Download "Проанализировал комментарии НЕЙРОСЕТЬЮ. И вот, что я узнал"

"videoThumbnail Проанализировал комментарии НЕЙРОСЕТЬЮ. И вот, что я узнал

Дождались ФОТОШОП в НЕЙРОСЕТИ Stable Diffusion! Photopea для a1111

Дождались ФОТОШОП в НЕЙРОСЕТИ Stable Diffusion! Photopea для a1111

Channel: Cg_Stas

Я ТОРГОВАЛ 6 ДНЕЙ С Помощью НОВОЙ НЕЙРОСЕТИ На ByBit И вот Что из этого вышло

Я ТОРГОВАЛ 6 ДНЕЙ С Помощью НОВОЙ НЕЙРОСЕТИ На ByBit И вот Что из этого вышло

Channel: NeuroArs

Ботоферма эльфов, Маск едет в Москву? Open AI уволила Сэма Альтмана. Климарев, Кушелев, Плющев

Ботоферма эльфов, Маск едет в Москву? Open AI уволила Сэма Альтмана. Климарев, Кушелев, Плющев

Channel: The Breakfast Show

СДЕЛАЛ ИГРУ ПОЛНОСТЬЮ С ПОМОЩЬЮ GPT-4 НА ЮНИТИ

СДЕЛАЛ ИГРУ ПОЛНОСТЬЮ С ПОМОЩЬЮ GPT-4 НА ЮНИТИ

ТОП нейросети для создания ВИДЕО и МУЗЫКИ

ТОП нейросети для создания ВИДЕО и МУЗЫКИ

Channel: Web3nity

Презентация OpenAI: Новая Умнейшая Модель GPT-4 Omni | Переведено и озвучено людьми!

Презентация OpenAI: Новая Умнейшая Модель GPT-4 Omni | Переведено и озвучено людьми!

Channel: Тест Тьюринга

Как ИИ-компании победили Google, Apple и Amazon | Искусственный интеллект: путь к завоеванию мира

Как ИИ-компании победили Google, Apple и Amazon | Искусственный интеллект: путь к завоеванию мира

Channel: PRO РОБОТОВ

CHATGPT - Как ЗАРАБОТАТЬ на Нейросети? Почему это ЛУЧШИЙ Инструмент

CHATGPT - Как ЗАРАБОТАТЬ на Нейросети? Почему это ЛУЧШИЙ Инструмент

Channel: ABRAMOV

Что такое нейросети, и кто из-за них может лишиться работы

Что такое нейросети, и кто из-за них может лишиться работы

Channel: 1на1

нейросеть

нейросети

gpt chat

chatgpt

эмбеддинги

машинное обучение

onigiri

онигири

анализ данных

большие данные

биг дата

big data

проанализировал комментрании

искусственный интеллект

ии

ai

визуализация нейросети

openai

00:00:02

I studied the statistics of comments on

00:00:04

YouTube, for this I downloaded 10

00:00:06

million comments from YouTube and analyzed

00:00:08

them using neural networks and many other

00:00:10

methods. In short, the best

00:00:13

chance of getting likes is for a comment

00:00:15

about 100 characters long that is written in

00:00:17

about 8 minutes after

00:00:19

the release of the video So if you set

00:00:21

the bell and immediately saw this video,

00:00:23

then you have a whole 8 minutes and you know

00:00:25

what to do And also with the help of neural networks I

00:00:28

built these clouds of comments

00:00:29

Where comments that are close in meaning

00:00:31

are nearby and they can be divided into

00:00:33

groups in which there are comments with

00:00:35

approximately the same meaning.

00:00:37

At first I thought that for my

00:00:38

statistics I would download somewhere around 100,000

00:00:40

comments or maybe even a million,

00:00:42

and then I looked at how you can

00:00:45

officially download comments from YouTube and it

00:00:47

says that you can make 10,000 requests per day.

00:00:48

yes it turns out If I

00:00:51

type only 10,000 comments per day,

00:00:53

then for 100,000 comments it will take 10 days

00:00:56

and for a million comments it will take

00:00:57

more than three months and I will have to

00:01:00

download these comments every day, but then it

00:01:02

turned out that one request produces

00:01:04

approximately 100 comments and if

00:01:07

you multiply 100 comments by 10 thousand

00:01:08

requests per day, it turns out that per

00:01:10

day we can receive a million

00:01:12

comments, even a little more

00:01:13

because on average it produces a little more than 100

00:01:15

comments per request, I downloaded all the

00:01:17

comments from almost a hundred channels, many

00:01:19

of these channels are similar in topic to

00:01:21

mine, but there are others I took some channels

00:01:23

simply from YouTube recommendations so that

00:01:25

the topics sometimes differed and there was a

00:01:27

greater variety of comments and

00:01:29

Let's see in what form YouTube

00:01:31

issues comments, here is a table with all the

00:01:33

properties of comments, firstly they

00:01:35

have a type of YouTube object, in this

00:01:37

case everyone will just be a comment,

00:01:39

then there is some other information

00:01:41

that I won’t need here, I

00:01:43

marked it all in red, which means that

00:01:45

I don’t save this data. What I save is, for

00:01:47

example, the ID of the comment itself, that

00:01:49

is, its unique number, I can

00:01:51

distinguish between comments and then the ID of the video

00:01:53

to which this was made comment and

00:01:55

then all sorts of information that I

00:01:57

don’t really need. But what I need is the

00:01:59

text of two comments and the display name of the

00:02:02

author of the comment, also the channel ID of the author of the comment, the

00:02:04

number of likes and when

00:02:07

the comment was published, for

00:02:09

example, there is such a field as is it possible to

00:02:10

evaluate comments and it everywhere True

00:02:12

I don’t know at all. Could there be a situation

00:02:14

where you can’t rate comments, that is, you ca

00:02:16

n’t give them a like? And here

00:02:18

the rating of the comment is stored; in fact, it

00:02:20

just means whether we liked it or

00:02:21

not; the number of

00:02:23

replies to comments and the replies themselves,

00:02:25

which are essentially more, are also stored. in one exactly the

00:02:27

same table in this table I

00:02:30

mentioned the display name of the author of the

00:02:32

comment And when this name is downloaded

00:02:34

it is displayed in the normal form in

00:02:36

which I am used to seeing names in

00:02:37

comments, new ones may have noticed that

00:02:39

instead of the usual names now in the

00:02:41

comments there are these tags and many

00:02:42

people I didn’t install them, it says User and there are

00:02:45

a lot of letters and numbers I at first thought

00:02:47

it was a bug but maybe it’s a feature but one way or

00:02:49

another in the downloaded comments the names are

00:02:51

still normal In this video I will show

00:02:53

different algorithms and approaches from Data Sans

00:02:55

but so far I’ve only just I’m studying all this.

