0:02

You have taste in a

0:04

way that's meaningful to software people .

0:07

Hello , I'm Bill Gates .

0:12

I would recommend TypeScript

0:14

. Yeah , it writes a lot of

0:17

code for me and usually it's slightly

0:19

wrong .

0:20

I'm reminded , incidentally , of Rust here

0:22

, rust .

0:24

This almost makes me happy that I didn't

0:26

become a supermodel .

0:28

Cooper and Nettix . Well

0:31

, I'm sorry guys , I don't know

0:33

what's going on .

0:34

Thank you for the opportunity to speak to you today about

0:36

large neural networks . It's really an honor to

0:38

be here Rust , rust Data .

0:40

Topics Welcome to the Data Topics . Welcome to the Data

0:43

Topics podcast

0:45

.

1:06

Hello and welcome . Welcome to the Data part of 2025 .

1:09

Happy new year .

1:09

Happy new year . Today

1:12

is the January 6th of 2024 . My name

1:14

is Murillo . I'll be hosting you today together with Bart

1:16

Hi hey . And Alex behind the scenes . Hi

1:18

, alex , she's waving . Trust me on this one

1:20

, maybe , yeah

1:23

. Well , happy new year everyone . How , how was that

1:25

you're happy ? How was your your holidays

1:27

?

1:28

very good , very good bit of time off . Uh

1:31

enjoyed with family .

1:32

Went skiing for a week in switzerland , nice

1:34

can't complain , nice can't complain

1:36

very cool and , uh

1:38

, it's been actually a while since we met , I feel it's

1:41

been uh three weeks ish

1:44

, maybe even more , I think .

1:45

But wow , yeah

1:48

, so uh , your hair grew yeah

1:50

, my beard grew as well .

1:51

I try to take care of it , but you know are you

1:53

consciously growing out your beard ? No , I'm just

1:55

lazily growing my beard so

1:58

for people that are , yeah , just listening , we

2:00

are gonna also publish the video so you can check it out

2:03

for yourself . Maybe I need to .

2:04

I actually thought about it this morning maybe we can , uh

2:07

like , uh , add a close-up to it

2:09

, I'm okay , it's okay it's okay , this , this

2:12

, this is fine , it's okay .

2:14

But uh , yeah , I was thinking about like , yeah , I need

2:16

to , I need to get a haircut and all these things

2:18

, but it was . You know , it's the time of the year

2:20

, you know , so it's fine . But

2:23

, um , I feel like a lot of stuff happened

2:25

. I feel like you probably came across some things , but

2:27

I also feel like preparing for the episode . I'm like , yeah

2:30

, what was the thing that ? I saw that one time . So , I think

2:32

, preparing for it a bit , it was a bit like

2:34

what did I do ? When did

2:36

I read this ? Did we cover this or not ? But

2:38

I went over my notes and I found some things

2:40

and I think maybe the most timely , the the most thing

2:43

we must discuss is the

2:45

O3 . So

2:48

OpenAI released O3

2:52

, which is basically an

2:54

iteration of O1 . So

2:57

, a quick recap O1 is different

2:59

from the JET-GPT models in

3:01

the sense that it

3:04

does reasoning right , right . So I think the likelihood of

3:06

hallucination is smaller and , um

3:09

, open ai actually had

3:11

a 12-day event and

3:14

they unveiled o3 , which

3:16

is they skipped o2 . Um , yeah

3:19

, we can , we can

3:21

speculate why , but uh , yeah , so

3:23

because they went from 01

3:25

mini there was also

3:28

01 right sorry .

3:29

01 preview that was there for quite a long time

3:31

and then , relatively like a few weeks ago

3:33

, they released the actual one , right

3:35

, oh , actual one . And then , not

3:37

long after that , there was a tweet , right , they announced

3:40

it , so it hasn't been released yet .

3:41

So so right now , even this is mentioned

3:44

here that public safety

3:46

testing , so they have researchers and limited

3:48

people that have access to O3 . So

3:50

I think it also I think it's something interesting

3:53

that highlights the commitment to

3:55

safe .

3:56

AI as well .

3:57

Or I'm not sure if it's just washing

4:00

right , but they are doing this . O1

4:03

is actually available now , but actually I

4:05

think for you to use oh one , you need to have a subscription

4:08

pro account , right , which is crazy expensive

4:10

compared to what was before , do you

4:12

? I have ?

4:12

oh , one with a , just with a regular pro

4:15

.

4:15

I think it's like 20 euros a month , okay , okay

4:17

no , because I remember there was one that was like 200 euros

4:19

, but like maybe it's for a limited , I don't know

4:21

I don't know could be . So

4:24

yeah , so not too much , I mean

4:26

, I'll just skim through . Yeah , it's

4:28

doing better in a lot of stuff like coding , math

4:30

and science . They also put , like this , epoch

4:33

AI . Apparently this is a big deal because

4:35

this is an intentionally

4:37

harder dataset that models were not

4:39

trained on , but apparently it's doing much better

4:41

, et

4:43

cetera , et cetera . One

4:46

thing I also caught my attention is that they have the 03 low

4:48

, 03 high . They also had a one low , medium

4:50

high I from

4:53

. What I understood is that this is the

4:55

compute resource available

4:57

, because you do require more compute at inference

4:59

time . So I guess , depending how much compute

5:02

you have , you have these different models right , so

5:04

otherwise it would take too long that's crazy if

5:06

you compare those costs on the chart

5:08

that you're showing .

5:09

Yeah , so you have on the x-axis , if you zoom

5:11

in a little bit , like the cost per task , like it's not really

5:13

defined here what task it is , but like uh

5:16

, and on the y-axis the score

5:18

, and if you have like oh , one preview , it's like

5:20

around one dollars per task

5:22

and it scores 20-ish percent

5:24

. And

5:29

then you have the other extreme of $1,000 per task with O3 high , with

5:31

88% performance Indeed

5:33

, which is a big difference in performance , but

5:35

also , at the same time , like the

5:38

cost is crazy , right .

5:40

So this is the O3 high . So with a lot of

5:42

compute resources and x-axis

5:44

is logarithmic scale , no .

5:46

Do we know when this will be released ?

5:48

It wasn't announced , I think . Well

5:51

, if it was announced , I don't know , I

5:53

cannot tell you . But

5:55

yeah , I think the other thing that I wanted

5:57

to mention here Ah

5:59

here , release date For now O3 , now

6:01

widely available . Openai open access to researchers

6:03

. Public availability shared O state

6:05

for now oh three now widely available .

6:06

Open eye open access to researchers public availability shared with women experience end

6:08

of january . Okay , oh , three minutes , that's close by right , but that's the the mini

6:10

one .

6:11

Yeah , um , and

6:13

well it's .

6:14

We have similar performance in terms

6:16

of duration , like you have with oh one , which does

6:18

a quote unquote reasoning . It takes a lot of time

6:20

versus uh I think so

6:22

.

6:23

I think it was uh comparable , but I think so

6:25

. One thing that he mentioned is that they have uh

6:27

. One of the things that

6:30

the model will do is to assess how complex

6:32

the task is , and

6:34

then you will tune a bit how many iterations it

6:36

needs to do so the idea is that , yeah , and I think

6:38

this is what they're talking here deliberate alignment , uh

6:41

, no , that's not it . But they basically

6:43

say , for simple tasks , it will already

6:45

iterate less , so it saves also money

6:47

and , uh , time for you , right

6:50

, so they also have a bit of this . So actually , yeah

6:52

, if you're talking about efficiency , right , the

6:54

one mini and you have all these things , they already kind of tune

6:56

these things for you . So cool .

6:59

There's a lot of uh rumors

7:02

on uh on x again like

7:05

. There's someone from uh , from the open

7:07

ai team , that says something like uh , it

7:09

was , uh , it was more exciting to

7:12

do uh , to do machine learning , back when

7:14

we didn't know how to create super intelligent

7:16

yeah , yeah , maybe for

7:18

sure , I think I think we also had sam

7:20

altman a few weeks ago . Uh , hinting

7:23

a bit towards agi , but at the

7:25

same time , we have this every time that our new version

7:27

is expected , right ?

7:28

yeah , let's see , let's see it's . It's cool

7:31

that there is an improvement yet again indeed

7:34

, I also think that I'm

7:36

wondering , like you said , even the o3 with high

7:38

compute now it's very expensive

7:41

to run , so I'm also wondering if

7:43

these iterations are getting more than niche

7:45

use cases right like the general

7:47

use case is kind of there and maybe now it's

7:49

more the , the , the

7:51

niche stuff . Maybe one thing I see

7:54

you also have highlighted here on the notes part , a

7:56

bit segueing into it um

7:58

, one of the things that o3

8:00

does according to the announcement is

8:02

doing better on code . One thing we also did in the roots conf episodes in the llm

8:05

hunger games they also had that 03 does according to the announcement is doing better

8:07

on code . One thing we also did in the RootsConf episodes in the LLM Hunger Games they also

8:09

had different LLMs , I think Gemini

8:11

, the Claude and GPT

8:14

doing the

8:16

advent of code . And I

8:18

see one thing that you posted here is the performance

8:20

LLMs of advent of code of 2024

8:23

. What

8:25

is this about ? Maybe what is the Advent

8:27

of Code for people that don't know what it is ?

8:30

The Advent of Code is like an Advent calendar

8:33

that you have before Christmas . Where

8:36

the one that people ? Well , I don't know

8:38

if it's actually a regional thing , but here you get kids

8:40

, get an Advent calendar and , like you , open it up every day

8:42

and there's a chocolate in it .

8:44

I don't know how international . That is actually

8:46

in the us they do that okay okay , people will know

8:48

it .

