1
00:00:03,154 --> 00:00:03,534
Hello.

2
00:00:05,414 --> 00:00:12,454
I want you to zoom out for a second and think about if somebody were to tell you today that

3
00:00:12,454 --> 00:00:18,254
we would take all the information on the internet and then we would train a model on it and

4
00:00:18,254 --> 00:00:22,094
we would use that in order to solve scientific problems that were not on the internet.

5
00:00:22,374 --> 00:00:23,954
I think a lot of people would tell you that that's crazy.

6
00:00:27,414 --> 00:00:31,314
And what we're doing has a lot of parallels to that story.

7
00:00:32,094 --> 00:00:38,574
So if you think about how intelligence works, imagine for a second that you're in the movie

8
00:00:38,574 --> 00:00:41,454
Frozen and you're Olaf, the snowman.

9
00:00:43,354 --> 00:00:48,954
If Olaf was an LLM, and there's a scene which is one of my favorites, which is where he rolls

10
00:00:48,954 --> 00:00:56,174
down a mountain and there's a little stone that pops up and Olaf spreads out, temporarily

11
00:00:56,174 --> 00:00:59,354
becomes a disassembled snowman and then becomes an assembled snowman.

12
00:00:59,794 --> 00:01:05,254
Olaf knows how to do that because Olaf can observe as he is acting.

13
00:01:05,694 --> 00:01:06,794
Olaf can, unlike

14
00:01:06,954 --> 00:01:16,734
LLMs, which autoregressively predict tokens, constantly part of that behavior is that it's constantly observing as it is acting.

15
00:01:17,074 --> 00:01:31,854
And so this is the big difference between, you know, when you see LLMs try to solve robotics problems or when you look at native architectures like we're working on, it's very, very easy to think of LLMs in the demos to see on Twitter as the thing that will

16
00:01:31,854 --> 00:01:37,334
Solve everything, there are fundamental limitations where purpose-built models are a much better

17
00:01:37,334 --> 00:01:37,614
fit.

18
00:01:38,614 --> 00:01:42,134
Today I'm going to talk a little bit about what the other side of that coin looks like.

19
00:01:45,334 --> 00:01:46,994
For us, it all started with Medal.

20
00:01:47,234 --> 00:01:53,374
We built the largest video game recorder in the world, and we think of that as the internet.

21
00:01:54,234 --> 00:01:58,754
People use Medal to record themselves playing tens of thousands of different video games,

22
00:01:58,754 --> 00:02:07,414
And if you think about it, as a human, we can teleoperate a robot with the exact same

23
00:02:07,414 --> 00:02:16,894
brain that we played at video game, and in fact, in many cases, we also can use the same

24
00:02:16,894 --> 00:02:17,314
controller.

25
00:02:17,654 --> 00:02:21,674
So what this means is that the action space of video games, often because of game controllers,

26
00:02:21,834 --> 00:02:25,674
and the action space of teleoperations or the real world, in many cases is the same.

27
00:02:26,774 --> 00:02:32,534
And so every single robot, below certain degrees of freedom, ships with the game controller.

28
00:02:32,734 --> 00:02:34,474
The majority of the supply chains already work this way.

29
00:02:34,714 --> 00:02:39,734
And so we thought, what if you just take all of the action data that we get through the

30
00:02:39,734 --> 00:02:45,394
video recorder, which associates the inputs and the actions with the actual frames, we

31
00:02:45,394 --> 00:02:50,394
label the actions so we understand semantic meaning behind every action, and we use it

32
00:02:50,394 --> 00:02:53,474
to give it to a vision-based model.

33
00:02:53,474 --> 00:02:57,734
What you see here on the left is just an intuitive way of how the model works.

34
00:02:58,394 --> 00:03:02,254
You give the model a set of keys to choose from, essentially a computer use model.

35
00:03:03,154 --> 00:03:06,374
It gives frames as the input, so you can give it a text prompt as well.

36
00:03:07,514 --> 00:03:11,714
And you send it through API, and you say somebody waves at you, wave back.

