AI Engineer World's Fair 2026

Physical AI's Next Bottleneck Is Finding the Right Video — Rafael Levi, Bright Data

Read the talk

Physical AI’s Next Bottleneck Is Finding the Right Video

Rafael Levi of Bright Data walks through the gap between abundant web video and useful robot training data, then demonstrates a workflow that searches for actions before collecting clips.

From a talk by Rafael Levi

At a glance

Ideas worth remembering

  • Collecting physical-action data has several constraints: instructed recordings can change behavior, simulation raises questions about physical fidelity, and teleoperation requires operator time.

  • Frame-to-frame changes can supply motion information, but relevant human video still needs processing before it becomes useful robot training material.

  • Bright Data’s proposed workflow searches an action index before collection, returning trimmed clips so customers can avoid downloading entire videos to find a useful moment.

  • Action search can find incidental events that titles miss, from dishwashing and brand appearances to driving maneuvers and game-level completions.

A robot needs examples of the world changing

A system that can see, understand and act needs examples of actions and their consequences. Rafael Levi, who works on data collection at Bright Data, opens with that requirement: improving the AI still leaves the question of what to feed it. His shorthand is memorable: “AI without data is just a box.” The presentation focuses on finding video that can help train robots and world models.

Source frame: A robot needs examples of the world changing
Source frame: A robot needs examples of the world changing

The opening history traces a progression from robots learning through images and real-world actions in 2022, to AI controlling multiple robots in 2023, to wider access through open source in 2024. As more labs gain access to models, collecting suitable experiences becomes a practical constraint. Levi also distinguishes the excitement around humanoids from their autonomy: many of the robots people encounter are still remotely controlled.

The scale comparison explains why he turns to video discovery. Levi puts language-model training material at trillions of words, image training material at billions of labeled images, and existing robot-action video at only about a million videos. These are his broad estimates rather than a like-for-like dataset census, but the distinction matters: a large supply of text or pictures does not automatically supply sequences showing how objects move when someone acts on them.

0:280:31
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:28 · section reference included

Why recording more actions has costs of its own

Opening a door seems like an easy action to collect. Pay someone to record it, then add the video to a training set. The complication is that the instruction changes the circumstances: demonstrating how to open a door for a camera may produce a different movement from walking into the house during an ordinary day. Camera awareness can change behavior too. Levi calls this instructed data and considers it biased; the recording does not establish how much that difference affects robot performance.

Source frame: Why recording more actions has costs of its own
Source frame: Why recording more actions has costs of its own

Other collection methods trade different constraints:

  • Simulation: Virtual environments are cheap to use, but Levi questions whether their physics is faithful enough for the robot behaviors he wants to train. The attraction is inexpensive experience; the concern is whether that experience carries over to the physical world.
  • Teleoperation: A person can control a real robot and record its actions, but collecting more hours requires more operator time. His example of recording eight hours a day makes the scaling problem concrete: a large dataset still depends on many people spending many days operating machines.
  • Existing datasets: Reusing a prepared dataset avoids fresh collection, but the available supply is limited in the account he gives.
3:093:16
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:09 · section reference included

Web video supplies motion, but still needs a path to robot control

The alternative is to look for actions people have already recorded. Web video contains people handling objects, opening doors and encountering the consequences of motion and gravity. That makes it a potential source of physical experience without commissioning every demonstration. It also creates a discovery problem: useful interactions sit among enormous amounts of unrelated footage.

Source frame: Web video supplies motion, but still needs a path to robot control
Source frame: Web video supplies motion, but still needs a path to robot control

Levi cites a Meta training result as motivation: about a million hours of real-world video followed by 62 hours of robot data, after which the system could control a real robot. The result is presented without the model identity, task or evaluation conditions, so the useful lesson is narrower than a universal recipe. A large video stage can supply experience that a smaller robot-specific stage then connects to control; the cited example still includes real robot data.

How can a video of a person pouring water help a robot? The explanation starts with change between frames. Compare two images, measure the movement between them, then repeat across the sequence. The changing positions provide information about motion; Levi describes extracting angles and distances from that progression. A still image captures a state, while the sequence exposes how the state changes during an action.