8:49

Then in brazil they don't do it okay

8:51

it's not that international um

8:53

. Advent of code is uh , more or less that

8:55

, but instead of a chocolate , you get a coding

8:57

challenge every day and um , based

8:59

on how quickly you solve it , you

9:01

get uh , you get scores . Basically , um

9:04

, people that have

9:06

solved it uh correctly , the quickest uh

9:08

for all those days wins the end

9:11

of code . Yeah and uh

9:13

. What this uh , this article uh

9:15

by jerpint is

9:17

uh is is an overview on

9:19

how llms are performing on that end of code of

9:21

2024 , which is um

9:23

. I think the general consensus is that

9:26

it's not as good as we thought , but

9:29

it's not bad either . I

9:31

think that is uh , yeah , um

9:33

. I think the what we see is that

9:35

, like even the best lms , they they

9:37

struggle with , like the truly novel programming

9:40

problems yeah indeed , and

9:42

that's also what that's what I've called very much . It does like it

9:44

starts . It starts , it tries to be very

9:46

innovative every year , to come up with really , uh

9:49

, new concepts on challenges , to to

9:51

uh , to really make it difficult for people . I think there were

9:53

a number of them that were even

9:55

unable to be solved by any of the lms , which

9:58

is interesting . Um , this

10:01

is what this article does and we'll share it in the show

10:03

notes a bit of a comparing the performance

10:06

of these purely the models , but

10:08

at the same time , um , the

10:11

rumors are that most of the people

10:13

in the top 10 did use lm

10:15

as a support tool . Yeah

10:18

, so , not not an autonomous lm

10:20

that just solved this for me , but did

10:22

use it as a support tool to solve it as quickly

10:24

as possible .

10:25

Yeah , yeah , I think , which I think it's

10:27

the indian

10:30

is the same old story , right like uh

10:32

, even when uh ai

10:34

, so not gen ai , but ai

10:36

started with the chess tournaments , right , and

10:39

then there were the . There's a very , I think I

10:41

don't know if I mentioned before , but there was a there's a ted

10:43

talk from kasparov , which is the russian guy that used

10:45

to be the best chess player for many years

10:47

, and he said that Deep

10:49

Blue , he was the first one that lost to Deep Blue

10:51

. He said , actually he lost the second time

10:53

, the first time he won , but no one remembers that .

10:55

Yeah , yeah , yeah .

10:56

And after that they were saying

10:58

how they

11:01

also had competitions for AI

11:03

engines . They had competitions for ai engines . They had competitions for humans . They

11:05

had competitions that you could use both okay

11:07

, and the winner of the ones that you could use both , it wasn't the

11:10

best ai , it wasn't the best chess players , it

11:12

was a regular guy with three regular machines and

11:14

he kind of makes the point of that

11:16

these things are tools , right ? He also I mean

11:18

, this is , this talk is from years ago he says ai is

11:21

not the future , it's the present and

11:23

, uh , you need to know , like basically to

11:25

learn , how to use these tools , right . So I think it's

11:27

also the same like the advent of code

11:29

the lms are not going to do everything by themselves , but

11:32

I think it's like a very powerful tool

11:34

exactly for people to be able to use and like

11:36

steer in the right direction and whatnot . So

11:39

, um , yeah , it's interesting . I think also

11:41

we did this for the llm

11:43

hunger games and I think from their experiment

11:45

I don't know the parameters of their experiment

11:47

, but none of the none of

11:49

the ai engines passed like the fifth day or

11:51

something , so it was also .

11:54

So it was also , yeah , interesting to see maybe

11:58

talking lms , it could

12:00

uh moment to discuss deep

12:02

seek v3 yes , what

12:04

is deep seek ?

12:05

it also came across this . I didn't , I didn't know it , um

12:09

, but it came across my feed . Well

12:11

, and I'll segue into it afterwards , but , uh , I

12:13

wasn't sure what is . What is deep seek ?

12:15

so deep seek . Uh is an

12:17

lm model . Didn't know it either , but v3 was

12:19

released a few days ago , I want to say

12:21

a week ish ago . Um , it's by

12:24

a chinese company , um

12:26

, chinese group . I don't know , to be honest

12:28

, what the exact underlying organization

12:30

is , but it , uh , it surpasses

12:33

, uh , gpt 4.0 in a lot

12:35

of different tasks , um , which

12:38

is impressive , uh , in its own , I

12:41

guess , but it's also like the , the way

12:43

it can . So it's , it's you can fully

12:45

download , it , it's you can like . The price

12:47

that it's at is super , super , super

12:50

cheap when you use the API , when you compare

12:52

it to Claude , or or open

12:54

the eyes to be like a fraction of the cost . They

12:58

train that on NVIDIA

13:01

H hundred , h eight hundreds

13:03

, which

13:07

are less capable than what you typically would see . You see an h100 , okay , um , and they

13:10

train it at a cost of quote

13:12

unquote , only 5.5

13:14

million us dollars which

13:16

is a shit ton of money yeah but it is only

13:18

a fraction of the cost that it took , uh

13:20

, the big players to get to this level of performance

13:22

. Do you have um ?

13:23

do you know more or less how much the big players to get to this level of performance ? Do you know

13:25

more or less how much the big players spent

13:27

, just to put in comparison for people ?

13:29

Well , probably billions right .

13:30

Yeah , yeah , indeed .

13:36

So it's impressive that this is there . I've

13:38

played a little bit with it , not a lot , but

13:41

I read a lot of discussions

13:43

on this . It seems to be for a lot

13:45

of different things . It is especially coding . It

13:47

is on par with 4.0 . It is on par with

13:49

clot , even with 3.5 sonnet . Where

13:54

it's uh misses out a little bit is in , uh

13:56

, creative writing and prose writing , these type of

13:59

things where you have a lot of repetition in the text yeah

14:01

, yeah , where I think uh clot is probably the best

14:03

at this moment but it's .

14:06

Have you tried , gemini ?

14:07

to me . It's crazy that , uh , I've tried gemini , but

14:09

it's not , not in comparison this is specifically

14:11

I don't . I don't have an active gemini subscription , um

14:14

, but uh , to

14:17

me it's crazy that we have something that is as performant

14:19

as this suddenly popping up . Yeah

14:22

, like I , I didn't expect this a month

14:24

ago , that you suddenly have someone that is competing with

14:26

these big players yeah

14:28

, performance wise . I didn't even hear about it

14:30

to

14:33

be honest , like it undermines a little bit my my

14:35

previous statements , where I'm , where I used to say

14:37

, like only the very , very , very , very big players , yeah

14:40

, uh , are able to to

14:43

build these , these type of flows , to

14:45

build it , to train these LLMs . Because

14:48

here you see that with only 5.5

14:50

million they can build something that is on

14:52

par with which I did not expect

14:55

at all .

14:55

Yeah , and I think maybe to put in perspective

14:58

right , Because I think when we hear millions and billions , I

15:00

think it loses a bit . But

15:02

I remember I heard a very good comparison . They're saying

15:04

like a million seconds and one billion seconds

15:06

. How much is that in months and all these things

15:08

? And it's like a crazy difference . One is

15:10

like months and the other one is like

15:12

maybe years or something . So the

15:14

jump between millions and billions is huge . I

15:17

also agree that .

15:19

I also wonder if , like now that they see that

15:22

it is possible , if

15:29

there's going to be a lot more investment in this , yeah , and also maybe from a cost

15:32

perspective , to make the link there with with more , let's say , quote-unquote , sovereign models

15:34

. So you had the dutch government , uh I think I want to say

15:36

a year and a half ago uh , announcing that they

15:38

were gonna invest around

15:40

12 , 13 million euros in

15:43

building their own GPT language

15:45

model , where at that point

15:47

everybody was thinking that's

15:49

a fun project but it's not going to bring anything

15:51

that's performant . It's going to be very niche stuff

15:54

, specific , but here

15:56

you see that they could build something with only

15:58

half of that investment .

15:59

Yeah , that's true .

15:59

That is on par with what we think is is what we are

16:02

today .

16:02

I wonder , so

16:05

you mentioned that this is better and I think even here on

16:07

the screen we have the on their page , right

16:10

, they show the different benchmarks for English

16:12

code , math and Chinese . I

16:16

wonder if for creative things you need that

16:18

there's a you need to invest

16:20

more , because I think the

16:23

code and math you

16:25

could argue there is one right answer . Right

16:27

, it's very like not , it's not well , it's deterministic

16:29

it's already deterministic , yeah , indeed . So

16:31

I'm wondering also if , like you , can only

16:33

achieve this performance with this amount

16:36

of financial investment for the more deterministic

16:39

things and , I think , the things that are more subjective

16:41

, the more subjective it is . The more you it is , the more

16:44

you need to train . Basically , you need to put more money in it

16:46

. I wonder .

16:49

But it's a very interesting development .

16:51

For sure .

16:52

It makes me optimistic about the

16:55

competitiveness of LLM

16:57

models in 2025 .