37
00:03:12,194 --> 00:03:15,454
In the input space, the model understands that you have a grid action available, it

38
00:03:15,454 --> 00:03:18,014
will output the action associated with that.

39
00:03:19,194 --> 00:03:25,874
One way of intuitive understanding how this works, I've obviously simplified here because

40
00:03:25,874 --> 00:03:28,774
it's predicting the next token, not the next word, but bear with me here.

41
00:03:30,854 --> 00:03:35,454
LLMs took large-scale internet data, then trained it to predict the next text token,

42
00:03:35,634 --> 00:03:37,274
and now use it for writing software.

43
00:03:38,294 --> 00:03:43,134
These are all workloads that are really, really well fit for futures that are not simultaneously

44
00:03:43,654 --> 00:03:45,174
unfolding as you are predicting.

45
00:03:46,074 --> 00:03:47,254
Code is turn-based.

46
00:03:47,254 --> 00:03:51,794
You can say something, the model will do something, you can say something again, the model will

47
00:03:51,794 --> 00:03:52,434
do something again.

48
00:03:53,794 --> 00:03:58,374
And you can use it for lots of valuable work, but at the end of the day, because it's turn-based,

49
00:03:59,034 --> 00:04:06,114
it's very, very difficult for it to generalize to environments, especially that include space

50
00:04:06,114 --> 00:04:06,474
and time.

51
00:04:06,634 --> 00:04:14,254
So our models, the way that they work is they're continuously receiving frames or receiving

52
00:04:14,254 --> 00:04:19,374
batches of information from the camera of the robot, or it could be a screen.

53
00:04:19,814 --> 00:04:24,894
And it predicts everything as if it were just a video game.

54
00:04:25,194 --> 00:04:30,254
So you can think of the way that Anthropic focused on just nailing coding.

55
00:04:30,634 --> 00:04:34,154
So they were never frontier and just base text generation.

56
00:04:34,314 --> 00:04:36,594
They always focused on being the best at generating code.

57
00:04:36,954 --> 00:04:40,594
And then the bet that they made is that more people would start bringing more problems

58
00:04:40,594 --> 00:04:41,454
into coding space.

59
00:04:41,454 --> 00:04:53,334
If you think of general intuition, our bet is that you can, as humans, we are generally pixel-to-action converters, plus senses, which in many cases you can convert into pixels as well.

60
00:04:53,754 --> 00:04:56,854
And you can output actions at the same time.

61
00:04:57,034 --> 00:05:00,014
Those actions are, in this case, either keyboard and mouse or game controller.

62
00:05:00,194 --> 00:05:08,774
So the output tokens of the model are essentially equivalent to that of enhanced computer use model with more actions on the output space to internalize it.

63
00:05:08,814 --> 00:05:10,374
So you send frames, and you get actions.

64
00:05:11,454 --> 00:05:18,694
This works, again, because every single thing that you see here can control all of these different embodiments.

65
00:05:18,834 --> 00:05:28,414
So, as humans, we do not know how to predict motor torque.

66
00:05:28,554 --> 00:05:31,674
We do not know what value to send to a machine.

67
00:05:32,474 --> 00:05:39,354
There's this beautiful device that has been made for many years, which actually does that work for us so we can control them.

68
00:05:39,354 --> 00:05:44,494
Most robotics companies today are trying to solve the problem at the end-to-end level,

69
00:05:44,654 --> 00:05:49,714
trying to predict exact poses or exact values in the motor,

70
00:05:49,854 --> 00:05:56,954
instead of just looking and picking up the thing that every robotics company has built

71
00:05:56,954 --> 00:06:05,834
to generalize our very limited abilities to have physics understanding and predict exact values,

72
00:06:06,114 --> 00:06:07,634
and solve for it.

73
00:06:07,634 --> 00:06:10,574
And so our approach is, let's just use that.

74
00:06:12,114 --> 00:06:19,414
And so Medal, at the moment, sees about a billion videos uploaded per year.

