WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=z1dqv74SpUs

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/z1dqv74SpUs.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.400 --> 00:00:14.440
My name is Sydney.

s3
00:00:14.440 --> 00:00:19.160
I'm the CTO and founder of Lemon Slice.

s4
00:00:19.160 --> 00:00:27.880
And Lemon Slice is on a mission to break the Avatar Turing test.

s5
00:00:27.880 --> 00:00:39.480
What we mean by this is making an Avatar that is indistinguishable from a human on a video call.

s6
00:00:39.480 --> 00:00:44.440
And all of this is of course making the Avatar photo-realistic.

s7
00:00:44.440 --> 00:00:53.560
But there's actually a long tail of technical problems that we care a lot about that I'll be talking about today that we're planning to solve.

s8
00:00:53.560 --> 00:01:08.120
And those problems are things like getting the emotions right, getting the object interactions right, getting the like micro-expressions right, and even figuring out an internal state for these Avatars

s9
00:01:08.120 --> 00:01:11.920
so that they feel like a real human being.

s10
00:01:11.920 --> 00:01:22.560
So what I'm going to do in this Oh, this What I'm going to do in this talk is talk about our technical bet to solving the Avatar Turing test,

s11
00:01:22.560 --> 00:01:33.800
what we've built today, and the technical the hard technical problems we've solved today, and our bet on like how we think this problem will be solved in the long term.

s12
00:01:33.800 --> 00:01:41.840
But before we dive into that I want to give you guys a sense for what our product is and how it works.

s13
00:01:41.840 --> 00:01:45.520
So actually yesterday was a was a pretty big and exciting day for us.

s14
00:01:45.520 --> 00:01:48.360
We had a big big launch yesterday.

s15
00:01:48.360 --> 00:01:57.080
We had a partnership with Microsoft that brought an ex-US president to life, uh Roosevelt.

s16
00:01:57.080 --> 00:02:07.560
And they built his presidential library and inside of the presidential library uh they did a replica of the Oval Office and you can walk in and you can have a full interaction

s17
00:02:07.560 --> 00:02:10.440
uh with Teddy Roosevelt in real time.

s18
00:02:10.440 --> 00:02:18.680
Uh so definitely not uh Turing test solved yet, but like a big milestone uh towards trying to solve solve that problem.

s19
00:02:18.680 --> 00:02:24.320
Uh and the coolest part of all of this was uh we had a special guest.

s20
00:02:24.320 --> 00:02:28.280
Uh but hate hate him or love him, uh he showed up.

s21
00:02:28.280 --> 00:02:36.880
Uh so we had President Trump uh who came to the opening uh of this ceremony uh and interacted uh with our avatar.

s22
00:02:36.880 --> 00:02:44.560
So uh we actually got some footage, real footage, uh not not fake uh AI-generated footage of Trump uh interacting.

s23
00:02:44.560 --> 00:02:48.400
So let me let me show you guys this real quick.

s24
00:02:50.480 --> 00:02:53.160
your greatest achievement?

s25
00:02:57.320 --> 00:02:59.840
Right then, yes.

s26
00:02:59.880 --> 00:03:01.680
What is that answer?

s27
00:03:01.680 --> 00:03:04.560
The canal stands as one of my proudest battles, no question.

s28
00:03:04.560 --> 00:03:06.959
But greatness is a strange thing.

s29
00:03:06.959 --> 00:03:10.680
It isn't always the biggest or boldest job that matters most.

s30
00:03:10.680 --> 00:03:17.680
Panama Canal showed what America could achieve if we held steady and acted fast when the world dragged its feet.

s31
00:03:17.680 --> 00:03:22.519
That said, I measure my greatest work by the lives improved.

s32
00:03:22.519 --> 00:03:29.000
Mark set aside, food and drugs made safe, the square deal given to all, not just to a few.