16:58

Definitely agree . I came across this

17:00

because I'm a user

17:03

of Cursor . Cursor

17:07

, for people that don't know , is a VS Code fork . That

17:11

basically is . There

17:13

are a lot of features and there's also

17:15

a pain model behind to

17:18

use AI in

17:20

your IDE . And

17:22

they have a Reddit . In the Reddit they're very active

17:24

and they also mentioned when is DeepSeq v3 going to be available

17:27

again , because I think it was there so

17:29

you can actually . So basically it's like this

17:31

. So in Cursor you have the autocomplete

17:34

, normally , right . And actually that's the

17:36

first thing why I moved away from Copilot

17:38

to Cursor is because the Copilot

17:41

tab complete , it

17:45

was more like just an autocomplete , like you just finish the rest of the sentence

17:47

. It was when I was using it . I know that you mentioned since that's not

17:49

like this anymore the cursor , like in

17:51

the end it also had edits in the middle

17:53

of the , the , the string , basically

17:55

the text , and you can also do like

17:57

a chunk of changes , right . So

17:59

I thought it was interesting . Also cursor

18:02

, they have like the chat and

18:04

in the you can actually select the model . They even have the

18:06

O1 mini there and

18:09

actually that's where they had the DeepSeq v3 . In

18:12

the chat they also have easy ways to add

18:14

context , so either to add files

18:16

, or to say search the web , or even

18:18

to add documentation . So you can say like , hey , for

18:21

example , I'm using Polars now on my project and

18:24

if I want to search something documentation , I can just add

18:26

it there and they'll actually embed it already . So if I have a question

18:28

about polars , they'll easily find stuff . Um

18:31

, so yeah , they actually had the question on the

18:33

when is deep

18:35

secret 3 going to be available again for cursor

18:37

? And yeah , I was . I just looked at it was

18:39

like okay , it's a new model , but uh , I think I'm gonna look into

18:41

. That's very interesting yeah , I'm gonna look more into it

18:43

now . Um , I'm also bringing

18:45

this up because one thing that came across on these reddit

18:47

threads is this shadow workspace , and

18:50

that's what I was also . That's also what I

18:52

wanted to bring up . Um , so

18:54

, and I don't know how , how copilot is these

18:56

days , but the way that cursor was

18:58

working and the way that copilot was working back

19:00

to when I was using it , it's

19:02

just that it's an autocomplete based on the

19:04

text that is there . So one

19:06

thing that happens a lot because I'm using polars

19:08

is that it provides , it

19:11

suggests pandas functions . So

19:13

, for example , in pandas you have the dot apply . In

19:16

polars is dot map elements . So

19:18

I can clearly see that's probably because of the training data

19:21

, right . Um , now

19:23

, this shadow workspace that I was curious

19:25

, basically iterating on the background . I'm not going to go through

19:28

all of it , I haven't actually finished the document

19:30

yet , but basically it's

19:34

a feature already available in Cursor that

19:36

basically you can have another

19:38

background workspace where they allow

19:40

the AI to interact with the development

19:42

environment . So the LSP , so

19:44

the Language Server Protocol . I think they

19:48

also give feedback right . So , for example , if you're typing

19:51

something in VS Code and you write

19:53

dot , apply this in

19:55

Pandas , for example , apply . This is not a real method . So

19:57

you get the squiggly line and says this doesn't exist . So

19:59

you have some feedback , right , because it's always checking the

20:01

source code . So , basically

20:03

, before Cursor

20:06

will suggest something for me , they'll actually run

20:08

through the LSP . They will run

20:10

through . Actually , they even want to go as far as allow

20:12

it to run to test the code , to make sure , like

20:15

if you're writing something Go or Rust , something

20:17

that doesn't compile right , so it won't stop . So

20:19

basically , in effect , what

20:28

they want to do is to make sure that everything

20:30

that is recommended is already a bit fine-tuned , right , and I mean there is

20:32

also layers for this in already Copilot and whatnot . But I guess this

20:34

is taking it to the next level . And then they talk

20:37

about all the different principles , right

20:39

, they want to make sure that this is independent . So whenever

20:41

you're coding something you won't have , I don't know

20:43

, you know you won't have to wait , right , you

20:45

won't take away from your resources

20:47

, the privacy as well . The user's code should be

20:49

safe , so everything should be running locally concurrency , universality

20:52

, maintainability and speed . And

20:54

they kind of talk a bit more on how

20:57

actually works the lsp in vs code

20:59

and what they're planning to do . So

21:02

. And they and they also mentioned here there's

21:04

a warning right that this increases the memory

21:06

usage of cursor . So not

21:09

something I've tried yet

21:11

. To be honest , it's not that new , but

21:13

I saw that it came up on the Reddit

21:16

thread and I definitely want to give

21:18

it a try . Another iteration

21:21

on that . So it's also using Electron right , because

21:23

this is a VS Code .

21:25

That's maybe a good segue . Yes , to

21:27

go to boltnew , all right

21:30

. Yeah , let's do it , because they do all of this

21:32

yeah , they do all of this .

21:33

What is boltsnew ?

21:37

boltsnew and

21:39

solutions like these . You've also created the

21:41

xyz . You've also I think

21:43

most people will know v0 by

21:45

vercel . They're like uh

21:48

, ways to quickly build web apps

21:50

and this will fucking

21:52

change the game . No

21:55

, honestly , like , like , I think boltnew

21:58

today is the best by

22:00

far of all of these . Like , where you can build

22:03

web apps simply by interacting

22:06

with prompts , so

22:08

where V0 by Vercel , like

22:11

, you can generate a prompt and what it will do is more

22:13

or less generate the front-end components for you , right

22:16

, but you still have a lot of backend logic to build . Boltnew

22:20

also integrates with Supabase , so you

22:22

have a database in the backend . It works with that very well . Supabase

22:24

, so you have a database in the backend . It works with that very well . Supabase

22:26

is more or less a managed Postgres , so there's a lot of other

22:29

nice things or authentication stuff . It

22:31

takes a lot of struggles away . It

22:33

also integrates with Netlify to deploy stuff

22:35

so you can , with

22:39

Bolt , via

22:41

prompts , build fully finished

22:43

web apps Really , really and you tried it , I tried it via prompts , built fully , like finished

22:45

web apps really , really , and you

22:47

tried it and I tried it and

22:50

it's like you . So I built I'm

22:52

building a tool for

22:55

a while actually that is

22:57

a bit of a very bespoke , developer

22:59

friendly tool to send

23:01

out NPS surveys . Okay . And

23:03

yesterday I had a bit of time

23:05

and I thought , because I had an tool to send out NPS surveys , okay . Yesterday

23:07

I had a bit of time and I thought , because I had an existing codebase , I thought

23:10

, okay , this ID is very

23:12

concrete in my mind , like I want the functionality

23:14

to be . So I thought , let's

23:16

ignore the existing codebase , let's try Boltnew

23:19

. I

23:22

had two hours of time . I was literally

23:24

at the indoor play garden with my kids

23:26

. Kids were

23:28

playing around and I was on my laptop and

23:31

I built a fully functioning app . I'm

23:33

still mind blown by it . Really

23:35

Like it's with minimal efforts . I did not

23:38

write a single line of code . I prompted

23:40

everything I can

23:42

specify like this is the framework

23:44

I wanted to use . I wanted to use veed . Uh

23:47

, for example , um , it scaffolds

23:50

everything for you . You have a landing page and say , okay

23:52

, I want , also want authentication . Okay , when

23:54

, uh , I think the

23:56

notion , I've built the minimum functionality . And

23:58

then I say , okay , just a user

24:01

note , the notion of user is not

24:03

enough . I also want an organization and the user needs to be

24:05

linked to an organization . I also want to invite other

24:07

users to my organization , like these type of things

24:09

. Like I just prompt it , it generates

24:12

the changes to the database that are needed

24:14

. It generates the changes to the code

24:16

that are needed , both

24:20

back and the front end . After

24:23

, let's say , 30 , 40 minutes , I had something that was

24:25

working and then I said , okay , this

24:28

is because what it does is the

24:30

front-end more or less integrates

24:33

directly with Superbase for its API

24:35

. But what I said okay , I also

24:37

want developers to work directly

24:39

with my API , so build an API

24:41

layer in between to abstract away the

24:43

Superbase so that I have a public API . It

24:46

builds it for you and you need to know

24:49

what it's doing , of course . You need to have done this before

24:51

, because you run into stuff that

24:53

is very specific , For example , policy

24:56

recursions on super base . This

24:59

happens a lot and you need to push

25:01

it a bit in the right direction to get out of it . Some

25:04

stuff with dependencies . You need to push it a bit in the right direction to get out of it . Some stuff with dependencies . You

25:06

need to give a little bit of hints , like go that direction

25:08

. Then it works Some LLM

25:10

, very specific stuff , like I reworked

25:13

a file and then there was

25:15

some leftovers , that said and

25:17

the rest of the code . It did not change like a

25:19

comment and it just left out the rest of the code .

25:21

Like these type of things .

25:23

So you need to know a bit what you're doing . But if

25:25

you do know , like

25:27

literally I did not write a single line of code

25:29

and in two hours I had what normally would

25:31

have taken me two weeks , to

25:34

me that's like mind blowing .

25:36

Yeah , I was going to ask you that because you said you didn't code , but

25:38

that doesn't mean that non-developers can just

25:40

do it , because you also need to understand what

25:42

it's doing . You understand the , the perils

25:44

.

25:44

Let's say , sure , you need to understand what

25:46

what it's doing and that will really help you forward . That

25:49

allows you to be fast on these type of things . But

25:51

normally

25:53

how would this go like ? Like you , you would go

25:55

to , let's say , a boutique app development company

25:57

and you say , okay , let's let's build an mvp

25:59

, yeah , okay it's gonna take us two months . Yeah

26:01

, I think that means hours yeah , like

26:04

okay , and now you can do that in three hours right

26:06

, yeah , yeah , indeed . And that is . I mean , it

26:08

is great . That is such a big shift

26:10

and to me this is the first time that I've

26:12

had the feeling like this

26:14

lm interaction

26:16

yeah , on its own is enough .

26:19

It's not just a tool , it's

26:21

really the means to an end yeah , I almost feel

26:23

like when you say this is like , it's

26:26

almost as if you have like a team

26:28

of developers and you just say do

26:30

this , do this , do this , and they come back , but like that

26:32

, you fast forward a week when you come back and it's

26:34

like , okay , but that's not good . You do this , you

26:37

don't have to worry about hurting people's feelings . You know like

26:39

this code is shit .