75
00:06:19,594 --> 00:06:21,254
This year we're on track for three billion,

76
00:06:22,614 --> 00:06:27,394
which lead to over a trillion and a half action tokens per month.

77
00:06:27,494 --> 00:06:29,314
So just for comparison, to think about the bet,

78
00:06:30,054 --> 00:06:33,514
the total amount of data that Frontier Labs use today to train robotics models,

79
00:06:33,534 --> 00:06:35,894
we collect in, at the moment, half a day.

80
00:06:36,854 --> 00:06:45,174
And so the type of bet, and again, think of the fact that humans can use the exact same

81
00:06:45,174 --> 00:06:49,374
controllers and their eyes to play video games that they can't help upgrade robots.

82
00:06:49,814 --> 00:06:50,934
It is the same thing.

83
00:06:51,174 --> 00:06:57,154
We just bet on scaling pixel space the same way that anthropics killed coding space, and

84
00:06:57,154 --> 00:07:01,314
betting that people will just, as we scale data compute into pixel space, that people

85
00:07:01,314 --> 00:07:03,094
will bring more problems into pixel space.

86
00:07:03,094 --> 00:07:06,154
And so what happens?

87
00:07:08,554 --> 00:07:18,454
of a sudden you get an incredibly general agent that can go and solve problems on computers

88
00:07:18,454 --> 00:07:19,694
in the real world.

89
00:07:21,414 --> 00:07:27,774
in addition, similarly to how Anthropic focuses on doing coding environments, we focus

90
00:07:27,774 --> 00:07:32,114
on world models, which most of you know is for, the Mirai release world models, to then

91
00:07:32,114 --> 00:07:34,474
generate additional environments for these models to learn in.

92
00:07:34,854 --> 00:07:40,334
One way of internalizing a big advantage of this approach against LLMs is if you're an

93
00:07:41,234 --> 00:07:46,734
LLM company and you have to bet on the ability to generate environments for your models to

94
00:07:46,734 --> 00:07:48,114
learn, you have two problems.

95
00:07:48,414 --> 00:07:52,394
One, you actually need to spend the compute to simulate environments, and the more realistic

96
00:07:52,394 --> 00:07:54,654
you try to simulate, the more expensive that gets.

97
00:07:55,634 --> 00:07:58,334
and so imagine you want to learn a new thing

98
00:07:58,334 --> 00:08:00,774
you want to learn that new thing

99
00:08:00,774 --> 00:08:04,894
in the exact input space of the human

100
00:08:04,894 --> 00:08:07,094
so if we are learning from pixels

101
00:08:07,094 --> 00:08:08,554
we're learning by looking around

102
00:08:08,554 --> 00:08:10,254
if I'm trying to communicate something

103
00:08:10,254 --> 00:08:13,954
I want my robot partner to observe that exact same thing

104
00:08:13,954 --> 00:08:18,014
and so what I don't want

105
00:08:18,014 --> 00:08:20,034
is for that to have to go through an LM

106
00:08:20,034 --> 00:08:21,654
for a digital twin to have to be created

107
00:08:21,654 --> 00:08:23,234
for the model to have to get deployed

108
00:08:23,234 --> 00:08:25,774
for it to be a bunch of verifier checks

109
00:08:25,774 --> 00:08:27,074
whether or not things are equivalent

110
00:08:27,074 --> 00:08:29,394
and then for the robot to learn much later.

111
00:08:29,634 --> 00:08:31,074
And so one way of thinking about it is

112
00:08:31,074 --> 00:08:33,554
when you can learn in the native modality of the human,

113
00:08:34,034 --> 00:08:38,374
you can all of a sudden start solving a lot more problems

114
00:08:38,374 --> 00:08:39,094
a lot faster.

115
00:08:39,234 --> 00:08:40,734
What you'll see when these models come out

116
00:08:40,734 --> 00:08:44,994
is they're significantly faster than any LLM

117
00:08:46,014 --> 00:08:47,554
at doing these things.