s33
00:03:29.000 --> 00:03:39.720
Still, when I stood in the mud watching those steam shovels, knowing ships would pass through, changing the world's map forever, I felt I'd left a mark that would last.

s34
00:03:39.720 --> 00:03:40.680
Perfect.

s35
00:03:40.680 --> 00:03:44.840
Uh so you guys get a sense uh of of the experience.

s36
00:03:44.840 --> 00:03:54.600
Um the funny thing is he was scheduled to be there for, you know, a quick minute, one interaction, and he actually stayed for 10 minutes.

s37
00:03:54.600 --> 00:04:01.240
He left He left the room and then came back again to talk to Teddy Teddy Roosevelt even more.

s38
00:04:01.240 --> 00:04:04.440
So, you can see there like the avatars there.

s39
00:04:04.440 --> 00:04:05.360
It's full body avatar.

s40
00:04:05.360 --> 00:04:06.040
It has hands.

s41
00:04:06.040 --> 00:04:06.760
It does actions.

s42
00:04:06.760 --> 00:04:08.480
It does movement.

s43
00:04:08.480 --> 00:04:12.800
And all of this is obviously done in real time.

s44
00:04:12.800 --> 00:04:20.600
I'll talk a little bit more about like the why we think this is an important problem later in the talk.

s45
00:04:20.600 --> 00:04:25.160
Um but a big part of it is actually related to this interaction.

s46
00:04:25.160 --> 00:04:30.640
Um humans like struggle to to to focus and pay attention.

s47
00:04:30.640 --> 00:04:35.440
It's It's often hard for a lot of people spend a bunch of time reading.

s48
00:04:35.440 --> 00:04:39.440
And even voice it's hard to really lock in and understand what's happening.

s49
00:04:39.440 --> 00:04:47.040
So, we think like we're biologically wired to best like understand things when there's a visual component as well.

s50
00:04:47.040 --> 00:04:58.040
So, our bet here is in the long term we think most interactions between AI and humans will have a visual visual layer.

s51
00:04:58.040 --> 00:05:01.080
And we're building that visual layer.

s52
00:05:02.160 --> 00:05:11.760
Um so, now I want to talk about our approach to building these avatars.

s53
00:05:11.760 --> 00:05:17.600
It's a very different approach than what most other avatar companies use.

s54
00:05:17.600 --> 00:05:25.080
Essentially what we do is we take these world models and we focus them on humans.

s55
00:05:25.080 --> 00:05:36.760
Um and the reason we take this bet is even though it's harder to get the initial model working, it's harder to train the model, it's harder to deploy the model.

s56
00:05:36.760 --> 00:05:47.640
Once you have a model, you get all of these nice emergent properties that we all know about where you just very easily can solve things like full body movement,

s57
00:05:47.640 --> 00:05:58.000
like object interactions, like movements in the scene, all the way down to the micro micro expressions as well and emotions.

s58
00:05:58.000 --> 00:06:06.000
And all of those kind of don't come for free, but come more for free than when you use the other approaches that other people use.

s59
00:06:06.000 --> 00:06:12.840
So, let me guys let me show you guys real quick the product in action, so you guys can get a sense.

s60
00:06:14.560 --> 00:06:20.640
Uh so Let me expand this a little bit.

s61
00:06:32.720 --> 00:06:33.919
How is it going?

s62
00:06:33.919 --> 00:06:36.640
I am here to listen and help with anything you need.

s63
00:06:36.640 --> 00:06:41.880
We can chat about something in particular or play a game if you would like.

s64
00:06:41.880 --> 00:06:48.240
So, the lipstick might be off here a little bit, uh but that's just because of the AV setup here.

s65
00:06:48.240 --> 00:06:55.400
Uh but as you can see, um we have actions, they can wave, they really can do any actions.

s66
00:06:55.400 --> 00:07:05.080
Uh we have movement, we have hands, um and the way the way this works is it it just can use any single image to create the avatar.