00:02:57

Therefore, I’m not an expert in this area,

00:02:59

but I know for sure that the demand for data

00:03:01

scientists is now frantic and

00:03:03

they earn on average 200 thousand

00:03:05

rubles and there is no ceiling on income, so if

00:03:07

you also want to work from anywhere in

00:03:09

the world, make good money and be in

00:03:11

great demand among large companies I

00:03:13

advise you to go study at skillfactory on the

00:03:15

data sentis course from scratch to pro, which was

00:03:17

developed jointly with a professor from

00:03:19

Moscow State University. As part of the course, you will learn the

00:03:22

very basics from scratch, including remembering

00:03:23

school mathematics and mastering the

00:03:25

advanced level necessary to

00:03:27

work as a data scientist at skillfactory this is an

00:03:30

online school that gives you not just

00:03:31

training, but guides you by the hand into the world of it,

00:03:34

training on real tasks from

00:03:36

customers at the end of the course gives you

00:03:38

a portfolio that stands out favorably and elevates

00:03:40

you in the list of candidates, especially since you

00:03:42

choose the time and pace of training yourself and,

00:03:44

if necessary, you can even freeze the

00:03:46

course by resuming it after a

00:03:48

while so as not to abandon it halfway,

00:03:50

there will be a mentor next to you who

00:03:52

will answer all your questions and the career center

00:03:54

will help with employment. And if after

00:03:57

training you do not find a job, the school

00:03:59

will return the money and this is stated in the contract.

00:04:01

So sign up for the course by the QR code

00:04:03

on the screen or by the link in the description and

00:04:05

start a career in datascience now and

00:04:08

with the promotional code onigiri you will receive a 45

00:04:10

percent discount on training skillfactory teaches

00:04:13

those who work I decided to start studying

00:04:15

comment statistics with something

00:04:17

simple, let's see for example how long are

00:04:19

comments under videos usually For example,

00:04:21

if I take my channel and plot the

00:04:23

distribution of the length of comments, I

00:04:25

will get a graph like this: on the X axis

00:04:27

there are different long comments, and

00:04:30

on the Y axis the number of comments of this

00:04:32

length, for example, comments of less

00:04:34

than 10 characters, is quite small. But the

00:04:37

more characters, the more the

00:04:38

number of comments grows. and then

00:04:40

such a maximum of the graph is formed

00:04:42

for approximately long comments from 40

00:04:44

to 60, here there are the most comments and

00:04:47

then the number of comments gradually

00:04:49

decreases, that is, there are much fewer comments with a length of 100

00:04:51

and comments with a length of 200

00:04:53

Even less than 300 Even less than 400 Even less

00:04:55

and 500 I have displayed in at the end of

00:04:57

the graph, in general, the maximum length of a

00:04:59

comment on YouTube is 10 thousand

00:05:01

characters, but there the graph will only

00:05:03

gradually decrease and decrease, so

00:05:04

I did not draw it further. And what’s

00:05:06

interesting is that each channel has a

00:05:08

slightly different graph. For example, the same

00:05:10

graph for Makar Light starts

00:05:13

very similar they are almost the same, but

00:05:15

then he goes a little higher. Which means

00:05:17

that he has a little more

00:05:19

longer comments. But here is the graph for

00:05:21

Dagon, you can see that at the beginning he has a lot

00:05:23

more comments, but then the graph

00:05:25

compares roughly with my channel. Which

00:05:28

means that he has more short

00:05:29

comments and a little less long, and

00:05:32

since on some channels there are

00:05:33

shorter comments on average, and on some

00:05:35

longer ones we can sort the

00:05:37

channels by this parameter. Here, for example, I

00:05:40

took from my sample popular science channels

00:05:42

or channels related to them in topic and

00:05:45

sorted them their medians for her

00:05:46

comments are in the first places

00:05:48

Anthropogenesis Reviews and Phelologist of All

00:05:51

Russia And on the last topless Audi Hot and

00:05:53

Utopia shows a long comment most

00:05:56

likely depends on many factors,

00:05:57

for example, it seems to me that they are strongly all

00:06:00

this middle-aged audience and

00:06:02

therefore if there are cholevar topics or Is

00:06:04

it possible to write a sheet of text in

00:06:06

response to a phrase from the video? Well, I

00:06:08

showed the distribution of

00:06:09

comment lengths on my channel on the

00:06:11

Makar of Light and Dagon channel. But what if

00:06:13

we consider such a general distribution across

00:06:15

all channels and here the graph turns out to be

00:06:17

much more even due to the large

00:06:19

number of comments but you can immediately

00:06:21

pay attention to this strange

00:06:23

detail, from somewhere a lot of

00:06:25

comments with a length of 142 characters came from, I started

00:06:28

looking for what it was and it turned out that these were

00:06:30

several tens of thousands of comments,

00:06:31

all these comments look about the

00:06:33

same Hello, please

00:06:35

watch my videos, I’m not asking you

00:06:37

to subscribe, you just need an opinion from

00:06:38

sides and so on And what’s interesting is that I

00:06:41

found more than a thousand accounts that

00:06:43

write these messages and this is only from

00:06:45

what I found in my statistics. And

00:06:47

I might not have noticed something and something might

00:06:49

not have been included in my statistics at all, that is,

00:06:51

someone- I somehow registered or

00:06:53

somehow got more than a thousand

00:06:55

accounts on YouTube and in all these

00:06:57

accounts there is the same request to

00:06:59

switch to another account. That is, all

00:07:01

these more than 1000 accounts lead to

00:07:04

one single account and by the way it is

00:07:06

blocked for violating the rules YouTube,

00:07:09

that is, what it was, it was some kind of

00:07:11

advertising or it was some kind of experiment,

00:07:13

now we most likely won’t know. But

00:07:15

let’s move on, I decided to take a look.

00:07:17

What do the most popular

00:07:19

comments under the video have in common? Firstly, the

00:07:22

largest number of likes in in my

00:07:23

statistics I got comments I want an

00:07:25

episode about sleeping under a video on the

00:07:27

topless channel in general, usually the most popular

00:07:29

comments are either supportive or

00:07:31

quoting something from the video and almost

00:07:33

always the most popular

00:07:35

comment on the channel is quite ordinary

00:07:36

except for my channel I have this in some at a

00:07:39

certain point, I began to realize

00:07:41

that this video may never

00:07:43

end and Yes, these are the comments under the

00:07:45

video with the largest numbers, firstly,

00:07:47

the video has the largest numbers, it has the

00:07:49

largest number of views on my channel,

00:07:51

and secondly, the comments under this video

00:07:53

have received the largest number of likes A

00:07:56

I also had a video making a neural network from scratch

00:07:58

where I showed how a neural network can

00:08:00

recognize handwritten numbers and the

00:08:02

comment under this video is that alarmists

00:08:04

are afraid that neural networks will get out of

00:08:06

control and enslave humanity. The

00:08:07

neural network thinks that the black screen is 5.