118
00:08:49,574 --> 00:08:51,594
Because games are so action-dense,

119
00:08:51,594 --> 00:08:57,934
So what you see, right, is one thing that's interesting is you saw, for instance, the model predicting multiple actions at the same time.

120
00:08:59,294 --> 00:09:06,074
We're able to take that action density from gamers and convert them into really complex robotics workloads.

121
00:09:06,234 --> 00:09:13,434
So, for example, if you are operating a drone and you need to make lots of little adjustments, you really, really, really want to start with games.

122
00:09:13,674 --> 00:09:18,494
You don't want to start with teleoperations because the human will just take a lot less actions.

123
00:09:18,654 --> 00:09:19,494
It has a lot less information.

124
00:09:19,494 --> 00:09:21,394
You want a model that can predict really high APM.

125
00:09:23,054 --> 00:09:28,794
And so the APM of the data set puts a lot of these problems all of a sudden in reach

126
00:09:28,794 --> 00:09:30,934
so that it can do a lot of these actions simultaneously.

127
00:09:31,214 --> 00:09:34,914
And as I mentioned, the model doesn't just understand the key it's predicting.

128
00:09:35,094 --> 00:09:38,514
It also understands the semantic meaning associated with that key it's predicting, which is really

129
00:09:38,514 --> 00:09:40,614
important for things like interpretability.

130
00:09:44,034 --> 00:09:46,094
So we will be releasing the models this year.

131
00:09:48,074 --> 00:09:50,254
we are very happy

132
00:09:50,254 --> 00:09:52,314
I think the correct word would be

133
00:09:52,314 --> 00:09:54,534
that we're stunned

134
00:09:54,534 --> 00:09:56,594
by the fact of what these models can do

135
00:09:56,594 --> 00:09:57,574
it is

136
00:09:59,714 --> 00:10:00,534
this is a

137
00:10:00,534 --> 00:10:02,394
live stream so I unfortunately couldn't share

138
00:10:02,394 --> 00:10:04,334
the release videos without

139
00:10:04,334 --> 00:10:06,194
breaking the surprise but if you're in New York

140
00:10:06,194 --> 00:10:07,394
we are also in New York

141
00:10:08,154 --> 00:10:10,274
we have started doing a few demos

142
00:10:10,274 --> 00:10:11,974
of the model so if you can

143
00:10:11,974 --> 00:10:14,054
know somebody on the team reach out and we can start

144
00:10:14,054 --> 00:10:16,734
It's showing you some of the capabilities live.

145
00:10:17,994 --> 00:10:21,794
But we have essentially one model which can play video games at the highest level

146
00:10:21,794 --> 00:10:27,294
and control over 10 types of robots using the exact same primitives.

147
00:10:27,414 --> 00:10:29,674
So pixels in, actions out, completely real time.

148
00:10:30,914 --> 00:10:35,094
And we'll be rolling out early access to these models over the next few months.

149
00:10:35,494 --> 00:10:39,114
We already have some customers that are on them, but it's still private,

150
00:10:39,734 --> 00:10:40,894
so you're not hearing too much about it.

151
00:10:40,894 --> 00:10:49,354
But if you scan this QR code, you'll be able to get the models because we go by who signs

152
00:10:49,354 --> 00:10:50,774
up on the wait list at which times.

153
00:10:51,074 --> 00:10:54,634
So this is a great moment to get in touch.

154
00:10:54,734 --> 00:10:55,714
As I mentioned, we are in New York.

155
00:10:55,834 --> 00:11:01,754
What you'll find is similar to how Jev and a lot of these decision-making models have

156
00:11:02,154 --> 00:11:06,174
kind of opened people's eyes on what you can do when you have a new architectural primitive.

157
00:11:06,174 --> 00:11:06,954
People are ready

158
00:11:07,094 --> 00:11:16,494
to do away with the unpredictability of an LLM and put something in that is more reliable and faster and you understand how it is going to react as opposed to having to work around it.

159
00:11:17,074 --> 00:11:28,234
If that resonates, these models are exactly what you're looking for. And all of a sudden you can start exploring the environments that you generate with LLMs.

