[
  {
    "startMs": 1309,
    "text": "[music]"
  },
  {
    "startMs": 12639,
    "text": "and welcome back for those on the stream"
  },
  {
    "startMs": 14799,
    "text": "and those those in person. um we take"
  },
  {
    "startMs": 17520,
    "text": "tend to basically take these longer"
  },
  {
    "startMs": 19199,
    "text": "sessions between uh all the sort of"
  },
  {
    "startMs": 21840,
    "text": "mainstage keynotes to reflect on things"
  },
  {
    "startMs": 25039,
    "text": "that um you know are particularly"
  },
  {
    "startMs": 27519,
    "text": "important but like don't have like a"
  },
  {
    "startMs": 29199,
    "text": "significant like sort of launch moments."
  },
  {
    "startMs": 31119,
    "text": "Today we're very lucky to have people"
  },
  {
    "startMs": 32960,
    "text": "working on Omni and VO Nano Banana like"
  },
  {
    "startMs": 36079,
    "text": "the you know the world's best generative"
  },
  {
    "startMs": 38239,
    "text": "models here with us. Uh, Demetrio, I I I"
  },
  {
    "startMs": 41760,
    "text": "first saw you when you were posting"
  },
  {
    "startMs": 43440,
    "text": "about your office. [laughter]"
  },
  {
    "startMs": 46719,
    "text": "Um, I think you're you're probably"
  },
  {
    "startMs": 48719,
    "text": "number one uh Google Google's number one"
  },
  {
    "startMs": 51520,
    "text": "office influencer at least in in San"
  },
  {
    "startMs": 53120,
    "text": "Francisco. I think you like you like to"
  },
  {
    "startMs": 54559,
    "text": "bike as well. You like to take photos of"
  },
  {
    "startMs": 56879,
    "text": ">> bike here."
  },
  {
    "startMs": 57840,
    "text": ">> Yeah. Um, but you know, but also you"
  },
  {
    "startMs": 60800,
    "text": "work on video models."
  },
  {
    "startMs": 61760,
    "text": ">> That's right."
  },
  {
    "startMs": 62239,
    "text": ">> Um, Shane, I I met you I think at like a"
  },
  {
    "startMs": 64960,
    "text": "dinner."
  },
  {
    "startMs": 65680,
    "text": ">> Yeah. Um and uh and uh and I I remember"
  },
  {
    "startMs": 70479,
    "text": "you were trying to get me invested in"
  },
  {
    "startMs": 72159,
    "text": "like one of the companies. I forget"
  },
  {
    "startMs": 73760,
    "text": "forget which one."
  },
  {
    "startMs": 74799,
    "text": ">> Forget about that. [laughter]"
  },
  {
    "startMs": 77759,
    "text": ">> But now but now you're um now you're"
  },
  {
    "startMs": 80159,
    "text": "working on Omni Thinking. Um and and"
  },
  {
    "startMs": 82640,
    "text": "just you know a bunch of other"
  },
  {
    "startMs": 84320,
    "text": ">> Gemini RL."
  },
  {
    "startMs": 85520,
    "text": ">> Yeah. Yeah. Uh and Nicole also uh the"
  },
  {
    "startMs": 88640,
    "text": "rest of the gen media models uh nano"
  },
  {
    "startMs": 91119,
    "text": "banana and uh all and everything you"
  },
  {
    "startMs": 93520,
    "text": "just launched actually even this week."
  },
  {
    "startMs": 94960,
    "text": "Uh,"
  },
  {
    "startMs": 95520,
    "text": ">> yeah. We launched some APIs."
  },
  {
    "startMs": 97360,
    "text": ">> Yeah. Yeah. Yeah."
  },
  {
    "startMs": 98479,
    "text": ">> And I haven't tried to convince you to"
  },
  {
    "startMs": 99840,
    "text": "invest in anything, but maybe I should."
  },
  {
    "startMs": 101600,
    "text": ">> I mean, so I try not to be an investor."
  },
  {
    "startMs": 104079,
    "text": "People just convince me anyway. I'm like"
  },
  {
    "startMs": 105680,
    "text": "just, okay, well, I'm not that rich, but"
  },
  {
    "startMs": 107360,
    "text": "know like you can't not try to invest in"
  },
  {
    "startMs": 110720,
    "text": "some of these things. And, you know, for"
  },
  {
    "startMs": 112240,
    "text": "those of us who are not working at a"
  },
  {
    "startMs": 113520,
    "text": "Frontier Lab, this is the best this"
  },
  {
    "startMs": 115200,
    "text": "closest we'll ever get. Um, so yeah,"
  },
  {
    "startMs": 117920,
    "text": "actually, let's kind of recap since"
  },
  {
    "startMs": 119200,
    "text": "you're closest to it and we just did it,"
  },
  {
    "startMs": 120719,
    "text": "like what was launched this week? What"
  },
  {
    "startMs": 122479,
    "text": "should people go try out?"
  },
  {
    "startMs": 123680,
    "text": ">> Yeah. Um so yesterday we had two launch"
  },
  {
    "startMs": 126479,
    "text": "moments. Uh one of them we launched"
  },
  {
    "startMs": 128399,
    "text": "NanoBanana 2 light uh which is our"
  },
  {
    "startMs": 131599,
    "text": "fastest, cheapest um image model in the"
  },
  {
    "startMs": 134319,
    "text": "nano banana model family. Um and it's"
  },
  {
    "startMs": 137120,
    "text": "better than the original NanoBanana. Um"
  },
  {
    "startMs": 139280,
    "text": "so really for most people um that model"
  },
  {
    "startMs": 141440,
    "text": "replaces what you you know used and love"
  },
  {
    "startMs": 143280,
    "text": "the original Nano Banana for across like"
  },
  {
    "startMs": 145120,
    "text": "generation and editing and it gets"
  },
  {
    "startMs": 147280,
    "text": "really close to the frontier quality of"
  },
  {
    "startMs": 150080,
    "text": "of the kind of mainland bigger models."
  },
  {
    "startMs": 152080,
    "text": "So that that's really exciting. I think"
  },
  {
    "startMs": 153840,
    "text": "if you look at some of the demos or like"
  },
  {
    "startMs": 155840,
    "text": "things that people have been trying like"
  },
  {
    "startMs": 157440,
    "text": "getting kind of that like 3 second"
  },
  {
    "startMs": 158959,
    "text": "latency just unlocks a whole bunch of"
  },
  {
    "startMs": 160879,
    "text": "things that you can do with like"
  },
  {
    "startMs": 161920,
    "text": "ideation and iteration and it's just"
  },
  {
    "startMs": 163920,
    "text": "really fun and the model's getting to a"
  },
  {
    "startMs": 165840,
    "text": "point where like the quality is really"
  },
  {
    "startMs": 167040,
    "text": "good um where um it you know you can use"
  },
  {
    "startMs": 170239,
    "text": "it for iteration but you can also use"
  },
  {
    "startMs": 171680,
    "text": "some of those outputs as just kind of"
  },
  {
    "startMs": 172959,
    "text": "like ready um production output. So"
  },
  {
    "startMs": 174879,
    "text": "that's really exciting. Um and then"
  },
  {
    "startMs": 176560,
    "text": "second launch we finally um launched the"
  },
  {
    "startMs": 179280,
    "text": "Gemini Omni Flash APIs um that we"
  },
  {
    "startMs": 181920,
    "text": "pre-announced at IO. So thank you for"
  },
  {
    "startMs": 183840,
    "text": "waiting. Um and that you know is the"
  },
  {
    "startMs": 188319,
    "text": "first time that we're making the APIs"
  },
  {
    "startMs": 190080,
    "text": "available for developers and it's"
  },
  {
    "startMs": 191440,
    "text": "basically really exciting kind of video"
  },
  {
    "startMs": 192959,
    "text": "generation and editing and we're pricing"
  },
  {
    "startMs": 195120,
    "text": "it the same as Y31 fast. So we're"
  },
  {
    "startMs": 197280,
    "text": "getting you kind of like really really"
  },
  {
    "startMs": 198560,
    "text": "good quality for a really awesome price"
  },
  {
    "startMs": 200560,
    "text": "hopefully. Um"
  },
  {
    "startMs": 202159,
    "text": ">> yeah, I mean that that's incredible. I'm"
  },
  {
    "startMs": 204560,
    "text": "actually really So when you guys"
  },
  {
    "startMs": 206400,
    "text": "launched Omni for the first time, you"
  },
  {
    "startMs": 208000,
    "text": "also did a podcast uh with Logan who"
  },
  {
    "startMs": 210159,
    "text": "couldn't be here today uh and you added"
  },
  {
    "startMs": 212000,
    "text": "like a sloth uh and and Ramen and all"
  },
  {
    "startMs": 214640,
    "text": "these all these things. I actually"
  },
  {
    "startMs": 216159,
    "text": "really want to do that to our videos. I"
  },
  {
    "startMs": 217840,
    "text": "just didn't have an API for it because"
  },
  {
    "startMs": 219200,
    "text": "obviously I have to automate the whole"
  },
  {
    "startMs": 220400,
    "text": "thing. So thank you for the API."
  },
  {
    "startMs": 221760,
    "text": ">> Uh that is my favorite use case."
  },
  {
    "startMs": 223440,
    "text": "Everybody should do that. Um I got a cat"
  },
  {
    "startMs": 225760,
    "text": "which is probably like the most boring"
  },
  {
    "startMs": 227440,
    "text": "of the animals. Um if you don't know"
  },
  {
    "startMs": 228879,
    "text": "what we're talking about, you should"
  },
  {
    "startMs": 229680,
    "text": "look it up. It's very funny. Feurer. um"
  },
  {
    "startMs": 231680,
    "text": "Furer who's um you know on on the team"
  },
  {
    "startMs": 234080,
    "text": "did that."
  },
  {
    "startMs": 234560,
    "text": ">> Furer is the number one guy you should"
  },
  {
    "startMs": 236480,
    "text": "follow for you should follow ideas on"
  },
  {
    "startMs": 238879,
    "text": "okay what can this thing do?"
  },
  {
    "startMs": 240239,
    "text": ">> Yes."
  },
  {
    "startMs": 240720,
    "text": ">> Right."
  },
  {
    "startMs": 241040,
    "text": ">> Yes. He he's he's amazing at that."
  },
  {
    "startMs": 243360,
    "text": ">> I've tried to get him for the last two"
  },
  {
    "startMs": 245280,
    "text": "years to come to AIE. He hasn't made it"
  },
  {
    "startMs": 247280,
    "text": "yet. He's actually come in person. He"
  },
  {
    "startMs": 249040,
    "text": "just didn't want to speak because he's"
  },
  {
    "startMs": 250159,
    "text": "anonymous."
  },
  {
    "startMs": 251040,
    "text": ">> I know."
  },
  {
    "startMs": 251920,
    "text": ">> I I want to say his real name but I"
  },
  {
    "startMs": 253200,
    "text": "can't say his real name."
  },
  {
    "startMs": 253840,
    "text": ">> No [laughter] no we won't we won't do"
  },
  {
    "startMs": 255200,
    "text": "that to him. But you should really"
  },
  {
    "startMs": 256320,
    "text": "follow him. He's amazing."
  },
  {
    "startMs": 257759,
    "text": ">> He did all that work. I actually met him"
  },
  {
    "startMs": 259759,
    "text": "uh in the office uh when we did the"
  },
  {
    "startMs": 262079,
    "text": "podcast I think and I didn't realize it"
  },
  {
    "startMs": 264080,
    "text": "was him. So his badge doesn't say"
  },
  {
    "startMs": 266720,
    "text": "Popers."
  },
  {
    "startMs": 267440,
    "text": ">> Yeah,"
  },
  {
    "startMs": 267759,
    "text": ">> I know."
  },
  {
    "startMs": 268400,
    "text": ">> So he used to be part of uh Replicate"
  },
  {
    "startMs": 270560,
    "text": "and Replicate had this joke where like"
  },
  {
    "startMs": 272639,
    "text": "everyone was Deep Fates. Deep Fates is"
  },
  {
    "startMs": 274479,
    "text": "this like kind of mysterious character"
  },
  {
    "startMs": 275840,
    "text": "and replicate. Replicate is very cool"
  },
  {
    "startMs": 277120,
    "text": "company and both was part of it. Um, so,"
  },
  {
    "startMs": 280320,
    "text": "okay, one thing I want to get on there"
  },
  {
    "startMs": 282560,
    "text": "before I go into like sort of the the"
  },
  {
    "startMs": 283840,
    "text": "the the sort of omniper is we added"
  },
  {
    "startMs": 287919,
    "text": "cats, we added sloths, very cool, very"
  },
  {
    "startMs": 290639,
    "text": "cute, very fun."
  },
  {
    "startMs": 291759,
    "text": ">> Uh, what are the, you know, inspire"
  },
  {
    "startMs": 293440,
    "text": "people as to like what are the more sort"
  },
  {
    "startMs": 294720,
    "text": "of workhorse use cases that maybe are"
  },
  {
    "startMs": 297360,
    "text": "not just demos, you know?"
  },
  {
    "startMs": 299360,
    "text": ">> Yeah. So, so obviously the hero"
  },
  {
    "startMs": 300960,
    "text": "capability of the model or maybe there's"
  },
  {
    "startMs": 302560,
    "text": "two like one is the ability to kind of"
  },
  {
    "startMs": 304880,
    "text": "take in anything as input and then get"
  },
  {
    "startMs": 306560,
    "text": "video on the other side. Obviously in"
  },
  {
    "startMs": 308320,
    "text": "the future and and we've kind of talked"
  },
  {
    "startMs": 309680,
    "text": "about this as a pre-announce like we"
  },
  {
    "startMs": 311039,
    "text": "want to get the other output modalities"
  },
  {
    "startMs": 312560,
    "text": "out as well but basically what that"
  },
  {
    "startMs": 314000,
    "text": "means is you know you can take a set of"
  },
  {
    "startMs": 315919,
    "text": "images that you have as maybe a"
  },
  {
    "startMs": 317440,
    "text": "storyboard. You can take like an audio"
  },
  {
    "startMs": 319120,
    "text": "track as a reference of you know like a"
  },
  {
    "startMs": 321440,
    "text": "voice that you want a character to speak"
  },
  {
    "startMs": 323120,
    "text": "and then you can get a video on the"
  },
  {
    "startMs": 324400,
    "text": "other side. So like that just unlocks a"
  },
  {
    "startMs": 326320,
    "text": "whole bunch of things that you can do in"
  },
  {
    "startMs": 327680,
    "text": "like you know short film production or"
  },
  {
    "startMs": 329919,
    "text": "you know shorts we've launched on"
  },
  {
    "startMs": 331120,
    "text": "YouTube as well um to help creators kind"
  },
  {
    "startMs": 333759,
    "text": "of like create um content more easily."
  },
  {
    "startMs": 336400,
    "text": "Um and then the other one is obviously"
  },
  {
    "startMs": 338240,
    "text": "video editing. Like that's another thing"
  },
  {
    "startMs": 339600,
    "text": "that we're really excited about that"
  },
  {
    "startMs": 340880,
    "text": "we're just making easier because now you"
  },
  {
    "startMs": 342639,
    "text": "can use natural language to take a"
  },
  {
    "startMs": 345280,
    "text": "video, you know, add something, remove"
  },
  {
    "startMs": 346639,
    "text": "something. Sloth is obviously like fun"
  },
  {
    "startMs": 348240,
    "text": "example. Um, but there there's obviously"
  },
  {
    "startMs": 351120,
    "text": "kind of there's consumer use cases that"
  },
  {
    "startMs": 353039,
    "text": "we kind of had in mind where, you know,"
  },
  {
    "startMs": 354560,
    "text": "you could take your beach vacation video"
  },
  {
    "startMs": 356880,
    "text": "that was too noisy and you want to clean"
  },
  {
    "startMs": 358400,
    "text": "up that noise. Maybe in the past you"
  },
  {
    "startMs": 360479,
    "text": "wouldn't have because you didn't have"
  },
  {
    "startMs": 361680,
    "text": "the tools or you didn't know what the"
  },
  {
    "startMs": 362880,
    "text": "tools were that you needed to go to. So,"
  },
  {
    "startMs": 365120,
    "text": "that's one use case that you can, you"
  },
  {
    "startMs": 366560,
    "text": "know, go to. We've seen a lot of folks"
  },
  {
    "startMs": 368800,
    "text": "use it for kind of marketing ad campaign"
  },
  {
    "startMs": 371199,
    "text": "creation and I'm excited to see more of"
  },
  {
    "startMs": 373199,
    "text": "those use cases as we launch the APIs."
  },
  {
    "startMs": 376479,
    "text": "um because obviously like we don't we"
  },
  {
    "startMs": 378000,
    "text": "don't see all of it in the first party"
  },
  {
    "startMs": 379440,
    "text": "products but I'm really excited for"
  },
  {
    "startMs": 381039,
    "text": "people to start to explore that um in"
  },
  {
    "startMs": 383120,
    "text": "the API. So those are just some of the"
  },
  {
    "startMs": 384639,
    "text": "kind of like high level um things that"
  },
  {
    "startMs": 386720,
    "text": "have come up. U people also use it to"
  },
  {
    "startMs": 389039,
    "text": "create like education materials. Yes. Um"
  },
  {
    "startMs": 391759,
    "text": "and like like that's really exciting. I"
  },
  {
    "startMs": 394000,
    "text": "think we're all we've all kind of talked"
  },
  {
    "startMs": 395840,
    "text": "about being excited about the future of"
  },
  {
    "startMs": 397759,
    "text": "education where like everything can be"
  },
  {
    "startMs": 399520,
    "text": "kind of customized to you and"
  },
  {
    "startMs": 400800,
    "text": "personalized to your knowledge level and"
  },
  {
    "startMs": 403759,
    "text": "the style that you prefer and and so"
  },
  {
    "startMs": 405680,
    "text": "this is kind of just like a step in that"
  },
  {
    "startMs": 407440,
    "text": "direction."
  },
  {
    "startMs": 407919,
    "text": ">> Yeah. I I I sort of actually used just"
  },
  {
    "startMs": 409680,
    "text": "none of yesterday, but my my parents are"
  },
  {
    "startMs": 411520,
    "text": "visiting and there was there was a very"
  },
  {
    "startMs": 412960,
    "text": "fun sort of use case. They I bought some"
  },
  {
    "startMs": 415199,
    "text": "gadget off from Amazon that they wanted"
  },
  {
    "startMs": 417360,
    "text": "and the instructions to use it was were"
  },
  {
    "startMs": 419599,
    "text": "only in English and there was plenty of"
  },
  {
    "startMs": 420800,
    "text": "diagrams or whatever and I took a"
  },
  {
    "startMs": 422000,
    "text": "picture of it and said, you know,"
  },
  {
    "startMs": 423039,
    "text": "translate this into Romanian. Yes."