That is a high-level account of extracting motion, not a complete method for turning human movement into robot commands. Levi explicitly acknowledges that downloaded video does not get a training system “100% there”: further processing is necessary. Keep that distinction in mind through the product demonstration. Finding a relevant clip supplies material for the training pipeline; it does not finish the pipeline.

5:165:20
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:16 · section reference included

Move action selection ahead of the download

Collecting video first and deciding what is useful afterward can be expensive. Levi reports that NVIDIA discards about 96% of video collected for Cosmos training, and that Stable Video Diffusion discards 74%. He uses a million-hour download as an illustration of the first percentage: only 4% would remain useful. The rates are cited claims rather than independently established measurements here, but they identify the cost at issue. Footage rejected after collection has already consumed bandwidth, storage and processing.

Source frame: Move action selection ahead of the download
Source frame: Move action selection ahead of the download

Bright Data’s proposed order is “search first collect second.” Describe the action you need, search an existing video index for that action, then receive the relevant snippets. The intended unit of collection becomes the action clip rather than the entire source video. Queries can describe a person washing dishes or folding a T-shirt, even when those actions are not the subject of the video’s title.

Where does selection happen, and what work remains afterward? The flow below follows the dishwashing query developed in the next demonstration. Searching and trimming happen before the customer collects the training material. Motion and distance processing still follow retrieval. The proposed savings come from collecting a smaller, more relevant slice of video; the index is the infrastructure that makes that earlier selection possible.

How it fits togetherFrom an action description to training material

Specify the action and visible hand interactions.

The video index moves relevance selection before collection. A trimmed result still needs downstream processing for robot training.

7:588:00
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:58 · section reference included

The dishwashing query describes a sequence, not a topic

The demonstration makes the query concrete: a person washing dishes by hand in a sink, removing grease from a plate with a sponge and soap, rinsing, and placing dishes in a drying rack, with close-up hand interactions. Each added detail narrows what useful footage should contain. A video broadly about kitchens is insufficient; the desired material shows the manipulation and the steps around it.

Source frame: The dishwashing query describes a sequence, not a topic
Source frame: The dishwashing query describes a sequence, not a topic

The search then returns clips of people washing dishes. The observable change is from a written description of the desired activity to candidate video snippets containing it. Those snippets may come from videos about cleaning a kitchen or another broader subject. That is the point of indexing actions: an incidental moment can become a search result even when the source video was named for something else.

Levi says the demonstration was recorded with about 100 million indexed videos and that the index had since reached 1.1 billion. The playback is sped up, so it does not establish search latency, and the examples do not quantify retrieval accuracy. A second query asks for a human folding different types of clothes and returns folding footage. Together, the examples illustrate the intended interface: describe a visible action in detail, then collect clips that can feed subsequent training work.

9:5710:00
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:57 · section reference included

The API returns clips and information about the match

The same workflow is available through an API, so repeated searches need not depend on a person using the demonstration interface. Levi describes returning a video snippet and its URL, along with a link to the original video. The result also carries information about where and how the query matched:

  • Timestamp: Locates the relevant moment in the source video.
  • Match score: Describes how close the result is to the query.
  • Frame count: Reports how many frames concern the requested content.
Source frame: The API returns clips and information about the match
Source frame: The API returns clips and information about the match

Brand discovery uses the same gap between a video’s title and its visible contents. Imagine the example Levi gives: a person applies makeup, and a particular brand is visible on the table, while the recording’s title concerns something else. Searching the brand name in titles could miss that appearance. Searching indexed video content is meant to find the moment when the product is present or used. Here the useful result is a brand appearance rather than a robot demonstration.

For training, the intended benefit is less irrelevant footage entering the collection and preparation process. Levi connects noisy data to model problems and describes the product as removing that noise. The demonstrated capability supports a more limited conclusion: action retrieval can select material before bulk collection. Match scores describe query relevance; the talk does not establish that retrieval eliminates all unsuitable training data or downstream model errors.