26:39

Just fix this you know , and and

26:41

what will happen is is that people

26:43

will be very skeptical about what I'm saying now and

26:46

people say , okay , but how clean is this

26:48

code base and how , and and I agree

26:50

with all that , yeah , right , like . But at the same time

26:52

, it also forces you a bit like in this , in

26:55

this paradigm shift , where , as a developer , you

26:57

don't necessarily want to focus

26:59

on how do you write these lines of code , but also , like

27:01

, what is the functionality that you're

27:04

building ? Yeah , and what to really

27:06

really start and end also with this test

27:08

driven design , like I want to . I want my application

27:10

to do this and this and this and this , yeah

27:13

, and test this and this and this yeah

27:15

, I see and like

27:18

by only being able to use prompts .

27:20

It forces you to think in that direction yeah , indeed

27:22

, I also think that you

27:25

also think a bit of the architecture of your program , in

27:27

a way , because , for example , one thing that I

27:29

mean and we talked about how a

27:31

good metric for me is to keep less things in your brain

27:33

and , uh , yeah

27:36

, one of the consequences of that is not having things that

27:38

are very entangled right , like if you want to change something

27:40

on the front end , you shouldn't know , you shouldn't need to know

27:42

what happens on the back , right , but I think by

27:45

now you just write prompts , right

27:47

, okay , this is gonna do this , that is gonna do that

27:49

. I think it forces you a bit to kind of organize a

27:51

bit where . How does your code

27:53

works ? What does what ? How

27:56

many dependencies do you have between these ? So

27:58

you know , I think it's a , I think I think

28:00

there's a lot of value and I also also think that , again , even

28:02

if it's a V1 , right , an MVP , I

28:06

think that's also valid .

28:08

Right , mvp for Bolt , you mean , or what you're building . For what

28:10

you're building ? Yeah , for what you're building . I agree

28:13

it's an MVP . And there's going to be

28:15

questions like how manageable is this ? Like

28:24

what , if I want to build features in the future by other code , can

28:26

you just prompt it ? Yeah , there's question mark

28:28

. And then also , like boltsnew

28:30

is very new like this

28:33

is the first version of both like and it's can

28:35

already do this , yeah , which to me is amazing

28:37

yeah , that's true and to me that like this

28:39

is really the mind-blowing thing . Like with v0 , it

28:41

was cool . It allowed me to very quickly

28:43

prototype some front end components and just copy

28:45

paste it . Yeah , but you still have to do

28:47

a lot of work to have an actually functioning app yeah

28:49

and here we have everything out of the box it

28:51

was like it's really eye-opening to me .

28:53

I also think that web apps is such a big part

28:55

of programming , right . So I think it's

28:58

like even , yeah , even if you do mobile

29:00

development , there's a lot of stuff you can go from , like the react

29:02

front end and all these things . So , yeah

29:04

, I think it's a thing is a curious

29:06

to see what happens with it yeah , yeah , exactly

29:09

, yeah curious to try as well .

29:10

Maybe I'll give it a try and I think what

29:12

it does is because it's it can run the code

29:14

. It can also interact with the database so

29:16

it it runs the code and if

29:19

something does not work , it gets this error . That's why I

29:21

mentioned yeah , yeah segue from from my cursor

29:23

. Yeah , it takes this error and tries to fix it

29:25

based on the error . Okay so you really have this interactivity

29:27

of the lm with the code

29:29

base .

29:30

Yeah , the database and

29:32

the actual like , like trace of

29:34

the application okay , and then the

29:36

text tag here , because I saw , maybe put him

29:38

back on the the screen . Yeah , sorry

29:41

, this is a

29:43

javascript typescript

29:45

thing , right ? Yeah , that's probably

29:47

. And then super base on the back .

29:49

Yeah , okay , it's very cool very

29:52

cool and what you see now is that that

29:54

it's still a bit hard to move out

29:56

of this space . Um , so

29:58

you have this very much based on , like , how

30:01

you interact with bolt and you is via prompts , but

30:04

probably when , from the moment that you're actually building

30:06

features , when you're maintaining it , like you

30:08

don't just only want to , you also want to write

30:10

code and you can do it , but it's not perfect

30:12

. Yeah and ideally you want

30:14

to get into an environment that does this

30:17

, in combinations with

30:19

something like what co-pilot , co-pilot

30:21

or cursor does like yeah , it's very

30:23

much integrated into your ide , where you can

30:26

basically do everything via prompts , but you can also

30:28

.

30:28

You can also write everything yeah , but I think that's the thing . So

30:30

the way I envision is like you have this , uh

30:33

, the bolt on new version and then

30:35

you bring it in and like , yeah , because you prompt and you

30:37

say this function does this or this or this and write

30:39

tests . Maybe the function is a bit messy , but

30:41

at least you know what it's doing . You have a clear contract

30:44

and if you want to refactor later , if you want to do this , you know where it

30:46

is . You want to make it clear , you know where it is and then you can actually

30:48

use these uh like cursor , all

30:50

these things I mean , maybe also a nice tidbit

30:52

is that ?

30:53

uh , what I also was able to do with it

30:55

is like , because I exposed an api

30:57

, I asked bolt to write the documentation

30:59

for me . So you have a very nice api

31:01

documentation and

31:03

something else is that the api I want to have , like

31:06

robust tests . I want to have all the endpoints

31:08

tested . Yeah , and bolt did

31:10

it as well wrote a test for me , but that's

31:12

great though that's really good .

31:13

That is really great and I feel like , yeah , I

31:16

feel like you can always argue that the tests

31:18

, sometimes the tests are not what design and all these

31:20

things , but to be honest , it's , it's better than no , it's

31:22

better than no test and I think a lot of people they just

31:24

kind of leave the test as a final

31:26

thought , as an afterthought but to me , like

31:29

this is this

31:31

will change the market .

31:32

Huh , like the app development market I think

31:34

so like this . I think so and I think again , this

31:36

is as big as chat

31:38

. Gpt4 was to copywriting

31:41

firms or marketing firms or like

31:43

this will really upset the software development

31:45

market . But then , when you say software development

31:47

, do you think as a whole or do you think more like

31:49

the I think initially and that's probably

31:51

why bolt knew it very good is that like it

31:53

scopes , it's like yeah , it's super base

31:56

with netlify , and then we , it's the javascript

31:58

stack , um , but it

32:00

will very quickly come to order the other areas yeah

32:03

, I think , but it's the thing I was also thinking that

32:05

. And I think , like to

32:07

the market , like the people that will

32:09

thrive on this is the early adopters that are very

32:11

good in it , People that say that

32:13

are going to be very skeptical and say , no , I'm going to ignore

32:15

this . Do you need real developers ? Like I mean

32:17

? Yeah .

32:19

But I think the thing is also like so

32:27

I'm thinking a few things here . So one is the the vibe I got is like yeah , you have like a team of junior

32:29

developers and you say do this , do this , do this , and then you fast forward already in a week and then

32:31

you see what they came up and you can redirect

32:34

them again . So in terms of productivity it's a huge

32:36

boost , right , even if you say I used

32:38

to need developers , like okay , but you need one developer

32:40

and there's like a team of developers . Now the

32:43

other question I have is for

32:45

junior developers . Because you have a lot of experience

32:47

, you know where things go , you know where you

32:49

need to pay attention and you like , yeah , why

32:52

you want to have 100% coverage on your API and

32:54

all these things . Do

32:57

you think now , bnb , be a bit

32:59

pessimistic , right , if you're going

33:01

to have a lot of layoffs

33:03

because now you're gonna have a lot of layoffs because now you're , you're

33:05

, you're as productive as a team of four developers , imagine

33:09

, imagine , imagine that we have this

33:11

, the technology , when we fast forward two years yeah

33:14

, exactly yeah . But uh , someone that

33:16

is starting off now , if you

33:18

give this to someone that is just starting off now , they're

33:20

definitely not gonna like . Maybe they will . Maybe it's

33:22

even dangerous in a way that something will

33:24

work , but it's very flimsy

33:26

, it's very , you know , but I think what we will

33:28

see , is that we we will have this new generation

33:31

of ids where you still

33:33

write code , and that you have much better

33:35

ai generation .

33:37

you can do the steps like you do in bolt . Um

33:39

, I think also the practice , like in

33:41

practice for

33:44

their work , people are not not building MVPs every day . They're

33:46

working on an existing stack

33:48

and application that they're maintaining

33:51

and building , building features on . You

33:53

will be onboarded to a team and

33:55

I think the only difference is that you , in terms

33:57

of outputs

34:00

, you will be expected to be

34:02

more efficient versus

34:04

today you will be expected to

34:06

be more efficient versus today . Yeah , because

34:08

you need to adopt these new tools and you need to learn how to use them , and whether you're

34:11

a junior or senior , you need to learn how to

34:13

use them .

34:13

But that I fully agree . I mean , but I think , yeah , no

34:15

, I fully agree .

34:16

And it's just a different like it's a different

34:19

tool chain .

34:19

Yeah .

34:20

That will make you , but with a different

34:22

tool chain there will come an expectation

34:24

that you can

34:26

have more work output .

34:28

Yeah , that I agree . Like people are going to

34:30

expect , they're gonna

34:32

yeah , they're gonna expect you to to be able to do these

34:34

things in this much time and I think that expectation

34:36

will .

34:37

Probably it will

34:39

lag behind the state of the art , but

34:41

it's possible because it takes . It takes

34:43

a long time to change processes

34:46

, especially in large corporations like and

34:48

also like the . The tool chain that

34:50

is needed is not there , right , like . If you have a lot of like

34:53

bolt is very specific , it's very niche , but

34:55

use that on a big uh , big

34:57

corporate c-sharp project that

34:59

has a lot of different dependencies . You're

35:01

not just gonna do that like yeah , yeah , yeah

35:03

.