  },
  {
    "startMs": 424880,
    "text": ">> And keep everything else the same,"
  },
  {
    "startMs": 426240,
    "text": "right? So it was amazing, right? Like it"
  },
  {
    "startMs": 428000,
    "text": "was just like, yeah, it looks identical"
  },
  {
    "startMs": 430319,
    "text": "and it has, you know, it's perfectly"
  },
  {
    "startMs": 432319,
    "text": "translated. I mean, more or less, right?"
  },
  {
    "startMs": 434160,
    "text": "But it's it's you know using Gemini"
  },
  {
    "startMs": 436160,
    "text": "under the hood obviously to kind of do"
  },
  {
    "startMs": 437599,
    "text": "the translation and so you can you can"
  },
  {
    "startMs": 439440,
    "text": "see this use case for video as well"
  },
  {
    "startMs": 441120,
    "text": "right like the the power of text"
  },
  {
    "startMs": 443280,
    "text": "rendering in in in Omni is is quite next"
  },
  {
    "startMs": 446240,
    "text": "level. So and you could you could you"
  },
  {
    "startMs": 448080,
    "text": "could think about plenty of use cases of"
  },
  {
    "startMs": 449680,
    "text": "like both text rendering translation"
  },
  {
    "startMs": 451280,
    "text": "internalization all sorts of things that"
  },
  {
    "startMs": 453360,
    "text": "would be actually genuinely useful to a"
  },
  {
    "startMs": 455280,
    "text": "lot of different people and sort of"
  },
  {
    "startMs": 456560,
    "text": "broader access to either you could like"
  },
  {
    "startMs": 459520,
    "text": "redub a video or whatever it is that you"
  },
  {
    "startMs": 461520,
    "text": "wanted to do. like there's plenty of"
  },
  {
    "startMs": 462880,
    "text": "different things that you could you"
  },
  {
    "startMs": 463759,
    "text": "could think about doing."
  },
  {
    "startMs": 465520,
    "text": ">> Yeah. Um one of the most enlightening"
  },
  {
    "startMs": 469280,
    "text": "conversations I have on my podcast is"
  },
  {
    "startMs": 471199,
    "text": "with uh just people researchers at the"
  },
  {
    "startMs": 473360,
    "text": "frontier of these things. Um I had one"
  },
  {
    "startMs": 475440,
    "text": "with um Ethan from the XAI video team,"
  },
  {
    "startMs": 477919,
    "text": "the Grock video team who was basically"
  },
  {
    "startMs": 480000,
    "text": "saying like you know the next trend is"
  },
  {
    "startMs": 482479,
    "text": "actually not just like single model,"
  },
  {
    "startMs": 483759,
    "text": "it's more like video agents. Um, and I"
  },
  {
    "startMs": 486960,
    "text": "don't know if that terminology resonates"
  },
  {
    "startMs": 489039,
    "text": "uh obviously for for very relevant for"
  },
  {
    "startMs": 491120,
    "text": "RL. Uh, but it was it was basically kind"
  },
  {
    "startMs": 493280,
    "text": "of like giving up on like trying to do"
  },
  {
    "startMs": 494800,
    "text": "everything in in effectively one pass."
  },
  {
    "startMs": 497120,
    "text": "Um, do you feel that same way or is it"
  },
  {
    "startMs": 500240,
    "text": "still an open research question which"
  },
  {
    "startMs": 502080,
    "text": "way the trends are going? [snorts]"
  },
  {
    "startMs": 504720,
    "text": ">> Yeah. So um what kind of excite me most"
  },
  {
    "startMs": 507840,
    "text": "is really when the symbolic kind of"
  },
  {
    "startMs": 509840,
    "text": "foundational models and this kind of"
  },
  {
    "startMs": 511440,
    "text": "like video foundational model can"
  },
  {
    "startMs": 513039,
    "text": "actually kind of really work together"
  },
  {
    "startMs": 514719,
    "text": "and u in a way the if you look at the"
  },
  {
    "startMs": 517039,
    "text": "beginning of the generative sort of like"
  },
  {
    "startMs": 518880,
    "text": "image generation video generation a lot"
  },
  {
    "startMs": 520880,
    "text": "of it kind of started when the language"
  },
  {
    "startMs": 522399,
    "text": "model got good enough to provide a very"
  },
  {
    "startMs": 524560,
    "text": "detailed captioning like from stable"
  },
  {
    "startMs": 526399,
    "text": "fusion days or kind of dowi 2 days. So"
  },
  {
    "startMs": 529600,
    "text": "um so basically like language is"
  },
  {
    "startMs": 532320,
    "text": "extremely u helpful representation uh"
  },
  {
    "startMs": 535040,
    "text": "one is that it's kind of universal but"
  },
  {
    "startMs": 536880,
    "text": "the other kind of more um technical"
  },
  {
    "startMs": 539279,
    "text": "thing like kind of my hypothesis is like"
  },
  {
    "startMs": 541440,
    "text": "um one very difficult thing about"
  },
  {
    "startMs": 542720,
    "text": "machine learning is um this sort of like"
  },
  {
    "startMs": 545519,
    "text": "spirious coordination. So you don't know"
  },
  {
    "startMs": 547760,
    "text": "you know if the if this kind of feature"
  },
  {
    "startMs": 550000,
    "text": "right that's kind of predictive is"
  },
  {
    "startMs": 551360,
    "text": "actually causal factor or not. There are"
  },
  {
    "startMs": 553279,
    "text": "two ways. One is we can have really"
  },
  {
    "startMs": 555360,
    "text": "diverse data training data like from"
  },
  {
    "startMs": 557600,
    "text": "every intervention of the causal graph."
  },
  {
    "startMs": 559680,
    "text": "The other is you condition the causal"
  },
  {
    "startMs": 561200,
    "text": "information and conditioning the"
  },
  {
    "startMs": 562959,
    "text": "language is kind of like conditioning"
  },
  {
    "startMs": 565040,
    "text": "like a coal information of the of the"
  },
  {
    "startMs": 567519,
    "text": "kind of world. So um"
  },
  {
    "startMs": 569360,
    "text": ">> which is a prompt or a concept what"
  },
  {
    "startMs": 572240,
    "text": ">> yeah exactly so if you look at like you"
  },
  {
    "startMs": 574080,
    "text": "know how we going to describe this video"
  },
  {
    "startMs": 575760,
    "text": "how this kind of image is actually very"
  },
  {
    "startMs": 578240,
    "text": "close to you know how would describe"
  },
  {
    "startMs": 579839,
    "text": "this kind of causality you know behind"
  },
  {
    "startMs": 581279,
    "text": "this like how this is kind of generated."
  },
  {
    "startMs": 582959,
    "text": "So one is like that can really allow for"
  },
  {
    "startMs": 585519,
    "text": "very rich generalization and then uh"
  },
  {
    "startMs": 588000,
    "text": "very kind of just like a good model. Um"
  },
  {
    "startMs": 591279,
    "text": "the other is so eight months ago uh we"
  },
  {
    "startMs": 594160,
    "text": "put the evaluation paper called video"
  },
  {
    "startMs": 596080,
    "text": "models zero shot learners and reasoners."
  },
  {
    "startMs": 598399,
    "text": ">> Yes. So that was a kind of you know it's"
  },
  {
    "startMs": 601600,
    "text": "it's a confirmed paper and then later on"
  },
  {
    "startMs": 603440,
    "text": "actually the N banana team followed up"
  },
  {
    "startMs": 605040,
    "text": "with a vision banana paper that"
  },
  {
    "startMs": 606640,
    "text": "basically used n banana to do but"
  },
  {
    "startMs": 608640,
    "text": "essentially the idea is uh video model"
  },
  {
    "startMs": 611279,
    "text": "is extremely good sort of a foundation"
  },
  {
    "startMs": 613519,
    "text": "model for space and time kind of"
  },
  {
    "startMs": 615120,
    "text": "information. So um classic computer"
  },
  {
    "startMs": 617839,
    "text": "vision tasks a lot of could be kind of"
  },
  {
    "startMs": 619920,
    "text": "zero shorted and when you like say feed"
  },
  {
    "startMs": 622959,
    "text": "in some like a visual quiz uh it can you"
  },
  {
    "startMs": 626160,
    "text": "know there's definitely like a lot to"
  },
  {
    "startMs": 627440,
    "text": "improve it can kind of solve and it can"
  },
  {
    "startMs": 630160,
    "text": "um like robotics kind of like seeing it"
  },
  {
    "startMs": 632800,
    "text": "has really good kind of physical"
  },
  {
    "startMs": 633920,
    "text": "intuitions like word model uh and I"
  },
  {
    "startMs": 636959,
    "text": "think the the key is really the kind of"
  },
  {
    "startMs": 639200,
    "text": "mix of the visual kind of reasoning and"
  },
  {
    "startMs": 641839,
    "text": "then the text kind of reasoning kind of"
  },
  {
    "startMs": 643600,
    "text": "all tied together Um obviously you know"
  },
  {
    "startMs": 646399,
    "text": "like whether doing it you know as kind"
  },
  {
    "startMs": 647839,
    "text": "of unified model versus like just kind"
  },
  {
    "startMs": 649519,
    "text": "of agent coation I think that's more"
  },
  {
    "startMs": 652240,
    "text": "like uh it's going to be more kind of"
  },
  {
    "startMs": 654560,
    "text": "incremental you know how it's going to I"
  },
  {
    "startMs": 656160,
    "text": "imagine everything's going to go into"
  },
  {
    "startMs": 657279,
    "text": "like a single model eventually"
  },
  {
    "startMs": 659120,
    "text": ">> but right now there's like a lot you can"
  },
  {
    "startMs": 660959,
    "text": "do if you uh basically take like really"
  },
  {
    "startMs": 663440,
    "text": "good video understanding image"
  },
  {
    "startMs": 664640,
    "text": "understanding Gemini agentically with"
  },
  {
    "startMs": 667519,
    "text": "anomy and that's actually gonna yeah our"
  },
  {
    "startMs": 669519,
    "text": "team is like exploring a lot"
  },
  {
    "startMs": 672000,
    "text": ">> yeah okay that there's a there's a lot"
  },
  {
    "startMs": 673680,
    "text": "in there um I I think uh one question I"
  },
  {
    "startMs": 677040,
    "text": "I am increasingly starting to wonder is"
  },
  {
    "startMs": 678959,
    "text": "does it all trend towards one product"
  },
  {
    "startMs": 680640,
    "text": "for you guys right like now you have"
  },
  {
    "startMs": 682399,
    "text": "multiple models out the naming of omni"
  },
  {
    "startMs": 685920,
    "text": "does imply that eventually everything"
  },
  {
    "startMs": 688079,
    "text": "will go away and it just goes into omnis"
  },
  {
    "startMs": 690320,
    "text": "um is that the plan"
  },
  {
    "startMs": 693920,
    "text": ">> is it [laughter] I don't know I I think"
  },
  {
    "startMs": 696720,
    "text": "I think uh maybe I mean I think"
  },
  {
    "startMs": 700560,
    "text": "eventually I I think there's sort of"
  },
  {
    "startMs": 702480,
    "text": "different trade-offs engineering"
  },
  {
    "startMs": 704079,
    "text": "research product trade-offs in like it's"
  },
  {
    "startMs": 707279,
    "text": "like for the same reason like the the"
  },
  {
    "startMs": 710079,
    "text": "sorry how is it called nano banana light"
  },
  {
    "startMs": 711680,
    "text": "I don't know what the product name"
  },
  {
    "startMs": 712800,
    "text": ">> nanob banana tite"
  },
  {
    "startMs": 713920,
    "text": ">> nano banana too light yeah right it's"
  },
  {
    "startMs": 716000,
    "text": "it's it's it serves a particular niche"
  },
  {
    "startMs": 718800,
    "text": "right and it probably doesn't"
  },
  {
    "startMs": 720959,
    "text": "necessarily fit immediately in the same"
  },
  {
    "startMs": 724160,
    "text": "model literally checkpoint as uh"
  },
  {
    "startMs": 727200,
    "text": "something that can do 4K you know uh 30"
  },
  {
    "startMs": 730079,
    "text": "secondond videos right like they're"
  },
  {
    "startMs": 731360,
    "text": "probably not like trainable in the same"
  },
  {
    "startMs": 734399,
    "text": "quite way, right? Like, so I I don't"
  },
  {
    "startMs": 736160,
    "text": "know. It depends on how how far into the"
  },
  {
    "startMs": 737839,
    "text": "future you look like. Sure, in five"
  },
  {
    "startMs": 739279,
    "text": "years from now, will they all be the"
  },
  {
    "startMs": 740880,
    "text": "same model? Probably. Uh but like, you"
  },
  {
    "startMs": 743760,
    "text": "know, six months from now, we'll we'll"
  },
  {
    "startMs": 745680,
    "text": "probably still have, you know, multiple"
  },
  {
    "startMs": 746880,
    "text": "different models doing different things"
  },
  {
    "startMs": 748639,
    "text": "because kind of from pragmatically the"
  },
  {
    "startMs": 751360,
    "text": "trade-offs are such that we we should"
  },
  {
    "startMs": 753519,
    "text": "have multiple different kinds of models."
  },
  {
    "startMs": 755200,
    "text": ">> Yeah, I"
  },
  {
    "startMs": 756160,
    "text": ">> I think that's right. And and just on"
  },
  {
    "startMs": 757440,
    "text": "that note, I mean, we did call it Gemini"
  },
  {
    "startMs": 760000,
    "text": "Omni because we wanted to hint at the"
  },
  {
    "startMs": 762639,
    "text": "future where Gemini just becomes fully"
  },
  {
    "startMs": 764639,
    "text": "multimodal in and out, right? And so so"
  },
  {
    "startMs": 766959,
    "text": "it's definitely a move in that"
  },
  {
    "startMs": 768480,
    "text": "direction. I think we'll probably see a"
  },
  {
    "startMs": 770240,
    "text": "move in the direction where Omni also"
  },
  {
    "startMs": 771839,
    "text": "generates images and edits images and"
  },
  {
    "startMs": 773680,
    "text": "all those kinds of things. But Doo is"
  },
  {
    "startMs": 775440,
    "text": "right that I think on the way there,"
  },
  {
    "startMs": 777600,
    "text": "there's a bunch of really really useful"
  },
  {
    "startMs": 779200,
    "text": "applications of some of these more"
  },
  {
    "startMs": 781040,
    "text": "specialized models. And so we we will"
  },
  {
    "startMs": 783600,
    "text": "probably continue to work on those as"
  },
  {
    "startMs": 784959,
    "text": "well because like that serves a certain"
  },
  {
    "startMs": 787279,
    "text": "need at this point in time that may not"
  },
  {
    "startMs": 789120,
    "text": "exist you know a year from now. There's"
  },
  {
    "startMs": 790480,
    "text": "also like a research question about like"
  },
  {
    "startMs": 792000,
    "text": "just how much transfer there is between"
  },
  {
    "startMs": 794639,
    "text": "different kinds of modalities, right? I"
  },
  {
    "startMs": 796560,
    "text": "think you may believe that there's some"
  },
  {
    "startMs": 799279,
    "text": "transfer between coding and video"
  },
  {
    "startMs": 800959,
    "text": "generation and I think most people don't"
  },
  {
    "startMs": 803279,
    "text": "necessarily believe that but they you"
  },
  {
    "startMs": 805040,
    "text": "know you could try to think that there"
  },
  {
    "startMs": 806720,
    "text": "is some some there something there or it"
  },
  {
    "startMs": 808720,
    "text": "could be a waste right to put them"
  },
  {
    "startMs": 810000,
    "text": "together to try to learn these both"
  },
  {
    "startMs": 811279,
    "text": "tasks at the same time right so I think"
  },
  {
    "startMs": 812959,
    "text": "it's it's it's interesting sort of"
  },
  {
    "startMs": 814399,
    "text": "question to which extent like image and"
  },
  {
    "startMs": 816399,
    "text": "video obviously kind of there's some"
  },
  {
    "startMs": 817839,
    "text": "transfer like kind of not that different"
  },
  {
    "startMs": 820000,
    "text": "there's value in in learning to output"
  },
  {
    "startMs": 822399,
    "text": "video and audio at the same time because"
  },
  {
    "startMs": 824000,
    "text": "joint audio visual is you know that's"
  },
  {
    "startMs": 826079,
    "text": "how that's how it is. Um and then"
  },
  {
    "startMs": 828800,
    "text": "there's you know other kind of"
  },
  {
    "startMs": 829839,
    "text": "intersections of modalities that are not"
  },
  {
    "startMs": 831279,
    "text": "super obvious right like 3D"
  },
  {
    "startMs": 832880,
    "text": "representation coding I don't know maybe"
  },
  {
    "startMs": 835519,
    "text": "uh things like that right so like I"
  },
  {
    "startMs": 837120,
    "text": "think it's worth sort of exploring the"
  },
  {
    "startMs": 838560,
    "text": "different corners there and we are"
  },
  {
    "startMs": 839839,
    "text": "actively doing that um with a focus"
  },
  {
    "startMs": 842480,
    "text": "towards like what people actually want"
  },
  {
    "startMs": 843920,
    "text": "to do with these models"
  },
  {
    "startMs": 845440,
    "text": ">> yeah um what one thing I feel I feel"
  },
  {
    "startMs": 847839,
    "text": "like uh I'm surprised by but also I feel"
  },
  {
    "startMs": 851600,
    "text": "like it's insufficiently answered is"
  },
  {
    "startMs": 853839,
    "text": "what is the correct intermediate"
  },
  {
    "startMs": 856000,
    "text": "representation Um, so captioning, right?"