s67
00:07:05.080 --> 00:07:09.880
Uh and so yes, you can have photorealistic, uh you also can have Pixar, you also can have cartoons.

s68
00:07:09.880 --> 00:07:18.840
Uh it just literally takes a single image and um Show you this.

s69
00:07:18.840 --> 00:07:27.280
Oh, no. Perfect.

s70
00:07:27.280 --> 00:07:33.080
Uh and I'll I'll show you kind of a quick interaction here as well and then and then we can talk about how we built this.

s71
00:07:33.080 --> 00:07:42.840
Um here I within the same video call and on the same kind of like inference setup, uh we can easily change clothes, we can change the scene she's in.

s72
00:07:42.840 --> 00:07:46.800
And I can talk to it, too.

s73
00:07:46.800 --> 00:07:47.960
I'm not going to do it here.

s74
00:07:47.960 --> 00:07:49.000
be trying to say story.

s75
00:07:49.000 --> 00:07:52.680
Would you like to tell me a story or hear one?

s76
00:07:52.680 --> 00:07:56.880
Yeah, so you can see here, too, like that there's physics on the earrings.

s77
00:07:56.880 --> 00:07:57.960
They move.

s78
00:07:57.960 --> 00:07:59.240
The water moves.

s79
00:07:59.240 --> 00:08:08.400
So, like the the video model has a good understanding of of physics and like can use that to like make the avatar feel more realistic.

s80
00:08:10.000 --> 00:08:12.000
Perfect.

s81
00:08:14.280 --> 00:08:20.040
Now, the way this is used, just for for kind of everybody's information, is we're mostly the API layer.

s82
00:08:20.040 --> 00:08:21.880
So, we provide an API.

s83
00:08:21.880 --> 00:08:23.760
People bring their own LLM.

s84
00:08:23.760 --> 00:08:27.640
People bring their own usually like voices.

s85
00:08:27.640 --> 00:08:32.760
And then we're the the visual layer on top of it.

s86
00:08:32.760 --> 00:08:36.000
So, people use this for language learning.

s87
00:08:36.000 --> 00:08:38.120
People use it for like AI sales call.

s88
00:08:38.120 --> 00:08:47.880
Basically, anything today where you have a voice agent, you now also kind of a video agent if it's on your screen or on your computer.

s89
00:08:48.440 --> 00:08:50.440
Great.

s90
00:08:50.560 --> 00:08:54.839
So, now let's talk about how we built this today.

s91
00:08:54.839 --> 00:08:57.040
This is our approach.

s92
00:08:57.040 --> 00:09:06.320
And can't talk about all the technical details, but I'm going to try to go give you enough information to make it interesting.

s93
00:09:06.320 --> 00:09:11.680
So, what we do is we get started with training our own video DIT model.

s94
00:09:11.680 --> 00:09:16.000
The big thing here for us that really matters is the audio.

s95
00:09:16.000 --> 00:09:23.680
The audio turns out to be very important for getting emotions right and the facial expressions right.

s96
00:09:23.680 --> 00:09:25.760
People care about that a lot.

s97
00:09:25.760 --> 00:09:31.160
So, we really focus on getting great audio data that we can train on.

s98
00:09:31.160 --> 00:09:37.440
And then actually also getting the the encoders very right.

s99
00:09:37.440 --> 00:09:47.480
Like most audio encoders today are trained on basically audiobooks, which is very monotone, very simple, don't have a lot of emotions.

s100
00:09:47.480 --> 00:09:59.280
So, if you want to have a very expressive model, you can't use those audio encoders and you actually have to spend a lot of time like getting the audio embeddings right so that the video model is super expressive.

s101
00:09:59.280 --> 00:10:01.760
So, so that's kind of like step one for us.

s102
00:10:01.760 --> 00:10:04.320
And this is basically just a video model.

s103
00:10:04.320 --> 00:10:11.640
You know, people can call this a world model, but but it's basically a video model that understands the physics of the world.