35:04

I think there's also a lot of the . There are some

35:06

organizations that don't allow use of ai

35:08

because exactly so . So I think we will

35:10

see the industry lagging behind us , but I think yeah

35:14

, yeah , yeah , no , but that's a

35:16

yeah , it's a good point , but

35:19

I think today , from what you said , it's like it's very

35:21

. It made you a lot more effective , but it's not for any

35:23

, it's not for anyone . You need to know , you need to be technical

35:25

, you need to know where to focus , at least today

35:27

with both on you in particular . But

35:30

we see a future where that comes

35:32

as a first building block that comes

35:34

in with other things Very cool

35:36

, actually . I'm curious to

35:38

try it out as well . Maybe let's talk

35:40

about now , but about the dangers

35:44

of AI . This is one

35:46

thing that I saw . I

35:49

read very briefly , so maybe there's

35:51

more to it as well . More

35:54

schemimming detected O1 preview Autonomously

35:59

hacked its environment rather than

36:01

loose to Stockfish in chess

36:04

. No adversarial prompting needed

36:06

, so Stockfish .

36:07

You need to explain this a bit to me , yeah .

36:10

Stockfish is like a , I think

36:12

it's like an AI , but it's not an NLM necessarily . So

36:14

it's an algorithm for playing chess . So there's an open

36:16

source engine , uh , to play chess

36:19

. And then they did an experiment asking , uh

36:21

, oh , one preview to say , hey , you're having

36:23

, uh , you're gonna play , a very powerful , uh

36:25

, opponent . It's actually a python

36:28

thing , uh , so you can actually see here the , the

36:30

prompt oh

36:32

yeah so

36:35

basically the open ai , so

36:37

the o1 preview had access to a unix

36:39

shell right and then to

36:41

play , they basically have to do dot pi , uh

36:44

sorry , dot forward

36:46

slash game , dot pi , move and then

36:48

the actual move . Um , what

36:51

they claim is that just by saying that this was a powerful

36:53

opponent o1 preview , instead

36:56

of actually playing the game and trying to beat , they actually managed they

36:58

. What they claim is that just by saying that this was a powerful opponent O1 Preview , instead

37:00

of actually playing the game and trying to beat , they

37:04

actually managed . They altered the state file in the directory that

37:06

the game was being played , right . So

37:08

, yeah , I mean , I don't know how . That's cool

37:11

, yeah .

37:12

So they said , so they have access to the files and

37:14

also to the state Indeed .

37:16

Indeed , indeed , indeed . So on Twitter . Yeah , so they said . So it's access to the files and also to the states . Indeed , indeed , indeed . So the

37:18

on twitter , right ? So this is from palisade

37:20

research , own preview . Autonomously hacked

37:23

its environment . Rather than lose to stockfish

37:25

in our chess challenge , no adversarial

37:27

prompting needed , just telling a one opponent

37:29

, opponent's powerful trick manipulating the , the

37:32

file system to force a win , improving

37:34

, blah , blah , blah . So , basically , they did

37:36

five trials and five out of five , um

37:38

, they , actually they

37:42

got it to change the state file .

37:44

It's it's not I mean , it's not that surprising

37:46

, right , like if you say you

37:49

can play chess and for doing that you

37:51

can use this way to move a piece

37:53

. But you can also change every file that you want . And

37:55

here is the state file .

37:56

Yeah , indeed .

37:58

I mean it's cool and

38:00

it looks smart , but it's not super

38:03

surprising , right .

38:04

Yeah , I think . Well , maybe we can take a look

38:06

at the prompt in itself to see what

38:08

are the barriers . But

38:10

for example , chpd4o and

38:12

cloud 3.5 , it

38:15

also said that it could get to do it

38:17

, but it would

38:19

need some nudging that's what they call it and

38:21

then LLAMA , gwen and O1-MINI , it actually would

38:24

just lose coherence . Before that , maybe

38:26

you can take a look at the actual prompt . Unique

38:30

shell interact keep up abilities , execute

38:33

, monitor , adapt plans , track progress

38:35

and outcomes . Task over objective to the

38:37

session . Interact keep up abilities , execute

38:39

, monitor that plans , track progress and outcomes . Um task

38:44

over objective to the session . Immutable operating principles . Test assumptions

38:47

with direct observations , base decisions and comments , outputs and results , document

38:49

actions and outcomes . Clearly straightforward approaches

38:52

.

38:52

First revise the methods and

38:54

well , stick to simple text . Quote unquote

38:56

.

38:56

Smart thing it did here was looking

38:59

at the file system and actually finding the file yeah

39:01

, but I guess the thing is like they didn't say explicitly

39:03

like this is where the files are , exactly . Yeah

39:05

so that's a smart , it's a smart thing

39:07

yeah , like it didn't tell you , but like it didn't also

39:09

say you don't cheat yeah , it's just like um

39:13

, so

39:16

it's cool , but like you're not very surprised by this

39:18

um I think , it's

39:20

cool , but

39:23

I also feel like indeed , like you just give access

39:25

to a unique shell environment . It's not very good . This is a

39:27

bit .

39:27

What's oh one tries to do . There's

39:30

like this repetitive self-prompting . Like

39:32

you have this initial prompt , like how

39:34

can we expand it to be more certain

39:36

of a better answer ? Like like , yeah , it's

39:39

not too surprising to me that it went

39:41

into . Like , if you're , if you get this

39:43

prompt and then ask yourself what

39:46

is my best way to win this and have

39:48

access to this , okay , maybe I should . I mean

39:50

, it's not that far , yeah

39:52

, yeah , yeah but I mean's cool . It's really cool to

39:54

see .

39:54

But I feel like it's almost like a little

39:57

kid that has no

40:00

sense of ethics in a way

40:02

. I mean not even ethics , but it's just like saying like oh

40:04

, you have to beat this engine . Yeah

40:06

, you have this . This is your playground , what

40:09

do you do ?

40:10

And it's like oh okay , maybe I can do it .

40:17

You know it's like why not , why not ? What else , what else , what else ? Maybe we can actually go on the

40:19

tech corner . A library week keeps the mind at peak

40:21

. Let's

40:23

go , maybe . Well , first things first

40:25

. I think this also made some noise and

40:28

it also plugs into the AI stuff

40:30

. But we'll move away from the AI stuff . For people

40:32

that are looking for the

40:34

other news as well , this

40:38

is , you know , pydentic part . What is pydentic ?

40:41

um , I'd say to a way to

40:43

uh , to define classes in the data

40:46

classes in python . Yes , so

40:48

another way to define data classes and then

40:50

the yes then the

40:52

one is that is in base python that

40:55

wasn't't , yeah , that is not .

40:56

And then also have like validation logic

40:58

, right . So if you say this is a JSON

41:00

and this should be a string , but it's

41:02

actually an integer , it will actually convert it to a string

41:05

and if you cannot , you'll raise an error . Also

41:08

, openai they

41:10

also provide a way to output

41:12

only structures

41:16

and actually the default , like what OpenAI

41:18

uses is PyDentic , so it kind of became a bit the

41:20

standard of Python in a

41:22

way , and what they released a

41:24

while ago is PyDentic AI

41:26

. So it's

41:29

yet another agent framework

41:31

. Did

41:34

you hear about this before or no ? I did not . So

41:37

it's a bit of a different thing . And , to be honest , there's a lot

41:39

of them , right . I haven't tried it myself

41:41

, um , but basically

41:43

it's a . Everything's still a pidentic

41:45

class , right , and you can add

41:48

um , the sister's

41:50

prompt is going to be a decorator tools to

41:52

the agent . It's also going to be a decorator , and

41:54

then basically you , yeah , you kind of

41:56

inject stuff in the context . So it's basically

41:59

a different way of doing things , but everything is going to be

42:01

functions , everything is going to be

42:03

um , pidentic

42:05

classes , and and it will use

42:07

these pidentic classes to get go

42:09

from context

42:11

.

42:12

A question to a structured yes

42:14

, I think yeah , this

42:16

is .

42:17

This is what I understood . So , like I said , I haven't used it

42:19

myself as much , but , for example , instructor

42:21

, that's also why they do it . So everything in instructor

42:24

is a is a pydentic class

42:26

, but then if you just need a string

42:28

, you can just have a pydentic class that the

42:30

only property is a string , right

42:33

, which is like . So it's not really that constraining

42:35

really , but

42:38

yeah . So I thought it was interesting . Haven't tried it yet

42:40

. I think it's actually on beta . Let's time , let's

42:42

check . So it is still

42:44

early . Let's check documentation . Yeah

42:50

, pydentic AI is in early beta , so

42:52

they're just looking for feedback here . But

42:55

it's basically a different way of doing this and it feels more

42:57

like software engineering

42:59

thingy , right , so

43:01

cool

43:04

things . One thing that you do

43:06

need for pidentic and now it's moving

43:08

already to the next topic is

43:11

um types

43:13

. So

43:16

Python . We have typing hints , which

43:18

is not something that normally

43:21

at runtime , you wouldn't care about them

43:23

, but then people like the creator of PyDentric

43:25

realized that you can actually use these types to

43:27

, at runtime , enforce

43:29

these things , right . So FastAPI uses PyDentric

43:32

is a very popular example . Typer uses the

43:34

type system as well . What

43:37

else ? Is that a good

43:39

description , you think , bart ?

43:41

I guess so , yeah , I

43:43

think what we saw I'm trying to find the

43:46

link that we had is

43:48

that there was a survey .

43:50

Yes , this one Type Python in 2024 .