  },
  {
    "startMs": 860079,
    "text": "XI does captioning. Omni does"
  },
  {
    "startMs": 862480,
    "text": "captioning. Um, and I I I understand how"
  },
  {
    "startMs": 866079,
    "text": "captioning works for images. Um, and I"
  },
  {
    "startMs": 868959,
    "text": "understand that you can extend it into"
  },
  {
    "startMs": 870560,
    "text": "to video and and sort of guide it across"
  },
  {
    "startMs": 873120,
    "text": "time. It just feels very inefficient. It"
  },
  {
    "startMs": 875760,
    "text": "there's got to be I feel like there"
  },
  {
    "startMs": 877199,
    "text": "should be something better. Uh maybe"
  },
  {
    "startMs": 878880,
    "text": "it's code and maybe we generate you know"
  },
  {
    "startMs": 881519,
    "text": "and obviously I think a lot of um ffmpeg"
  },
  {
    "startMs": 885600,
    "text": "and mapplot um what's the three blue one"
  },
  {
    "startMs": 888399,
    "text": "brown one manm um a lot of like video is"
  },
  {
    "startMs": 891600,
    "text": "generated through code and maybe that's"
  },
  {
    "startMs": 893120,
    "text": "like the optimal representation uh any"
  },
  {
    "startMs": 896000,
    "text": "hypothesis as to like is is it better or"
  },
  {
    "startMs": 899040,
    "text": "is just English all you need"
  },
  {
    "startMs": 901279,
    "text": ">> well as so I'm in the Gemini and they"
  },
  {
    "startMs": 903680,
    "text": "know we do like a lot of RL agent and of"
  },
  {
    "startMs": 905920,
    "text": "course kind of coding so yeah We we're"
  },
  {
    "startMs": 908240,
    "text": "definitely exploring the coding"
  },
  {
    "startMs": 909360,
    "text": "representations."
  },
  {
    "startMs": 910240,
    "text": ">> Yeah."
  },
  {
    "startMs": 910720,
    "text": ">> As kind of better kind of way to"
  },
  {
    "startMs": 912560,
    "text": "represent. Yeah."
  },
  {
    "startMs": 913440,
    "text": ">> But you know like do you what's your"
  },
  {
    "startMs": 915920,
    "text": "probability estimate on like [laughter]"
  },
  {
    "startMs": 918000,
    "text": "if we just output binaries like we just"
  },
  {
    "startMs": 920079,
    "text": "you know like just it's just ones and"
  },
  {
    "startMs": 921839,
    "text": "zeros."
  },
  {
    "startMs": 923440,
    "text": ">> Um I I guess maybe a kind of similar"
  },
  {
    "startMs": 927360,
    "text": "discussion was like um basically is the"
  },
  {
    "startMs": 931120,
    "text": "language the right representation like"
  },
  {
    "startMs": 933279,
    "text": "right. So uh one kind of question for"
  },
  {
    "startMs": 935440,
    "text": "example uh professor you know like some"
  },
  {
    "startMs": 937839,
    "text": "ask is like you know why why does the"
  },
  {
    "startMs": 939839,
    "text": "channel of thought need to be in the"
  },
  {
    "startMs": 941519,
    "text": "natural language?"
  },
  {
    "startMs": 942320,
    "text": ">> Yes."
  },
  {
    "startMs": 942639,
    "text": ">> Can it just be the kind of any kind of"
  },
  {
    "startMs": 944480,
    "text": "like continuous tokens just any amount"
  },
  {
    "startMs": 946639,
    "text": "of you know additional computations."
  },
  {
    "startMs": 948800,
    "text": ">> Um so one is like obviously the test"
  },
  {
    "startMs": 952959,
    "text": "like adaptive compute is going to give"
  },
  {
    "startMs": 955040,
    "text": "like you know better results. So it's"
  },
  {
    "startMs": 956560,
    "text": "that but what really kind of made CH"
  },
  {
    "startMs": 958720,
    "text": "thought so you know like four years ago"
  },
  {
    "startMs": 960240,
    "text": "I wrote you know the larger model zero"
  },
  {
    "startMs": 962560,
    "text": "sort reasoner and then self-improvement."
  },
  {
    "startMs": 964079,
    "text": "So I kind of know from the very early"
  },
  {
    "startMs": 965759,
    "text": "day but the reason like it works really"
  },
  {
    "startMs": 968079,
    "text": "well is um right now the recipe that"
  },
  {
    "startMs": 971519,
    "text": "works is the pre-training that scales a"
  },
  {
    "startMs": 973920,
    "text": "lot and then that basically like learns"
  },
  {
    "startMs": 975680,
    "text": "a lot of intelligence. there are a lot"
  },
  {
    "startMs": 977440,
    "text": "of you know scaling RL but those are"
  },
  {
    "startMs": 979920,
    "text": "still like extremely kind of comput"
  },
  {
    "startMs": 981360,
    "text": "incent intensive to extract the"
  },
  {
    "startMs": 983360,
    "text": "information and um you really want to"
  },
  {
    "startMs": 986320,
    "text": "rely the intelligence on that so"
  },
  {
    "startMs": 988560,
    "text": "basically by tying the sort of like a"
  },
  {
    "startMs": 991360,
    "text": "reasoning in the natural language you"
  },
  {
    "startMs": 992800,
    "text": "basically directly use the intelligence"
  },
  {
    "startMs": 994240,
    "text": "of the pre-training to it while if you"
  },
  {
    "startMs": 996720,
    "text": "remove that kind of constraints then"
  },
  {
    "startMs": 998639,
    "text": "you're not um and these days uh I feel"
  },
  {
    "startMs": 1002079,
    "text": "the a lot of advancements in the texts"
  },
  {
    "startMs": 1005920,
    "text": "but also doing this kind of multimodal"
  },
  {
    "startMs": 1007519,
    "text": "space is very driven by this uh kind of"
  },
  {
    "startMs": 1010800,
    "text": "text as a kind of great uh sort of"
  },
  {
    "startMs": 1013360,
    "text": "representation."
  },
  {
    "startMs": 1014480,
    "text": ">> Yeah, it's a good backbone."
  },
  {
    "startMs": 1016480,
    "text": ">> Yeah,"
  },
  {
    "startMs": 1017199,
    "text": ">> I I think to me it's even simpler than"
  },
  {
    "startMs": 1018959,
    "text": "that. It's text is is how we"
  },
  {
    "startMs": 1020880,
    "text": "communicate. So I think fundamentally if"
  },
  {
    "startMs": 1022160,
    "text": "you're building kind of products that"
  },
  {
    "startMs": 1023839,
    "text": "humans will be interfacing with um like"
  },
  {
    "startMs": 1027280,
    "text": "like that we will be using text somehow"
  },
  {
    "startMs": 1029919,
    "text": "if it's a text interface, right? Not not"
  },
  {
    "startMs": 1031919,
    "text": "for everything. So I think it's it's"
  },
  {
    "startMs": 1033360,
    "text": "natural to default to that."
  },
  {
    "startMs": 1035280,
    "text": ">> Yeah. Obviously there's like a conf"
  },
  {
    "startMs": 1037280,
    "text": "discussion you know some arrow like"
  },
  {
    "startMs": 1038959,
    "text": "arrow maximalists is like oh we don't"
  },
  {
    "startMs": 1040880,
    "text": "care about you know kind of channel"
  },
  {
    "startMs": 1042000,
    "text": "those kind of like stuff it's just just"
  },
  {
    "startMs": 1044000,
    "text": "additional compute"
  },
  {
    "startMs": 1044880,
    "text": ">> sure but I personally yeah"
  },
  {
    "startMs": 1046880,
    "text": ">> ro maximalists I wonder I wonder who who"
  },
  {
    "startMs": 1050240,
    "text": "qualifies in that description David"
  },
  {
    "startMs": 1051919,
    "text": "silver"
  },
  {
    "startMs": 1052720,
    "text": ">> ah okay yeah I mean they they've just"
  },
  {
    "startMs": 1055600,
    "text": "left to to start their thing um"
  },
  {
    "startMs": 1058559,
    "text": "interesting okay so uh I I mean I think"
  },
  {
    "startMs": 1061919,
    "text": "I'm very interested in just like better"
  },
  {
    "startMs": 1063520,
    "text": "representations because I think that's"
  },
  {
    "startMs": 1064799,
    "text": "one of our themes that we're curating"
  },
  {
    "startMs": 1066720,
    "text": "today uh at the world fair is world"
  },
  {
    "startMs": 1068559,
    "text": "models. You mentioned the word world"
  },
  {
    "startMs": 1070000,
    "text": "models but it's not something that's"
  },
  {
    "startMs": 1071360,
    "text": "like super well defined. I think"
  },
  {
    "startMs": 1072960,
    "text": "everyone's like sort of converging on"
  },
  {
    "startMs": 1074960,
    "text": "some version of it that is like the"
  },
  {
    "startMs": 1076799,
    "text": "ideal."
  },
  {
    "startMs": 1077760,
    "text": ">> Sure. Everything is a world model now."
  },
  {
    "startMs": 1079600,
    "text": "It's sort of a"
  },
  {
    "startMs": 1080640,
    "text": ">> it's not it's not that useful, right?"
  },
  {
    "startMs": 1082880,
    "text": ">> So I just gave a keynote at the IER"
  },
  {
    "startMs": 1085440,
    "text": "world model workshop. Yeah. And then uh"
  },
  {
    "startMs": 1087600,
    "text": "yeah essentially uh I definitely"
  },
  {
    "startMs": 1089120,
    "text": "encourage to check out the definition by"
  },
  {
    "startMs": 1090960,
    "text": "Jandra Matic. He's like the you know OG"
  },
  {
    "startMs": 1093200,
    "text": "computer vision professor UC Berkeley."
  },
  {
    "startMs": 1095200,
    "text": ">> Uh he has pretty you know bit of word to"
  },
  {
    "startMs": 1097360,
    "text": "say about world model"
  },
  {
    "startMs": 1098559,
    "text": ">> but also kind of Schmidt Herburver's"
  },
  {
    "startMs": 1099919,
    "text": "kind of how he defined the world model"
  },
  {
    "startMs": 1101360,
    "text": "from 2019 like 1990 sort of uh uh you"
  },
  {
    "startMs": 1106160,
    "text": "know like Wayne was just basically just"
  },
  {
    "startMs": 1107600,
    "text": "that kind of model base. Uh for me the"
  },
  {
    "startMs": 1109120,
    "text": "word model is basically just the model"
  },
  {
    "startMs": 1110480,
    "text": "in the model based RL and I feel that"
  },
  {
    "startMs": 1112320,
    "text": "has sufficient to describe but obviously"
  },
  {
    "startMs": 1114559,
    "text": "you know there are like a lot of uh FE"
  },
  {
    "startMs": 1116559,
    "text": "had a kind of nice blog post about what"
  },
  {
    "startMs": 1119039,
    "text": "about yeah this kind of broken down"
  },
  {
    "startMs": 1121600,
    "text": ">> um but yeah"
  },
  {
    "startMs": 1123440,
    "text": ">> yeah I mean so you know I I'll end this"
  },
  {
    "startMs": 1127679,
    "text": "part of the conversation but like I I do"
  },
  {
    "startMs": 1129760,
    "text": "think that language to me relying on"
  },
  {
    "startMs": 1132160,
    "text": "language as like the sort of like the"
  },
  {
    "startMs": 1133360,
    "text": "narrow pipe through which everything"
  },
  {
    "startMs": 1134799,
    "text": "goes through um still is like a lossy"
  },
  {
    "startMs": 1137679,
    "text": "compression. No, no, no. But we're not"
  },
  {
    "startMs": 1139360,
    "text": "seeing that, right? We're basically"
  },
  {
    "startMs": 1140400,
    "text": "saying the video model and the language"
  },
  {
    "startMs": 1142000,
    "text": "together."
  },
  {
    "startMs": 1142960,
    "text": ">> So, so I think the language alone is uh"
  },
  {
    "startMs": 1145520,
    "text": "not sufficient. That's why we feel like"
  },
  {
    "startMs": 1147520,
    "text": "the video is a very complement model."
  },
  {
    "startMs": 1149679,
    "text": "Right? Now the um you know kind of v"
  },
  {
    "startMs": 1151840,
    "text": "omni many people feel as uh you know"
  },
  {
    "startMs": 1154000,
    "text": "generating kind of pretty videos but I"
  },
  {
    "startMs": 1156160,
    "text": "think our vision it's it's much more"
  },
  {
    "startMs": 1158000,
    "text": "than that. It's a missing foundational"
  },
  {
    "startMs": 1159600,
    "text": "model that's absolutely required if you"
  },
  {
    "startMs": 1161679,
    "text": "want to make the AGI that match to"
  },
  {
    "startMs": 1163120,
    "text": "humans not just a jacked one."
  },
  {
    "startMs": 1165120,
    "text": ">> Yeah. Um okay. So one one other thing"
  },
  {
    "startMs": 1168640,
    "text": "you know you you mentioned on the vision"
  },
  {
    "startMs": 1170320,
    "text": "side um and I'm kind of curious how sort"
  },
  {
    "startMs": 1174400,
    "text": "of uh parallel you know in terms of your"
  },
  {
    "startMs": 1177120,
    "text": "research careers um this development is"
  },
  {
    "startMs": 1180480,
    "text": "like I think basically a lot of vision"
  },
  {
    "startMs": 1181919,
    "text": "people have crossed over into more model"
  },
  {
    "startMs": 1184240,
    "text": "people um a lot of vision people also"
  },
  {
    "startMs": 1186799,
    "text": "become generative video and image people"
  },
  {
    "startMs": 1190480,
    "text": "and is it just as simple as you know"
  },
  {
    "startMs": 1193120,
    "text": "reversing uh image to text and then now"
  },
  {
    "startMs": 1195520,
    "text": "it's text to image like"
  },
  {
    "startMs": 1197353,
    "text": ">> [laughter]"
  },
  {
    "startMs": 1198559,
    "text": ">> is is that if I mean that effectively"
  },
  {
    "startMs": 1200400,
    "text": "was the diffusion process. Um"
  },
  {
    "startMs": 1203760,
    "text": "I I just you know I I just see the"
  },
  {
    "startMs": 1206240,
    "text": "career paths of the people that I talked"
  },
  {
    "startMs": 1207760,
    "text": "to and and see and I I I see this"
  },
  {
    "startMs": 1210320,
    "text": "overall trend of research directions and"
  },
  {
    "startMs": 1212720,
    "text": "I just wanted you to guys to sort of"
  },
  {
    "startMs": 1214640,
    "text": "reflect on on that."
