Generative Video at the Speed of Light — Keegan McCallum, uRun

AI Engineer · 8 min · 79 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

  1. 00:01[music]
  2. 00:12I am Keegan.
  3. 00:14I'm the founder of U Run, um, a new kind of inference provider focused around, uh, interactive media.
  4. 00:24And I'm here to talk about generative video.
  5. 00:25So, we hear a lot about generative video improving along the quality axis at the frontier.
  6. 00:32We have the classic Will Smith eating spaghetti from 2023.
  7. 00:36It is nightmare fuel and not something you would ever mistake for reality.
  8. 00:41In 2024, we got Sora and it gets a little better.
  9. 00:46It still has a bit of, you know, an AI feel to it, but it it's getting there.
  10. 00:50And Sora 2, you know, even better.
  11. 00:54But SeeDance this year, um, absolutely incredible.
  12. 00:58So photorealistic.
  13. 00:59And 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,
  14. 01:12the long horizon generations.
  15. 01:16So, what you're watching here is a demo for a model called Helios that we serve at U Run.
  16. 01:22Um, the generation in the bottom right corner, you'll see, is a long continuous generation.
  17. 01:28And the other video, um, is a bunch of clips, um, that have been generated faster than you can consume them.
  18. 01:35Uh, and they're about at the same quality as the frontier models were last year.
  19. 01:41They're Helios is a distill of Juan 2.1 14B.
  20. 01:47Um, 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
  21. 02:00and capabilities.
  22. 02:03Um 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
  23. 02:13a number of minutes.
  24. 02:15And the one on the right is and arguably a bit better.
  25. 02:20It's got better motion and it was generated for about a 100th of the cost.
  26. 02:28And these are just some of the charts showing the quality bar for both long and short video generation.
  27. 02:35Helios came out in March and it's it's pretty incredible to see how fast these are improving.
  28. 02:42But these are techniques that are being applied all over the place, not just the one model.
  29. 02:47There's world models which can keep consistency over long horizons and you can control in a fine-grained way, the camera and the viewport.
  30. 02:57There'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.
  31. 03:10There's actually been an explosion of innovation.
  32. 03:12There's been at least 40 models with real-time capabilities and long horizon generation capabilities released this year.
  33. 03:23Show of hands, who here has burned 10 or even $50 worth of tokens in an hour with Clockwork?
  34. 03:33A lot of people.
  35. 03:34And 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
  36. 03:47with an AI in a visual medium.
  37. 03:4915 hours.
  38. 03:52And so I want to talk a little bit about the different things this enables in terms of the way that we interact with computers.
  39. 03:58Um 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.
  40. 04:07So, one such use case would be a magic mirror.
  41. 04:11You
  42. 04:11[snorts]
  43. 04:11could have your webcam and you could ask to see yourself in any outfit.
  44. 04:17You could ask to see yourself in a car you like or with a haircut you're considering.
  45. 04:23Um a lot of different possibilities because these are open-ended models that can transform what they're seeing on a webcam in real time.
  46. 04:31I also think about accessibility a lot with these models.
  47. 04:35Um you know, working with AI involves a lot of reading and a lot of text.
  48. 04:39For some people that's more difficult.
  49. 04:42Um for some people they just don't think in in text.
  50. 04:46They think visually and learn better that way.
  51. 04:50Uh 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.
  52. 05:02And I'm excited about content creation.
  53. 05:05Um 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.
  54. 05:17But 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.
  55. 05:26Maybe you're piloting an agent that you're able to look over its shoulder and see what it's generating in real time.
  56. 05:33Um 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.
  57. 05:47And 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
  58. 05:57the horizons of what we can do with this technology.
  59. 06:01And so, what does it look like to actually build an application like this?
  60. 06:07So, 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.
  61. 06:14You're going to need to think
  62. 06:15[snorts]
  63. 06:15about where you're connecting the users to, what GPUs you're going to use to serve them.
  64. 06:22You're going to need to set up probably WebRTC and ICE and TURN.
  65. 06:28And for the most interesting use cases, you're going to want a model wire multiple models together in continuous streaming workflows.
  66. 06:37Building 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.
  67. 06:53And 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
  68. 07:09with any model.
  69. 07:13And 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
  70. 07:30transformation models, you can experiment and and build whatever you can really imagine on top of these.
  71. 07:39And I argue that in 2026, don't just need platforms, we need software factories and ways for agents interact with these.
  72. 07:49And so we've actually built one that will let folks hook into a CLI or an MCP server and build these kinds of applications.
  73. 07:58[gasps and sighs]
  74. 07:58And so the models are here and the frontier is really in how we serve them.
  75. 08:06Uh I went way over I went way under time.
  76. 08:11Um
  77. 08:11[laughter]
  78. 08:12but we are looking for design partners who want to push the boundaries of human human-computer interaction and we're hiring at YURUN.
  79. 08:20Um so come see me after the talk if uh if you're interested in chatting more.