s104
00:10:11.640 --> 00:10:17.280
And now the big challenge is how do you take that video model and how do you make it real time and interactive?

s105
00:10:17.280 --> 00:10:22.680
And so, the way that works is the interactivity part is interesting.

s106
00:10:22.680 --> 00:10:31.200
So, usually video models are bidirectional, so they can look into the past, but they actually also can look into the future and like what's about to come.

s107
00:10:31.200 --> 00:10:39.560
And they basically generate videos all at the same time by looking at all of the the latents that that are being generated.

s108
00:10:39.560 --> 00:10:41.400
For us, you can't do that.

s109
00:10:41.400 --> 00:10:43.600
You only can look into the past.

s110
00:10:43.600 --> 00:10:54.120
And so, here on this like make interactive, you can see that we basically train a model with an attention mask so that the model can only look into the past.

s111
00:10:54.120 --> 00:11:01.360
So, when you do inference, it never can see the future because the future doesn't exist because like you haven't given it those inputs yet.

s112
00:11:01.360 --> 00:11:06.880
So, for example, it doesn't see audio in the future, it only see what has has been said in the past.

s113
00:11:06.880 --> 00:11:11.440
And then, once you have that, we make it real time.

s114
00:11:11.440 --> 00:11:25.000
And there's a bunch of things that goes in here, but the biggest thing to speed up video generation is usually you spend like a bunch of steps denoising these video models.

s115
00:11:25.000 --> 00:11:26.160
So, like let's say 30 steps.

s116
00:11:26.160 --> 00:11:32.120
You spend 30 steps like removing the noise to generate the beautiful beautiful videos.

s117
00:11:32.120 --> 00:11:36.320
And what we need to do is go from like 30 steps, bring it out to one step.

s118
00:11:36.320 --> 00:11:42.240
So you basically just in a single step go from like pure noise to a pure video model to make it make it real time.

s119
00:11:42.240 --> 00:11:46.600
So those are the the two big things that like make it interactive and real time.

s120
00:11:46.600 --> 00:11:58.280
Now the hard part here, like what makes this difficult is this problem here on the left, which is called error accumulation, which is a problem anybody in like real-time video generation

s121
00:11:58.280 --> 00:12:00.480
and world model is familiar with.

s122
00:12:00.480 --> 00:12:04.560
I'll show you this video and then we'll talk about it briefly.

s123
00:12:04.760 --> 00:12:08.440
And ophthalmology at Stanford School of Medicine.

s124
00:12:08.440 --> 00:12:18.960
So you can see I mean that's a very bad example of error accumulation, but you can see as the video continues to generate more errors introduced.

s125
00:12:18.960 --> 00:12:33.200
And the reason for that is because since you only can look backwards, you're looking actually backwards at at videos that you've previously generated, but each video you generate or each video block you generate has some error in it.

s126
00:12:33.200 --> 00:12:38.680
So now you're looking in the past, you're looking at the error, you're adding more error to it, and then just the error compounds over time.

s127
00:12:38.680 --> 00:12:45.240
And this is especially hard because I mean ideally these these video models are endless.

s128
00:12:45.240 --> 00:12:56.200
Like the teddy avatar is generating continuously non-stop frame by frame for 8 hours straight with like no reset throughout the entire process.

s129
00:12:56.200 --> 00:12:58.720
We have another one that's going to be generating for 16 hours straight.

s130
00:12:58.720 --> 00:13:03.040
So it's a very hard problem to not have any error accumulation over long periods of time.

s131
00:13:03.040 --> 00:13:14.360
And um we came up with a new way to solve this problem that is different to the best of our knowledge than what everybody else does today.

s132
00:13:14.360 --> 00:13:21.560
And so I can't talk about that yet, but because of that we can basically generate these very long videos.

s133
00:13:21.560 --> 00:13:28.280
I mean you guys can can try it out that essentially have no error accumulation, no noticeable error accumulation in them.