  },
  {
    "startMs": 1216160,
    "text": ">> I mean I certainly went that way right I"
  },
  {
    "startMs": 1217919,
    "text": "I started long time ago uh doing"
  },
  {
    "startMs": 1220720,
    "text": "computer vision sort of object detection"
  },
  {
    "startMs": 1223039,
    "text": "recognition things like that. Uh I think"
  },
  {
    "startMs": 1224960,
    "text": "just that's just simpler problem right"
  },
  {
    "startMs": 1226720,
    "text": "just generation is just harder like it's"
  },
  {
    "startMs": 1228640,
    "text": "a it's a different kind of mapping right"
  },
  {
    "startMs": 1230400,
    "text": "you map from the the inverse mapping is"
  },
  {
    "startMs": 1232880,
    "text": "not as simple as just inverting the the"
  },
  {
    "startMs": 1234640,
    "text": "kind of network you use right it's it's"
  },
  {
    "startMs": 1236320,
    "text": "a it's it's more ambiguous right to go"
  },
  {
    "startMs": 1238400,
    "text": "from cat to image of a cat and in some"
  },
  {
    "startMs": 1241039,
    "text": "ways it's also a loop because your"
  },
  {
    "startMs": 1242720,
    "text": "vision work creates the synthetic labels"
  },
  {
    "startMs": 1244960,
    "text": "that then continues"
  },
  {
    "startMs": 1246640,
    "text": ">> I mean sure [laughter]"
  },
  {
    "startMs": 1248400,
    "text": "I don't know I don't know I try to"
  },
  {
    "startMs": 1249679,
    "text": "validate my my sort of theories about"
  },
  {
    "startMs": 1251360,
    "text": "how fields develop how how careers has"
  },
  {
    "startMs": 1253919,
    "text": "progressed through this"
  },
  {
    "startMs": 1255360,
    "text": ">> I mean for like the the the better the"
  },
  {
    "startMs": 1257840,
    "text": "understanding side gets like we have"
  },
  {
    "startMs": 1259760,
    "text": "seen that the generation side also gets"
  },
  {
    "startMs": 1262000,
    "text": "better right so like like"
  },
  {
    "startMs": 1263919,
    "text": ">> it's completely bootstrapping yeah it's"
  },
  {
    "startMs": 1265840,
    "text": ">> and so like like like there's definitely"
  },
  {
    "startMs": 1267760,
    "text": "they're there to that thesis and I think"
  },
  {
    "startMs": 1269600,
    "text": "yeah I think a lot of people have kind"
  },
  {
    "startMs": 1270880,
    "text": "of like I I definitely worked with a lot"
  },
  {
    "startMs": 1272720,
    "text": "of um image understanding people who"
  },
  {
    "startMs": 1274559,
    "text": "became image generation people you know"
  },
  {
    "startMs": 1276400,
    "text": "and then some of them have moved on to"
  },
  {
    "startMs": 1277840,
    "text": "video because it's kind of like the next"
  },
  {
    "startMs": 1279360,
    "text": "thing where you have so many more"
  },
  {
    "startMs": 1280799,
    "text": "dimensions to work with so yeah I'm"
  },
  {
    "startMs": 1282400,
    "text": "curious about you spec as your"
  },
  {
    "startMs": 1284960,
    "text": ">> so I definitely like recommend start"
  },
  {
    "startMs": 1286559,
    "text": "with understanding recognition because"
  },
  {
    "startMs": 1288240,
    "text": "that's basically discriminator and then"
  },
  {
    "startMs": 1289840,
    "text": "that's going to lead to better"
  },
  {
    "startMs": 1290960,
    "text": "generation and that's what the bridge is"
  },
  {
    "startMs": 1292880,
    "text": "basically reinforcement learning so my"
  },
  {
    "startMs": 1295280,
    "text": "um my kind of journey is I initially"
  },
  {
    "startMs": 1296880,
    "text": "kind of worked on the algorithmic"
  },
  {
    "startMs": 1298159,
    "text": "research in the gent model against some"
  },
  {
    "startMs": 1300480,
    "text": "like you know amnest kind of generation"
  },
  {
    "startMs": 1302640,
    "text": "and then I worked on like RL and"
  },
  {
    "startMs": 1304559,
    "text": "robotics um and then like six years ago"
  },
  {
    "startMs": 1307280,
    "text": "I was like leading like a moonshot on"
  },
  {
    "startMs": 1309120,
    "text": "the dexterity it was pretty early but I"
  },
  {
    "startMs": 1311440,
    "text": "see now everyone's kind of doing"
  },
  {
    "startMs": 1313360,
    "text": "uh four years ago I basically kind of"
  },
  {
    "startMs": 1314880,
    "text": "figured out that this like symbolic AGI"
  },
  {
    "startMs": 1318000,
    "text": "is going to accelerate much faster than"
  },
  {
    "startMs": 1320240,
    "text": "the kind of physical AGI kind of"
  },
  {
    "startMs": 1321679,
    "text": "counterpart. So uh I decided to kind of"
  },
  {
    "startMs": 1324080,
    "text": "like language models and then those"
  },
  {
    "startMs": 1326000,
    "text": "things. Um and then recently kind of"
  },
  {
    "startMs": 1328480,
    "text": "work with Doomi and then like omni team"
  },
  {
    "startMs": 1330480,
    "text": "I quite enjoy kind of collaboration"
  },
  {
    "startMs": 1332159,
    "text": "there. the what I quite enjoy uh what I"
  },
  {
    "startMs": 1335600,
    "text": "recommend definitely to the researcher"
  },
  {
    "startMs": 1337039,
    "text": "is to uh definitely kind of explore or"
  },
  {
    "startMs": 1340159,
    "text": "at least like get exposure to what the"
  },
  {
    "startMs": 1342640,
    "text": "top people in each of the community are"
  },
  {
    "startMs": 1344480,
    "text": "like looking at how they kind of think"
  },
  {
    "startMs": 1346000,
    "text": "about problems. So when I look at the"
  },
  {
    "startMs": 1348240,
    "text": "video model to me it kind of reminds me"
  },
  {
    "startMs": 1350159,
    "text": "like pretty early on sort of like"
  },
  {
    "startMs": 1352799,
    "text": "language model where like very early"
  },
  {
    "startMs": 1354799,
    "text": "language model was a kind of creative"
  },
  {
    "startMs": 1356640,
    "text": "sort of demo right you kind of like try"
  },
  {
    "startMs": 1359120,
    "text": "to write like a story like mobile and"
  },
  {
    "startMs": 1361360,
    "text": "then like you know GBD2 and then those"
  },
  {
    "startMs": 1363039,
    "text": "kind of days like LTM kind of days right"
  },
  {
    "startMs": 1365440,
    "text": "and then you know uh instruction tuning"
  },
  {
    "startMs": 1368559,
    "text": "you actually kind of make it usable as a"
  },
  {
    "startMs": 1370000,
    "text": "chatbot but then at the chatbot stage it"
  },
  {
    "startMs": 1372320,
    "text": "still had so much hallucinations and"
  },
  {
    "startMs": 1374000,
    "text": "instruction for wasn't good enough so it"
  },
  {
    "startMs": 1376159,
    "text": "couldn't use for reasoning and when I"
  },
  {
    "startMs": 1378320,
    "text": "got good enough um in pre-training and"
  },
  {
    "startMs": 1380799,
    "text": "post- trainining for reasoning then you"
  },
  {
    "startMs": 1382880,
    "text": "know this kind of test time scaling the"
  },
  {
    "startMs": 1384240,
    "text": "RL really took off to like many of the"
  },
  {
    "startMs": 1386559,
    "text": "kind of best performing models and right"
  },
  {
    "startMs": 1388400,
    "text": "now I think the video model is as we"
  },
  {
    "startMs": 1390799,
    "text": "mentioned it's it is a complimentary"
  },
  {
    "startMs": 1392400,
    "text": "foundational model and I can imagine"
  },
  {
    "startMs": 1393919,
    "text": "it's going to follow a similar path it's"
  },
  {
    "startMs": 1395919,
    "text": "going to be very uh it's going to"
  },
  {
    "startMs": 1397600,
    "text": "improve a lot instruction following a"
  },
  {
    "startMs": 1399520,
    "text": "lot of uh this it's going to improve a"
  },
  {
    "startMs": 1401120,
    "text": "lot in reducing coordinations to extend"
  },
  {
    "startMs": 1403440,
    "text": "that it become a very reliable world"
  },
  {
    "startMs": 1405120,
    "text": "model so we can kind of like intermixed"
  },
  {
    "startMs": 1407600,
    "text": "video like space-time simulation was a"
  },
  {
    "startMs": 1410080,
    "text": "text simulation to solve like arbitrary"
  },
  {
    "startMs": 1411919,
    "text": "AGI problems. Also like I think the"
  },
  {
    "startMs": 1414320,
    "text": "difference still is between sort of text"
  },
  {
    "startMs": 1416080,
    "text": "models and like image video models is"
  },
  {
    "startMs": 1418080,
    "text": "that like we haven't quite unified"
  },
  {
    "startMs": 1419840,
    "text": "understanding and generation in in"
  },
  {
    "startMs": 1421760,
    "text": "multimedia I'd say yet like I mean I"
  },
  {
    "startMs": 1424080,
    "text": "think I think without going to the"
  },
  {
    "startMs": 1425360,
    "text": "details of course there's like it"
  },
  {
    "startMs": 1426880,
    "text": "depends on on at which level you're"
  },
  {
    "startMs": 1428799,
    "text": "thinking about this but generally like"
  },
  {
    "startMs": 1431200,
    "text": "there's not that many as far as I know"
  },
  {
    "startMs": 1432960,
    "text": "models sot kind of you know frontier"
  },
  {
    "startMs": 1435679,
    "text": "models that are genuinely"
  },
  {
    "startMs": 1438720,
    "text": "kind of good at both understanding and"
  },
  {
    "startMs": 1441280,
    "text": "generation of of let's videos, right?"
  },
  {
    "startMs": 1444000,
    "text": "Like it's a it's a it's an interesting"
  },
  {
    "startMs": 1445520,
    "text": "challenge. I'm not saying that we should"
  },
  {
    "startMs": 1447120,
    "text": "do this. Uh but but I think uh it kind"
  },
  {
    "startMs": 1450559,
    "text": "of stands to reason that like you know"
  },
  {
    "startMs": 1452320,
    "text": "understanding and generation are two"
  },
  {
    "startMs": 1453679,
    "text": "sides of the same coin. So they they"
  },
  {
    "startMs": 1455120,
    "text": "kind of should be in the same model in"
  },
  {
    "startMs": 1456640,
    "text": "some ways. Uh but we don't necessarily"
  },
  {
    "startMs": 1458480,
    "text": "always do that. So yeah."
  },
  {
    "startMs": 1460480,
    "text": ">> Uh you mentioned audio as well, right?"
  },
  {
    "startMs": 1462960,
    "text": "Yeah. Uh is that as hard as video or"
  },
  {
    "startMs": 1468400,
    "text": "qualitatively different? If if so, in"
  },
  {
    "startMs": 1470720,
    "text": "what way? Uh, one of the interesting"
  },
  {
    "startMs": 1473840,
    "text": "directions three years ago was people"
  },
  {
    "startMs": 1476320,
    "text": "using um, I guess diffusion to do audio"
  },
  {
    "startMs": 1481440,
    "text": "uh, as in like the the sort of refusion"
  },
  {
    "startMs": 1484159,
    "text": "approach. I don't know if you you guys"
  },
  {
    "startMs": 1485520,
    "text": "saw that. Um, and I just think it's like"
  },
  {
    "startMs": 1487840,
    "text": "very interesting if a modality that we"
  },
  {
    "startMs": 1491120,
    "text": "perceive which is audio is different"
  },
  {
    "startMs": 1492720,
    "text": "than video actually two machines is"
  },
  {
    "startMs": 1494799,
    "text": "exactly the same like there's they see"
  },
  {
    "startMs": 1497440,
    "text": "no difference."
  },
  {
    "startMs": 1499279,
    "text": "I mean I think on a technical level"
  },
  {
    "startMs": 1501520,
    "text": "there are some differences but I think"
  },
  {
    "startMs": 1502559,
    "text": "they're like relatively minor. I think"
  },
  {
    "startMs": 1504159,
    "text": "from my perspective audio came into into"
  },
  {
    "startMs": 1507440,
    "text": "my life when we shipped V3 which was I"
  },
  {
    "startMs": 1510720,
    "text": "believe the first model that did like a"
  },
  {
    "startMs": 1512400,
    "text": "joint"
  },
  {
    "startMs": 1512720,
    "text": ">> with the slicing of the"
  },
  {
    "startMs": 1514240,
    "text": ">> Yeah. Yeah. the gold bars or whatever."
  },
  {
    "startMs": 1516480,
    "text": "Um it it was the first model that did"
  },
  {
    "startMs": 1518559,
    "text": "this sort of joint audiovisisual"
  },
  {
    "startMs": 1519760,
    "text": "generation. Yes. uh like in the in the I"
  },
  {
    "startMs": 1522400,
    "text": "mean there are there were other models"
  },
  {
    "startMs": 1523600,
    "text": "that did kind of you know kind of kind"
  },
  {
    "startMs": 1525120,
    "text": "of agentic hacking under the hood but"
  },
  {
    "startMs": 1526960,
    "text": "this one was truly sort of you know"
  },
  {
    "startMs": 1529120,
    "text": "generating everything at once and we the"
  },
  {
    "startMs": 1532159,
    "text": "reason we did that is because we felt"
  },
  {
    "startMs": 1534400,
    "text": "and I think was the right choice we felt"
  },
  {
    "startMs": 1536559,
    "text": "that like uh it only makes sense to"
  },
  {
    "startMs": 1539039,
    "text": "generate them at the same time because"
  },
  {
    "startMs": 1540159,
    "text": "there sort of kind of like from a"
  },
  {
    "startMs": 1542159,
    "text": "machine learning perspective there's one"
  },
  {
    "startMs": 1543360,
    "text": "latent kind of you know causal kind of"
  },
  {
    "startMs": 1544960,
    "text": "you know generative process right like"
  },
  {
    "startMs": 1546320,
    "text": "there's something that generates you"
  },
  {
    "startMs": 1548000,
    "text": "speaking it's not the pixels and then"
  },
  {
    "startMs": 1550320,
    "text": "the the audio or somehow somehow"
  },
  {
    "startMs": 1552080,
    "text": "generated by some other process like the"
  },
  {
    "startMs": 1553520,
    "text": "lips have to move in sync with with the"
  },
  {
    "startMs": 1555520,
    "text": "with the audio, right? So, I think that"
  },
  {
    "startMs": 1557520,
    "text": "that solved a lot of the issues that"
  },
  {
    "startMs": 1559039,
    "text": "previous models had or the way that"
  },
  {
    "startMs": 1560480,
    "text": "people did video generation before where"
  },
  {
    "startMs": 1562559,
    "text": "it was like, okay, we generate the"
  },
  {
    "startMs": 1563760,
    "text": "pixels and then we're going to hack"
  },
  {
    "startMs": 1565120,
    "text": "something on top of it that like moves"
  },
  {
    "startMs": 1566799,
    "text": "the lips with the audio that we"
  },
  {
    "startMs": 1568559,
    "text": "generate. And that's was very bad."
  },
  {
    "startMs": 1570708,
    "text": "[laughter]"
  },
  {
    "startMs": 1571440,
    "text": "And so, I think I think that was that's"
  },
  {
    "startMs": 1573120,
    "text": "to me that's the the I mean after V3"
  },
  {
    "startMs": 1576000,
    "text": "like you know people were like what do"
  },
  {
    "startMs": 1577360,
    "text": "you mean like there's no audio in your"
  },
  {
    "startMs": 1578640,
    "text": "model? like that makes no sense like"
  },
  {
    "startMs": 1580159,
    "text": "once it's there like you you have to"
  },
  {
    "startMs": 1581919,
    "text": "have it. So I think that was that was"
  },
  {
    "startMs": 1583440,
    "text": "the right choice and doing it to one"
  },
  {
    "startMs": 1585279,
    "text": "single generative model I think was was"
  },
  {
    "startMs": 1587200,
    "text": "the right choice."
  },
  {
    "startMs": 1588720,
    "text": ">> One thing I kind of want to also can ask"
  },
  {
    "startMs": 1590559,
    "text": "you guys an opinion as well once one"
  },
  {
    "startMs": 1592480,
    "text": "difference I find the audio and then"
  },
  {
    "startMs": 1594159,
    "text": "against the image and video is like the"
  },
  {
    "startMs": 1595840,
    "text": "audio information is less verbalized. I"
  },
  {
    "startMs": 1598799,
    "text": "mean of course the TTS and stuff is"
  },
  {
    "startMs": 1600400,
    "text": "trivial right but the when you get her"
  },
  {
    "startMs": 1602799,
    "text": "outside like how to describe music how"
  },
  {
    "startMs": 1605200,
    "text": "do you describe this like this person's"
  },
  {
    "startMs": 1607520,
    "text": "tone kind of pitch I feel the sort of"
  },
  {
    "startMs": 1610240,
    "text": "the verbalization is insufficient and"
  },
  {
    "startMs": 1612960,
    "text": "the interesting thing is that you kind"
  },
  {
    "startMs": 1614240,
    "text": "of see that in two other things like"
  },
  {
    "startMs": 1616000,
    "text": "taste taste sense and also uh say um"
  },
  {
    "startMs": 1620559,
    "text": "touch"
  },
  {
    "startMs": 1620960,
    "text": ">> like smell and then the another"
  },
  {
    "startMs": 1622880,
    "text": "interesting thing is the skin color so"
  },
  {
    "startMs": 1625760,
    "text": "skin color the the language is pretty"
  },
  {
    "startMs": 1628559,
    "text": "limited to describe the skin color and"
  },
  {
    "startMs": 1630720,
    "text": "the reason is that we're extremely uh"
  },
  {
    "startMs": 1632480,
    "text": "sensitive to the small difference"
  },
  {
    "startMs": 1633919,
    "text": "perturvations or not skin color because"
  },
  {
    "startMs": 1635919,
    "text": "that basically shows us is this person"
  },
  {
    "startMs": 1637600,
    "text": "going to kill me or is can I befriend"
  },
  {
    "startMs": 1639840,
    "text": "this person kind of those kind of"
  },
  {
    "startMs": 1641279,
    "text": "information and then I feel the smell"
  },
  {
    "startMs": 1642960,
    "text": "tastes um skin color and like sound kind"
  },
  {
    "startMs": 1646480,
    "text": "of stuff is very very tied into"
  },
  {
    "startMs": 1648960,
    "text": "primitive a like survival kind of stuff"
  },
  {
    "startMs": 1652320,
    "text": "and so our sort of sensory system is so"
  },
  {
    "startMs": 1654960,
    "text": "sensitive that it's intractable to Um,"
  },
  {
    "startMs": 1658640,
    "text": "so for example, I asked like one the"
  },
  {
    "startMs": 1660559,
    "text": "wine sort of taster and then like"
  },
  {
    "startMs": 1662559,
    "text": "professional and then he basically said"
  },
  {
    "startMs": 1663840,
    "text": "he kind of use like a language from like"
  },
  {
    "startMs": 1665840,
    "text": "a dating, you know, describing like a,"
  },
  {
    "startMs": 1667600,
    "text": "you know, partner as a way to describe"
  },
  {
    "startMs": 1670159,
    "text": "the taste because there's no sufficient"
  },
  {
    "startMs": 1672640,
    "text": "vocab to describe. Um, so I'm kind of"
  },
  {
    "startMs": 1676240,
    "text": "curious. Yeah. Do you guys feel that?"
  },
  {
    "startMs": 1678480,
    "text": ">> I think well to some extent I think the"
  },
  {
    "startMs": 1680799,
    "text": "same is true for visual information,"
  },
  {
    "startMs": 1683840,
    "text": "right? when you think about like a"
  },
  {
    "startMs": 1686399,
    "text": "certain style or a certain aesthetic,"
  },
  {
    "startMs": 1688159,
    "text": "right? Like like there are some people"
  },
  {
    "startMs": 1690080,
    "text": "who just have a much more kind of"
  },
  {
    "startMs": 1691840,
    "text": "developed like whether it's palette or"
  },
  {
    "startMs": 1693840,
    "text": "kind of visual taste and aesthetic,"
  },
  {
    "startMs": 1695679,
    "text": "right? Like I I think language just"
  },
  {
    "startMs": 1698399,
    "text": "tends to be a bit of a limiting factor"
  },
  {
    "startMs": 1700559,
    "text": "when you are trying to describe any of"
  },
  {
    "startMs": 1702159,
    "text": "these things that like we experience"
  },
  {
    "startMs": 1704240,
    "text": "with sensory information. And to your"
  },
  {
    "startMs": 1706559,
    "text": "point earlier, I think that is the kind"
  },
  {
    "startMs": 1709360,
    "text": "of the reason why we are investing in"
  },
  {
    "startMs": 1711600,
    "text": "world models and why we are pushing on"
  },
  {
    "startMs": 1713360,
    "text": "kind of the like perception and like"
  },
  {
    "startMs": 1714960,
    "text": "generation side of things because it it"
  },
  {
    "startMs": 1717760,
    "text": "is such a large part of how we as humans"
  },
  {
    "startMs": 1720399,
    "text": "navigate the world. It's a large part of"
  },
  {
    "startMs": 1722960,
    "text": "how like embodied AI navigates the"
  },
  {
    "startMs": 1725760,
    "text": "world. Um, and and I do I do think"
  },
  {
    "startMs": 1728320,
    "text": "language like does have a lot of it's"
  },
  {
    "startMs": 1730080,
    "text": "it's gotten us very far and it can"
  },
  {
    "startMs": 1731840,
    "text": "probably get us really far, but it it"
  },
  {
    "startMs": 1733840,
    "text": "feels limiting in a lot of these kind of"
  },
  {
    "startMs": 1735760,
    "text": "areas. And yeah, I don't I don't really"
  },
  {
    "startMs": 1737279,
    "text": "know how to describe, you know, like"
  },
  {
    "startMs": 1738799,
    "text": "sense and taste. Um, but yeah, I'm"
  },
  {
    "startMs": 1741840,
    "text": "curious to me."