00:08:10

A. It’s funny that the comments are about the number 5

00:08:12

is the fifth largest number of likes on

00:08:15

my channel, then I thought, so I made a

00:08:17

distribution of comment lengths. But what

00:08:19

if comments of a certain

00:08:21

length get more likes? Well, for example,

00:08:24

if a comment consists of one letter, it is

00:08:25

unlikely that anyone will like it, or

00:08:28

if it is a sheet of text It’s also unlikely that

00:08:30

anyone will like it. But if

00:08:31

the comments are easy to read, most likely

00:08:33

then they will like it and probably such a

00:08:35

comment will not be too long and

00:08:37

not too short. If you build the

00:08:39

same distribution as I did only for

00:08:41

likes, then of course there are more

00:08:43

there will be more likes in the comments,

00:08:44

so of course you need to take this into account and

00:08:46

divide the number of likes by the number of

00:08:48

comments for each length of

00:08:49

the comment, I tried to make this

00:08:52

graph and this is what it turned out in general.

00:08:54

Looks like she’s just random numbers, as if there is

00:08:56

no dependence here, but

00:08:58

when I I made this graph in my

00:09:00

statistics there were only 100 thousand

00:09:02

comments And now I already have

00:09:04

much more Therefore the graph can be

00:09:06

made much smoother I made a

00:09:08

new graph for 10 million

00:09:09

comments and this is how it looks and

00:09:12

now it is clearly visible that if

00:09:13

the comment is too short then

00:09:15

it has few likes; if it is longer, then there are more of them and

00:09:18

then they gradually subside. Yes, there are

00:09:20

very strong fluctuations and to

00:09:22

remove them and make the graph truly

00:09:23

smooth you probably need 100 million

00:09:25

comments on the statistics or even

00:09:27

a billion, but I can imagine that

00:09:29

if I had scored more comments, then

00:09:30

the graph would look something like this: here the

00:09:33

most likes are received by comments

00:09:34

from about 90 to 15 characters and then the

00:09:37

number of likes gradually decreases. And

00:09:40

here you can also notice such a

00:09:42

sharp dip in likes and there is

00:09:44

even a reason for it, I’ll tell you about it now, but

00:09:46

first Let's look at two graphs: the

00:09:48

blue graph is the same. This is the number of

00:09:50

likes adjusted for the number of

00:09:51

comments depending on the length of the

00:09:53

comment. But the red graph is

00:09:56

just the number of comments of different

00:09:57

damn things that I showed a few ago.

00:10:00

Firstly, here you can notice that the

00:10:02

majority write rather short

00:10:03

comments and comments of this length

00:10:05

get few likes, that is, typical

00:10:07

comments under a video on YouTube are somewhere from

00:10:09

20 to 35 characters, but a typical

00:10:12

comment under a video on YouTube that

00:10:14

gets a lot of likes is somewhere from 90

00:10:16

to 105 characters and then the longer

00:10:19

comment grows the number of likes under it

00:10:20

gradually decreases, but not too

00:10:22

much; rather, it is the lowest when there are

00:10:24

very few characters in the comment, despite the fact

00:10:25

that there are a lot of such comments,

00:10:27

and here is the

00:10:29

same dip in likes that I was

00:10:32

talking about; there is a sharp rise in the

00:10:34

number of comments and remember that

00:10:36

same spammer I talked about

00:10:37

earlier who left tens of thousands of

00:10:39

comments, so these sharp rises

00:10:41

and falls in the graph were caused by him, he wrote

00:10:43

a lot of comments but they got

00:10:45

few likes, but let’s move away from the graphs for now

00:10:47

and look at something more

00:10:48

interesting, for example then how can you

00:10:50

group comments and for this

00:10:52

you need a special algorithm that will

00:10:54

somehow break the comments into

00:10:56

main groups, that is,

00:10:58

clustering algorithms and

00:11:00

there are very, very many such clustering algorithms, here

00:11:03

in the table the most basic ones are shown,

00:11:05

probably one of the simplest and most

00:11:07

effective is the Dibiscan algorithm is

00:11:09

what I will use, it sounds

00:11:11

like a Database, that is, some kind of database.

00:11:13

But in fact, it is a density Base,

00:11:15

that is, based on density, it

00:11:17

finds the main groups that are

00:11:19

densely located like that. And those Points

00:11:20

that do not fall into any group They are

00:11:22

shown here in Gray and since comments

00:11:24

can also be represented as points in

00:11:26

space, we can also apply the

00:11:28

following clustering algorithm to them. So

00:11:30

stop, how can comments be represented

00:11:32

as a point in space? I’ll

00:11:35

also tell you about this, but let’s tell you everything in order

00:11:36

first about clustering, I

00:11:39

decided to write a program that

00:11:40

performs debiscant with visualization

00:11:41

in order to better understand how it

00:11:43

works, let’s say we have a set of points

00:11:45

and we want to split it into clusters, take

00:11:48

any point and see if it has at

00:11:50

least three other points in the radius. Well, for

00:11:52

example 100 pixels If there is, then

00:11:54

mark it as the main point, let’s say

00:11:56

I have it darker, the number of

00:11:58

neighboring points for each point so that

00:12:00

it becomes the main point and the radius in which we are

00:12:03

looking for these neighboring points, these are the two main

00:12:05

parameters of the dibiscan and accordingly

00:12:07

we can adjust these parameters, but

00:12:09

let’s let's go further than the main one, we

00:12:11

may have more than one point, but all the points

00:12:13

that have a sufficient number of

00:12:14

neighbors, so we look at all the points and

00:12:17

if they have a

00:12:19

certain number of neighbors within a certain radius, then

00:12:20

we mark them as main points, after

00:12:22

we have checked all the points, we move on to the

00:12:24

second stage again we take any point,

00:12:26

for example in my case I took

00:12:28

the Central one and assign it the first

00:12:30

cluster. For example, I will mark it in

00:12:31

red and now the cluster is spreading.

00:12:34

But other main points in its radius

00:12:36

they spread it further, that is,

00:12:38

they seem to be infected with this cluster of

00:12:39

now the main points in there are

00:12:42

no more points left in the radius of the cluster, then we take a new

00:12:44

random point and mark it with the next

00:12:46

cluster, for example green, when it

00:12:48

also has no points left, we start the blue

00:12:50

class, and so on, when we

00:12:52

have assigned all the main points to the cluster, we begin to

00:12:53

assign non-basic ones there, after which the

00:12:55

algorithms end, that is, the main ones

00:12:57

the points seem to spread the cluster. Then the

00:12:59

main points just fall into

00:13:02

it and in this picture you can see

00:13:04

that the green cluster is located somehow

00:13:05

strangely, it could relate to the red

00:13:07

cluster and indeed we can

00:13:09

change the parameters of this algorithm and, for

00:13:11

example, set the radius to be larger. Then

00:13:13

more points will fall into each

00:13:15

cluster and in this situation it looks

00:13:17

much better, only one point did not

00:13:19

fall into any cluster. Although it could be

00:13:21

attributed to the red cluster.

