WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=Xln-On3syJk

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/Xln-On3syJk.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
00:00:12.680 --> 00:00:14.080
I am Keegan.

s3
00:00:14.080 --> 00:00:23.960
I'm the founder of U Run, um, a new kind of inference provider focused around, uh, interactive media.

s4
00:00:24.480 --> 00:00:25.960
And I'm here to talk about generative video.

s5
00:00:25.960 --> 00:00:32.040
So, we hear a lot about generative video improving along the quality axis at the frontier.

s6
00:00:32.040 --> 00:00:36.480
We have the classic Will Smith eating spaghetti from 2023.

s7
00:00:36.480 --> 00:00:41.800
It is nightmare fuel and not something you would ever mistake for reality.

s8
00:00:41.880 --> 00:00:46.440
In 2024, we got Sora and it gets a little better.

s9
00:00:46.440 --> 00:00:50.840
It still has a bit of, you know, an AI feel to it, but it it's getting there.

s10
00:00:50.840 --> 00:00:54.840
And Sora 2, you know, even better.

s11
00:00:54.840 --> 00:00:58.360
But SeeDance this year, um, absolutely incredible.

s12
00:00:58.360 --> 00:00:59.960
So photorealistic.

s13
00:00:59.960 --> 00:01:12.640
And it's it's no wonder that we talk a lot about quality, but I'm here to talk about another axis which models are improving along, which is efficiency and, uh,

s14
00:01:12.640 --> 00:01:15.560
the long horizon generations.

s15
00:01:16.040 --> 00:01:22.160
So, what you're watching here is a demo for a model called Helios that we serve at U Run.

s16
00:01:22.160 --> 00:01:28.400
Um, the generation in the bottom right corner, you'll see, is a long continuous generation.

s17
00:01:28.400 --> 00:01:35.640
And the other video, um, is a bunch of clips, um, that have been generated faster than you can consume them.

s18
00:01:35.640 --> 00:01:41.840
Uh, and they're about at the same quality as the frontier models were last year.

s19
00:01:41.840 --> 00:01:47.360
They're Helios is a distill of Juan 2.1 14B.

s20
00:01:47.360 --> 00:02:00.560
Um, and I'll talk a bit about the techniques that are used in the various models that are hitting the scene right now, but there's been an explosion in just the last year uh, terms of efficiency

s21
00:02:00.560 --> 00:02:02.720
and capabilities.

s22
00:02:03.080 --> 00:02:13.680
Um so like looking at this, I kind of ruined it with the last club, but you can guess which one is real time and which one was generated in

s23
00:02:13.680 --> 00:02:15.280
a number of minutes.

s24
00:02:15.280 --> 00:02:20.040
And the one on the right is and arguably a bit better.

s25
00:02:20.040 --> 00:02:27.400
It's got better motion and it was generated for about a 100th of the cost.

s26
00:02:28.680 --> 00:02:35.280
And these are just some of the charts showing the quality bar for both long and short video generation.

s27
00:02:35.280 --> 00:02:42.080
Helios came out in March and it's it's pretty incredible to see how fast these are improving.

s28
00:02:42.080 --> 00:02:47.120
But these are techniques that are being applied all over the place, not just the one model.

s29
00:02:47.120 --> 00:02:57.160
There's world models which can keep consistency over long horizons and you can control in a fine-grained way, the camera and the viewport.

s30
00:02:57.160 --> 00:03:10.480
There's avatar models like we just talked about with lemon slice and there's video-to-video models that can can transform what you're seeing in in real time, almost like a magic mirror.

s31
00:03:10.480 --> 00:03:12.519
There's actually been an explosion of innovation.

s32
00:03:12.519 --> 00:03:23.120
There's been at least 40 models with real-time capabilities and long horizon generation capabilities released this year.

s33
00:03:23.880 --> 00:03:32.840
Show of hands, who here has burned 10 or even $50 worth of tokens in an hour with Clockwork?

s34
00:03:33.160 --> 00:03:34.680
A lot of people.

s35
00:03:34.680 --> 00:03:47.239
And so we're at a place right now where $10 can get you 3 hours worth of generated video continuously with most of these models and $50 would give you an entire day interacting

s36
00:03:47.239 --> 00:03:49.720
with an AI in a visual medium.

s37
00:03:49.720 --> 00:03:52.160
15 hours.

s38
00:03:52.239 --> 00:03:58.880
And so I want to talk a little bit about the different things this enables in terms of the way that we interact with computers.

s39
00:03:58.880 --> 00:04:07.480
Um and I'll talk a little bit about what we're doing at You Run to try and make it easier for folks to experiment and build out applications like this.

s40
00:04:07.480 --> 00:04:11.240
So, one such use case would be a magic mirror.