  },
  {
    "startMs": 1743279,
    "text": ">> Um, I I yeah, I don't know that I have"
  },
  {
    "startMs": 1746159,
    "text": "thought that deeply about this yet. So,"
  },
  {
    "startMs": 1748320,
    "text": "uh, yeah, I mean yeah, I don't have a"
  },
  {
    "startMs": 1751039,
    "text": "good answer about audio. I mean like I"
  },
  {
    "startMs": 1754720,
    "text": "don't know the limit because I'm"
  },
  {
    "startMs": 1756159,
    "text": "thinking about like well what is what is"
  },
  {
    "startMs": 1757760,
    "text": "Omni bad at in terms of audio but"
  },
  {
    "startMs": 1759360,
    "text": "they're all like solvable problems I"
  },
  {
    "startMs": 1761120,
    "text": "find uh so like with more data or better"
  },
  {
    "startMs": 1764480,
    "text": "data or whatever it is so I don't know"
  },
  {
    "startMs": 1766399,
    "text": "like that we have pushed the frontier so"
  },
  {
    "startMs": 1768320,
    "text": "much that like we are have hit some sort"
  },
  {
    "startMs": 1771279,
    "text": "of limits that are rooted in"
  },
  {
    "startMs": 1773200,
    "text": "evolutionary uh kind of you know limits"
  },
  {
    "startMs": 1776240,
    "text": "imposed by humans. I don't know. He's"
  },
  {
    "startMs": 1778720,
    "text": "feeling the limits of captioning which"
  },
  {
    "startMs": 1780080,
    "text": "is the the thing I was"
  },
  {
    "startMs": 1781039,
    "text": ">> Yeah, exactly. [laughter] There there's"
  },
  {
    "startMs": 1782480,
    "text": "a lot of information in the world and it"
  },
  {
    "startMs": 1784080,
    "text": "connects to basically why we do work"
  },
  {
    "startMs": 1785679,
    "text": "modeling you mentioned. You just need"
  },
  {
    "startMs": 1787840,
    "text": "srefs sref476"
  },
  {
    "startMs": 1790320,
    "text": "and then that's your that's what your"
  },
  {
    "startMs": 1791760,
    "text": "journey does, right? I guess maybe I"
  },
  {
    "startMs": 1793120,
    "text": "can't describe this vibe but"
  },
  {
    "startMs": 1794720,
    "text": ">> well well I think that that's kind of"
  },
  {
    "startMs": 1796159,
    "text": "the point of providing some of these"
  },
  {
    "startMs": 1797760,
    "text": "references, right? Because because like"
  },
  {
    "startMs": 1799919,
    "text": "even just describing how someone talks"
  },
  {
    "startMs": 1802000,
    "text": "and like their tone and and like procity"
  },
  {
    "startMs": 1804320,
    "text": "and all of these things like I think I"
  },
  {
    "startMs": 1805919,
    "text": "think some of these terms even like I"
  },
  {
    "startMs": 1808480,
    "text": "didn't used to know what they mean,"
  },
  {
    "startMs": 1809600,
    "text": "right? Well, now"
  },
  {
    "startMs": 1810880,
    "text": ">> yes. Dispuencuencies ex like like there"
  },
  {
    "startMs": 1814320,
    "text": "there's kind of an entire vocabulary"
  },
  {
    "startMs": 1816080,
    "text": "that even if you're not kind of steeped"
  },
  {
    "startMs": 1817919,
    "text": "in a domain, which is true for actually"
  },
  {
    "startMs": 1819600,
    "text": "like most human domains that like you"
  },
  {
    "startMs": 1821360,
    "text": "don't even know what it means. Um and"
  },
  {
    "startMs": 1823520,
    "text": "sometimes it's also a question of like"
  },
  {
    "startMs": 1824880,
    "text": "if we haven't focused on those things,"
  },
  {
    "startMs": 1826399,
    "text": "you know, with the large language models"
  },
  {
    "startMs": 1827840,
    "text": "that they may also have gaps in those"
  },
  {
    "startMs": 1829279,
    "text": "areas, right? And then we feel them on"
  },
  {
    "startMs": 1830799,
    "text": "the other side with generation because"
  },
  {
    "startMs": 1832480,
    "text": "we're like fundamentally relying on on"
  },
  {
    "startMs": 1834720,
    "text": "the language models understanding of the"
  },
  {
    "startMs": 1836240,
    "text": "world to then be able to like represent"
  },
  {
    "startMs": 1838000,
    "text": "it. Um, so I yeah, it all kind of goes"
  },
  {
    "startMs": 1840320,
    "text": "back to your question about like the the"
  },
  {
    "startMs": 1842320,
    "text": "language as an intermediary. Um, but"
  },
  {
    "startMs": 1844799,
    "text": "yeah, I think to De's point like some of"
  },
  {
    "startMs": 1846080,
    "text": "these might just be like focus areas and"
  },
  {
    "startMs": 1848399,
    "text": "things that we haven't necessarily"
  },
  {
    "startMs": 1849840,
    "text": "pushed on as much as we can and like as"
  },
  {
    "startMs": 1851679,
    "text": "we will we will discover what the actual"
  },
  {
    "startMs": 1854240,
    "text": "ceiling is."
  },
  {
    "startMs": 1855919,
    "text": ">> Yeah, as a podcaster I think a lot about"
  },
  {
    "startMs": 1858880,
    "text": "sound."
  },
  {
    "startMs": 1859600,
    "text": ">> Um, and and I I'll just offer a couple"
  },
  {
    "startMs": 1862320,
    "text": "things for discussion in case in case it"
  },
  {
    "startMs": 1863919,
    "text": "triggers anything with you guys. Um I"
  },
  {
    "startMs": 1866159,
    "text": "have three domains of rough audio which"
  },
  {
    "startMs": 1867840,
    "text": "is like a music voice SFX you know is"
  },
  {
    "startMs": 1870960,
    "text": "that rough okay covers everything and"
  },
  {
    "startMs": 1873279,
    "text": "then also even within voice let's just"
  },
  {
    "startMs": 1875039,
    "text": "let's just focus on voice forget the"
  },
  {
    "startMs": 1876399,
    "text": "other two um room sound like the the"
  },
  {
    "startMs": 1878960,
    "text": "echoiness of like big room small room in"
  },
  {
    "startMs": 1881760,
    "text": "person in a car over a phone all these"
  },
  {
    "startMs": 1885200,
    "text": "like are labelable but we experience"
  },
  {
    "startMs": 1887520,
    "text": "them very differently and I I often"
  },
  {
    "startMs": 1889279,
    "text": "think like one of the tells of a AI"
  },
  {
    "startMs": 1891600,
    "text": "video is that it is studio quality"
  },
  {
    "startMs": 1893840,
    "text": "because it was recorded in a studio"
  },
  {
    "startMs": 1895120,
    "text": "video because that's your training data"
  },
  {
    "startMs": 1896880,
    "text": "and like and and to me that's one thing"
  },
  {
    "startMs": 1899519,
    "text": "actually like the most interesting thing"
  },
  {
    "startMs": 1900880,
    "text": "is just uh when I tell this is how I"
  },
  {
    "startMs": 1903679,
    "text": "convince people who are kind of"
  },
  {
    "startMs": 1904960,
    "text": "skeptical about the need for world"
  },
  {
    "startMs": 1906399,
    "text": "models because you need it even for"
  },
  {
    "startMs": 1908080,
    "text": "audio about well I'm further away from"
  },
  {
    "startMs": 1911279,
    "text": "you so I should sound a little bit"
  },
  {
    "startMs": 1912640,
    "text": "softer or more diffused and like the the"
  },
  {
    "startMs": 1915440,
    "text": "video models need to pick that up"
  },
  {
    "startMs": 1916799,
    "text": "because if they're going to do immersive"
  },
  {
    "startMs": 1918559,
    "text": "video and audio you need that"
  },
  {
    "startMs": 1921440,
    "text": ">> I I I love that example of basically"
  },
  {
    "startMs": 1923440,
    "text": "like studio quality or not in a way like"
  },
  {
    "startMs": 1925440,
    "text": "we don't have enough language to really"
  },
  {
    "startMs": 1927600,
    "text": "describe like like this kind of echoing"
  },
  {
    "startMs": 1930399,
    "text": "or like some kind of noise kind of"
  },
  {
    "startMs": 1932159,
    "text": "happening we just like don't have"
  },
  {
    "startMs": 1933600,
    "text": "precise enough and uh if you um you know"
  },
  {
    "startMs": 1936399,
    "text": "basically the reason that I think it's"
  },
  {
    "startMs": 1938159,
    "text": "quite important to have like relatively"
  },
  {
    "startMs": 1940480,
    "text": "information rich like kind of captioning"
  },
  {
    "startMs": 1942159,
    "text": "is that we kind of rely on the natural"
  },
  {
    "startMs": 1943519,
    "text": "language as a representation but if you"
  },
  {
    "startMs": 1945760,
    "text": "basically don't have enough uh"
  },
  {
    "startMs": 1947360,
    "text": "representation that basically means the"
  },
  {
    "startMs": 1949039,
    "text": "condition on the language the generation"
  },
  {
    "startMs": 1950799,
    "text": "is very multimodal and if you anything"
  },
  {
    "startMs": 1953039,
    "text": "can learn from the BAE kind of like you"
  },
  {
    "startMs": 1954880,
    "text": "know very old you know GMBA kind of"
  },
  {
    "startMs": 1956480,
    "text": "research the idea is we really want to"
  },
  {
    "startMs": 1958399,
    "text": "capture most of the stoasticity in the"
  },
  {
    "startMs": 1960720,
    "text": "later representation and then the the X"
  },
  {
    "startMs": 1963519,
    "text": "given the Z should be kind of like"
  },
  {
    "startMs": 1964960,
    "text": "deterministic so yeah"
  },
  {
    "startMs": 1966720,
    "text": ">> yeah yeah um well I hope I hope there's"
  },
  {
    "startMs": 1969200,
    "text": "more uh progress there and I'm sure you"
  },
  {
    "startMs": 1970960,
    "text": "guys are doing"
  },
  {
    "startMs": 1971760,
    "text": ">> I even actually like facial expressions"
  },
  {
    "startMs": 1973919,
    "text": "right and maybe this gets to your point"
  },
  {
    "startMs": 1975279,
    "text": "about like things that we're very"
  },
  {
    "startMs": 1976480,
    "text": "sensitive to right I think you can tell"
  },
  {
    "startMs": 1978559,
    "text": "a lot of AI content also just by from"
  },
  {
    "startMs": 1981519,
    "text": "like people's facial expressions"
  },
  {
    "startMs": 1982960,
    "text": "stressful."
  },
  {
    "startMs": 1984672,
    "text": "[laughter]"
  },
  {
    "startMs": 1984720,
    "text": ">> Yes."
  },
  {
    "startMs": 1986080,
    "text": ">> Yes. And we try not to contribute to it,"
  },
  {
    "startMs": 1988240,
    "text": "but you know, um and or or like skin"
  },
  {
    "startMs": 1991200,
    "text": "textures, right? Like like the things"
  },
  {
    "startMs": 1992720,
    "text": "that kind of make things look real in"
  },
  {
    "startMs": 1995039,
    "text": "real life. Like I you know, I can tell"
  },
  {
    "startMs": 1996720,
    "text": "from the way you're nodding or from the"
  },
  {
    "startMs": 1998240,
    "text": "way like your micro expressions are kind"
  },
  {
    "startMs": 1999919,
    "text": "of changing of like how you're reacting"
  },
  {
    "startMs": 2001279,
    "text": "to what I'm saying. Like we haven't"
  },
  {
    "startMs": 2003360,
    "text": "quite crossed that chasm. I think like"
  },
  {
    "startMs": 2006080,
    "text": "we're we're so much better than we were"
  },
  {
    "startMs": 2007760,
    "text": "a year ago."
  },
  {
    "startMs": 2008399,
    "text": ">> Yeah. Um, but there's so much more"
  },
  {
    "startMs": 2010399,
    "text": "headroom kind of in a lot of those"
  },
  {
    "startMs": 2012000,
    "text": "things that like we as humans are super"
  },
  {
    "startMs": 2013760,
    "text": "sensitive to. And like I think image"
  },
  {
    "startMs": 2015840,
    "text": "arguably probably is there because"
  },
  {
    "startMs": 2018399,
    "text": "there's there's a lot of kind of images"
  },
  {
    "startMs": 2020240,
    "text": "that I will see that like really do look"
  },
  {
    "startMs": 2022399,
    "text": "indistinguishable from reality and I"
  },
  {
    "startMs": 2024320,
    "text": "can't tell if they're generated or not."
  },
  {
    "startMs": 2026240,
    "text": ">> Better than reality"
  },
  {
    "startMs": 2027360,
    "text": ">> um or well that's a different"
  },
  {
    "startMs": 2029919,
    "text": ">> No, I I think that one of the parad"
  },
  {
    "startMs": 2031833,
    "text": "[laughter] better than what I would take"
  },
  {
    "startMs": 2032799,
    "text": "on my vacation as a photo. Yes. One of"
  },
  {
    "startMs": 2035039,
    "text": "the one of the fun experiments that we"
  },
  {
    "startMs": 2036399,
    "text": "did a while ago in the team is is like"
  },
  {
    "startMs": 2038799,
    "text": "can we generate videos that are better"
  },
  {
    "startMs": 2040240,
    "text": "than than real videos, right? So you"
  },
  {
    "startMs": 2041919,
    "text": "just take the same caption from like oh"
  },
  {
    "startMs": 2043679,
    "text": "yeah some video and then"
  },
  {
    "startMs": 2046640,
    "text": ">> recycle it. Yeah. Just just try to like"
  },
  {
    "startMs": 2048480,
    "text": "describe a real video and then generate"
  },
  {
    "startMs": 2050399,
    "text": "the equivalent version with omni and"
  },
  {
    "startMs": 2052079,
    "text": "then do a human eval. How does how does"
  },
  {
    "startMs": 2054399,
    "text": "it do? And then humans largely prefer AI"
  },
  {
    "startMs": 2057040,
    "text": "generated"
  },
  {
    "startMs": 2058960,
    "text": ">> margin."
  },
  {
    "startMs": 2060079,
    "text": ">> But because it's because it's the RL"
  },
  {
    "startMs": 2061919,
    "text": "process,"
  },
  {
    "startMs": 2062800,
    "text": ">> that's the RL process working."
  },
  {
    "startMs": 2063839,
    "text": ">> It's however you want to rationalize it."
  },
  {
    "startMs": 2065119,
    "text": "It's not necessarily the old process."
  },
  {
    "startMs": 2066320,
    "text": "It's just like I think it's just I'm not"
  },
  {
    "startMs": 2068720,
    "text": "saying this is a good result. I'm just"
  },
  {
    "startMs": 2070079,
    "text": "saying is we have optimized in a way"
  },
  {
    "startMs": 2073040,
    "text": "that like kind of potentially sort of,"
  },
  {
    "startMs": 2075440,
    "text": "you know, triggers something in the"
  },
  {
    "startMs": 2076639,
    "text": "human brain that like, oh, it looks it"
  },
  {
    "startMs": 2078480,
    "text": "looks all a lot of the videos just look"
  },
  {
    "startMs": 2081440,
    "text": "better. Like I'm not Yeah. Yeah. Yeah."
  },
  {
    "startMs": 2083359,
    "text": "on on inspection on on deeper inspection"
  },
  {
    "startMs": 2085760,
    "text": "they they would not actually be more"
  },
  {
    "startMs": 2087599,
    "text": "useful or whatever but like if you just"
  },
  {
    "startMs": 2090399,
    "text": "say side by side random YouTube video"
  },
  {
    "startMs": 2092639,
    "text": "versus"
  },
  {
    "startMs": 2093760,
    "text": ">> generated version of it will you will"
  },
  {
    "startMs": 2096079,
    "text": "just have a it will just look better"
  },
  {
    "startMs": 2097760,
    "text": "because it's more it's a sharper more"
  },
  {
    "startMs": 2099599,
    "text": "HDR uh you know the skin tone is is is"
  },
  {
    "startMs": 2103359,
    "text": "better it's not again it's not more"
  },
  {
    "startMs": 2104960,
    "text": "realistic"
  },
  {
    "startMs": 2106320,
    "text": ">> uh it doesn't solve your problem"
  },
  {
    "startMs": 2107599,
    "text": "necessarily but it it looks better"
  },
  {
    "startMs": 2109839,
    "text": ">> I I since also depend on the sensitivity"
  },
  {
    "startMs": 2112400,
    "text": "of the people. Uh I was born raised in"
  },
  {
    "startMs": 2114560,
    "text": "Japan and I think one thing I kind of"
  },
  {
    "startMs": 2116240,
    "text": "know is like they're extremely extremely"
  },
  {
    "startMs": 2117839,
    "text": "like sensitive about like you know"
  },
  {
    "startMs": 2119680,
    "text": "that's why you know like architecture"
  },
  {
    "startMs": 2121040,
    "text": "like food and stuff like they have."
  },
  {
    "startMs": 2123040,
    "text": ">> Um so I talked to like a mangar like"
  },
  {
    "startMs": 2125440,
    "text": "like artist there and he's like he's"
  },
  {
    "startMs": 2126880,
    "text": "kind of disgusted by like the generation"
  },
  {
    "startMs": 2128960,
    "text": "AI and one kind of thing he mentioned is"
  },
  {
    "startMs": 2131040,
    "text": "like the eye gaze"
  },
  {
    "startMs": 2132560,
    "text": ">> eye gaze that slight difference"
  },
  {
    "startMs": 2135200,
    "text": ">> makes me makes him kind of feel creepy"
  },
  {
    "startMs": 2137359,
    "text": "about like unnatural"
  },
  {
    "startMs": 2138640,
    "text": ">> like if you're looking a little bit off."
  },
  {
    "startMs": 2140480,
    "text": ">> Yeah. It's just uh Yeah. just like uh it"
  },
  {
    "startMs": 2143200,
    "text": "looks too fake. Yeah. To the point. So"
  },
  {
    "startMs": 2145200,
    "text": "So I think it does depend on the"
  },
  {
    "startMs": 2146720,
    "text": "sensitivity and"
  },
  {
    "startMs": 2147440,
    "text": ">> Yeah. Yeah. All I'm saying is like you"
  },
  {
    "startMs": 2149440,
    "text": "know human preferences are like a not"
  },
  {
    "startMs": 2151200,
    "text": "particularly like uh reliable barometer"
  },
  {
    "startMs": 2154480,
    "text": "of like what you should be optimizing"
  },
  {
    "startMs": 2156000,
    "text": "for like if you just ask people do you"
  },
  {
    "startMs": 2157520,
    "text": "like this or not you not necessarily get"
  },
  {
    "startMs": 2159599,
    "text": "what you wanted."