00:13:22

But if we increase

00:13:25

the radius even a little, the red cluster will unite with

00:13:27

green and here specifically in In this

00:13:29

situation, the clusters can be easily separated,

00:13:31

this is even visually visible, but this is not

00:13:33

always the case. Therefore, sometimes if we start

00:13:35

increasing the radius for a point, the clusters

00:13:38

will simply stick together into one large cluster.

00:13:40

Also, the problem with clustering is that

00:13:42

it has quadratic complexity and

00:13:44

sometimes it takes a very long time to calculate it,

00:13:45

especially here because you

00:13:47

need to cluster tens and sometimes more than hundreds of

00:13:49

thousands of objects, yes And there is

00:13:52

far more than one clustering here and in addition to clustering

00:13:53

you need to use other time-consuming

00:13:55

methods, but there are algorithms that

00:13:57

can speed up the speed of implementation

00:13:59

using a video card, so I

00:14:01

decided to transfer these calculations to a remote server

00:14:03

for example, here I will use

00:14:04

selectel because I have already

00:14:06

used it in previous videos, it is

00:14:07

familiar and you can create a server

00:14:09

very quickly. I am creating a new server

00:14:11

on Windows because it is more familiar to me and I choose

00:14:13

a video card and you can start the

00:14:16

server even for 1 hour with a special

00:14:18

video card for training neural networks and

00:14:20

it will cost about 50 rubles,

00:14:22

I click create a server and it starts,

00:14:24

I connected to it via a remote

00:14:27

desktop and now my magician

00:14:28

flips to Windows, two computers

00:14:30

in one for the test, I tried

00:14:32

to generate clusters for several

00:14:33

videos and overall it turned out very well

00:14:36

not bad, for example, this was the first one that

00:14:38

came across 280

00:14:40

comments and basically all the

00:14:42

comments say thank you, probably

00:14:44

this cluster could be given

00:14:46

the name Thank you or gratitude And then

00:14:49

I thought the name of the clusters could be

00:14:51

asked from gpc just like in the

00:14:53

previous video where I had neural networks

00:14:55

playing mafia; the program could

00:14:57

connect via IP to gpt automatically

00:14:59

receive a response and I added it and it

00:15:02

even worked well. But in some

00:15:04

cases it broke. For example, here I give

00:15:06

gpt the first three messages from the cluster who is

00:15:09

from 2023 funny And why am I watching this is at two

00:15:13

o’clock in the morning, and not for the first time,

00:15:14

it’s not clear what unites these three

00:15:16

messages and why they even ended up in

00:15:18

one cluster, but the whole point is that I

00:15:20

decided to take the first few

00:15:22

comments I came across from the cluster in order to

00:15:24

make a headline based on them. Because if, for

00:15:26

example the cluster consists of thousands of

00:15:27

messages, I won’t send them all

00:15:29

to gpt and it may turn out that

00:15:31

the comments that I send to gpt

00:15:33

are located, as it were, at different ends of

00:15:34

the cluster and then the comments at different

00:15:37

ends of this cluster can be

00:15:38

quite different from each other. Therefore, you

00:15:40

just need to find the center of the cluster and

00:15:43

take the comments closest to the center of the cluster,

00:15:44

they will best

00:15:46

show the meaning of the comments in the cluster,

00:15:48

but how to understand where the center is and where the

00:15:51

cluster is located in general, I was going to

00:15:52

tell you how comments can be

00:15:54

turned into a point in space. And

00:15:56

now is probably the time for this,

00:15:58

comments need to be turned in imbidding,

00:16:00

imbidding is like a numerical

00:16:02

representation of the meaning of something and all

00:16:04

modern generative models that

00:16:06

create text, pictures, videos, sounds, they

00:16:09

all use imbidings, there are

00:16:11

special models, but conditionally neural networks

00:16:13

that generate these same bedings, that is,

00:16:16

as an input we give it, for example, text

00:16:18

and as an output it gives us a set of numbers and

00:16:21

these numbers mean the meaning of a picture or a

00:16:23

phrase or something else, if these numbers are

00:16:26

represented as some coordinates, then

00:16:27

accordingly there will be some point in

00:16:29

space with these coordinates, but this

00:16:31

point will not be on a two-dimensional plane or

00:16:33

even in a three-dimensional one space

00:16:35

because there are not two or three numbers. But for

00:16:37

example, here I use a model from Open

00:16:39

ai and it gives us 1.536 numbers,

00:16:42

it turns out that this point is in 1.536 dimensional

00:16:46

space. And in fact, this model of

00:16:48

Medics and not even the best is a

00:16:51

rating of such models and it is only

00:16:53

in eighth place there But this is a general rating And

00:16:56

here I use it for a specific task

00:16:57

for clustering and in terms of clustering

00:17:00

it is already in first place Moreover, our

00:17:02

comments here are in Russian and almost

00:17:04

all of these models work only with

00:17:06

English So I get all these

00:17:07

imbidities and since they are, in a

00:17:09

sense, points in a multidimensional space,

00:17:10

you can apply the same Debiscan to them

00:17:12

and accordingly divide

00:17:14

it all into clusters and get the center of each

00:17:16

cluster, find the closest points to the centers of

00:17:18

each cluster and send them to

00:17:20

gpt and thus for each cluster

00:17:22

you can get a name I’ll try to find

00:17:25

clusters of comments in my video about

00:17:27

geometric shapes in four-dimensional

00:17:29

space and the first one gives me a cluster

00:17:31

of only four comments guy You’re a

00:17:33

cool guy you’ve confused everyone the guy is a

00:17:35

prodigy cool continue and the neural network

00:17:38

titled it as support and

00:17:39

achievement the next cluster consists from

00:17:42

the comments And what are the figures in 5D

00:17:43

next video 5D figures And what does

00:17:46

5D look like Or maybe 5D and the neural network correctly

00:17:48

titled this cluster cocktail

00:17:50

geometry and I didn’t even notice that under

00:17:53

this video there are so many comments

00:17:54

about 5D the next cluster is small it

00:17:57

consists of comments Hello everyone nonsense

00:17:59

nonsense another nonsense and the neural network

00:18:02

titled it as problems in

00:18:04

statements there is also a cluster

00:18:06

called Overcoming Difficulties and it

00:18:08

consists of comments difficult Difficult

00:18:09

complex Difficult There

00:18:11

is also a cluster called

00:18:13

evaluation and praise and comments there,

00:18:15

well, you are smart young smart Kick you are smart

00:18:18

so young and already so stupid and a

00:18:21

cluster of negative reactions What kind of

00:18:23

nonsense What nonsense God what kind of nonsense some kind of

00:18:25

nonsense It’s interesting that this cluster is

00:18:27

different from the other cluster Brad

00:18:29

clusters emoji that the neural network

00:18:31

named positive reviews and emotions

00:18:33

cluster where comments like

00:18:35

comments for promoting the channel got

00:18:36

Thank you comment for promotion

00:18:38

video and so on, and for some reason the neural network

00:18:40

named this topic and the marketing video and

00:18:42

also a cluster where they discuss the name of the

00:18:44

four-dimensional figure binder from ferender.