s134
00:13:28.280 --> 00:13:31.760
Um Great.

s135
00:13:31.760 --> 00:13:42.760
Um and so the last two hard problems uh we worked on to enable this is a hardware optimization.

s136
00:13:42.760 --> 00:13:46.640
Um There's There's a bunch of stuff that goes into this.

s137
00:13:46.640 --> 00:13:51.800
The big thing here is uh making it low cost enough.

s138
00:13:51.800 --> 00:13:54.360
Uh so you actually can use this.

s139
00:13:54.360 --> 00:13:59.000
Like usually when you generate an AI video, it's 5 seconds and you just share that.

s140
00:13:59.000 --> 00:14:05.000
Uh we have to generate minutes and hours of AI AI video and still have it be cost-effective.

s141
00:14:05.000 --> 00:14:14.720
Uh and so the cool cool thing here is we've been able to make the models small enough and efficient enough so that the costs are about the same as a voice model.

s142
00:14:14.720 --> 00:14:16.400
Which is crazy to me.

s143
00:14:16.400 --> 00:14:22.560
Cuz like think about the a voice model, how much data is streamed there compared to a video model.

s144
00:14:22.560 --> 00:14:27.440
Like it it's much more pixel heavy in a video model than with voice.

s145
00:14:27.440 --> 00:14:29.880
Uh and the costs costs are about similar.

s146
00:14:29.880 --> 00:14:35.400
Um so the cool thing there is this enables us to do uh consumer use cases.

s147
00:14:35.400 --> 00:14:43.920
And a lot of our customers are consumer companies that uh basically use the video for consumer and like entertainment applications.

s148
00:14:43.920 --> 00:14:48.360
And then the final thing to mention here is the model hardness.

s149
00:14:48.360 --> 00:14:57.200
Um I feel like the model hardness is something that is often overlooked but is actually super important and super hard.

s150
00:14:57.200 --> 00:15:01.600
Uh like getting the model hardness right uh is a huge technical challenge for us.

s151
00:15:01.600 --> 00:15:06.320
And I feel like a lot of our value actually in like productizing this is in the model hardness.

s152
00:15:06.320 --> 00:15:11.200
And the way to think about it is you just have a bunch of separate threads.

s153
00:15:11.200 --> 00:15:15.760
Uh all of it is managing real data streaming through our system.

s154
00:15:15.760 --> 00:15:21.800
And you have basically a bunch of stuff you do on a GPU and a bunch of stuff you do on a CPU.

s155
00:15:21.800 --> 00:15:26.680
And you have to orchestrate this perfectly in a way that like the video always remains real time.

s156
00:15:26.680 --> 00:15:29.800
There is never any stutter that happens inside of the video.

s157
00:15:29.800 --> 00:15:35.560
Um and this is especially hard when you have things like interrupts, you have queues, you're buffering data, you have to clean the queues.

s158
00:15:35.560 --> 00:15:40.280
It's all like getting this orchestration right at production at scale has been a ton of work.

s159
00:15:40.280 --> 00:15:49.280
And honestly, I feel like over time a lot more of the value uh of like the things we build will be in like figuring out the model hardness.

s160
00:15:49.280 --> 00:15:53.080
I think it's especially true for like any real-time applications.

s161
00:15:53.080 --> 00:15:55.320
Um cool.

s162
00:15:55.320 --> 00:16:00.960
So, this is This is our technical approach of like commercializing these uh world models for avatars.

s163
00:16:00.960 --> 00:16:05.120
Uh let me talk a little bit about like what we're actually working on actively now.

s164
00:16:05.120 --> 00:16:15.920
Um so, the big thing we want to enable next is having more of an emotional engine uh that's driving these avatars.

s165
00:16:15.920 --> 00:16:23.280
Uh having them be more more aware of what's actually happening inside of the conversation, and then be able to react to what's happening.

s166
00:16:23.280 --> 00:16:30.600
So, react means uh the right the right emotions at the right time for the right duration, and same same for the actions.