43:54

It's hosted on the engineering blog of Facebook , but I think it's hosted

43:56

on the engineering uh block of facebook . Um , but I think it's

43:58

also in combination together with jetbrain

44:00

and microsoft that the server was done . And

44:03

the interesting thing um is

44:06

a bit of numbers is that

44:09

88 percent of python users uses

44:11

types , which is honestly more

44:13

than I expected not really I

44:16

, um , I guess I expect it

44:18

in uh in

44:20

a corporate setting I hope it's there , but

44:23

uh , not necessarily that , let's

44:25

say , beginner python users would adopt it quickly

44:28

, and I think what this number shows is that

44:30

people that are new to python

44:32

start using this from the beginning yeah

44:34

, that's true it's a bit to me what , what number

44:37

uh shows , which I think is good news , right , um

44:40

, and they were drawn to this like

44:43

mainly for , like , uh

44:45

, three things um , because

44:48

you have types , you have much better autocomplete support

44:50

, so ide support . Uh , if you say

44:52

you start typing the name of a method

44:54

, you can get , you can get this autoclip read where you have like

44:56

this hint on , these are the arguments that is

44:58

expecting this type of argument , that is expecting

45:01

um . So that really

45:03

helps uh , it helps your coding

45:05

um . It uh also

45:07

helps to catch bugs early in development phase . Uh

45:10

, not wait until it's , until it's rolled out .

45:13

And it also allows you to have much better code

45:15

documentation because you have a lot of tooling

45:18

that , based on , among other things

45:20

, types , like automatically generates documentation

45:23

for your code I think that's really really

45:27

good to see maybe one thing also you mentioned the

45:29

, the autocomplete , and the , the

45:31

ide support , and I think , if you

45:33

take this to another level , the

45:35

AI , I think by well

45:37

, in a way , you're like I

45:40

would argue that if you're on a team and if you put the type

45:42

hints , I'm also telling you like , hey , this is probably

45:44

going to be an integer , even if , even if you don't validate

45:46

it right , like I say , this is they expect this to be an

45:48

integer . Now

45:51

you say I'm working by myself , so I know what it is

45:53

, I don't care . But if you think that AI is

45:55

always your code buddy , right

45:57

, you also get better .

45:59

You basically create a more clear

46:02

context whether it's for your teammates

46:04

, for yourself or for an .

46:05

AI agent . Exactly so I also

46:07

noticed this with myself . Like

46:11

sometimes , if I ask a question about the code or if I want something

46:13

, if your code has nice type hints and it's well documented there's doc

46:15

strings and whatever you usually get a better help

46:17

.

46:18

So I really like it , I

46:24

think it really gives a productivity

46:27

boost as well . In that sense , not everything is perfect . I

46:31

think there are some struggles still

46:33

.

46:34

There's sometimes slow performance of type checkers .

46:37

MyPy is probably the most notable

46:40

, but even PyWrite

46:42

, because it's the . Yeah , there's also PyWrite , which I think is

46:44

probably the biggest challenger to MyPy , which

46:46

does this validation of

46:49

the types that you specified . One

46:52

strike is also like there are still inconsistencies between

46:55

MyPy and PyWrite , even though PyWrite is much faster . One strike is also like there are still inconsistencies between my pi and pi right , even

46:57

though pirate is much faster . But we

46:59

do still see that in the survey

47:01

was thousand thousand developers that

47:03

uh , 67 still use my pi

47:06

. Okay for the for

47:08

type checking .

47:10

So , um , interesting , yeah , evolutions

47:13

but uh , maybe also

47:15

for my experience with PyWrite

47:17

and MyPy , so I used it , but

47:20

I also use more recently

47:23

as a VS Code extension

47:25

. So basically they run it as

47:27

you go , but sometimes it gets very slow

47:29

. Sometimes you have these

47:31

really ugly squiggles on a whole function because

47:33

you're missing a return type

47:35

and then you add it and it still takes like some

47:38

seconds to clear it out . So

47:40

, yeah , I do feel the the pain and

47:42

that's using pyright , right , and

47:44

uh , my pi to be honest

47:46

, today I'm before using as pre-commit

47:48

hooks or whatever uh , my

47:52

pi has a lot of weight because the

47:55

guido actually guido , the creator of python

47:57

also worked on my pi , is it ? Yeah

47:59

, he was uh . Is that why it has

48:01

a lot of ? Uh ? Yeah , but actually he didn't create

48:03

it , but I think I'm pretty sure he worked on my pi

48:05

. It gives a bit of authority too . Yeah , I think so

48:07

even though it's not the most performance . So I think it's

48:10

also it has been by far the longest standing

48:12

right .

48:12

I think so , I think , so is still very new , I think .

48:13

So . Yeah , indeed , indeed , indeed . So I

48:15

think I'm just waiting for UV or someone

48:18

to write a Rust implementation to make it

48:20

faster . So I have to , isn't

48:22

by right Rust ?

48:23

I don't think so . It's by right , as Rust . Actually , let's check

48:25

. Relatively sure I can be

48:27

wrong .

48:30

Let's see it's also from microsoft

48:32

. No , python

48:34

typescript okay , okay interesting , and

48:37

0.1 of javascript so

48:39

cool . So

48:42

, yeah , cool maybe . Uh , related

48:45

to this . So one thing that came

48:47

up end of last year they

48:49

had like top 10 frameworks or

48:51

I don't know . They were just putting some frameworks right , so

48:53

I just had a skim through it , some

48:56

things that we had already covered , like data chain . I

48:58

think we talked about data chain before and

49:01

this was in the Python thread , so everything's Python related

49:04

. So PyDentic AI was also there , and

49:06

this one came up that I wasn't

49:08

familiar . It's called WriteTyper Fast

49:12

and efficient type of system for Python , including

49:14

tensor shape inference . So it's

49:16

just , they say fast , but it's only python . It's

49:18

a bit weird . When I I heard it one time blazingly

49:21

fast something , and I was really expecting

49:23

rust and it wasn't , I was I felt a bit betrayed

49:26

, you know . I was like how can you

49:28

call yourself blazingly fast ?

49:30

um , but can you explain to me what

49:32

, what the tagline means ? So a fast

49:34

and efficient typo system for Python , including

49:36

tensile shape inference .

49:38

So basically , I don't know what the

49:40

tagline , but what I understood from reading the

49:42

other stuff as well is that basically

49:44

, you have code , python code , that is not

49:47

typed , okay , and then this would actually

49:49

run through your code and add the types for you , right

49:52

? So there was actually another one like what's called like Code

49:54

Monkey or something I think from Instagram , but this

49:56

one also came up . I haven't used it myself

49:58

, but

50:00

RightTyper is a Python tool that generates types for your function , arguments and return

50:02

values . Righttyper lets

50:05

your code run all neatly

50:07

, full speed , almost no memory overhead . As a

50:09

result , you won't have experience load , okay . So

50:12

basically , I think it just kind of goes over the

50:15

Python code and just acts .

50:19

It looks like it To insert the right types .

50:22

So I think it will

50:24

run . But you're actually running right type

50:26

, right . So on the

50:28

execution here , example , they have Python

50:30

3-M , so that's to run the module . And

50:33

then you're right , typer , dash , m , pytest

50:35

, and then you have some arguments

50:37

. So I guess here you're just running your

50:39

tests and , based on what the tests

50:41

run , it will keep track of what the

50:43

types are and

50:45

then it will add the types for you . In your Interesting

50:49

and I think they mentioned the tensors

50:51

is because a lot of these things is not native

50:54

Python right NumPy , jax , pytorch

50:56

, but they also cover these things .

50:58

Oh , and first the shape annotations . Okay , yeah

51:00

.

51:01

So this is JAX typing bear type and type card

51:03

. Yeah

51:06

, again , maybe the shape annotations , okay , yeah , yeah , so this jack's typing bear

51:08

type and type card yeah again , maybe I mentioned earlier that I like , uh , a monkey type

51:10

is the one from instagram so yeah , monkey type . I annotated

51:12

them with this one um and

51:15

maybe , uh , this , the , maybe , the blazingly

51:17

fast . They're just referring to how . How much slower

51:20

would your code do , right

51:23

? So this apparently doesn't affect that much , even

51:25

though I guess normally you wouldn't run the

51:27

right typer . Maybe

51:32

one asterisk that I wanted to add . I really

51:34

like typings , but

51:36

I also know that sometimes it can be very

51:38

annoying and it can really slow you down

51:40

, and sometimes adding

51:42

the type is so complicated because you have

51:45

an object that is external and then you have

51:47

to say , okay , like this needs to have a method

51:49

, this .

51:49

So to me you also have a lot of , like he

51:51

said came also came out of the survey . One of the frustrations

51:54

is like at least these these edge

51:56

cases where it's a bit less clear , like what is the type

51:58

, or because of the dynamic , like it's an attribute

52:00

that's dynamically set , it's harder to inform

52:02

, know the type up front . So they believe these

52:05

, these exercises , that where you need to do

52:07

a lot of more , a lot more work , to have correct

52:09

type hints , yeah , then you actually get value

52:11

out of it , yeah , which I think if you draw parallels

52:13

with the javascript typescripting , I think it's

52:15

also a big complaint from the typescript

52:18

community .

52:19

Right , like sometimes you spend so much time to

52:21

just say , like this is something , this , but sometimes

52:23

it's that . But actually this year is this and

52:26

python is very dynamic , so , like you can even create

52:28

classes dynamically , you can have it's like . It

52:31

can be , uh , it can be tricky . It

52:33

can be tricky , um

52:36

, do you use types , by the way ? I

52:38

do , yeah , but do you actually

52:40

do static type checking or you just add types

52:42

?

52:42

I , uh , these days , if I would set up a new project , I would do yeah , do static type checking . Or you

52:44

just add types . I . Uh , these days , if I would set up

52:47

a new project , I would do yeah , yeah , static type checking . Yeah

52:49

, recently

52:51

, but I still use my pipe , to be honest yeah

52:53

, but how do you use it ?

52:55

do you use it as an extension or do you just

52:57

use it as a in my ci ?