00:18:46

It’s interesting that the neural network understood this and

00:18:49

called the cluster the name of the geometric

00:18:50

shapes. Everywhere here for clustering I

00:18:53

used the method But this is far from the

00:18:55

only method clustering As

00:18:56

you can see from the table earlier,

00:18:58

for example, there is also a fairly

00:18:59

common method called keymins, here from the

00:19:02

pictures it seems that it does not work

00:19:03

very well, but in most cases

00:19:05

the clusters will look something like in the

00:19:07

third picture or 5. And

00:19:09

it copes well with such cases. In

00:19:11

addition it works very quickly, but its disadvantage

00:19:13

And in some situations it may even be a

00:19:15

plus, this is that this method needs to

00:19:17

indicate the number of clusters in advance,

00:19:19

let’s say I want to split the comments

00:19:21

into exactly 10 clusters, then I run

00:19:22

keymins With development into 10 clusters

00:19:25

if I want to split for 30 clusters, then I’ll

00:19:27

indicate 30, it works relatively. It’s just that

00:19:29

it

00:19:30

randomly scatters the center of the clusters in space and

00:19:33

then tries to adjust them as much as possible

00:19:34

to the points. In fact, with this method I did

00:19:36

n’t always get a good

00:19:38

result, but if you select a good

00:19:40

number of clusters, it works very

00:19:42

good. Besides, and very quickly, for

00:19:43

example, like this, he broke me up

00:19:45

in the comments and what’s interesting is that one of the

00:19:47

commentators writes that none of the

00:19:48

four-dimensional crazy stuff is used anywhere,

00:19:50

otherwise say that Well, then he has

00:19:53

some complaints about the theory of

00:19:54

relativity, which are even in this video

00:19:55

was not discussed, but he writes that

00:19:57

four-dimensional space is not

00:19:59

used, and I’m now reading

00:20:00

the comments thanks to the fact that I found it

00:20:02

among a group of points in

00:20:04

1536-dimensional space and it would be

00:20:06

nice to visualize all these points in

00:20:09

multidimensional space, I already had

00:20:10

videos about visualizing multidimensional

00:20:12

objects but I mostly showed 4D and

00:20:15

showed a little 5d6d, but there were

00:20:17

mostly three-dimensional slices of these figures. And

00:20:20

here you need a projection and projection of a

00:20:22

1536-dimensional space. But in

00:20:25

fact, this is not as difficult as it seems if

00:20:27

we had some more complex

00:20:28

figures, such as cubes balls, they would

00:20:30

block each other, but the points

00:20:32

can be arbitrarily small and

00:20:34

block each other. They will be smaller,

00:20:36

but they will still be there because we need to

00:20:38

draw them. Which means they will have a

00:20:40

size, so there are special

00:20:42

algorithms to sort of move these

00:20:43

points so that they block less than each other,

00:20:45

probably the most popular of them

00:20:47

is the algorithm with this name,

00:20:49

which is a semi-acronem and

00:20:51

is pronounced like teni or in Russian,

00:20:53

stochastic nesting of neighbors

00:20:55

by distribution and it just projects

00:20:57

the points And then either moves it to better

00:21:00

maintain the distance between them, I

00:21:02

experimented with Debiscan and

00:21:03

emboss and it turned out that they work fine on the processor

00:21:05

but not on the video card And

00:21:08

all because I installed Windows and I

00:21:10

had to install Linux, all these

00:21:11

algorithms were already made by NVIDIA for their

00:21:13

video card and for some reason they made them

00:21:16

only for Linux, I tried to look for

00:21:18

alternative libraries but It turns out

00:21:20

they all use the same

00:21:21

implementation from NVIDIA and Yes,

00:21:24

I associate this company with gaming

00:21:25

video cards and games are usually for Windows and

00:21:28

I thought that these algorithms would also be for

00:21:30

Windows But it turned out that not, but

00:21:32

fortunately you can simply copy it from that Windows server

00:21:33

delete all the files

00:21:35

and create a new one in a minute on Linux and

00:21:38

why didn’t I immediately notice that there is

00:21:39

a version here that is literally called the

00:21:40

Learning machine,

00:21:42

almost everything you need for machine

00:21:44

learning is already installed on it, where is

00:21:45

Python installed here and the libraries for it,

00:21:47

I tried to project points onto a

00:21:49

two-dimensional plane and it turned out like this

00:21:51

picture here I used my

00:21:53

latest video of a neural network playing mafia

00:21:55

But there are not many comments in it so the

00:21:57

clusters are not so clearly visible Let's

00:21:59

better take my video in which the

00:22:01

most comments are this video that

00:22:04

looks more like infinity this is it so here

00:22:06

the points are clearly not randomly scattered; there

00:22:09

are some patterns in them;

00:22:10

the projection is calculated using embossing; and

00:22:13

clustering using Dibiscan and

00:22:15

different clusters are marked with different colors,

00:22:17

but here it seems that all the points are the same

00:22:20

color, that is, they seem to have all stuck together

00:22:22

into one cluster and the first problem here

00:22:24

is that the first cluster is the

00:22:26

largest and the points that did not fall into

00:22:28

any cluster are approximately the same color,

00:22:30

but I can simply say that

00:22:32

approximately 2000 points did not fall into more than one cluster,

00:22:34

but 4600 points ended up in the first cluster;

00:22:37

sudden connection before the release

00:22:39

video, I finally found how

00:22:41

to put better colors on the dots and now you can

00:22:43

clearly see that this blue

00:22:45

cluster is the same main cluster into

00:22:47

which most of the dots stuck together. But

00:22:49

these orange dots are those that did

00:22:51

not fall into more than one class, the author

00:22:53

can still It seems strange that the main

00:22:54

cluster is kind of scattered throughout this

00:22:56

picture, there is some part here there is a

00:22:58

part here here it is in different places.