11:3411:36
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:34 · section reference included

Search for the event hidden inside the video

Driving footage extends the idea from household manipulation to road events. Dashcam videos can contain someone running a red light, stopping at one, turning left or turning right. These are distinct behaviors within a larger recording, and action search could collect examples of each. Levi proposes them as useful material for self-driving systems and other world models; the recording does not demonstrate a driving model trained on the returned clips.

Source frame: Search for the event hidden inside the video
Source frame: Search for the event hidden inside the video

The ending returns to physics and everyday movement: sitting, walking and moving through the world. Levi’s frustration is with commissioning huge numbers of recordings while relevant actions already exist online. His practical objection is discovery. A title search asks what the uploader called the video; a training-data search needs to ask what happened inside it.

The proposed product brings discovery across public video providers, including YouTube and Vimeo, into one place. Its usefulness can extend beyond training and brand appearances. The final example is a gamer looking for footage of someone beating a particular level, even when no video is titled for that specific event. The same mechanism applies: find the relevant action within a longer recording, collect that portion, and process it for the task at hand.

13:3413:36
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:34 · section reference included

Read the complete timestamped transcript
  1. 0:01

    [music]

  2. 0:13

    Hi everybody.

  3. 0:15

    Welcome.

  4. 0:16

    For those that came to actually see me,

  5. 0:19

    thank you. For those that are just

  6. 0:21

    hanging out here, also thank you. I'm

  7. 0:23

    going to try to entertain you guys,

  8. 0:25

    teach you something new.

  9. 0:28

    Um again my name is Rafael. I work at

  10. 0:31

    Bright Data. I've been with Brighta for

  11. 0:33

    over eight years. And in Bright Data we

  12. 0:35

    are collecting data and uh we're trying

  13. 0:37

    to innovate. So today I want to talk

  14. 0:40

    about video discovery for agentic world

  15. 0:42

    models training. And uh we're going to

  16. 0:45

    explain what that is in a little bit.

  17. 0:48

    So

  18. 0:49

    what does it take to take to build a

  19. 0:52

    system that can see understand and act

  20. 0:56

    right? I mean, we figured out what an AI

  21. 1:00

    is. We're constantly improving it. Like,

  22. 1:02

    you know, that's kind of handled. But

  23. 1:04

    now, the hard part is what data do you

  24. 1:07

    provide to it? Because everybody knows

  25. 1:10

    AI without data is just a box, right?

  26. 1:15

    So, let me take you a little bit back in

  27. 1:17

    history, right? In 2022,

  28. 1:20

    Google taught a robot by images, real

  29. 1:25

    world actions, right? Then going

  30. 1:28

    forward, there was a leap uh in 2023. Uh

  31. 1:32

    we had uh an AI actually control

  32. 1:35

    multiple robots.

  33. 1:37

    In 2024, the open source came out. Now,

  34. 1:40

    this where the game changed. Once we had

  35. 1:42

    an open source, all the labs had access

  36. 1:45

    to AI and they started already putting

  37. 1:48

    AI into robots, right? And so now these

  38. 1:52

    models are already driving humanoid

  39. 1:55

    robots. I'm you seeing some of course

  40. 1:58

    most here are remote controlled, but I

  41. 2:01

    mean in San Francisco you got weo

  42. 2:03

    driving without any drivers. How cool is

  43. 2:06

    that? First time I got in the car, I

  44. 2:09

    thought I was going to die, but it was

  45. 2:10

    actually very nice. Right now, how do

  46. 2:14

    you think it learned how to drive? By

  47. 2:17

    actually teaching itself. By by learning

  48. 2:20

    on dashboard cameras, right? Every car

  49. 2:23

    has a dashboard. So, if you provide that

  50. 2:25

    to an AI, it can figure out how to

  51. 2:28

    actually operate a machine.