s167
00:16:30.600 --> 00:16:36.040
Um and um today when you're when you interact with these avatars, you can try.

s168
00:16:36.040 --> 00:16:42.520
It still feels a little bit I mean, it definitely still feels um uh awkward.

s169
00:16:42.520 --> 00:16:46.800
Uh and I think a big part of that is uh they don't emotionally react to you.

s170
00:16:46.800 --> 00:16:48.240
They they're not listening to you.

s171
00:16:48.240 --> 00:16:52.800
They're not like emoting in the right Um and so, this is what we're trying to solve.

s172
00:16:52.800 --> 00:17:05.319
And so, let me show you guys some of um kind of the the new videos we're able to create uh with with basically basically better emotional control.

s173
00:17:05.319 --> 00:17:07.520
Um so, this is a real-time video.

s174
00:17:07.520 --> 00:17:09.120
Oh my god.

s175
00:17:09.120 --> 00:17:10.240
Goal!

s176
00:17:10.240 --> 00:17:12.079
Goal!

s177
00:17:12.079 --> 00:17:13.760
We won the game.

s178
00:17:13.760 --> 00:17:15.600
I can't believe it.

s179
00:17:15.600 --> 00:17:16.839
We won!

s180
00:17:16.839 --> 00:17:18.280
Oh my god.

s181
00:17:18.280 --> 00:17:20.160
Oh my god.

s182
00:17:20.160 --> 00:17:21.800
Oh my god.

s183
00:17:21.800 --> 00:17:24.920
So you can see like a lot like the emotions just matter a lot.

s184
00:17:24.920 --> 00:17:28.439
They make a huge difference in like connecting with this avatar.

s185
00:17:28.439 --> 00:17:31.120
You feel you feel much more connected with that avatar.

s186
00:17:31.120 --> 00:17:39.680
So that's one thing and then the neck this is the next generation model not in real time yet but there's just a lot more

s187
00:17:39.680 --> 00:17:43.080
I think the most important thing is just to take a moment and breathe.

s188
00:17:43.080 --> 00:17:45.080
Life gets so busy, you know.

s189
00:17:45.080 --> 00:17:46.840
Do you think you'll stay there for a while?

s190
00:17:46.840 --> 00:17:50.040
Yeah, I really think I need to stay right here for a long time.

s191
00:17:50.040 --> 00:17:55.160
So there there'll be a lot more basically natural interaction.

s192
00:17:56.400 --> 00:17:57.280
Sorry.

s193
00:17:57.280 --> 00:18:02.880
There'll be a lot more natural interactions with themselves and then also with their environment.

s194
00:18:02.880 --> 00:18:07.280
And the big thing here is like our model can already do this today.

s195
00:18:07.280 --> 00:18:09.240
It has the capabilities to do this.

s196
00:18:09.240 --> 00:18:17.720
It's just not controllable enough to like make it real time with the conversation and not deterministic enough to to make it useful with the conversation.

s197
00:18:17.720 --> 00:18:29.760
So a big part of our work here is like making these actions actually controllable at the right right time and to do that we're building this emotion engine that is just literally predicting based on like the audio input

s198
00:18:29.760 --> 00:18:37.880
from the avatar and the text input for what the avatar is going to say what the action they should be doing at this moment in time.

s199
00:18:37.880 --> 00:18:49.000
So target launch for this next model is and and basically one to two months is what we're targeting.

s200
00:18:49.720 --> 00:18:50.360
Awesome.

s201
00:18:50.360 --> 00:18:54.480
Oh All right, great.

s202
00:18:54.480 --> 00:18:56.720
So that's where we are today.

s203
00:18:56.720 --> 00:19:00.720
Um I just want to talk briefly about let's see what the time is.

s204
00:19:00.720 --> 00:19:02.960
Almost done.

s205
00:19:03.560 --> 00:19:08.680
I want to talk briefly about where we see this heading.