160
00:11:28,234 --> 00:11:32,934
All of a sudden, one of the use cases that I'm really, really excited about is, well,

161
00:11:33,434 --> 00:11:37,694
these models, they can take actions as if a human was on the interface.

162
00:11:37,974 --> 00:11:42,374
And so if you're doing UX work and you're trying to understand, is this an intuitive

163
00:11:42,374 --> 00:11:42,854
interface?

164
00:11:43,054 --> 00:11:45,054
Well, this is a very interesting way of solving the problem.

165
00:11:45,294 --> 00:11:49,514
So all these problems that LLMs can't solve because they can't actually explore like a

166
00:11:49,514 --> 00:11:53,794
human, you can use these problems and start using these models for.

167
00:11:55,074 --> 00:12:01,594
So the way that they're going to be released is we still expect you to use an orchestrator model.

168
00:12:01,754 --> 00:12:06,294
So you may use something like Gemini or Astra in order to do long-term planning and everything.

169
00:12:06,454 --> 00:12:08,534
And you'll be able to plug these into your workloads.

170
00:12:09,294 --> 00:12:11,734
And we're actually, the models are so general.

171
00:12:12,054 --> 00:12:19,494
It is genuinely going to, I think, feel like the early LLM days when you have a new toy

172
00:12:19,494 --> 00:12:22,414
and you just want to figure out what it does and all the things you can use it for.

173
00:12:23,454 --> 00:12:25,934
my reaction was like that

174
00:12:25,934 --> 00:12:28,214
when I first started using the checkpoints

175
00:12:28,214 --> 00:12:30,514
and I think you'll have the same reaction

176
00:12:31,694 --> 00:12:33,654
and I think a lot of you will

177
00:12:33,654 --> 00:12:35,094
just like with LLMs

178
00:12:35,094 --> 00:12:37,554
if you're a bit early and you start understanding these primitives

179
00:12:38,114 --> 00:12:39,454
I think you'll understand

180
00:12:40,614 --> 00:12:43,554
that you can even build lots of businesses on top of this

181
00:12:43,554 --> 00:12:45,134
think about for instance

182
00:12:47,874 --> 00:12:49,054
a science lab

183
00:12:50,054 --> 00:12:51,774
all of a sudden if you can reliably

184
00:12:51,774 --> 00:12:56,634
orchestrate robots in space and time, and they can take text instructions and do things.

185
00:12:57,014 --> 00:13:01,874
All of a sudden, it's a recursive problem, right?

186
00:13:01,994 --> 00:13:05,494
Robotics is interesting because one of the big problems of solving robotics is making

187
00:13:05,494 --> 00:13:06,074
more robots.

188
00:13:06,634 --> 00:13:11,234
And so there are all these businesses that, because it's so bottlenecked by intelligence,

189
00:13:11,574 --> 00:13:15,954
are going to emerge just from solving this one simple problem, which is pixels to actions.

190
00:13:17,354 --> 00:13:22,934
And I think if you're early on getting your hands on these, it will have a similar effect

191
00:13:22,934 --> 00:13:25,174
on getting your hands on early LLMs.

192
00:13:27,374 --> 00:13:32,954
I am most excited also, one of the things that we're working on which is interesting

193
00:13:32,954 --> 00:13:41,054
is if you have any friends or family who are in some way need accessibility help for using

194
00:13:41,054 --> 00:13:44,894
their computer, because these models can use computers like a human, all of a sudden, for

195
00:13:44,894 --> 00:13:50,674
For instance, if you only can use one arm, all of a sudden the model can allow you to

196
00:13:50,674 --> 00:13:56,654
play video games because you can just control the mouse and the model is predicting the

197
00:13:56,654 --> 00:13:57,594
actions for your keyboard.

198
00:13:58,194 --> 00:14:02,514
Because the model also understands which actions you're predicting, because that's the input,

199
00:14:02,654 --> 00:14:07,054
it's observing as it's acting, it will actually just not predict the things that you're doing

200
00:14:07,054 --> 00:14:08,974
and it will just predict the keys.