  52. 2:31

    So, as I said, the AI is no longer the

  53. 2:33

    hard part. The data is. And this is what

  54. 2:36

    I want to talk about. Right? So for chat

  55. 2:42

    LLMs, there's trillions of words and

  56. 2:44

    texts, right? So we have so much text

  57. 2:47

    data to train the LLMs. For image

  58. 2:50

    generation, we got billions of images,

  59. 2:53

    labeled images. So that's also a huge

  60. 2:55

    database.

  61. 2:57

    But for robotics, there's only about a

  62. 2:59

    million videos. It's a very small

  63. 3:02

    limited data sets of robots doing

  64. 3:05

    things, right?

  65. 3:09

    And so the idea here is that of course

  66. 3:14

    you a lot of companies out there what

  67. 3:16

    they do is they pay people to record

  68. 3:19

    actions, right? So for example, hey

  69. 3:22

    record me how you open a door or record

  70. 3:25

    to me how you're sitting on the chair.

  71. 3:28

    But then the problem becomes is that

  72. 3:30

    when somebody is told to do something,

  73. 3:33

    they do not do it naturally. So I call

  74. 3:35

    it instructed, right? If you're told to

  75. 3:38

    record how you open the door and you're

  76. 3:40

    doing it for somebody, it's not going to

  77. 3:41

    be the same as if you just walk into the

  78. 3:43

    house. The movement is totally

  79. 3:45

    different. So is that data good for

  80. 3:48

    robotic training? In my assumption, no.

  81. 3:50

    It's biased data and it doesn't deliver

  82. 3:52

    the same results. So obviously the robot

  83. 3:55

    is not going to be as effective as if

  84. 3:57

    it's intuitive. Right? So I just threw

  85. 4:00

    in some examples on the slide, right? So

  86. 4:03

    uh

  87. 4:06

    the built by hand data is not the same

  88. 4:08

    as the data that is kind of on the site,

  89. 4:11

    right? So and also the fact is that

  90. 4:14

    people behave very differently when they

  91. 4:16

    on the camera. Some of you know some

  92. 4:17

    people are camera shy. So you know as

  93. 4:20

    soon as you put a camera on them you

  94. 4:22

    they change but in the natural world

  95. 4:25

    it's a totally different thing. So this

  96. 4:27

    is what we are trying to solve in bright

  97. 4:29

    data right

  98. 4:32

    um

  99. 4:35

    again so why is the usual data sources

  100. 4:37

    are not enough right simulations

  101. 4:39

    uh virtual it's it's cheap but it's um

  102. 4:42

    you know everybody's playing video games

  103. 4:44

    the physics in video games are almost

  104. 4:46

    there but they're not good enough where

  105. 4:48

    to train a robot on it right a hand

  106. 4:51

    control the person controlling a robot

  107. 4:53

    it's doable but how many hours a day can

  108. 4:56

    you record the robot, how many people

  109. 4:58

    you need to actually get huge scale

  110. 5:00

    data, right? I mean, you can record

  111. 5:02

    maybe eight hours a day every day. How

  112. 5:05

    many hours are you going to get in a

  113. 5:06

    year? It's not really scalable.

  114. 5:09

    And of course, there's already pre-built

  115. 5:11

    data sets, but as I said before, they're

  116. 5:13

    small. There's only about a million

  117. 5:14

    videos at this point. So, what is the

  118. 5:16

    alternative? The alternative is the web.

  119. 5:20

    Let's just think about YouTube. How many

  120. 5:22

    videos are there on YouTube? I don't

  121. 5:25

    know, five billion videos. How many

  122. 5:28

    actions are in those videos that a robot

  123. 5:32

    can replicate?

  124. 5:34

    I mean, let's think about it. How many

  125. 5:36

    do you think videos are there of a

  126. 5:37

    person opening the door firsterson view?

  127. 5:42

    Millions of hours. You'll be surprised.

