is directly

45:13

a response to deep seek

45:15

, to be honest .

45:16

Yeah , I mean , I'm I'm not sure that all

45:19

the all these companies do need to open

45:21

source . I don't know what the benefit to open AI , open

45:23

sourcing , their publicity

45:25

models are , but like yeah

45:30

, and their um publicity models are , but like yeah , maybe publicity like I don't

45:32

. In some ways , I don't really mind whether they are or not . I think for me , the argument

45:35

of open open source is almost as much about

45:37

having competition

45:39

as it is about having

45:42

open source models , and my

45:44

fear , like if you went back like a year

45:46

or two ago , like 18

45:48

months ago , my fear was that

45:50

everything was going to be concentrated

45:52

in like one or two big tech companies

45:55

and it feels like

45:57

with the release of deep seek

45:59

, that's just blown that out of the water I

46:01

still don't buy that the open air model that they

46:03

are giving to , you know , the dod

46:05

is not a better model , or the

46:07

deep seat model they're giving to the pla in

46:09

china is not a better model so when we

46:12

say when we say that these are , like , you

46:14

know , they're open sourcing instead

46:16

of closed source , like they're not

46:19

open sourcing . The best models even grok , you

46:21

know have said that they will open source six

46:23

months after they've released it . So by that point

46:25

they're basically waiting to release a new one . So it

46:28

does feel like , even officially

46:30

, they are not open sourcing from day

46:32

one . And I think we know now you

46:34

know we talked to what I'd go about how it

46:37

was the only technology in history where the kind of commercial

46:39

public operation was

46:42

ahead of the military one . I'm pretty sure we're

46:44

not in that space anymore no , no

46:46

, I don't think so .

46:47

I think it'd be it think so . I think it'd

46:49

be quite delusional

46:51

to think that really .

46:57

So let's finish off with an honourable mention

46:59

Jimmy .

47:01

Yeah , so Runway , I think this came out

47:03

. This is the most recent drop actually , it

47:05

was in the last 24 hours or so but

47:08

the Runway

47:10

ml4 has come

47:12

out , which is the latest

47:14

version of a video

47:16

generation um generating

47:19

generation model , um

47:21

, which , uh , so runway

47:24

has always been like one of the best and

47:27

the latest one . I mean I'll

47:29

be honest , like video generation is still

47:31

one of the areas where

47:33

it's I mean , it's obviously really difficult to do

47:35

and very computationally

47:37

expensive If you imagine creating

47:39

an image and then multiplying that by , you know , 30

47:42

frames a second , 60 frames a second but

47:45

it is making

47:48

really good progress . Um , I think

47:50

the the best way

47:52

to check this out is definitely not listening to

47:54

a podcast , um , because it's video generation

47:56

. So I think , if you want to have a look , check

47:59

out some of the latest stuff that can be done

48:01

with um runway 4 um

48:03

, it does look really cool and I think

48:05

you have to pay for it . But if you want to have a go with

48:07

it , some of

48:09

the stuff it can do now .

48:11

Is it the best ?

48:12

So it looks . It's

48:15

really weird because Sora came out ages ago

48:17

but then never actually came out . This was a

48:20

sort of side project of OpenAI

48:22

. It has come out now . Runway

48:26

is definitely comparable , if not

48:28

better , than sora . Um , it

48:30

still can only generate like 20

48:33

second clips . The people I've seen talking

48:35

about it are doing that whole . You know

48:37

this is the worst it will ever be thing , and they're , and they're

48:39

right . Um , I don't know how excited

48:42

to get about this because they're still .

48:43

It's only moved from 10 seconds to 20

48:45

in a year , which like okay , that's a year , it's doubled

48:47

. But it also is like yeah , I'm pretty sure

48:50

we talked to one point about . Oh . I think

48:52

you said at one point like in two years time

48:54

you'll be able to make a whole star wars film it

48:56

might be a little bit it might be a bit longer yeah

48:59

, and that's what people are still talking about .

49:00

So I think , I think I think there will

49:02

crack video generation at some point and

49:05

then you will be able to do things like that

49:07

. Um , I mean just to elaborate on

49:09

a little bit . So one of the things that previous

49:11

video generation models were bad

49:13

at was like the consistency between

49:15

frames . So you would have like

49:17

. So , for example , with the new model

49:20

you can have something like if you've got your

49:22

, if you're , if you're , if your main subject

49:24

, um is in the background and you've got

49:26

stuff passing in front of them in the foreground , like

49:29

imagine you've like got um , it's like

49:31

panning and you've got trees or , or

49:33

lampposts or other stuff like

49:35

passing in the foreground . Um , the

49:38

new model can retain

49:40

like the consistency . The example I saw

49:42

the other day was like it was a bloke who

49:44

had a like a crease on his shirt

49:46

in a specific pattern and

49:49

like a lamppost passed between you , passed

49:51

between the camera and the subject

49:53

, and after it

49:55

passed there was like total , total

49:58

consistency with the before and after

50:00

. And again

50:02

, this is much easier to like watch a

50:04

video of it online to see what I'm talking about . But

50:07

that kind of thing was something that video

50:09

generation models previously really struggled

50:11

with that kind of temporal consistency

50:14

, and so it's

50:16

a . It's just another step in the right direction

50:18

. I'll be honest , like if you watch the videos

50:20

, it's still . There's still obvious

50:23

mistakes in places , there's still

50:25

like of things that are just like clearly wrong

50:27

and messed up , a bit like the six

50:29

fingers in photos like a year ago

50:31

, um , but I want

50:33

to keep an eye on still cool

50:36

.

50:36

Well , uh , less than an hour

50:38

. That's good going for us . I think

50:40

we thought this would be half an hour

50:42

but it never is so . Uh

50:45

, thanks for listening everyone . Uh , as usual

50:47

, take care , have a good week there's

50:53

nobody like her .

50:55

She drives me crazy . Yeah

50:59

, she's my baby

51:02

. She's my baby

51:04

. There's

51:07

nobody like her . She drives

51:09

me crazy . Yeah

51:13

, she's my Gemini baby

51:16

. She's my Gemini

51:18

baby . There's

51:21

nobody like her . She drives

51:23

me crazy

51:25

. Yeah , she's

51:27

my Gemini baby . She's my Gemini baby

51:29

. There's

51:34

no party like her . She drives

51:37

me crazy

51:39

. Yeah , she's

51:41

my Gemini baby . But

51:44

remember , it's only

51:46

a matter of time until she

51:48

gets her dreams . But

52:18

remember , it's only a matter of time until she

52:20

gets replaced . But remember , it's only a matter of

52:22

time until she gets replaced . But

52:26

remember , it's only a matter

52:28

of time until she gets replaced

52:31

. But remember

52:33

, it's only a matter of time

52:36

until she gets replaced . But

52:39

remember , it's

52:45

only a matter of time until she gets her place . Thank

52:48

you .