s206
00:19:08.680 --> 00:19:21.360
So I strongly believe that in the end uh there'll be a single model um that is the EQ layer for AI.

s207
00:19:21.360 --> 00:19:29.040
Uh the way you can think about this model is the model will literally take in directly uh the user video and the user audio.

s208
00:19:29.040 --> 00:19:39.720
So, directly the use the you know, the user talking to you like you are in a FaceTime uh and put out the avatar video and audio and do all of that in an end-to-end model

s209
00:19:39.720 --> 00:19:41.880
a in a single model end-to-end.

s210
00:19:41.880 --> 00:19:58.520
Um internally inside of the model it will do the audio understanding generate what it's going to say uh model its own internal state like own internal emotional state uh and then based on all of that

s211
00:19:58.520 --> 00:20:03.160
produce an output video and audio.

s212
00:20:03.160 --> 00:20:04.880
And all of this will be trained in one model.

s213
00:20:04.880 --> 00:20:06.160
Like this will happen.

s214
00:20:06.160 --> 00:20:11.120
Uh we're already seeing like early papers that are like doing proof proofs of concepts around this.

s215
00:20:11.120 --> 00:20:17.360
What we're not saying is that this EQ model will be very intelligent.

s216
00:20:17.360 --> 00:20:30.440
Uh it'll be very it'll have very high EQ and it'll be very good at like interacting with people uh but there'll be a separate model uh that will drive the IQ that will basically give this EQ model input

s217
00:20:30.440 --> 00:20:33.800
to like do all the magical things that AI can do today.

s218
00:20:33.800 --> 00:20:38.000
Like do the tool calling, do the deep thinking, do all the intelligent stuff.

s219
00:20:38.000 --> 00:20:40.240
Uh but there will be an end-to-end layer up top.

s220
00:20:40.240 --> 00:20:55.920
And so, that's that's our bet like we're working towards towards building that uh and um we can talk more about like some of the benefits of this approach but I feel strongly that within two or three years you'll be seeing these kinds of end-to-end

s221
00:20:55.920 --> 00:20:58.720
EQ models coming on the market.

s222
00:20:58.720 --> 00:21:01.120
Um that's it.

s223
00:21:01.120 --> 00:21:05.920
Um Well, maybe take some questions if there's any questions.

s224
00:21:05.920 --> 00:21:08.320
Um yeah, go ahead.

s225
00:21:08.320 --> 00:21:18.480
Um so Mhm.

s226
00:21:31.160 --> 00:21:32.160
In the future?

s227
00:21:32.160 --> 00:21:40.520
Well, I think in the future in the long term, uh it's going to be more of an internal state uh where the model will will have an internal state in latent space

s228
00:21:40.520 --> 00:21:50.000
that like defines what the emotional state is or the goals are of the model and they will be more implicit.

s229
00:21:50.000 --> 00:21:53.320
So, they won't be human interpretable necessarily.

s230
00:21:53.320 --> 00:21:59.640
Um today today basically the way we think about actions and emotions is basically words.

s231
00:21:59.640 --> 00:22:04.760
If you can describe it in words, you can generate those actions and emotions.

s232
00:22:17.000 --> 00:22:18.080
Exactly.

s233
00:22:18.080 --> 00:22:20.160
Exactly and track that over time.

s234
00:22:20.160 --> 00:22:23.160
Right, like have an internal state that tracks that over over some history of time.

s235
00:22:23.160 --> 00:22:25.440
But yes, exactly.

s236
00:22:31.080 --> 00:22:33.080
Yes.

s237
00:22:36.520 --> 00:22:38.240
[laughter]

s238
00:22:38.240 --> 00:22:41.000
Uh there's a big debate about that.

s239
00:22:41.000 --> 00:22:42.480
Uh Twit- Twitter is debating.