  },
  {
    "startMs": 2161040,
    "text": ">> Yeah. Let let me just kind of add one"
  },
  {
    "startMs": 2162480,
    "text": "thing but like four years ago there was"
  },
  {
    "startMs": 2164000,
    "text": "a like debate that if the prompt"
  },
  {
    "startMs": 2165440,
    "text": "engineering is going to disappear and uh"
  },
  {
    "startMs": 2167920,
    "text": "my my like you know some very powerful"
  },
  {
    "startMs": 2170400,
    "text": "people say you know it's going to"
  },
  {
    "startMs": 2171359,
    "text": "disappear but I basically said like it"
  },
  {
    "startMs": 2173440,
    "text": "shouldn't because the prompt engineering"
  },
  {
    "startMs": 2176320,
    "text": "like sort of you know specifying that is"
  },
  {
    "startMs": 2178000,
    "text": "like the the only way you can sort of"
  },
  {
    "startMs": 2180480,
    "text": "control the output sort of you know when"
  },
  {
    "startMs": 2182400,
    "text": "you have like sort of control by the AI"
  },
  {
    "startMs": 2185280,
    "text": "and what allows you to prompt engineer"
  },
  {
    "startMs": 2187680,
    "text": "is really that sensitivity. So sure"
  },
  {
    "startMs": 2190480,
    "text": "maybe like right now the AI can do a lot"
  },
  {
    "startMs": 2192400,
    "text": "of autoprompting and that and it can"
  },
  {
    "startMs": 2194320,
    "text": "generate something that's sufficient but"
  },
  {
    "startMs": 2197359,
    "text": "uh if it's like that never be satisfied"
  },
  {
    "startMs": 2199359,
    "text": "like never be satisfied with the AI's"
  },
  {
    "startMs": 2201200,
    "text": "generated content always fine-tune your"
  },
  {
    "startMs": 2203680,
    "text": "sensitivity and always kind of keep"
  },
  {
    "startMs": 2205920,
    "text": "prompting the differences. I I think to"
  },
  {
    "startMs": 2208079,
    "text": "the there's also a big difference"
  },
  {
    "startMs": 2209520,
    "text": "between like the average human untrained"
  },
  {
    "startMs": 2213119,
    "text": "eye which I I would put myself in that"
  },
  {
    "startMs": 2215359,
    "text": "bucket you know like I have I have some"
  },
  {
    "startMs": 2216720,
    "text": "aesthetic sensibilities and I've done"
  },
  {
    "startMs": 2218400,
    "text": "this long enough that you know like I"
  },
  {
    "startMs": 2220000,
    "text": "have I have a preference um but you know"
  },
  {
    "startMs": 2222960,
    "text": "like your example of a manga artist like"
  },
  {
    "startMs": 2225119,
    "text": "that's somebody who has honed a craft"
  },
  {
    "startMs": 2227280,
    "text": "like over possibly many decades. Um, and"
  },
  {
    "startMs": 2231040,
    "text": "anybody who does that, whether it's like"
  },
  {
    "startMs": 2232480,
    "text": "design, architecture, right? Like you"
  },
  {
    "startMs": 2234160,
    "text": "you you just have a very different level"
  },
  {
    "startMs": 2236960,
    "text": "of like expertise and you see things"
  },
  {
    "startMs": 2239440,
    "text": "that like the average human will not"
  },
  {
    "startMs": 2241119,
    "text": "see. But Doom is right. Like when we"
  },
  {
    "startMs": 2242480,
    "text": "look at if you were to just, you know,"
  },
  {
    "startMs": 2245520,
    "text": "um, poll 10 people on the street, they"
  },
  {
    "startMs": 2248240,
    "text": "would probably prefer the like overly"
  },
  {
    "startMs": 2251040,
    "text": "smooth like very saturated kind of"
  },
  {
    "startMs": 2253839,
    "text": ">> It's called the Instagram filter."
  },
  {
    "startMs": 2255520,
    "text": ">> It is. It is the Yeah. [laughter] Um,"
  },
  {
    "startMs": 2258079,
    "text": "and you know, and and so there's also a"
  },
  {
    "startMs": 2259839,
    "text": "little bit of a question of like what"
  },
  {
    "startMs": 2261520,
    "text": "does your default aesthetic look like if"
  },
  {
    "startMs": 2263920,
    "text": "you don't specify? But then to Shane's"
  },
  {
    "startMs": 2265760,
    "text": "point, one of the things we always try"
  },
  {
    "startMs": 2267760,
    "text": "to get these models better at is"
  },
  {
    "startMs": 2269200,
    "text": "instruction follow so that like when you"
  },
  {
    "startMs": 2271359,
    "text": "want to get them to a different outcome"
  },
  {
    "startMs": 2273359,
    "text": "like you should be able to whether"
  },
  {
    "startMs": 2274800,
    "text": "that's through language or whether"
  },
  {
    "startMs": 2276240,
    "text": "that's through references because"
  },
  {
    "startMs": 2277680,
    "text": "language is sometimes too limiting. Um,"
  },
  {
    "startMs": 2280079,
    "text": "and so like these models continue to get"
  },
  {
    "startMs": 2282079,
    "text": "better at it but they so much at work."
  },
  {
    "startMs": 2283839,
    "text": "Do do you feel pressure as a as a"
  },
  {
    "startMs": 2285760,
    "text": "product director to set the default for"
  },
  {
    "startMs": 2287680,
    "text": "the world like I mean [laughter]"
  },
  {
    "startMs": 2290720,
    "text": ">> kind of"
  },
  {
    "startMs": 2291280,
    "text": ">> maybe I should I don't know I haven't"
  },
  {
    "startMs": 2292960,
    "text": "thought about this"
  },
  {
    "startMs": 2294079,
    "text": ">> you know you know it's like someone has"
  },
  {
    "startMs": 2296400,
    "text": "to have a default the default has to"
  },
  {
    "startMs": 2298079,
    "text": "exist"
  },
  {
    "startMs": 2298640,
    "text": ">> actually I will say like we have thought"
  },
  {
    "startMs": 2300240,
    "text": "about this um and I I think one of the"
  },
  {
    "startMs": 2303520,
    "text": "so for example actually like if you look"
  },
  {
    "startMs": 2305200,
    "text": "at nanobanana generations we had like an"
  },
  {
    "startMs": 2307599,
    "text": "explosion of nanobanana infographics"
  },
  {
    "startMs": 2309680,
    "text": "when nanobanana pro came out"
  },
  {
    "startMs": 2311280,
    "text": ">> I tried it yeah"
  },
  {
    "startMs": 2312240,
    "text": ">> um yeah"
  },
  {
    "startMs": 2313440,
    "text": "I think Nurb's papers were like all you"
  },
  {
    "startMs": 2315920,
    "text": "know so so many had like infographics"
  },
  {
    "startMs": 2317920,
    "text": "generated. Can you run your uh"
  },
  {
    "startMs": 2319680,
    "text": "watermarking on it and see how many"
  },
  {
    "startMs": 2321839,
    "text": ">> uh we pro we probably could we have we"
  },
  {
    "startMs": 2324160,
    "text": "haven't done that but I saw so like my"
  },
  {
    "startMs": 2325599,
    "text": "Twitter was maybe this is just also like"
  },
  {
    "startMs": 2327200,
    "text": "the bias of my algorithm but they were"
  },
  {
    "startMs": 2329520,
    "text": "everywhere um and it was actually very"
  },
  {
    "startMs": 2331920,
    "text": "painful because um I think our default"
  },
  {
    "startMs": 2334960,
    "text": "aesthetic was a little bit too it was"
  },
  {
    "startMs": 2337119,
    "text": "too cluttered like I think that the the"
  },
  {
    "startMs": 2339200,
    "text": "model is like a bit of an overeager"
  },
  {
    "startMs": 2341119,
    "text": "student that just like learned you know"
  },
  {
    "startMs": 2343680,
    "text": "it was like oh I know all these like I"
  },
  {
    "startMs": 2345520,
    "text": "know all this information about this"
  },
  {
    "startMs": 2346800,
    "text": "concept let me like shove into the same"
  },
  {
    "startMs": 2349040,
    "text": "image. Japanese infographics 5x that"
  },
  {
    "startMs": 2352132,
    "text": "[laughter]"
  },
  {
    "startMs": 2353440,
    "text": ">> or maybe it was you know um but it just"
  },
  {
    "startMs": 2356880,
    "text": "and and"
  },
  {
    "startMs": 2357280,
    "text": ">> wait so same prompt same content if it's"
  },
  {
    "startMs": 2359839,
    "text": "in Japanese it's"
  },
  {
    "startMs": 2360720,
    "text": ">> density density"
  },
  {
    "startMs": 2361920,
    "text": ">> oh wow"
  },
  {
    "startMs": 2362880,
    "text": ">> because that's the style in Japan"
  },
  {
    "startMs": 2364880,
    "text": ">> yeah some like very you know bureaucrat"
  },
  {
    "startMs": 2367520,
    "text": "and [laughter] there's a famous word for"
  },
  {
    "startMs": 2369040,
    "text": "it yeah"
  },
  {
    "startMs": 2369839,
    "text": ">> no but we do do go through this process"
  },
  {
    "startMs": 2371280,
    "text": "with Omni we did it together right like"
  },
  {
    "startMs": 2372720,
    "text": "where like we had like a bunch of like"
  },
  {
    "startMs": 2374400,
    "text": "we like at the very end okay like this"
  },
  {
    "startMs": 2376320,
    "text": "is we did some tuning and like okay what"
  },
  {
    "startMs": 2378480,
    "text": "kind of style do we prefer right like"
  },
  {
    "startMs": 2380400,
    "text": "you know"
  },
  {
    "startMs": 2381040,
    "text": ">> is it more muted more saturated"
  },
  {
    "startMs": 2383280,
    "text": ">> we had a lot of saturation"
  },
  {
    "startMs": 2384560,
    "text": ">> yeah there was there were I think Nicole"
  },
  {
    "startMs": 2386960,
    "text": "just has PTSD so has forgotten about it"
  },
  {
    "startMs": 2388640,
    "text": "but she was very much involved in this"
  },
  {
    "startMs": 2390160,
    "text": "of like okay which which kind of color"
  },
  {
    "startMs": 2392320,
    "text": "palette do we basically prefer right and"
  },
  {
    "startMs": 2394240,
    "text": "it's you know it's it's it's not"
  },
  {
    "startMs": 2395839,
    "text": "something that like you have to make a a"
  },
  {
    "startMs": 2398079,
    "text": "trade-off there like uh"
  },
  {
    "startMs": 2400000,
    "text": ">> and and and it's because it ends up"
  },
  {
    "startMs": 2401440,
    "text": "being us right like actually it is true"
  },
  {
    "startMs": 2403040,
    "text": "like it it ends up being the modeling"
  },
  {
    "startMs": 2404720,
    "text": "teams and you could ask the question"
  },
  {
    "startMs": 2406160,
    "text": "legitimately of like are we the best"
  },
  {
    "startMs": 2407839,
    "text": "people to do that or should we actually"
  },
  {
    "startMs": 2410079,
    "text": "work with someone who like has a really"
  },
  {
    "startMs": 2412320,
    "text": "creative point of view and is more of"
  },
  {
    "startMs": 2414000,
    "text": "like you know an art director and like"
  },
  {
    "startMs": 2415760,
    "text": "has like and we kind of go back and"
  },
  {
    "startMs": 2417280,
    "text": "forth on this um"
  },
  {
    "startMs": 2419280,
    "text": ">> we have the trusted testers I'm on"
  },
  {
    "startMs": 2421040,
    "text": ">> we do we have trusted testers who give"
  },
  {
    "startMs": 2423280,
    "text": "us a lot of feedback and we take that"
  },
  {
    "startMs": 2424640,
    "text": "serious"
  },
  {
    "startMs": 2425040,
    "text": ">> very well organized by the way to have"
  },
  {
    "startMs": 2426400,
    "text": "these like weekly calls and stuff like"
  },
  {
    "startMs": 2427839,
    "text": "it's it's amazing"
  },
  {
    "startMs": 2429359,
    "text": ">> um Logan's team does a lot of that so"
  },
  {
    "startMs": 2431440,
    "text": "kudo kuda kudos kudos to Logan um who"
  },
  {
    "startMs": 2434400,
    "text": "couldn't be here today um and we have a"
  },
  {
    "startMs": 2436720,
    "text": "lot of people actually internally at"
  },
  {
    "startMs": 2438000,
    "text": "Google like Fulfur who give us like a"
  },
  {
    "startMs": 2440160,
    "text": "ton of No, no, no. Truly like who give"
  },
  {
    "startMs": 2442240,
    "text": "us a ton of feedback on like when we"
  },
  {
    "startMs": 2444400,
    "text": "when we release new checkpoints and like"
  },
  {
    "startMs": 2446240,
    "text": "sometimes it will be stuff that we like"
  },
  {
    "startMs": 2447839,
    "text": "don't see right like we would be like oh"
  },
  {
    "startMs": 2450240,
    "text": "yeah this optimization seems okay and"
  },
  {
    "startMs": 2452560,
    "text": "then they would come back what have you"
  },
  {
    "startMs": 2453839,
    "text": "done like you completely ruined my grass"
  },
  {
    "startMs": 2455760,
    "text": "you know because now the detail is all"
  },
  {
    "startMs": 2457119,
    "text": "blurry."
  },
  {
    "startMs": 2457359,
    "text": ">> I think he just noticed not not a super"
  },
  {
    "startMs": 2459520,
    "text": "secret at this point but like that our"
  },
  {
    "startMs": 2461359,
    "text": "model tends to put rings wedding rings"
  },
  {
    "startMs": 2463359,
    "text": "on on on hand. That's yeah"
  },
  {
    "startMs": 2464880,
    "text": ">> very strange. I had never noticed that"
  },
  {
    "startMs": 2466319,
    "text": "but he's like he I just saw it and"
  },
  {
    "startMs": 2467920,
    "text": "there's a faux fur channel basically."
  },
  {
    "startMs": 2469839,
    "text": ">> Uh he posted I was like why is there"
  },
  {
    "startMs": 2472240,
    "text": "wedding ring in every hand? I'm like"
  },
  {
    "startMs": 2473520,
    "text": "that's strange."
  },
  {
    "startMs": 2474160,
    "text": ">> That sounds very common reward hacking."
  },
  {
    "startMs": 2475839,
    "text": ">> Yeah. Yeah. Yeah. So but you know"
  },
  {
    "startMs": 2477280,
    "text": "something that we would not have we"
  },
  {
    "startMs": 2478800,
    "text": "would not have noticed necessarily while"
  },
  {
    "startMs": 2480079,
    "text": "while developing this right is an oral"
  },
  {
    "startMs": 2481520,
    "text": "artifact or"
  },
  {
    "startMs": 2482319,
    "text": ">> I I don't know you do have like a lot of"
  },
  {
    "startMs": 2485119,
    "text": "preference based and then you know you"
  },
  {
    "startMs": 2486960,
    "text": "may can prefer that sperious correlation"
  },
  {
    "startMs": 2489760,
    "text": "reward hacking it can happen like in"
  },
  {
    "startMs": 2491839,
    "text": "many weird ways. Yeah"
  },
  {
    "startMs": 2493040,
    "text": ">> it does. It is"
  },
  {
    "startMs": 2494240,
    "text": ">> uh this was related to another topic"
  },
  {
    "startMs": 2496319,
    "text": "that again I I try to use these"
  },
  {
    "startMs": 2498400,
    "text": "mainstage things as introductions or"
  },
  {
    "startMs": 2500960,
    "text": "ties in. Uh we have the eval track we"
  },
  {
    "startMs": 2503680,
    "text": "have character AI and YouTube talking"
  },
  {
    "startMs": 2505599,
    "text": "about how they evaluate videos. Um how"
  },
  {
    "startMs": 2508880,
    "text": "do you evaluate videos"
  },
  {
    "startMs": 2511040,
    "text": ">> apart from furer [laughter]"
  },
  {
    "startMs": 2513359,
    "text": ">> not everyone has a fauxur but also you"
  },
  {
    "startMs": 2515119,
    "text": "know I think there needs to be something"
  },
  {
    "startMs": 2516319,
    "text": "more quantitative"
  },
  {
    "startMs": 2517440,
    "text": ">> well I mean it's you improve Gemini"
  },
  {
    "startMs": 2520720,
    "text": "to improve the evaluation for video. Um"
  },
  {
    "startMs": 2523440,
    "text": "that's that's no no that's that's"
  },
  {
    "startMs": 2525599,
    "text": "definitely one way uh it's actually very"
  },
  {
    "startMs": 2527760,
    "text": "hard."
  },
  {
    "startMs": 2528240,
    "text": ">> It's very hard. It's very hard um to get"
  },
  {
    "startMs": 2530560,
    "text": "like you know audators to evaluate"
  },
  {
    "startMs": 2533280,
    "text": "things in a video like including"
  },
  {
    "startMs": 2535359,
    "text": "especially things like aesthetics right"
  },
  {
    "startMs": 2537200,
    "text": "like that it's like there are some"
  },
  {
    "startMs": 2538880,
    "text": "things that are a little bit more"
  },
  {
    "startMs": 2540400,
    "text": "objective like especially when we talk"
  },
  {
    "startMs": 2542000,
    "text": "like let's say we talk about images and"
  },
  {
    "startMs": 2543440,
    "text": "we look at like infographics text"
  },
  {
    "startMs": 2544800,
    "text": "rendering that's actually fine right"
  },
  {
    "startMs": 2546640,
    "text": "because like you can kind of OCR things"
  },
  {
    "startMs": 2548960,
    "text": "out and then you can look at like okay"
  },
  {
    "startMs": 2551040,
    "text": "this letter is like messed up and then"
  },
  {
    "startMs": 2552400,
    "text": "the whole thing is actually useless"
  },
  {
    "startMs": 2553680,
    "text": "because if like literally if a letter is"
  },
  {
    "startMs": 2555599,
    "text": "off in render text you just can't use"
  },
  {
    "startMs": 2557839,
    "text": "that asset. Right. So th those things"
  },
  {
    "startMs": 2560000,
    "text": "are like a little bit more auto ratable."
  },
  {
    "startMs": 2562400,
    "text": "Um from what we found we do rely a lot"
  },
  {
    "startMs": 2564960,
    "text": "on humans looking at things and so we do"
  },
  {
    "startMs": 2567760,
    "text": "do a lot of human evals. We do a lot of"
  },
  {
    "startMs": 2569760,
    "text": "human evals."