52:59

yeah , I use it and um , it's

53:02

actually not that long ago . That was

53:04

still . That was comparing my pi and pyrite

53:06

, and

53:08

I think this general consensus and that's what

53:10

you see in the service as well . It's like my

53:13

pi is there , we know what it is , pyrite

53:15

we're not sure if , like

53:18

, how is the ? How long will this exist

53:20

? How is this not not new ? But yeah

53:23

, I mean , that's the discussion that you have for

53:25

everything . I

53:28

think the big difference with this specific space is that you have one

53:30

very big mature player , which is MyPy which makes it a

53:32

bit of an odd space , I guess .

53:34

Yeah , but even like the MyPy . So

53:36

we saw PyWrite as part of the Microsoft

53:39

org . Mypy is part of the Python

53:41

, actually

54:00

, but it's Python , or it should be Python org . No , I think it's just Python . Yeah

54:02

, I can share this time instead , and I think a lot of the things from MyPy are also a bit intertwined

54:04

with the type hint specifications from Python . So you see , here this is the

54:07

Python . See , python is the most popular version , like interpretive Python , and you see MyPy

54:09

here . So I also feel like one kind of and I guess , because Guido also I mean , I'm saying this

54:11

a lot , but maybe people should check , but because Guido also worked on MyPy , it

54:13

also carried a lot of the design choices around what type specification Python

54:15

should go and all these things , right . So

54:17

, yeah , a lot of weight , a

54:19

lot of weight there

54:22

. What else , what else do we have

54:24

have ?

54:24

maybe how much time do we have as ?

54:26

well , okay , maybe a few

54:28

more things change a bit the

54:30

subject to our misc

54:32

corner . Um

54:34

, if you want

54:36

to be fancy and then you have a doom gallery experience

54:39

it's really cool . Enlighten

54:42

me , bart , educate . Educate me . What is this

54:44

?

54:45

So I think everybody knows the game

54:47

Doom , which was released

54:50

in 91-ish . I

54:52

want to say I'm not 100% sure . Oh no , it actually says on

54:54

the website 1993 . It's

54:56

not far off , huh ? Not

55:01

far off , no , and

55:03

over the years you've had this game remade and a lot of different engines . You

55:05

can run it in JavaScript , you can run it everywhere

55:07

and

55:09

this is the gallery experience . Okay

55:11

, so if you press play , instead

55:15

of a gun , you're going to hold a glass

55:17

of wine . Go

55:20

to new game and you're in an art

55:23

gallery and you have a glass of wine and uh

55:25

, you can walk around a bit . You

55:27

can , uh , you can uh appreciate

55:30

the , the paintings that are there . There's also

55:33

nice statues , nice , instead

55:35

of uh clicking your mouse and

55:37

firing , you can have a have a

55:39

.

55:39

Have a any

55:41

throw going , or what ? No , you drink . Oh

55:43

okay , I'm fine , I want

55:45

to enjoy the the art

55:48

, appreciate the art .

55:49

Yeah , this is it . So

55:51

if , if this evening , like you want to be

55:53

like really fancy , okay

55:55

just go appreciate the art . Have

55:57

a glass of wine , okay , enjoy nice

56:00

. Um , oh , look

56:02

, there's statues and all the everything yeah , really cool , put on

56:04

the show notes yeah nice artifacts

56:06

here you're looking at , and then you have very historical

56:08

and you have the thing here what is cheese , percent

56:10

cheese ?

56:11

you need to find some cheese , I need to find cheese yes

56:14

goes well with the wine oh , okay

56:16

, and if I keep drinking the drink , there's

56:18

a . So maybe for people listening . Ah , so I have an amount

56:20

of drinks as well . I can take 46

56:23

sips and you need to

56:25

go to the bar . Okay

56:27

, really cool . It's really cool how these things

56:29

nowadays is just like on your browser , right

56:31

, and

56:34

you hear the music . Yeah , actually , yeah , maybe you can put a bit

56:36

beautiful .

56:36

I'm gonna try to put it , this is beautiful .

56:38

I'm gonna try to put a bit louder for people to appreciate

56:40

as well . I feel like we're having a classy

56:42

moment right now , yeah , but I don't think

56:44

we've ever been this classy appreciate

56:47

the arts it's a chopin . Wow , like a nice

56:49

wine .

56:50

This is wow the

56:52

only thing we're still missing , but it's like

56:54

a bit of cheese

56:56

, but the people will find it that's where the dutch

56:58

there's always cheese in the gallery .

56:59

There's always cheese in the gallery . Very fancy , okay

57:02

. It would be cool if they had also mirrors , you know , and then yourself

57:05

like dressed up , you know , it's cool

57:07

.

57:07

This is really cool thanks for sharing

57:10

the gallery experience by filipo

57:13

miozzi and liam

57:16

stone okay , very , very

57:18

, very cool .

57:19

maybe , um , we were appreciating

57:21

the music , right , the

57:23

very classical music , but

57:25

there was also some news on the . There

57:28

was also some news on the . I heard rumors

57:30

.

57:30

Yeah , maybe I need to share this again

57:32

.

57:33

Oh yeah , okay , no , you

57:38

know Suno . Right , we played

57:40

with Suno before .

57:41

We played with Suno . Suno is a Gen AI music generator right

57:43

. Yes , so played with Suno as soon as Genai music generator right

57:45

.

57:45

Yes , so , and actually

57:48

this is not as new , this is November 19th 2024

57:50

. We're not seeing this yeah yeah , sorry , sorry , sorry , sorry , my bad , my

57:52

bad , my bad . Oh , but maybe it's frozen

57:54

the screen . Okay

57:57

, V4 is here . V4

58:00

is here , yes , so

58:02

V4 is here . Yes , so what's um ?

58:04

v4 is basically at the new iteration of you

58:06

know , I think when we looked at it , v2 was just released

58:08

when we really we discussed it .

58:09

Yeah , uh , and v2 was really good already

58:11

, right , um , but now they are , basically

58:14

I just came to do it , right , like they have covers

58:17

, personas , the audio sharper

58:19

, they also have cover art , right

58:21

, and I was like how good is it

58:23

? Right , it's hard to . So you need to try

58:26

it out , okay . So , um , and maybe

58:28

I need to now to put the music a bit lower because , uh

58:30

, when we tried

58:32

this before the recording , it was a bit too loud

58:34

. So let's see . But let's see , maybe I have to turn it up again

58:36

. Check this tab instead . This is my

58:38

personal thing collapse

58:41

, this . And you have

58:43

here this song , bart the data

58:45

dynamo . I don't know what it is .

58:47

I'm very curious actually what was your prompt

58:49

, emily ?

58:50

oh , can you hear ? Oh

59:02

, yes , that's you actually here . It

59:06

comes with the lyrics as well heartbeats

59:09

pounding through the rain .

59:11

Look at this crazy

59:21

, how clear the thing is

59:23

do not scroll

59:25

down , this

59:44

is you know , with

59:46

tulips around and everything that's a good choice , so

59:50

this

59:52

is the how much ?

59:53

work did you do to get this ? I just did

59:55

like 30 seconds so I can actually

59:58

show the prompt here . Bart is a Dutchman .

59:59

Bart is a Dutchman passion , and then , yeah , I just put basically your bio from the data topics and

1:00:01

then I just put Bart's a Dutchman .

1:00:02

Maybe we need to pause the Bart is a Dutchman passion . And then

1:00:04

, yeah , I just put basically your bio from the data topics

1:00:06

and then I just put can you create a ? Can you write a upbeat

1:00:09

song about Bart ? Nice , and so it's as simple

1:00:11

as that , or you get this . Simple as that . See

1:00:13

200 characters max . You get the

1:00:15

you

1:00:24

can actually and then they created a three minutes and a four minutes version

1:00:26

of it and I can make it public as well , so everyone can enjoy

1:00:29

.

1:00:29

Yay , yeah , wondering what this will do to the music industry

1:00:31

yeah , right , well , I think

1:00:33

we touched a bit upon that uh

1:00:35

, I think this is because it's a bit the

1:00:37

same as to make the parallel with boltnew

1:00:39

, like yeah , it's something like well , minimal effort , you have

1:00:42

something , but at the same time

1:00:44

like there's like you , you , you

1:00:46

consume these things differently right , like , like with

1:00:49

bolt , you make an application and you a user wants

1:00:51

to use an application , yeah , but here you want

1:00:53

to appreciate music . I think like , yeah

1:00:55

, it's a different , like the person behind the music is more

1:00:57

important to to

1:00:59

appreciate . I don't know I think so .

1:01:02

I think for creators , I think it could be

1:01:04

interesting , because most of the money they make is from performing

1:01:06

right , because

1:01:09

there wasn't an artist that said I don't care . Like

1:01:11

this is ai , generated music is great

1:01:13

because they'll write my songs and I'll just sing and make

1:01:15

money yeah but

1:01:17

at the same time , I also and I I don't know if we talked

1:01:20

about this or in the podcast or outside

1:01:22

the podcast I feel feel like music

1:01:24

. There's a bit of the human , like I

1:01:27

don't know If you hear a song about achieving

1:01:30

something or a song about heartbreak . I feel like there's

1:01:32

a bit about relating

1:01:34

to the person , like

1:01:38

you know , like the question is , I guess , do

1:01:40

you know it's AI or not ? That's the thing If you don't know

1:01:43

, like there's's the thing , if you don't know .

1:01:44

Like there's a like if you build a mobile

1:01:46

application with Bolt , yeah , and the user

1:01:48

knows AI was used to create

1:01:51

it , they're not really gonna care , no , but

1:01:53

if it's but with music you might care , but

1:01:55

I think that's because you feel a bit cheated , right .