s240
00:22:42.480 --> 00:22:53.480
Gavin Newsom said uh that he definitely doesn't know uh and you know, uh he got tricked by like the ghost of of uh of Teddy Roosevelt.

s241
00:23:06.720 --> 00:23:11.080
You're asking about what the way method is or the test is?

s242
00:23:11.080 --> 00:23:12.840
Yeah, that's a really good point.

s243
00:23:12.840 --> 00:23:23.600
We haven't done this yet, but we're uh in the process of figuring out our own version of the Turing test for these avatars, which will just include real people.

s244
00:23:23.600 --> 00:23:27.000
Uh and uh we're planning to run that this year.

s245
00:23:27.000 --> 00:23:36.600
Um and we won't pass it this year, but it'll just be cool for us to track it over time to see when we can actually solve the Turing test

s246
00:23:36.600 --> 00:23:42.120
uh and ideally allow other people to also go through the test as well.

s247
00:23:42.120 --> 00:23:47.400
But yeah, we plan to we've planned to create a test and then publish around around the test.

s248
00:23:50.000 --> 00:23:52.200
Oh, yeah.

s249
00:24:08.560 --> 00:24:11.440
Uh not for us.

s250
00:24:11.440 --> 00:24:13.760
Like that's not what we plan to do.

s251
00:24:13.760 --> 00:24:17.040
Um maybe people can use our technology to do that.

s252
00:24:17.040 --> 00:24:19.440
Uh but that's not our our goal.

s253
00:24:19.440 --> 00:24:33.000
Uh but you could imagine uh people using our technology, giving it the right context, having the right harness around it, uh and then having the nice like IQ brain around it to do exactly what you're saying.

s254
00:24:33.000 --> 00:24:36.280
But like we we don't want to solve that problem.

s255
00:24:53.880 --> 00:24:54.880
Yeah.

s256
00:24:54.880 --> 00:24:58.560
Um it's still we would still like it to be cheaper.

s257
00:24:58.560 --> 00:25:01.360
Um, the good news is that it is Okay, I'm almost done.

s258
00:25:01.360 --> 00:25:02.560
Last question.

s259
00:25:02.560 --> 00:25:07.480
Um, the good news is it is surprisingly like going into this we didn't know how cheap this would be.

s260
00:25:07.480 --> 00:25:09.720
I've been very surprised at how inexpensive it is.

s261
00:25:09.720 --> 00:25:16.600
Again, like the cost of this is at the same level as an audio model in terms of what we charge for it.

s262
00:25:16.600 --> 00:25:20.240
Um, and we're all hoping for this costs to go down.

s263
00:25:20.240 --> 00:25:24.360
Par- part of it is for consumer you just need very low costs.

s264
00:25:24.360 --> 00:25:29.840
Uh, and then part of it is that as the cost goes down we can do higher resolution which will help us a lot.

s265
00:25:29.840 --> 00:25:36.600
Um, but I guess the way we're thinking about it is there's like algorithm improvements and hardware improvements.

s266
00:25:36.600 --> 00:25:49.680
Um, and the what has been true for some time now is things have been improving on all vectors faster than expected and we're just hoping that continues basically.

s267
00:25:49.680 --> 00:26:00.280
I think there'll also be very cool architectural updates to to move to more of like a token approach instead of a diffusion approach that will make video like this type of video generation way cheaper.

s268
00:26:00.280 --> 00:26:02.560
Um, that's it for me.

s269
00:26:02.560 --> 00:26:05.320
Uh, thank you guys for showing up.

s270
00:26:05.320 --> 00:26:07.680
Um, we if you guys are we are hiring.

s271
00:26:07.680 --> 00:26:13.360
If you guys are interested in like solving these types of technical problems come talk to me afterwards.

s272
00:26:13.360 --> 00:26:16.600
Uh, or if you're just interested in this at all come talk to me afterwards.

s273
00:26:16.600 --> 00:26:19.000
Appreciate it guys.

s274
00:26:32.978 --> 00:26:34.978
[music]