  },
  {
    "startMs": 2570400,
    "text": ">> Do a lot of human ev and every time Jane"
  },
  {
    "startMs": 2573520,
    "text": "is like um and every time we have a new"
  },
  {
    "startMs": 2576400,
    "text": "model we like want to do more things and"
  },
  {
    "startMs": 2578480,
    "text": "we want to like gem in more capabilities"
  },
  {
    "startMs": 2580480,
    "text": "and then we have like more emails that"
  },
  {
    "startMs": 2582480,
    "text": "we have to run. Um, and then at some"
  },
  {
    "startMs": 2584560,
    "text": "point you do get two models that are"
  },
  {
    "startMs": 2587760,
    "text": "like kind of close to each other and"
  },
  {
    "startMs": 2590240,
    "text": "then like we literally make decisions"
  },
  {
    "startMs": 2591760,
    "text": "based on like looking at outputs side by"
  },
  {
    "startMs": 2593680,
    "text": "side. Sometimes like in a room like I've"
  },
  {
    "startMs": 2597440,
    "text": "been in rooms where there's like 10 of"
  },
  {
    "startMs": 2599200,
    "text": "us and we're just like looking at video"
  },
  {
    "startMs": 2601200,
    "text": "side by side and we're like do you"
  },
  {
    "startMs": 2602720,
    "text": "prefer this or do you prefer that? like"
  },
  {
    "startMs": 2604480,
    "text": "oh wow it's"
  },
  {
    "startMs": 2605680,
    "text": ">> I mean but it is it is genuinely very"
  },
  {
    "startMs": 2607359,
    "text": "complicated the more capabilities you"
  },
  {
    "startMs": 2608640,
    "text": "add like you know even just the one"
  },
  {
    "startMs": 2610240,
    "text": "capability but it's like almost AGI"
  },
  {
    "startMs": 2612079,
    "text": "complete capabilities like video editing"
  },
  {
    "startMs": 2613599,
    "text": "right like think about video editing as"
  },
  {
    "startMs": 2615760,
    "text": "a and like editing with audio and"
  },
  {
    "startMs": 2618640,
    "text": ">> my editor will be very happy to hear"
  },
  {
    "startMs": 2619920,
    "text": "this"
  },
  {
    "startMs": 2620400,
    "text": ">> edit the hardest problem in g media"
  },
  {
    "startMs": 2623040,
    "text": ">> I mean I don't know if it's the hardest"
  },
  {
    "startMs": 2624480,
    "text": "but it's definitely there right like uh"
  },
  {
    "startMs": 2627839,
    "text": "in terms of like complexity of of"
  },
  {
    "startMs": 2630000,
    "text": "evaluation like free form video editing"
  },
  {
    "startMs": 2633680,
    "text": "is you can do anything like"
  },
  {
    "startMs": 2635280,
    "text": ">> yes uh and like I I spent a lot of money"
  },
  {
    "startMs": 2637680,
    "text": "on that and it's very hard to tell me"
  },
  {
    "startMs": 2639839,
    "text": ">> like adding those we don't have like add"
  },
  {
    "startMs": 2641680,
    "text": "a sloth eval right like uh that we"
  },
  {
    "startMs": 2644560,
    "text": ">> well now we should"
  },
  {
    "startMs": 2645200,
    "text": ">> now we should yeah yeah yeah but like"
  },
  {
    "startMs": 2646800,
    "text": "things like that like it's it's it's not"
  },
  {
    "startMs": 2649040,
    "text": "that easy to track"
  },
  {
    "startMs": 2650240,
    "text": ">> I think I'm just surprised at the sample"
  },
  {
    "startMs": 2652240,
    "text": "size that you have right like to to test"
  },
  {
    "startMs": 2654720,
    "text": "the entire surface of your models you"
  },
  {
    "startMs": 2657200,
    "text": "still rely on a magnitude of hundreds"
  },
  {
    "startMs": 2660160,
    "text": ">> no no no so we do like yeah well we do"
  },
  {
    "startMs": 2662319,
    "text": "we do a ton of human evals on like on"
  },
  {
    "startMs": 2664480,
    "text": "like you know thousands of things. Um I"
  },
  {
    "startMs": 2667440,
    "text": "I think there's also like an element of"
  },
  {
    "startMs": 2669119,
    "text": "you know we can talk about things like"
  },
  {
    "startMs": 2670560,
    "text": "live experiments right like which which"
  },
  {
    "startMs": 2672720,
    "text": "is also where you get signal on like"
  },
  {
    "startMs": 2674960,
    "text": "like some of these more minute"
  },
  {
    "startMs": 2676640,
    "text": "differences at like much larger scale"
  },
  {
    "startMs": 2678480,
    "text": "then there's auto raers which is"
  },
  {
    "startMs": 2679920,
    "text": "definitely kind of a more it's a very"
  },
  {
    "startMs": 2682160,
    "text": "well defined space I think for LLMs much"
  },
  {
    "startMs": 2685760,
    "text": "more nent for media models and then like"
  },
  {
    "startMs": 2689040,
    "text": "sometimes you still do rely on human"
  },
  {
    "startMs": 2691119,
    "text": "judgment and we do rely on things like"
  },
  {
    "startMs": 2693200,
    "text": "feedback from people who just like have"
  },
  {
    "startMs": 2695760,
    "text": "a very owned like aesthetic and and"
  },
  {
    "startMs": 2699440,
    "text": "people who just like use these models in"
  },
  {
    "startMs": 2700960,
    "text": "their workflows dayto-day, right?"
  },
  {
    "startMs": 2702160,
    "text": "Because we could also like you could"
  },
  {
    "startMs": 2704000,
    "text": "have a model that does really well on"
  },
  {
    "startMs": 2705359,
    "text": "some slice of human evals, but then it"
  },
  {
    "startMs": 2707680,
    "text": "like really breaks a workflow for"
  },
  {
    "startMs": 2709200,
    "text": "somebody. And so this is why we do like"
  },
  {
    "startMs": 2710720,
    "text": "early access programs and we try to get"
  },
  {
    "startMs": 2712480,
    "text": "feedback and then we like try to"
  },
  {
    "startMs": 2713599,
    "text": "incorporate it before we release"
  },
  {
    "startMs": 2714800,
    "text": "something more broadly. I feel like"
  },
  {
    "startMs": 2716400,
    "text": "Shane had a hot take based on his"
  },
  {
    "startMs": 2719359,
    "text": ">> expression always when we were talking"
  },
  {
    "startMs": 2720880,
    "text": "about this"
  },
  {
    "startMs": 2721599,
    "text": ">> every kind of human sort of you know"
  },
  {
    "startMs": 2723680,
    "text": "work should be gradually kind of"
  },
  {
    "startMs": 2725040,
    "text": "amortized and then the interesting thing"
  },
  {
    "startMs": 2726960,
    "text": "is the video understanding especially"
  },
  {
    "startMs": 2729040,
    "text": "like against like AI gener like"
  },
  {
    "startMs": 2731040,
    "text": "detecting air stuff is extremely"
  },
  {
    "startMs": 2733200,
    "text": "interesting uh visual task"
  },
  {
    "startMs": 2735359,
    "text": ">> and then like some of it kind of"
  },
  {
    "startMs": 2737040,
    "text": "aesthetics or this kind of visual"
  },
  {
    "startMs": 2738400,
    "text": "quality but for some of the kind of"
  },
  {
    "startMs": 2740480,
    "text": "cases like semantically doesn't make"
  },
  {
    "startMs": 2742240,
    "text": "sense for example you're taking like"
  },
  {
    "startMs": 2744319,
    "text": "some like a famous scene from a movie"
  },
  {
    "startMs": 2746240,
    "text": "and try to sort of um construct that and"
  },
  {
    "startMs": 2749520,
    "text": "then if you kind of generate it uh it"
  },
  {
    "startMs": 2751520,
    "text": "can generate something there but at some"
  },
  {
    "startMs": 2753520,
    "text": "point some of the semantic information"
  },
  {
    "startMs": 2755839,
    "text": "doesn't make sense like it's actually"
  },
  {
    "startMs": 2757440,
    "text": "inconsistent. So can the AI actually"
  },
  {
    "startMs": 2760480,
    "text": "detect that? So when I evaluate the AI"
  },
  {
    "startMs": 2763359,
    "text": "videos like oh I feel I'm so smart you"
  },
  {
    "startMs": 2765520,
    "text": "know like like AI is still kind of"
  },
  {
    "startMs": 2768160,
    "text": "behind but we should make like a lot of"
  },
  {
    "startMs": 2770480,
    "text": "effort. I think the video understanding"
  },
  {
    "startMs": 2771920,
    "text": "is extremely uh important intelligence"
  },
  {
    "startMs": 2774480,
    "text": "task uh beyond just the pure aesthetics"
  },
  {
    "startMs": 2776720,
    "text": "or the preference. Um and yeah we we"
  },
  {
    "startMs": 2780400,
    "text": "should always try to amatize the human"
  },
  {
    "startMs": 2782720,
    "text": ">> human label. Yeah."
  },
  {
    "startMs": 2784480,
    "text": ">> Yeah. Um, what data do you need? A lot"
  },
  {
    "startMs": 2789119,
    "text": "of people I talked to wanted to get in"
  },
  {
    "startMs": 2792240,
    "text": "front of you actually. Uh, they I mean"
  },
  {
    "startMs": 2794960,
    "text": "they want to be nice about it. They have"
  },
  {
    "startMs": 2796640,
    "text": "a lot of video data. They have gaming"
  },
  {
    "startMs": 2798160,
    "text": "data. They have real world video data."
  },
  {
    "startMs": 2799839,
    "text": "They have images. They have labelers."
  },
  {
    "startMs": 2802319,
    "text": "What do you want?"
  },
  {
    "startMs": 2804880,
    "text": ">> Are you like offering?"
  },
  {
    "startMs": 2806079,
    "text": ">> I'm just like this is your request for"
  },
  {
    "startMs": 2808480,
    "text": "like Okay. Okay. We get I'm sure you get"
  },
  {
    "startMs": 2810960,
    "text": "a lot of pitches, right? You get a lot"
  },
  {
    "startMs": 2812720,
    "text": "of people want to talk to you. what's"
  },
  {
    "startMs": 2814560,
    "text": "like I think actually it's the signal is"
  },
  {
    "startMs": 2817520,
    "text": "this problem this sorting out signal"
  },
  {
    "startMs": 2819119,
    "text": "from noise is the main problem so"
  },
  {
    "startMs": 2821359,
    "text": "creating a nice API of like okay if you"
  },
  {
    "startMs": 2824400,
    "text": "actually do a b and c we are interested"
  },
  {
    "startMs": 2826640,
    "text": "in that"
  },
  {
    "startMs": 2829680,
    "text": ">> um"
  },
  {
    "startMs": 2831200,
    "text": "loaded question there so uh I don't know"
  },
  {
    "startMs": 2833200,
    "text": "that there's like an easy like you know"
  },
  {
    "startMs": 2834880,
    "text": "did you do I think we we do already have"
  },
  {
    "startMs": 2836880,
    "text": "a lot of data I think it's it's"
  },
  {
    "startMs": 2839280,
    "text": ">> hard to talk about this"
  },
  {
    "startMs": 2841280,
    "text": ">> you know you want to talk about the"
  },
  {
    "startMs": 2842560,
    "text": "public I don't want to get you in"
  },
  {
    "startMs": 2843359,
    "text": "trouble Yeah,"
  },
  {
    "startMs": 2843839,
    "text": ">> but like I think"
  },
  {
    "startMs": 2844640,
    "text": ">> No, no. What I just want to say is like"
  },
  {
    "startMs": 2846480,
    "text": "hard to talk about this in a sort of you"
  },
  {
    "startMs": 2848480,
    "text": "know without trying to without I have to"
  },
  {
    "startMs": 2851119,
    "text": "think about the what I am revealing"
  },
  {
    "startMs": 2853040,
    "text": "about our project and what where we're"
  },
  {
    "startMs": 2855119,
    "text": "going. Um generally high quality data I"
  },
  {
    "startMs": 2858240,
    "text": "think maybe maybe let's just put it this"
  },
  {
    "startMs": 2859359,
    "text": "way right it's not not the secret"
  },
  {
    "startMs": 2860640,
    "text": ">> embodied I'm sorry"
  },
  {
    "startMs": 2862319,
    "text": ">> embodied data"
  },
  {
    "startMs": 2863760,
    "text": ">> I mean"
  },
  {
    "startMs": 2864720,
    "text": ">> yeah sure I mean we have sort of"
  },
  {
    "startMs": 2867119,
    "text": "announced I think publicly right that we"
  },
  {
    "startMs": 2869040,
    "text": "we have some sort of robotics"
  },
  {
    "startMs": 2870240,
    "text": "collaboration right like so I think it's"
  },
  {
    "startMs": 2871760,
    "text": "like a like or or but you because we"
  },
  {
    "startMs": 2875440,
    "text": "have a robotics team at GDM so you know"
  },
  {
    "startMs": 2877200,
    "text": "they're always interested in things like"
  },
  {
    "startMs": 2878319,
    "text": "that um I mean for Omni specifically I"
  },
  {
    "startMs": 2881680,
    "text": "think we're just quite interested just"
  },
  {
    "startMs": 2883040,
    "text": "high quality data right like you know it"
  },
  {
    "startMs": 2884880,
    "text": "it's not some sort of not necessarily"
  },
  {
    "startMs": 2887520,
    "text": "like oh random YouTube video but like"
  },
  {
    "startMs": 2889920,
    "text": "you know some a some more professional"
  },
  {
    "startMs": 2891359,
    "text": "shop things like that right the things"
  },
  {
    "startMs": 2892800,
    "text": "that those are those are things that"
  },
  {
    "startMs": 2895119,
    "text": "we're always on the lookout for like uh"
  },
  {
    "startMs": 2897680,
    "text": "and yeah"
  },
  {
    "startMs": 2898800,
    "text": ">> and I think for you know maybe this is"
  },
  {
    "startMs": 2900880,
    "text": "easier to some extent to answer for like"
  },
  {
    "startMs": 2903040,
    "text": "some of the agentic work as well like"
  },
  {
    "startMs": 2905839,
    "text": "like like actual kind of like what are"
  },
  {
    "startMs": 2908720,
    "text": "the tests that people are trying to do"
  },
  {
    "startMs": 2910480,
    "text": "right these things are actually kind of"
  },
  {
    "startMs": 2912480,
    "text": "difficult to manufacture if you're doing"
  },
  {
    "startMs": 2915119,
    "text": "it yourself or if you're like doing it"
  },
  {
    "startMs": 2916319,
    "text": "with a vendor, like what is the actual"
  },
  {
    "startMs": 2918319,
    "text": "like if you're creating a marketing"
  },
  {
    "startMs": 2919760,
    "text": "campaign, like what does that look like,"
  },
  {
    "startMs": 2921920,
    "text": "right? Like do do you start from here's"
  },
  {
    "startMs": 2924800,
    "text": "like a picture of my new product and"
  },
  {
    "startMs": 2926640,
    "text": "then I want to turn that into a video ad"
  },
  {
    "startMs": 2928640,
    "text": "and I want to turn that into a bunch of"
  },
  {
    "startMs": 2930079,
    "text": "assets that like fit fit all these"
  },
  {
    "startMs": 2931839,
    "text": "different ad formats that I need to push"
  },
  {
    "startMs": 2933520,
    "text": "onto various platforms to promote and"
  },
  {
    "startMs": 2935920,
    "text": "then like so you kind of go from this to"
  },
  {
    "startMs": 2938000,
    "text": "that and like what is that kind of"
  },
  {
    "startMs": 2939599,
    "text": "trajectory of tasks that you're that"
  },
  {
    "startMs": 2941680,
    "text": "you're like you know experiencing along"
  },
  {
    "startMs": 2943599,
    "text": "the way like that is really useful and"
  },
  {
    "startMs": 2946240,
    "text": "that is actually kind of difficult to"
  },
  {
    "startMs": 2948480,
    "text": "get right u because like we don't always"
  },
  {
    "startMs": 2951839,
    "text": "have the right firstparty surface where"
  },
  {
    "startMs": 2954079,
    "text": "people are actually doing some of these"
  },
  {
    "startMs": 2955599,
    "text": "things or like you might work with"
  },
  {
    "startMs": 2958000,
    "text": "someone who's a vendor but they don't"
  },
  {
    "startMs": 2959440,
    "text": "also don't have that product surface"
  },
  {
    "startMs": 2961359,
    "text": "right like like a lot of this kind of"
  },
  {
    "startMs": 2963200,
    "text": "information lives in the places where"
  },
  {
    "startMs": 2964640,
    "text": "people are doing these tasks and so"
  },
  {
    "startMs": 2966000,
    "text": "that's kind of difficult to get like if"
  },
  {
    "startMs": 2967359,
    "text": "anyone's figured that out you should"
  },
  {
    "startMs": 2969280,
    "text": "reach out to us"
  },
  {
    "startMs": 2971280,
    "text": ">> every channel of thought yeah every"
  },
  {
    "startMs": 2973079,
    "text": "[laughter] thought"
  },
  {
    "startMs": 2973839,
    "text": ">> every thought yeah and maybe the data"
  },
  {
    "startMs": 2976160,
    "text": "the Chinese lab is using"
  },
  {
    "startMs": 2978559,
    "text": ">> yes yeah uh you know"
  },
  {
    "startMs": 2981520,
    "text": "as a media person myself, right? Like"
  },
  {
    "startMs": 2983680,
    "text": "there's so many podcasters and people in"
  },
  {
    "startMs": 2986000,
    "text": "in marketing departments and all these"
  },
  {
    "startMs": 2987599,
    "text": "like they would happy to be your data"
  },
  {
    "startMs": 2989760,
    "text": "like you know just like put a BCI on my"
  },
  {
    "startMs": 2991760,
    "text": "head"
  },
  {
    "startMs": 2992319,
    "text": ">> and podcast [laughter] watch my things"
  },
  {
    "startMs": 2994800,
    "text": "uh because like you know there's just"
  },
  {
    "startMs": 2996400,
    "text": "endless amount of work to do like"
  },
  {
    "startMs": 2998160,
    "text": "there's so much work and this is all"
  },
  {
    "startMs": 3000640,
    "text": "like this needs to somewhat be commodity"
  },
  {
    "startMs": 3002720,
    "text": "like obviously you can be an art like an"
  },
  {
    "startMs": 3005200,
    "text": "artisan like you can be Hollywood for"
  },
  {
    "startMs": 3007040,
    "text": "like the really high quality stuff but"
  },
  {
    "startMs": 3008720,
    "text": "actually a lot of work is commodity and"
  },
  {
    "startMs": 3010640,
    "text": "like should be modelable and we want you"
  },
  {
    "startMs": 3012640,
    "text": "to do"
  },
  {
    "startMs": 3013689,
    "text": ">> [laughter]"
  },
  {
    "startMs": 3013839,
    "text": ">> And but we we want the high quality to"
  },
  {
    "startMs": 3016000,
    "text": "Demi's point right like we do want we"
  },
  {
    "startMs": 3017440,
    "text": "want the high quality."