00:23:01

But in fact this happens

00:23:03

because initially it is multidimensional, namely

00:23:05

1536-dimensional and it is projected onto

00:23:08

two-dimensional space accordingly

00:23:09

there are some inaccuracies and the second

00:23:11

problem with this sample, no matter how I

00:23:13

set up the clustering parameters,

00:23:15

either most of the points do not fall

00:23:16

into any cluster, or vice versa, they all

00:23:18

stick together into one large cluster, but

00:23:21

nevertheless some clusters

00:23:22

stand out here, I’ll get closer, for example, to

00:23:24

this cluster and there are relatively

00:23:25

large clusters with blue dots. Well, they are

00:23:28

large compared to other

00:23:29

small clusters. Because there is

00:23:31

generally one supergiant cluster

00:23:32

that occupies most of the dots and a

00:23:34

comment in this class is like why the hell did

00:23:36

I enter this video at 2:00 am another

00:23:38

comment 2 o'clock in the morning so that I can still

00:23:40

watch Why am I watching this video at

00:23:42

2:00 am why am I watching this at night why am

00:23:45

I watching this at nine o'clock why am I

00:23:48

watching this and why are you watching It can be seen that with

00:23:50

distance the comments in the cluster

00:23:52

change a little, there are also

00:23:54

comments nearby ideal videos to

00:23:55

fall asleep ideal video before going to bed

00:23:58

I advise you to watch it before going to bed and don’t even

00:24:00

think about watching this video before going to bed

00:24:02

sounds like a clickbait title for the

00:24:04

video next to it there are also comments about the fact that

00:24:06

this video is watched at 5 o’clock in the morning and at

00:24:08

00:57, that is, it is clear that all the comments

00:24:10

around they are approximately about sleep or about the night,

00:24:13

but at the same time they did not unite into one

00:24:15

cluster, if you move away and look at the

00:24:17

clusters as they were projected

00:24:18

using embossing, and not how they were

00:24:20

clustered using Debiscan, then

00:24:22

this

00:24:24

large round cluster is visually visible here all

00:24:25

the comments are about the fact that no one

00:24:27

understood anything, but it’s very interesting. It’s interesting that there

00:24:29

is also one small

00:24:31

cluster nearby, they write in it: Scary, very

00:24:34

scary, we don’t know what it is and

00:24:36

these two meme clusters are nearby, and

00:24:38

these points can also be projected onto a

00:24:40

three-dimensional space, but it seems to me that

00:24:42

this only makes it more difficult, and then

00:24:45

I thought that my Debiscan does

00:24:46

n’t seem to cluster comments very effectively,

00:24:48

but at the same time the brushes project

00:24:50

them so that they immediately seem to

00:24:51

visually turn into some kind of

00:24:53

cool ones, and what if

00:24:55

I project the comments first on a 2D

00:24:57

plane and only then cluster them

00:24:59

using Debiscan and it seems that it

00:25:01

turned out much better. Let's

00:25:03

look at some cluster,

00:25:04

for example this green one. Here they

00:25:06

basically write that there is an infinity of infinities greater than infinity,

00:25:08

and next to this

00:25:10

pink class they write that there is

00:25:12

greater than infinity two infinities And

00:25:15

in this cluster they write that the letter Ali,

00:25:16

which is used to denote an

00:25:18

infinite set, is the first letter

00:25:20

of Hebrew in this pink cluster.

00:25:22

The question is, why are all these

00:25:23

infinities needed? And here they write that this

00:25:25

video blew up and of course I

00:25:27

googled my brain, is it possible to use

00:25:29

clustering after projection and

00:25:31

it turns out that many have already asked this

00:25:33

question and usually everyone answers that it’s better

00:25:35

not to because after projection the

00:25:37

distances are not fully

00:25:39

preserved and clusters will then be

00:25:41

searched for at distorted distances. But

00:25:43

it seems that in my case it worked

00:25:45

very well. Well, either I did

00:25:47

something... that’s not true, but now I want to show a

00:25:49

video with one of the largest numbers

00:25:52

of comments from those that were included in my

00:25:54

statistics, this video about sleep paralysis

00:25:56

on the channel Utopia show house 49 thousand

00:25:59

comments and quite interesting

00:26:01

clusters and this is how it looks

00:26:03

interesting what is it the first such video with a

00:26:05

huge number of comments

00:26:06

that came across statistics And it was on

00:26:08

it that the program crashed. At first I didn’t even

00:26:10

understand why, but then it turned out that

00:26:13

this was due to the fact that I

00:26:14

first converted the array of all imbidings into a string and

00:26:16

then saved this string to a file and

00:26:18

it turns out that note.gs simply cannot

00:26:20

create such a giant string and instead

00:26:23

I just decided to append each

00:26:24

imbiding to the end of the file. And to store

00:26:26

all these huge bedings on the server

00:26:28

we need more memory. Initially there was

00:26:31

40 GB, I increased it to 64 but it is

00:26:34

very easy to do just go to the

00:26:36

selectel server settings and press

00:26:37

the plus sign, let's start with the cluster on the

00:26:39

left side, here they talk about the dance at

00:26:41

the end, here about the splash screen And here they write

00:26:44

about the fact that they liked But these were

00:26:46

quite ordinary clusters Let's

00:26:48

look at more interesting ones here,

00:26:49

for example they write about coronavirus And here

00:26:52

they ask to make a video about lucid

00:26:53

dreams, here they write about the fact that they

00:26:56

had sleep paralysis, and in the largest

00:26:59

cluster they share stories of sleep

00:27:01

paralysis, usually they are quite

00:27:02

long, here they write that they have never

00:27:04

had sleep paralysis, and here they write that they

00:27:07

really want sleep and paralysis here are

00:27:09

already classic comments why am I

00:27:11

watching this at two in the morning And in this

00:27:13

class they write about who sleeps on their back on their

00:27:15

side or on their stomach here they write about

00:27:17

demons that come during

00:27:18

sleep paralysis And here they write about a

00:27:21

succubus here they discuss what they

00:27:23

call this in different cultures Demona But

00:27:25

this is a rather interesting cluster Let's get

00:27:27

closer to it, here they write no one

00:27:28

absolutely no one nearby they write about an episode from

00:27:31

Dr. House and on the other hand they write

00:27:34

about [ __ ] sofa pillar And next to the

00:27:37

cluster where they write no one absolutely

00:27:38

no one nearby there is such a small

00:27:39

cluster where they write no one absolutely no one

00:27:41

and then they write about the Witcher from the

00:27:44

center of the cluster where no one is absolutely

00:27:45

no one this cluster about the Witcher, it

00:27:47

seems to be pointing somewhere in the direction if

00:27:49

we look in this direction and go here then

00:27:51

here it will be Just the cluster where

00:27:53

they discuss the Witcher here they write timecodes

00:27:56

here are different plus symbols other

00:27:58

question marks And here are emojis And I also

00:28:02

decided to make statistics not only on

00:28:03

the comments but also on the videos themselves and this is what it

00:28:06

looks like here I didn’t use

00:28:08

clustering because in essence

00:28:10

we already know the clusters, these are the channels themselves

00:28:12

Some videos clearly form separate

00:28:14

clusters But for example, my videos are

00:28:16

scattered among these statistics, but it’s

00:28:18

not surprising because mainly for

00:28:20

these statistics I was looking for a channel and

00:28:22

there are at least 80 percent of such channels similar to mine.