  128. 5:46

    So,

  129. 5:50

    every day the web video shows gravity

  130. 5:52

    motions, right? So um cause and effect

  131. 5:56

    obviously accidents big the biggest

  132. 5:59

    cause and effect right how handle how

  133. 6:01

    people handle objects and billions of

  134. 6:03

    hours great training material for robots

  135. 6:07

    but what is the problem the problem is

  136. 6:09

    that there's a lot of noise right oh and

  137. 6:13

    for example right so one of the AI

  138. 6:15

    models that meta trained they gave it

  139. 6:17

    about a million hours of real world

  140. 6:20

    videos

  141. 6:21

    and then all it took is 62 hours of real

  142. 6:24

    robotics data to actually control to

  143. 6:27

    control a real robot. So if you think

  144. 6:30

    about it,

  145. 6:31

    the data was collected publicly, right?

  146. 6:33

    So it trained the the the robot on just

  147. 6:37

    random things and then in 62 robotic

  148. 6:40

    hours, it was already moving. So there

  149. 6:42

    no simulations needed, autonomous

  150. 6:45

    robotics delivered fast.

  151. 6:52

    But the videos don't show how the robots

  152. 6:54

    move, right? So you might think, well,

  153. 6:56

    listen, how useful is it a person

  154. 6:58

    pouring water into a cup for a robot. So

  155. 7:02

    there is methods actually out there how

  156. 7:04

    an AI can distinguish and actually learn

  157. 7:07

    from the actions that are in a video. If

  158. 7:09

    you take two frames frame by frame and

  159. 7:12

    the AI measures the movement this

  160. 7:14

    difference, right? So you have an image

  161. 7:16

    A and image two and there's a small

  162. 7:18

    movement that is changing between the

  163. 7:20

    two images and AI can actually measure

  164. 7:23

    that change. So if you keep doing that

  165. 7:26

    through the whole video and AI can

  166. 7:28

    actually figure out angles, distance and

  167. 7:31

    everything that it needs to train and

  168. 7:33

    robot right. So we don't necessarily

  169. 7:36

    need the sensor's data but of course the

  170. 7:39

    video itself that you download from

  171. 7:41

    video from YouTube or the video that uh

  172. 7:44

    you extract the data from doesn't

  173. 7:46

    necessarily get you 100% there. You do

  174. 7:48

    need to have some processing on top of

  175. 7:50

    that but

  176. 7:52

    it's available. It's out there and all

  177. 7:55

    you need to do is process it.

  178. 7:58

    So a few more examples right? So for

  179. 8:00

    example, Nvidia when they train in the

  180. 8:03

    Cosmos robot, they're throwing out about

  181. 8:05

    96% of the video. So what does that

  182. 8:08

    mean? They download a million hours of

  183. 8:11

    video and the only part is 4% is

  184. 8:14

    actually useful. Everything else gets

  185. 8:16

    thrown out. That's wasted compute.

  186. 8:18

    That's wasted bandwidth. That's wasted

  187. 8:20

    storage. That's just a lot of wasted

  188. 8:22

    money. Okay. Stable video diffusion is

  189. 8:25

    throwing out 74% of the videos that

  190. 8:28

    they're downloading.

  191. 8:30

    So in bright data we are trying to solve

  192. 8:33

    that problem and we're trying to take it

  193. 8:36

    a step forward. So what we propose is

  194. 8:39

    search first collect second. What we do

  195. 8:42

    is we are actually indexing videos and

  196. 8:45

    we are allowing you to search for

  197. 8:48

    specific actions. Since we're talking

  198. 8:51

    about a person opening a door, let's

  199. 8:52

    keep using that uh example. What you can

  200. 8:56

    do on our platform is actually input

  201. 8:59

    detailed information, right? A query,

  202. 9:02

    person washing dishes, person folding a

  203. 9:04

    t-shirt and so on. And we will provide

  204. 9:08

    the snippets from the videos of these

  205. 9:11

    actions.

  206. 9:14

    So what does that mean?

  207. 9:16

    That means that

  208. 9:19

    there's less waste of data collection.

  209. 9:22

    there's less waste of storage and

  210. 9:24

    bandwidth, right? So, a little more on

  211. 9:26

    how it works. Obviously, you define what

  212. 9:28

    you're looking for.

  213. 9:32

    We search through billions and billions

  214. 9:34

    of pre-indexed videos, not by keywords,

  215. 9:38

    but by actions, and we provide you with

  216. 9:41

    ready to use clips.