  },
  {
    "startMs": 3018480,
    "text": ">> We want commodity. Yeah. Yes. Yes. You"
  },
  {
    "startMs": 3020160,
    "text": "want on both sides."
  },
  {
    "startMs": 3021599,
    "text": ">> Um I I just"
  },
  {
    "startMs": 3024319,
    "text": ">> Thank you for the solicitation."
  },
  {
    "startMs": 3026789,
    "text": "[laughter]"
  },
  {
    "startMs": 3027440,
    "text": ">> Uh I you know we we we also I also added"
  },
  {
    "startMs": 3030400,
    "text": "a data quality track. I I think that uh"
  },
  {
    "startMs": 3032640,
    "text": "people want to understand like what uh"
  },
  {
    "startMs": 3035520,
    "text": "at AI like how to raise the bar, right?"
  },
  {
    "startMs": 3038079,
    "text": "like like the and a lot of it is just"
  },
  {
    "startMs": 3040720,
    "text": "educating the market and educating"
  },
  {
    "startMs": 3042800,
    "text": "researchers and engineers and founders"
  },
  {
    "startMs": 3044400,
    "text": "on like this is where we're going a lot"
  },
  {
    "startMs": 3047119,
    "text": "of this is stop doing that do this do"
  },
  {
    "startMs": 3049839,
    "text": "this instead and I'm like people will"
  },
  {
    "startMs": 3051839,
    "text": "listen"
  },
  {
    "startMs": 3053359,
    "text": "yeah I don't know uh to that extent you"
  },
  {
    "startMs": 3055680,
    "text": "know"
  },
  {
    "startMs": 3056000,
    "text": ">> but I think to that to that point like"
  },
  {
    "startMs": 3057200,
    "text": "there's a lot of again just like craft"
  },
  {
    "startMs": 3059200,
    "text": "that goes into this right and there's a"
  },
  {
    "startMs": 3060880,
    "text": "lot of process like you even to the"
  },
  {
    "startMs": 3062559,
    "text": "marketing campaign example you don't"
  },
  {
    "startMs": 3064000,
    "text": "create that in like five minutes right"
  },
  {
    "startMs": 3065839,
    "text": "you like go you go through a process and"
  },
  {
    "startMs": 3067920,
    "text": "you iterate and you like pick something"
  },
  {
    "startMs": 3070079,
    "text": "over something else because you liked it"
  },
  {
    "startMs": 3072000,
    "text": "for whatever reason like maybe the eye"
  },
  {
    "startMs": 3073680,
    "text": "gaze was correct right like we just we"
  },
  {
    "startMs": 3075520,
    "text": "don't know these things right because"
  },
  {
    "startMs": 3077200,
    "text": "none of us are marketing directors and"
  },
  {
    "startMs": 3079440,
    "text": "like the models don't know these things"
  },
  {
    "startMs": 3081119,
    "text": ">> I even kind of say this for the natural"
  },
  {
    "startMs": 3082880,
    "text": "like a language as well like I I always"
  },
  {
    "startMs": 3084720,
    "text": "kind of say 99% of information is inside"
  },
  {
    "startMs": 3087119,
    "text": "people you can only extract it through"
  },
  {
    "startMs": 3089440,
    "text": "active dialogue and befriending them so"
  },
  {
    "startMs": 3091599,
    "text": "most of the stuff on the internet is"
  },
  {
    "startMs": 3093599,
    "text": "like sort of the outcome the output of"
  },
  {
    "startMs": 3095680,
    "text": "that yes but you know what are what are"
  },
  {
    "startMs": 3097599,
    "text": "all the trajectories you know how did"
  },
  {
    "startMs": 3098960,
    "text": "this person have this inspiration to"
  },
  {
    "startMs": 3100559,
    "text": "write this paper"
  },
  {
    "startMs": 3101680,
    "text": ">> what is the starting point what is the"
  },
  {
    "startMs": 3102960,
    "text": "inspiration what are the dialogue that"
  },
  {
    "startMs": 3104160,
    "text": "sparked it those kind of stuff is kind"
  },
  {
    "startMs": 3105920,
    "text": "of inside people so even you know those"
  },
  {
    "startMs": 3108160,
    "text": "kind of like even the language space is"
  },
  {
    "startMs": 3109920,
    "text": "kind of that I think the creative is"
  },
  {
    "startMs": 3111520,
    "text": "kind of similar as well there's a lot of"
  },
  {
    "startMs": 3112880,
    "text": "dark knowledge"
  },
  {
    "startMs": 3113599,
    "text": ">> yeah it's like when you write a novel"
  },
  {
    "startMs": 3114880,
    "text": "right like a novel speaks to you because"
  },
  {
    "startMs": 3117040,
    "text": "like usually there's some sort of like a"
  },
  {
    "startMs": 3119040,
    "text": "personal connection that you feel to"
  },
  {
    "startMs": 3120800,
    "text": "like the story or the trajectory or the"
  },
  {
    "startMs": 3122720,
    "text": "characters right like if you read most"
  },
  {
    "startMs": 3124720,
    "text": "of the stuff that's written by LLM's"
  },
  {
    "startMs": 3126559,
    "text": "today like it's, you know, it's it's it"
  },
  {
    "startMs": 3129440,
    "text": "starts it falls into these like default"
  },
  {
    "startMs": 3131440,
    "text": "par patterns and like the language"
  },
  {
    "startMs": 3133280,
    "text": "starts to feel really similar and all"
  },
  {
    "startMs": 3134800,
    "text": "the descriptions sound really similar."
  },
  {
    "startMs": 3136160,
    "text": "You can kind of like quickly read it as"
  },
  {
    "startMs": 3138480,
    "text": "like, oh, this is not that interesting"
  },
  {
    "startMs": 3139839,
    "text": "because like I can't connect to it,"
  },
  {
    "startMs": 3141680,
    "text": "right? Um, and again, that's that's kind"
  },
  {
    "startMs": 3143920,
    "text": "of like a human expertise."
  },
  {
    "startMs": 3146240,
    "text": ">> One nice thing recently is the Google"
  },
  {
    "startMs": 3148000,
    "text": "Cloud and the Google Deep Mind are kind"
  },
  {
    "startMs": 3149760,
    "text": "of starting to invest a lot more in the"
  },
  {
    "startMs": 3151119,
    "text": "FTEEs for the product engineers. And I"
  },
  {
    "startMs": 3153680,
    "text": "also kind of saw some uh recruiting for"
  },
  {
    "startMs": 3155440,
    "text": "the creative you know gem media kind of"
  },
  {
    "startMs": 3157040,
    "text": "space as well. So I think those are kind"
  },
  {
    "startMs": 3159599,
    "text": "of really the effort because we we kind"
  },
  {
    "startMs": 3161200,
    "text": "of feel that you know what we can kind"
  },
  {
    "startMs": 3162319,
    "text": "of do with a lot of public data there's"
  },
  {
    "startMs": 3164240,
    "text": "limits but really you know partnering"
  },
  {
    "startMs": 3166000,
    "text": "with that we can provide kind of better"
  },
  {
    "startMs": 3167760,
    "text": "models and products and yeah we kind of"
  },
  {
    "startMs": 3169680,
    "text": "feedback"
  },
  {
    "startMs": 3170480,
    "text": ">> uh we have an FD track here for the"
  },
  {
    "startMs": 3172319,
    "text": "first time every lab is announcing it."
  },
  {
    "startMs": 3174160,
    "text": "It's it's crazy. Um, one thing I'm"
  },
  {
    "startMs": 3176559,
    "text": "actually very keen on doing and I push I"
  },
  {
    "startMs": 3179520,
    "text": "push for this at Cognition as well is to"
  },
  {
    "startMs": 3181520,
    "text": "turn the FDES not just into sales and"
  },
  {
    "startMs": 3184640,
    "text": "solutions but also to EVAL's uh eval"
  },
  {
    "startMs": 3187599,
    "text": "workers."
  },
  {
    "startMs": 3188319,
    "text": ">> FD is not the sales FD is way way bigger"
  },
  {
    "startMs": 3191280,
    "text": "than that. How do you frame FDs then?"
  },
  {
    "startMs": 3194960,
    "text": "Because [laughter] I do think about it"
  },
  {
    "startMs": 3196000,
    "text": "as sales like you're you know the more"
  },
  {
    "startMs": 3198000,
    "text": "the more you customize the solution for"
  },
  {
    "startMs": 3200480,
    "text": ">> so I define post training as anything"
  },
  {
    "startMs": 3202960,
    "text": "between the pre-training and the final"
  },
  {
    "startMs": 3205040,
    "text": "user experience anything anything is a"
  },
  {
    "startMs": 3207599,
    "text": "post training"
  },
  {
    "startMs": 3208480,
    "text": ">> and to me when I first sort of you know"
  },
  {
    "startMs": 3210319,
    "text": "learned a lot about I mean FD kind of I"
  },
  {
    "startMs": 3212400,
    "text": "guess originally you know came from like"
  },
  {
    "startMs": 3213839,
    "text": "path here and then that so I guess the"
  },
  {
    "startMs": 3216000,
    "text": "kind of history is different but yeah I"
  },
  {
    "startMs": 3218000,
    "text": "think the key is really that um you know"
  },
  {
    "startMs": 3220319,
    "text": "the key is like not only to kind of work"
  },
  {
    "startMs": 3222240,
    "text": "uh with them and ensure that they kind"
  },
  {
    "startMs": 3223760,
    "text": "of know how to"
  },
  {
    "startMs": 3224800,
    "text": "but also to sort of code like derive"
  },
  {
    "startMs": 3228079,
    "text": "kind of insights that can basically kind"
  },
  {
    "startMs": 3229920,
    "text": "of help both parties. They can put the"
  },
  {
    "startMs": 3231680,
    "text": "like a lot of harness how they use the"
  },
  {
    "startMs": 3233359,
    "text": "model. We can improve like very"
  },
  {
    "startMs": 3234720,
    "text": "upstream. So how to get the customer"
  },
  {
    "startMs": 3236880,
    "text": "feedback to the modeling I feel is the"
  },
  {
    "startMs": 3239520,
    "text": "kind of more the the role I I kind of"
  },
  {
    "startMs": 3241680,
    "text": "want for the fds. Yeah."
  },
  {
    "startMs": 3243440,
    "text": ">> Yeah. Yeah. and and even for sorry just"
  },
  {
    "startMs": 3245280,
    "text": "on that like if you want to talk to us"
  },
  {
    "startMs": 3246960,
    "text": "or at least me um I I'm not going to"
  },
  {
    "startMs": 3250000,
    "text": "offer up your time um but I it's really"
  },
  {
    "startMs": 3253280,
    "text": "helpful for us to actually talk to"
  },
  {
    "startMs": 3254640,
    "text": "people who are using our models and like"
  },
  {
    "startMs": 3256319,
    "text": "understand where they're struggling uh"
  },
  {
    "startMs": 3258240,
    "text": "because again that just like it's it's"
  },
  {
    "startMs": 3260480,
    "text": "the real world task that you're actually"
  },
  {
    "startMs": 3262240,
    "text": "trying to use them for right like I will"
  },
  {
    "startMs": 3263839,
    "text": "talk to people who do kind of interior"
  },
  {
    "startMs": 3266559,
    "text": "inter interior design with some of our"
  },
  {
    "startMs": 3268880,
    "text": "image models um you know and they will"
  },
  {
    "startMs": 3271280,
    "text": "say hey like I really want to take this"
  },
  {
    "startMs": 3273520,
    "text": "pattern pattern, but then I want to"
  },
  {
    "startMs": 3274640,
    "text": "scale it across like 10 different ruck"
  },
  {
    "startMs": 3276640,
    "text": "sizes and sometimes I have like a very"
  },
  {
    "startMs": 3278640,
    "text": "custom ruck size and then the model"
  },
  {
    "startMs": 3280480,
    "text": "fails at like replicating the pattern"
  },
  {
    "startMs": 3282319,
    "text": "the same way or you know I want to do a"
  },
  {
    "startMs": 3284800,
    "text": "try on for these earrings and then the"
  },
  {
    "startMs": 3287119,
    "text": "earrings have a certain size and then"
  },
  {
    "startMs": 3288640,
    "text": "like my head has a certain size like it"
  },
  {
    "startMs": 3290720,
    "text": "has to make sense if you're actually"
  },
  {
    "startMs": 3292480,
    "text": "trying to try things on and like the"
  },
  {
    "startMs": 3294880,
    "text": "models kind of fail at a bunch of these"
  },
  {
    "startMs": 3296400,
    "text": "things that like actually happen in the"
  },
  {
    "startMs": 3298400,
    "text": "real world, right? Um and so that that's"
  },
  {
    "startMs": 3300960,
    "text": "like useful for us because for some of"
  },
  {
    "startMs": 3302559,
    "text": "these things like we don't think about"
  },
  {
    "startMs": 3304079,
    "text": "because we don't you know we don't use"
  },
  {
    "startMs": 3305440,
    "text": "the models for those tasks"
  },
  {
    "startMs": 3306559,
    "text": ">> or like um you know I think to your"
  },
  {
    "startMs": 3308960,
    "text": "point about ad campaigns or whatever"
  },
  {
    "startMs": 3310160,
    "text": "like people have like notions of brand"
  },
  {
    "startMs": 3311520,
    "text": "languages or whatever like which is"
  },
  {
    "startMs": 3313040,
    "text": ">> yes"
  },
  {
    "startMs": 3313440,
    "text": ">> like a a bunch of images or PDFs saying"
  },
  {
    "startMs": 3316880,
    "text": "things you know it's a pretty kind of"
  },
  {
    "startMs": 3318720,
    "text": "you know ambiguous question as well what"
  },
  {
    "startMs": 3320880,
    "text": "is the IKEA brand language you know is"
  },
  {
    "startMs": 3322960,
    "text": "it is it blue and yellow I mean that's"
  },
  {
    "startMs": 3325040,
    "text": "that's not a very like"
  },
  {
    "startMs": 3326000,
    "text": ">> but like what shade of blue you know."
  },
  {
    "startMs": 3327280,
    "text": ">> Yeah. Yeah. Yeah. So there there's like,"
  },
  {
    "startMs": 3328400,
    "text": "you know, and the brands are pretty"
  },
  {
    "startMs": 3329599,
    "text": "spec, you know, pretty, you know, like"
  },
  {
    "startMs": 3330960,
    "text": "they they do care about the shade of"
  },
  {
    "startMs": 3332480,
    "text": "blue. It's not shouldn't just be a"
  },
  {
    "startMs": 3333680,
    "text": "random blue and a random yellow. That's"
  },
  {
    "startMs": 3335359,
    "text": "not going to be IKEA, right? I'm just"
  },
  {
    "startMs": 3336880,
    "text": "thinking about an example. But like this"
  },
  {
    "startMs": 3338240,
    "text": "is the kind of stuff that, you know,"
  },
  {
    "startMs": 3339359,
    "text": "it's not necessarily part of our like,"
  },
  {
    "startMs": 3341200,
    "text": "you know, developing frontier models"
  },
  {
    "startMs": 3342880,
    "text": "kind of, you know, necessarily mandate,"
  },
  {
    "startMs": 3344240,
    "text": "but it's something that we do want to we"
  },
  {
    "startMs": 3345680,
    "text": "do want to fundamentally like build"
  },
  {
    "startMs": 3347760,
    "text": "products that people will use to solve"
  },
  {
    "startMs": 3350079,
    "text": "concrete tasks, not just not just"
  },
  {
    "startMs": 3352160,
    "text": "research artifacts, right? So I think"
  },
  {
    "startMs": 3353680,
    "text": "it's useful to understand what people do"
  },
  {
    "startMs": 3355359,
    "text": "care about. Uh well, I'm sure a lot of"
  },
  {
    "startMs": 3358400,
    "text": "people are very grateful for your work"
  },
  {
    "startMs": 3360319,
    "text": "and there's a lot more to do that you've"
  },
  {
    "startMs": 3361839,
    "text": "made so much progress over the last like"
  },
  {
    "startMs": 3364319,
    "text": "even just couple years of like Nano"
  },
  {
    "startMs": 3366319,
    "text": "Banana and Theo and Omni and uh I don't"
  },
  {
    "startMs": 3369119,
    "text": "know what else you got cooking but we're"
  },
  {
    "startMs": 3370880,
    "text": "very excited like you this is one of"
  },
  {
    "startMs": 3372720,
    "text": "those things where like I was very"
  },
  {
    "startMs": 3374480,
    "text": "disappointed you know when Sora shut"
  },
  {
    "startMs": 3376559,
    "text": "down and and I think like there needs to"
  },
  {
    "startMs": 3378720,
    "text": "be more general exploration of uh you"
  },
  {
    "startMs": 3381440,
    "text": "know generative models and not just you"
  },
  {
    "startMs": 3384079,
    "text": "know coding. [laughter] I think I think"
  },
  {
    "startMs": 3385920,
    "text": "that is"
  },
  {
    "startMs": 3386400,
    "text": ">> we obviously like this."
  },
  {
    "startMs": 3387680,
    "text": ">> We love coding. Love coding and and uh"
  },
  {
    "startMs": 3390400,
    "text": "yes uh but thank you so much for your"
  },
  {
    "startMs": 3392000,
    "text": "time. Uh it's been a real pleasure and I"
  },
  {
    "startMs": 3393839,
    "text": "can't wait to see what this looks like"
  },
  {
    "startMs": 3394880,
    "text": "next."
  },
  {
    "startMs": 3395200,
    "text": ">> Thank you for having us. Great question."
  },
  {
    "startMs": 3396559,
    "text": ">> Thank you everyone. [applause]"
  }
]