00:28:24

Let’s look at those

00:28:27

videos that are clearly separated into

00:28:28

separate clusters for example, this chemistry is

00:28:30

simple and quite easy to understand why it

00:28:32

ended up in such a separate cluster. The fact is

00:28:34

that the neural network looks at the title of the

00:28:36

video and every time it has

00:28:39

chemistry added to the title at the end, it’s simple and therefore

00:28:41

easy to understand which channel

00:28:43

these videos belong to, there’s a cluster nearby

00:28:45

anthropogenesis and verdider it is interesting that the

00:28:48

verdider is scattered across the statistics and it

00:28:51

is, for example, here it is also

00:28:53

here and here it is a very compact

00:28:55

cluster. But the fact is that on the

00:28:57

water verdider channel there are mainly English-language

00:28:59

videos and at the end of each title the

00:29:02

channel from which the video is added was translated, there

00:29:03

is also a cluster of Makar Light with a

00:29:06

section and science show, it is located

00:29:08

quite densely But the rest of the videos of

00:29:09

Makar Light are scattered according to statistics, there

00:29:11

is a huge cluster of Boris

00:29:14

Trushin and among his videos there is a video of

00:29:16

Dagon Impossible task binary code in

00:29:19

action, which in general is suitable under this

00:29:21

cluster there is a super compact cluster of

00:29:23

one of the headings Utopia show here the

00:29:26

Space cluster is simple and next to it there is a DS Astra video and

00:29:29

around them are scattered videos from the

00:29:31

Zemlyakov channel, there is also Shklovsky street nearby

00:29:33

And below there are game

00:29:35

channels here many channels make videos

00:29:37

using Unity and some does Let's Play or

00:29:40

reviews of games, here for example there is Trishka boom

00:29:42

and a cluster with undertail below there is a

00:29:45

cluster with Detroitbike Human and then

00:29:48

suddenly there is one video from another

00:29:49

channel And this sitar is also about Detroit

00:29:52

Became Human there is a clusterinj And

00:29:55

if you get closer to it, there

00:29:57

is also one here The video here in the center is a fairly

00:29:59

compact cluster with videos from the

00:30:01

topless channel. And here at the top there is a very

00:30:03

separate cluster with the Domioz channel, it is

00:30:05

so separate because in each

00:30:06

video title there is a phrase 10 interesting

00:30:09

facts, but another interesting fact is that

00:30:11

two more videos from other channels are included here.

00:30:13

Let's let's get closer to this cluster and this is

00:30:15

suddenly a video from the DS Astra channel 10

00:30:18

most beautiful lunar craters and videos from the

00:30:20

Cosmos channel just 10 fakes about the Moon

00:30:23

I also tried to find the users

00:30:25

who leave the most

00:30:26

comments under the video and in first place

00:30:29

was the one who was in my statistics

00:30:31

hit

00:30:32

28,000 comments and It turned out to be chemistry It’s

00:30:35

just that he simply responded to comments

00:30:37

under his videos and thus scored

00:30:40

28,000 comments in the first place

00:30:42

chemistry is simple and in the second place Space just

00:30:46

got 20,000 comments in the same way

00:30:48

Then three closet wildbers

00:30:50

stubborn paleontologist Yes, I’m with Astra Boris

00:30:52

Rushin it turns out that the Seven first places in

00:30:55

comments are the bloggers and

00:30:57

only then the Subscriber who

00:30:59

wrote more than 4000 comments to a

00:31:01

variety of channels. Now let’s

00:31:03

see when comments are written after

00:31:05

the video is released, here you can see that

00:31:07

most comments are written at the very

00:31:08

beginning and then their number

00:31:10

drops sharply, but then it seems to rise a little

00:31:12

and then drops again,

00:31:14

rises as if there are some

00:31:16

waves here, and if you get closer to this graph, the

00:31:19

waves are visible even better and as you can

00:31:21

guess, the length of this wave is equal to a day

00:31:23

and most comments are

00:31:25

probably written during the day, but Wait, this is not an

00:31:27

absolute time, this is a time relative to the

00:31:29

release of the video, and since everyone uploads

00:31:32

videos at different times, these waves

00:31:34

could be smoothed out, but as you can see, they are there,

00:31:36

probably because the majority

00:31:37

upload videos at similar times. That

00:31:39

is, somewhere in the evening, but since all the channels

00:31:41

in this sample are Russian-language, and then the evening

00:31:43

takes place in approximately the

00:31:45

same time zones and Let's also

00:31:47

try to look at the number of likes

00:31:49

by time, it also, of course, needs to be

00:31:51

divided by the number of comments

00:31:52

so that there are no so that where there are more

00:31:54

comments there are more likes I thought

00:31:56

that the comments

00:31:58

that were written immediately after the release of the

00:31:59

video get more likes and it seems that this is the case then the

00:32:01

statistics become much

00:32:03

noisier and nothing is clear here, but

00:32:05

at the beginning it is much smoother because

00:32:07

there are simply more comments there

00:32:09

were the number of comments by the hour and

00:32:11

now let's by minute and it is clear that in the

00:32:13

first minutes after the release of the video a

00:32:15

very large number of

00:32:16

comments appear and it passes approximately in the

00:32:19

seventh minute after the release of the video this

00:32:21

graph shows exactly the first day

00:32:23

after the release of the video if you remove the first 7

00:32:25

minutes then the graph will go like this and

00:32:27

now let's see in terms of the number of

00:32:28

likes by minute, here it is of course

00:32:31

much more noisy and, again,

00:32:33

most likes receive comments

00:32:34

that are written at the very beginning or not

00:32:37

at all. Look, there is such a

00:32:39

small stripe because of the noise. It’s

00:32:41

not very visible, but in fact

00:32:42

the graph is not very good at first high And it

00:32:44

goes up and reaches the Peak

00:32:46

Approximately at the eighth minute.