  217. 9:44

    And obviously, they're already trimmed,

  218. 9:46

    prepared for your training. So all you

  219. 9:47

    got to do is process them a little bit

  220. 9:49

    for maybe motion distance sensors and so

  221. 9:52

    on and it's ready to be ingested into a

  222. 9:55

    robot.

  223. 9:57

    I have a small example of a video of how

  224. 10:00

    it actually works on our platform. So

  225. 10:02

    let me just hit play here if I can find

  226. 10:05

    it.

  227. 10:07

    So here we see person washing dishes by

  228. 10:09

    hand in sink removing grease from the

  229. 10:11

    plate using sponge and soap including

  230. 10:14

    rinse and placing dishes into drying

  231. 10:16

    rack. close-up hand interactions.

  232. 10:19

    So, right now what is happening is it's

  233. 10:21

    searching through

  234. 10:23

    billions of videos. Now, on this video,

  235. 10:25

    it's a bit old. We only got about 100

  236. 10:27

    million index there, but now we're up to

  237. 10:29

    1.1 billion videos. And it will

  238. 10:31

    literally like this is a little sped up,

  239. 10:33

    so I didn't want to waste your guys

  240. 10:34

    time, but it will provide to you clips

  241. 10:38

    from videos of people washing dishes.

  242. 10:41

    And you can see right here all the

  243. 10:43

    videos in

  244. 10:46

    That would be fine. Now, some of these

  245. 10:48

    videos might not have to do anything

  246. 10:49

    with washing dishes. They might have

  247. 10:51

    cleaning kitchen. They might have to do

  248. 10:53

    something else. And here's another

  249. 10:55

    example. For example, human folding

  250. 10:57

    different types of clothes,

  251. 10:59

    right? So, a little more description.

  252. 11:01

    The more description you give, the more

  253. 11:03

    you actually get back.

  254. 11:06

    So, again, a little speed up. Let's see.

  255. 11:14

    I want you to understand it could be

  256. 11:15

    anything. A person putting on makeup.

  257. 11:17

    Let's say you have a brand and you want

  258. 11:18

    to find videos where your brand is being

  259. 11:20

    found. All that can be also provided. So

  260. 11:24

    you see people folding clothes. So now

  261. 11:27

    you can train a robot on how to fold

  262. 11:29

    clothes.

  263. 11:34

    Of course everything is available via

  264. 11:36

    API. We don't expect people to actually

  265. 11:37

    do anything manually, right? So you can

  266. 11:40

    trigger it by API. you get a returned

  267. 11:43

    snippet of the video, the URL. We give

  268. 11:46

    you the link to the original video if

  269. 11:48

    you want to watch it. But the main thing

  270. 11:50

    here is that

  271. 11:52

    we can give you the snippets of the

  272. 11:54

    video to train your robotics. Now, I'm

  273. 11:58

    not sure if any of you training any

  274. 11:59

    robots, so I'm going to give you another

  275. 12:02

    example of how useful it could be. Let's

  276. 12:05

    say you have a brand and you want to

  277. 12:07

    know where your brand is being demoed on

  278. 12:11

    videos.

  279. 12:13

    The video titles doesn't matter because

  280. 12:15

    it doesn't actually consume. It's not

  281. 12:17

    about your product. It's just a person

  282. 12:20

    doing some podcast recording something.

  283. 12:23

    But you see a woman putting on makeup

  284. 12:25

    and she you see on the table that this

  285. 12:27

    is your brand of a makeup. So now all of

  286. 12:30

    a sudden you can find all videos for

  287. 12:32

    your brand where it's being used

  288. 12:34

    whatever it is right by searching name

  289. 12:37

    of the video you won't find it by by

  290. 12:40

    video indexing you can actually find

  291. 12:42

    specific things that you are interested

  292. 12:44

    in apple falling from the tree and so

  293. 12:48

    on. So what comes back? We provide to

  294. 12:51

    you this time stamp. We provide to you

  295. 12:53

    the score matching, how close it is to

  296. 12:54

    your query, and of course the frame

  297. 12:57

    count. How many frames are about what

  298. 13:00

    you're looking for?