201
00:14:09,694 --> 00:14:13,174
It has all these downstream effects of things that you can build on top.

202
00:14:13,654 --> 00:14:14,414
I'm an engineer.

203
00:14:15,154 --> 00:14:20,234
If you're interested in building in this space, this is going to be the most interesting toy

204
00:14:20,234 --> 00:14:21,034
that comes out this year.

205
00:14:23,194 --> 00:14:24,114
I'm fairly certain.

206
00:14:25,534 --> 00:14:27,894
I also think a lot of the initial usage will sit in simulation.

207
00:14:28,194 --> 00:14:32,034
If a lot of you are into generating environments with Astra or Gemini, it's perfect.

208
00:14:32,194 --> 00:14:34,814
Now all of a sudden you can have something that can explore it and you don't have to

209
00:14:34,814 --> 00:14:35,294
do it yourself.

210
00:14:35,294 --> 00:14:42,394
itself, and it's perfect timing because of all the environment generation capabilities

211
00:14:42,394 --> 00:14:43,514
going into these models.

212
00:14:45,334 --> 00:14:48,534
So really quickly before I wrap up, I want to talk about the team.

213
00:14:49,774 --> 00:14:51,354
We are based in New York City.

214
00:14:53,154 --> 00:14:56,414
The Medal team and the general intuition team is here in New York City, so if you're here

215
00:14:56,414 --> 00:15:02,074
you should reach out to us and we can both show you the models, especially if you want

216
00:15:02,074 --> 00:15:04,854
to join the team, I think we'll prioritize even higher.

217
00:15:07,834 --> 00:15:12,474
But we really are looking for people who, A,

218
00:15:12,634 --> 00:15:14,814
want to understand these models as much as we do.

219
00:15:14,974 --> 00:15:16,514
So if you're really, really good at generating

220
00:15:16,514 --> 00:15:19,514
novel environments, typically if you have some sort

221
00:15:19,514 --> 00:15:21,994
of research background, but I don't have a research

222
00:15:21,994 --> 00:15:22,874
background, so that's OK.

223
00:15:23,254 --> 00:15:25,814
If you come from an engineering background

224
00:15:25,814 --> 00:15:28,314
and you're very happy to learn, this is fine too.

225
00:15:29,034 --> 00:15:38,314
And ideally what we want is somebody who has experienced shipping end-to-end AI products,

226
00:15:38,974 --> 00:15:40,194
but it doesn't have to be on the research side.

227
00:15:40,274 --> 00:15:43,294
So we just want to hire people that understand the entire thing.

228
00:15:45,554 --> 00:15:46,194
Let's see.

229
00:15:47,654 --> 00:15:48,654
We have three teams.

230
00:15:48,794 --> 00:15:52,994
So one is role modeling, the other is the policy, so the agent, and the other one is

231
00:15:52,994 --> 00:15:57,174
working on essentially the data generation and the environment generation.

232
00:15:57,174 --> 00:16:01,214
If you are an engineer, the best one to join is the environment generation team.

233
00:16:02,214 --> 00:16:05,054
And then you grow into getting access to everyone.

234
00:16:07,574 --> 00:16:09,074
I'm very, very excited for November.

235
00:16:09,414 --> 00:16:13,254
I really hope that you'll be on the lookout.

236
00:16:13,994 --> 00:16:16,994
As I mentioned, it will be either late November or early December is the current planning.

237
00:16:18,414 --> 00:16:19,274
One more thing.

238
00:16:19,594 --> 00:16:22,174
We are going to organize three hackathons.

239
00:16:22,874 --> 00:16:25,474
The first is in New York City, it will be mid-November.

240
00:16:29,854 --> 00:16:32,994
If you sign up for the partner wait list, we will send out some emails so everybody

241
00:16:32,994 --> 00:16:33,654
can get access.

242
00:16:34,754 --> 00:16:36,834
We hope to see everybody at the hackathons.