00:32:48

Remember, I showed before that a

00:32:49

huge number of comments are written

00:32:51

in the first 7 minutes, but at the same time They do

00:32:53

n’t get that many likes. It happens at the

00:32:56

eighth minute and it is the

00:32:59

comments written then that get

00:33:00

the majority of likes, that is the comment that is most

00:33:03

likely to receive more likes is the one

00:33:04

written eight

00:33:06

minutes after the video is released, and

00:33:08

remember that its length should be from 90

00:33:10

to 105 characters. I also tried to find the

00:33:13

most popular words in the comments on

00:33:15

different channels. And in fact, they are

00:33:17

the same on all channels. and in and not Well,

00:33:19

simply because these are the most

00:33:21

common words in the Russian language,

00:33:23

so I tried to find not just those

00:33:25

words of which there are a lot, but those words of which there are

00:33:27

more on a specific channel than in the general

00:33:29

sample, and then on my channel, for example, the

00:33:31

top 10 words turn out to be They are weights

00:33:33

colon dog onigirirayo this is connected

00:33:37

with the number of paradise from the video about the largest

00:33:39

numbers they are a weight with a comma and not weights without a

00:33:41

comma they are weights in English letters with a

00:33:44

comma Nigeria without a comma also

00:33:46

this 3-3 from the video about the largest

00:33:48

numbers all Finnish this is the

00:33:50

Sierpinski triangle apparently because that the word

00:33:52

triangle is often found on

00:33:54

other channels, but not the word of Serfinsky and

00:33:56

Google plex, also from the video, the largest

00:33:58

numbers seem to me to be the most interesting top

00:34:01

10 words in chemistry, it’s just that his first one is

00:34:03

bullum bloom comma, then a link to a

00:34:06

deleted video on YouTube dog chemistry

00:34:08

bullum comma broma bulm will come out

00:34:12

colon exhaust bul bul with a

00:34:14

question mark and a spirit lamp Well,

00:34:16

here you can see that a lot of words

00:34:18

are repeated It’s just that they are followed by a

00:34:20

comma dot Because I

00:34:21

separate words simply by spaces You can

00:34:23

try not to take into account any commas,

00:34:25

dots and other symbols and then the top 10

00:34:28

words For example, it looks like this here,

00:34:30

first the place where the word is,

00:34:31

then the number of its mentions and the

00:34:33

word itself and the rating

00:34:35

haven’t changed much, then the word arrow was added,

00:34:37

which is interesting, this is not even the game about

00:34:39

arrows that I make, but arrows from a

00:34:42

video about the largest numbers and

00:34:43

a rating was also added. For example, chemistry is

00:34:46

simply found more than a thousand times as a

00:34:48

word and more than a thousand times.

00:34:50

It’s interesting that it occurs only

00:34:53

twice more than just the boom of a stubborn

00:34:56

paleontologist; the very words are basilosaurus,

00:34:59

mosasaurs, akylosaurus, and so on, and on the

00:35:03

DS Astra channel it is in fourth place in

00:35:05

popularity the word goes milking And I also

00:35:07

made this site with voting for the

00:35:09

best thing, two completely

00:35:11

random things are given to choose from and we choose what

00:35:14

we like best. I already posted this

00:35:16

on my Telegram Channel, by the way,

00:35:17

subscribe to it and there were

00:35:19

more than 100,000 votes but after this

00:35:22

video you can make sure that this is

00:35:23

not enough for good statistics And

00:35:26

this test will be indirectly related to the

00:35:27

next video and a link to it will be in the

00:35:29

description So go vote and

00:35:32

see what a beautiful animation of

00:35:33

cards I made there And if you want to

00:35:35

support the channel then Subscribe and

00:35:38

bye everyone

00:35:43

[music]

Description:

Начни карьеру в Data Science: https://go.skillfactory.ru/jperFg Возврат денег за курс, если не нашли работу после обучения + скидка 45% по промокоду ONIGIRI до 31.07.2023 Сервер с GPU в аренду: https://selectel.ru/services/gpu/?section=cloud Телеграм https://t.me/onigiriScience Boosty https://boosty.to/onigiriscience Сайт с голосованием: https://www.bestthingever.fun/ В этом видео я анализирую комментарии с разных каналов на ютубе. Применяю методы кластеризации и делаю визуализацию эмбеддингов Реклама: ООО "Селектел", ИНН: 7842393933, токен: 2VtzqupnNU4 ООО "СКИЛФЭКТОРИ" ИНН 9702009530, токен: 2VtzquWXGBn

Preparing download options

Popular

HD video

Only sound

All

* — If the video is playing in a new tab, go to it, then right-click on the video and select "Save video as..."

** — Link intended for online playback in specialized players

Questions about downloading video

How can I download "Проанализировал комментарии НЕЙРОСЕТЬЮ. И вот, что я узнал" video?

http://unidownloader.com/ website is the best way to download a video or a separate audio track if you want to do without installing programs and extensions.
The UDL Helper extension is a convenient button that is seamlessly integrated into YouTube, Instagram and OK.ru sites for fast content download.
UDL Client program (for Windows) is the most powerful solution that supports more than 900 websites, social networks and video hosting sites, as well as any video quality that is available in the source.
UDL Lite is a really convenient way to access a website from your mobile device. With its help, you can easily download videos directly to your smartphone.

Which format of "Проанализировал комментарии НЕЙРОСЕТЬЮ. И вот, что я узнал" video should I choose?

The best quality formats are FullHD (1080p), 2K (1440p), 4K (2160p) and 8K (4320p). The higher the resolution of your screen, the higher the video quality should be. However, there are other factors to consider: download speed, amount of free space, and device performance during playback.

Why does my computer freeze when loading a "Проанализировал комментарии НЕЙРОСЕТЬЮ. И вот, что я узнал" video?

The browser/computer should not freeze completely! If this happens, please report it with a link to the video. Sometimes videos cannot be downloaded directly in a suitable format, so we have added the ability to convert the file to the desired format. In some cases, this process may actively use computer resources.

How can I download "Проанализировал комментарии НЕЙРОСЕТЬЮ. И вот, что я узнал" video to my phone?

You can download a video to your smartphone using the website or the PWA application UDL Lite. It is also possible to send a download link via QR code using the UDL Helper extension.

How can I download an audio track (music) to MP3 "Проанализировал комментарии НЕЙРОСЕТЬЮ. И вот, что я узнал"?

The most convenient way is to use the UDL Client program, which supports converting video to MP3 format. In some cases, MP3 can also be downloaded through the UDL Helper extension.

How can I save a frame from a video "Проанализировал комментарии НЕЙРОСЕТЬЮ. И вот, что я узнал"?

This feature is available in the UDL Helper extension. Make sure that "Show the video snapshot button" is checked in the settings. A camera icon should appear in the lower right corner of the player to the left of the "Settings" icon. When you click on it, the current frame from the video will be saved to your computer in JPEG format.

What's the price of all this stuff?

It costs nothing. Our services are absolutely free for all users. There are no PRO subscriptions, no restrictions on the number or maximum length of downloaded videos.