s41
00:04:11.240 --> 00:04:11.522
You

s42
00:04:11.522 --> 00:04:11.600
[snorts]

s43
00:04:11.600 --> 00:04:17.480
could have your webcam and you could ask to see yourself in any outfit.

s44
00:04:17.480 --> 00:04:23.680
You could ask to see yourself in a car you like or with a haircut you're considering.

s45
00:04:23.680 --> 00:04:31.880
Um a lot of different possibilities because these are open-ended models that can transform what they're seeing on a webcam in real time.

s46
00:04:31.880 --> 00:04:35.080
I also think about accessibility a lot with these models.

s47
00:04:35.080 --> 00:04:39.919
Um you know, working with AI involves a lot of reading and a lot of text.

s48
00:04:39.919 --> 00:04:42.320
For some people that's more difficult.

s49
00:04:42.320 --> 00:04:46.960
Um for some people they just don't think in in text.

s50
00:04:46.960 --> 00:04:50.040
They think visually and learn better that way.

s51
00:04:50.040 --> 00:05:02.120
Uh so, there's more opportunities to have companions or visual mediums that are going to allow more people to experience the things a lot of us have with coding models.

s52
00:05:02.120 --> 00:05:05.880
And I'm excited about content creation.

s53
00:05:05.880 --> 00:05:17.320
Um so far we've very much had a slot machine type approach where you're setting up a prompt and maybe some key frames and spending about $10 a minute to try and get the shot that you want.

s54
00:05:17.320 --> 00:05:26.440
But with these models you can actually steer them in real time in under a second while they're generating and get the actual shots that you want.

s55
00:05:26.440 --> 00:05:33.440
Maybe you're piloting an agent that you're able to look over its shoulder and see what it's generating in real time.

s56
00:05:33.440 --> 00:05:47.160
Um but you're able to more granularly control the content you're generating and with modern models like Google Gemini Omni, you can actually render these out as a more full fidelity clip.

s57
00:05:47.320 --> 00:05:57.960
And of course, we all are thinking about world models, but I want to take the the focus off of just kind of the the the basic world models that we talk a lot about and just try to expand

s58
00:05:57.960 --> 00:06:01.480
the horizons of what we can do with this technology.

s59
00:06:01.680 --> 00:06:07.240
And so, what does it look like to actually build an application like this?

s60
00:06:07.240 --> 00:06:14.720
So, you're going to need GPUs all over the world potentially if you've got a global audience that's going to be using these.

s61
00:06:14.720 --> 00:06:15.319
You're going to need to think

s62
00:06:15.319 --> 00:06:15.520
[snorts]

s63
00:06:15.520 --> 00:06:22.760
about where you're connecting the users to, what GPUs you're going to use to serve them.

s64
00:06:22.760 --> 00:06:28.400
You're going to need to set up probably WebRTC and ICE and TURN.

s65
00:06:28.400 --> 00:06:37.040
And for the most interesting use cases, you're going to want a model wire multiple models together in continuous streaming workflows.

s66
00:06:37.040 --> 00:06:53.800
Building those real-time harnesses, and you're going to want things synchronized with your controls that you're providing to your end users with every frame and continually providing a smooth streaming experience.

s67
00:06:53.919 --> 00:07:09.400
And so, our idea is what if there was just a React component that you could drop into your application to make it easy to provide video interactively inside your applications

s68
00:07:09.400 --> 00:07:11.760
with any model.

s69
00:07:13.240 --> 00:07:30.600
And behind the scenes, there's a programmable Python runtime that lets you easily build these complex pipelines generating asynchronously so that you can build avatar models, you can build these video-to-video

s70
00:07:30.600 --> 00:07:38.480
transformation models, you can experiment and and build whatever you can really imagine on top of these.

s71
00:07:39.280 --> 00:07:49.160
And I argue that in 2026, don't just need platforms, we need software factories and ways for agents interact with these.

s72
00:07:49.160 --> 00:07:58.040
And so we've actually built one that will let folks hook into a CLI or an MCP server and build these kinds of applications.

s73
00:07:58.155 --> 00:07:58.400
[gasps and sighs]

s74
00:07:58.400 --> 00:08:06.000
And so the models are here and the frontier is really in how we serve them.

s75
00:08:06.000 --> 00:08:11.200
Uh I went way over I went way under time.

s76
00:08:11.200 --> 00:08:11.910
Um

s77
00:08:11.910 --> 00:08:12.440
[laughter]

s78
00:08:12.440 --> 00:08:20.680
but we are looking for design partners who want to push the boundaries of human human-computer interaction and we're hiring at YURUN.

s79
00:08:20.680 --> 00:08:26.760
Um so come see me after the talk if uh if you're interested in chatting more.