243
00:16:36,974 --> 00:16:40,134
I think it will be the first time to get your hands on them the way we are playing with

244
00:16:40,134 --> 00:16:40,834
them internally now.

245
00:16:42,574 --> 00:16:44,534
And there will be one in New York City and one in London.

246
00:16:44,794 --> 00:16:47,774
So if you are in either of those cities or want to travel to either of those cities,

247
00:16:47,774 --> 00:16:50,174
those would be good times to meet the team.

248
00:16:52,594 --> 00:16:53,434
Yeah, let's see.

249
00:16:53,954 --> 00:16:55,914
I have time for one or two questions,

250
00:16:55,974 --> 00:16:58,214
and then I think we can wrap.

251
00:17:01,094 --> 00:17:01,694
Let's see.

252
00:17:07,174 --> 00:17:08,634
How is it different from JEPA?

253
00:17:09,894 --> 00:17:10,994
Yeah, how is it different from JEPA?

254
00:17:11,094 --> 00:17:13,614
So JEPA doesn't train alongside the action labels,

255
00:17:13,654 --> 00:17:15,894
so they bet on inferred action labels for the most part,

256
00:17:16,314 --> 00:17:18,214
which I think, one of the ways of thinking about it

257
00:17:18,214 --> 00:17:20,374
is JEPA I think in the limit is the more correct approach

258
00:17:20,374 --> 00:17:22,074
but it's just a lot less practical in the short term

259
00:17:22,074 --> 00:17:24,014
because you need to bet on a separate

260
00:17:24,014 --> 00:17:26,094
scaling law on the action inference itself

261
00:17:27,254 --> 00:17:27,594
and

262
00:17:29,314 --> 00:17:30,294
we really like

263
00:17:30,294 --> 00:17:32,214
JEPA as a concept, we think it's a

264
00:17:33,014 --> 00:17:34,294
you know, we use some of the

265
00:17:34,294 --> 00:17:34,954
things in the models

266
00:17:36,174 --> 00:17:36,514
and

267
00:17:37,654 --> 00:17:39,614
but at the end of the day it's an approach

268
00:17:39,614 --> 00:17:42,274
not a product and I think we'll

269
00:17:42,274 --> 00:17:44,334
you know, borrow

270
00:17:44,334 --> 00:17:45,394
concepts both ways

271
00:17:47,934 --> 00:17:49,334
one way of thinking about it is

272
00:17:50,694 --> 00:17:52,194
because we have so much action data

273
00:17:53,114 --> 00:17:54,374
let's say you're flying a plane

274
00:17:54,374 --> 00:17:56,374
and you're moving the rudder in the back

275
00:17:56,374 --> 00:17:58,954
there's no way of being able to

276
00:17:58,954 --> 00:18:00,634
tell that from the video frames

277
00:18:00,634 --> 00:18:01,894
it's very very difficult

278
00:18:01,894 --> 00:18:04,014
so our mode on having the action data

279
00:18:04,014 --> 00:18:05,774
really transfers better to

280
00:18:06,374 --> 00:18:07,994
action sequences than it does

281
00:18:07,994 --> 00:18:10,114
to world models and to generating

282
00:18:10,114 --> 00:18:12,054
the pixels and that's why

283
00:18:12,054 --> 00:18:13,934
we just went all in on training these general

284
00:18:13,934 --> 00:18:15,274
foundation models for action predictions

285
00:18:15,274 --> 00:18:17,734
and we use all sorts of techniques to do world models.

286
00:18:18,014 --> 00:18:19,254
You know, we are own.

287
00:18:19,534 --> 00:18:21,794
We really like other companies building them as well.

288
00:18:24,034 --> 00:18:25,994
So, great.

289
00:18:26,314 --> 00:18:27,834
One more and then I've got to run, I think.

290
00:18:28,534 --> 00:18:28,774
Great.

291
00:18:32,434 --> 00:18:32,934
All right.

292
00:18:33,194 --> 00:18:33,654
Thank you, everyone.

293
00:18:34,374 --> 00:18:34,774
Have a good one.