  299. 13:04

    So what does this change?

  300. 13:06

    This changes a lot. Okay, less waste.

  301. 13:09

    You don't need to download million of

  302. 13:11

    hours of videos. you can actually get

  303. 13:13

    exactly the the specific actions that

  304. 13:17

    you are interested in.

  305. 13:21

    Again, it's very crucial for robotic

  306. 13:23

    training because noise creates problems,

  307. 13:27

    hallucinations, right? This removes the

  308. 13:30

    noise.

  309. 13:34

    What is it useful for? Not only robots,

  310. 13:36

    but obviously self-driving for example,

  311. 13:38

    Wimo, right? trained on dash cam video

  312. 13:42

    cameras and there's millions, hundreds

  313. 13:45

    of millions of hours on YouTube of dash

  314. 13:47

    cams. How many of you like to watch

  315. 13:49

    accidents? Dash cam c accidents. I'm

  316. 13:52

    sure everybody watched them. Crazy

  317. 13:54

    driving in Russia, right? So, this is

  318. 13:57

    what we're talking about, right? The

  319. 13:59

    videos that are out there, okay? A

  320. 14:01

    person running a red light, person

  321. 14:03

    stopping on red light, making a left

  322. 14:05

    turn, making a right turn, and so on.

  323. 14:07

    All of that can be a very useful data

  324. 14:10

    for training and any other world models.

  325. 14:13

    Physics, right? Robots needs to

  326. 14:15

    understand physics. What is the gravity?

  327. 14:18

    How to sit? How to walk? How to move?

  328. 14:21

    Again, of course, you can pre-record it.

  329. 14:24

    Um, a lot of times right now I speak to

  330. 14:27

    some of the people and they have

  331. 14:28

    millions of people recording videos for

  332. 14:30

    them. Hey, can you guys record me how

  333. 14:32

    you open doors? That's crazy. Why do you

  334. 14:35

    need a million people doing it when

  335. 14:37

    everything is available online? Online

  336. 14:40

    is one of the hugest databases out

  337. 14:42

    there. It's just people don't really

  338. 14:44

    like it's it's not that easy to use it,

  339. 14:46

    right? I mean, right now the only search

  340. 14:48

    available to us is by the keyword of the

  341. 14:51

    name of the video.

  342. 14:53

    And of course, one source, the whole

  343. 14:56

    public video web in one place. I mean,

  344. 15:00

    we got YouTube, we got Vimeo, we got so

  345. 15:02

    many different uh video providers out

  346. 15:05

    there, billions and billions of hours

  347. 15:09

    of training data just there waiting for

  348. 15:12

    you to grab it, collect it and process

  349. 15:14

    it. So basically this is a new product

  350. 15:16

    that we're doing a bright data right

  351. 15:18

    indexing videos so that you guys can

  352. 15:21

    find specific actions anything you can

  353. 15:24

    think about you know again it could be

  354. 15:26

    useful for brands uh as well it's not

  355. 15:29

    only for training data anything you can

  356. 15:32

    think of could be useful

  357. 15:35

    maybe you are a gamer and you want to

  358. 15:38

    know how do you beat this level but

  359. 15:41

    there's no exactly video like that so

  360. 15:43

    you can search okay people beating the

  361. 15:45

    level on the video game, right?

  362. 15:47

    Anything.

  363. 15:49

    The world is yours.

  364. 15:52

    So,

  365. 15:53

    I'm going to wrap it up. I don't know if

  366. 15:55

    you have any questions, but uh if you

  367. 15:57

    want to connect with me, talk about it

  368. 15:58

    more, feel free to to add me on

  369. 16:00

    LinkedIn. And um yeah, I'm going to

  370. 16:04

    thank you guys for your attention. And

  371. 16:07

    uh I'm at a bright data booth if you

  372. 16:09

    want to talk more about it. And thanks

  373. 16:12

    for coming.