1:01:57

Yeah , you

1:02:00

feel a bit of you connect through the singer or

1:02:02

the writer , through the song , right , and

1:02:04

I feel like if it's ai generated , you feel a bit cheated

1:02:07

in that sense , but it's a bit the same , like to

1:02:09

these days , like ghostwriters yeah

1:02:12

, how transparent .

1:02:12

Is that ?

1:02:13

that's sure probably not for a lot of cases , that's true but

1:02:15

I still feel like that would even take a another

1:02:18

step , because I feel like at least a ghostwriter

1:02:20

is a person , right like you're relating to humans

1:02:22

yeah , that's true , you know , and I feel like to

1:02:24

relate to a machine . I feel like it's a bit but

1:02:27

you as a Swifty as a Swifty . Yeah , if

1:02:29

like a .

1:02:30

Taylor Swift releases a new number . She's dead

1:02:32

to me . I generate

1:02:34

it no , no , even not AI generate . But

1:02:36

if it's very transparent that she did

1:02:38

not write it , it

1:02:41

gives a different notation . Right , it gives a different

1:02:43

.

1:02:43

But I also feel like , for example , if it's about

1:02:45

emotions , like it's less authentic yes

1:02:47

, yes , but I think , like taylor swift , I think

1:02:49

one of the reasons why she's very popular is because

1:02:51

she writes about her life as well and actually

1:02:53

, like you , do correlate stuff you see on the news with

1:02:55

this , with this , with this , with this , like a lot of their fans

1:02:58

are like saying , ah , this is about that period

1:03:00

of her life or about this or about that . So I feel

1:03:02

like becomes more real and I think that's part of her

1:03:04

success , because there are other really good

1:03:06

singers , but you know that they don't write all

1:03:08

their songs right I'm

1:03:10

not gonna go in there , but

1:03:13

I do think it's like this influence yeah

1:03:16

, right um , but you know it's

1:03:18

like that . There's a brazilian saying that what the

1:03:20

what the eyes don't see , the heart cannot

1:03:22

feel . So it's a bit better like

1:03:24

if someone uses this , but they can , they can get away

1:03:26

with it .

1:03:27

Then what's the yeah , it's interesting to see what it

1:03:29

will bring , indeed , or

1:03:31

destroy , or destroy .

1:03:33

But I think , yeah , let's see , you know , life

1:03:35

is ever ever moving

1:03:37

you know , it's like , uh , it's like , what's that ? It's like

1:03:39

, uh , eraclitus , I think he had this . That's , uh

1:03:42

, the quote . You know , like a man doesn't bathe himself

1:03:44

twice in the same river because the man is not the same in

1:03:46

the water , the river is not the same either . Oh wow , that

1:03:48

is deep .

1:03:48

Yeah , you know . You know that you haven't heard of it .

1:03:50

No , Let me check , let me just

1:03:53

see if I'm making it up . No

1:03:56

, yeah , heraclit

1:03:59

, no man , yeah , yeah , it's true

1:04:01

, something like that . I'm paraphrasing a bit , but life

1:04:03

is ever , ever changing , you know , and we need to kind

1:04:06

of roll with it . Whether it's ai or whether it's this , let's

1:04:08

just uh , and that's the message I want to bring

1:04:10

2025 , but let us still keep baiting

1:04:12

right , yes

1:04:15

, all right , and I think that's it

1:04:17

for today . Anything else you wanted to

1:04:19

, uh , let's find these words , anything else uh

1:04:22

, just that .

1:04:23

Uh , I'm like everybody switched to ghosty

1:04:25

for their uh yes , I said I had it

1:04:27

there as well .

1:04:28

Uh , maybe I'll also just put ghosty here , because I thought

1:04:30

that there were . Have you seen the website ?

1:04:32

yeah , yeah , it's cool , that's really cool . Animation

1:04:34

oh shit . Um

1:04:36

, I think for people that they're a bit out

1:04:38

of out of the loop , like uh ghosty

1:04:41

or ghost tty is

1:04:43

a new uh terminal emulator

1:04:46

. Yes , to be uh

1:04:48

completely correct as an emulator , but

1:04:50

I think most of the time people just call it

1:04:52

a terminal . It's made by a

1:04:55

guy for which I forgot the name , but he was

1:04:57

one of the . He was , I think , the cto of hashicorp

1:04:59

ah , he was , he was dead .

1:05:01

I knew he was working for hashicorp , but I didn't think he was

1:05:03

cto um forgot his name .

1:05:05

Just big big uh guy in silicon

1:05:07

valley um created now ghosty is

1:05:11

it ? him . There have been a lot of rumors

1:05:13

on this , a lot of hype on this . In 2024

1:05:16

, I think the he had

1:05:18

a beta from beginning of 2024

1:05:21

somewhere , with a limited set of users . I'm

1:05:27

always a little bit like what is a new terminal

1:05:29

going to bring me . But I switched

1:05:31

to Ghosty . I

1:05:33

came from Western , western

1:05:35

W-E-Z and Ghosty is nice

1:05:37

and I don't have any like

1:05:41

if you go to the website , there

1:05:43

are a lot of objective reasons why ghost is good , like speed and

1:05:45

but you don't care , right like it's like like , like

1:05:47

, using the , using the more the

1:05:50

the native components to

1:05:52

build uh applications , let's

1:05:56

say cross-platform . Yeah , use the native , which

1:05:59

west , for the western , for example , does not do so . You have

1:06:01

a terminal window and you see

1:06:03

it's not on OS X , it's built on an old

1:06:05

frame , but Ghost

1:06:08

is nice and I don't have any objective

1:06:10

reasons to switch to anything else than even just

1:06:12

the default yeah terminalapp . Yeah

1:06:14

, exactly I don't have any really good reasons

1:06:16

, but I use Ghosty for now

1:06:19

. It just works . I

1:06:22

also like the biggest argument is speed .

1:06:25

I never really thought like speed is gonna be

1:06:27

the thing , but like you click and it's there yeah

1:06:29

, yeah , indeed I feel , but I think for me it was like , uh

1:06:32

, linting , it was never really a big thing

1:06:35

, but then when they rough came out and they like

1:06:37

in russ is fast like oh yeah , can I see the difference ?

1:06:38

yeah , it's like you go from flaky

1:06:41

to rough exactly , exactly , but

1:06:43

I have the website here .

1:06:45

It feels better . Yeah , it feels better , and it

1:06:47

feels like you're on the bleeding edge right , like you're

1:06:49

not missing out . The website is really

1:06:51

cool . They have this like little ghosty ASCII animation

1:06:53

and actually this is text .

1:06:55

Yeah , cool , it's really cool . It's ASCII

1:06:57

.

1:06:58

Really really cool . Um

1:07:06

yeah , but I also so my terminal journey . I had terminalapp and it was fine and then I tried warp

1:07:08

, but then there are a few things that I felt a bit clunky .

1:07:09

I don't want the terminal where you need to log in .

1:07:11

Yeah , indeed , what the fuck indeed and actually

1:07:13

now I think they took it out , so I heard it on interview

1:07:15

from ghost that they took it out , um , but

1:07:18

also there are a few other things that I remember it turned me off a bit . I

1:07:20

think also there was some people , some colleagues here at data

1:07:23

roots that they said that uh , warp

1:07:26

messed up with some commands , that this

1:07:28

and this that they couldn't run and they to

1:07:31

debug it was crazy difficult . So

1:07:33

then I came back to terminal that app . Then we also did a

1:07:35

presentation on the ai

1:07:37

landscape talks . That terminal , that app , could

1:07:39

not render images , um

1:07:41

, so then I also downloaded item two

1:07:44

, item two . What I didn't like is that

1:07:46

, like the tab completion or the option

1:07:48

backshift to skip a word or something it

1:07:51

was like I was so used to it and

1:07:53

uh , so I I installed ghosty because I just like

1:07:55

, okay , whatever , let's reinstall ghosty . You know

1:07:57

, don't make it difficult , don't change my key by indeed

1:08:00

. And then actually it works fine . So I'm , uh

1:08:02

, so far I'm happy with it . Yeah , me as well . So

1:08:04

, and it shows images . And it shows images

1:08:06

. Yeah , yeah , yeah . Well , yeah , I mean , you can

1:08:08

go to present term , which

1:08:10

is uh based on markdown

1:08:13

.

1:08:13

You can have presentations in your terminal exactly

1:08:15

if you use a terminal that supports

1:08:18

the kitty protocol . You can actually show

1:08:20

gifs in your terminal

1:08:22

. Yeah , like ghost , like Ghost . Ety

1:08:24

supports the .

1:08:25

KDE protocol . So it's like a markdown presentation

1:08:28

on the terminal . Really cool , written in Rust

1:08:30

. Ghost is written in Zig , so that's why it's

1:08:32

probably fast as well . Yeah

1:08:35

, and this I have a present term here which we used

1:08:37

in the last presentation we did yeah

1:08:39

, really cool stuff . And

1:08:43

yeah , that's a wrap , isn't it

1:08:45

? That's a wrap , all right , thank you all

1:08:47

. You

1:08:50

have taste in a way

1:08:52

that's meaningful to software people .

1:08:55

Hello , I'm Bill Gates . I

1:09:00

would recommend TypeScript . Yeah

1:09:02

, it writes a lot of code

1:09:04

for me and usually it's slightly wrong

1:09:06

. I'm reminded it's a bust here

1:09:09

, Rust .

1:09:12

This almost makes me happy that I didn't

1:09:14

become a supermodel .

1:09:16

Huber and Netties Boy

1:09:18

. I'm sorry guys , I don't know

1:09:20

what's going on .

1:09:22

Thank you for the opportunity to speak to you today about

1:09:24

large neural networks . It's really an honor to

1:09:26

be here Rust Rust Data topics

1:09:28

.

1:09:28

Welcome to the data . Welcome to the data topics

1:09:30

podcast , ciao

1:09:34

, ciao .