[
  {
    "startMs": 1309,
    "text": "[music]"
  },
  {
    "startMs": 12880,
    "text": "Okay. Hi everyone. [applause]"
  },
  {
    "startMs": 16880,
    "text": "Um, and thanks a lot for making it down"
  },
  {
    "startMs": 19920,
    "text": "to the second floor and finding my"
  },
  {
    "startMs": 21680,
    "text": "session here. Okay. So I'm Chris Manning"
  },
  {
    "startMs": 24320,
    "text": "and what I want to do today is present"
  },
  {
    "startMs": 26720,
    "text": "something about Moonlakes's approach to"
  },
  {
    "startMs": 28640,
    "text": "producing a simulation infrastructure"
  },
  {
    "startMs": 30800,
    "text": "for practical physical AI. The overall"
  },
  {
    "startMs": 33760,
    "text": "goal here is that a north star for AI"
  },
  {
    "startMs": 37200,
    "text": "and actually cognitive science as well"
  },
  {
    "startMs": 39520,
    "text": "has always been to understand and work"
  },
  {
    "startMs": 42160,
    "text": "out how to build embodied intelligence."
  },
  {
    "startMs": 45120,
    "text": "So today I'm going to tell us about some"
  },
  {
    "startMs": 47120,
    "text": "of the recent work that we've been doing"
  },
  {
    "startMs": 49280,
    "text": "at approaching a practical form of"
  },
  {
    "startMs": 52079,
    "text": "embodied artificial general"
  },
  {
    "startMs": 53600,
    "text": "intelligence."
  },
  {
    "startMs": 55440,
    "text": "But you know I'm not actually going to"
  },
  {
    "startMs": 58399,
    "text": "start there because uh you know um Swick"
  },
  {
    "startMs": 63520,
    "text": "said no you shouldn't just do that. Um,"
  },
  {
    "startMs": 67520,
    "text": "you should have a double slot. And first"
  },
  {
    "startMs": 70799,
    "text": "of all, since I've got you here today, I"
  },
  {
    "startMs": 73040,
    "text": "should have people tell you about the"
  },
  {
    "startMs": 75119,
    "text": "history. You should tell people about"
  },
  {
    "startMs": 76880,
    "text": "the history of AI and your journey"
  },
  {
    "startMs": 79280,
    "text": "through it and how you ended up here."
  },
  {
    "startMs": 81759,
    "text": "Um, so, uh, sit back, um, get ready for"
  },
  {
    "startMs": 86799,
    "text": "story time and we're going to have the"
  },
  {
    "startMs": 89280,
    "text": "really slow build and we're going to"
  },
  {
    "startMs": 91439,
    "text": "hear about all about the history of AI"
  },
  {
    "startMs": 93600,
    "text": "for 20 minutes and then we're going to"
  },
  {
    "startMs": 95759,
    "text": "hear the details about what Moon Lake is"
  },
  {
    "startMs": 98479,
    "text": "doing. Um, he looks very persuasive"
  },
  {
    "startMs": 101680,
    "text": "there, doesn't he? Whereas at the time"
  },
  {
    "startMs": 103759,
    "text": "we were having this conversation, I was"
  },
  {
    "startMs": 105600,
    "text": "having a very bad hair day. So I had"
  },
  {
    "startMs": 108240,
    "text": "very little um choice um to agree to"
  },
  {
    "startMs": 111200,
    "text": "this task. Um so here I am with some"
  },
  {
    "startMs": 114399,
    "text": "long-term remarks on how did I and all"
  },
  {
    "startMs": 117680,
    "text": "of us get to this moment from the far"
  },
  {
    "startMs": 120399,
    "text": "ago days um when AI didn't work. Um so"
  },
  {
    "startMs": 124719,
    "text": "the very beginning of AI was the"
  },
  {
    "startMs": 127840,
    "text": "Dartmouth summer research project in"
  },
  {
    "startMs": 130160,
    "text": "1956. It was for the holding of this um"
  },
  {
    "startMs": 133760,
    "text": "kind of summer group project was when um"
  },
  {
    "startMs": 138080,
    "text": "John McCarthy coined the term AI and got"
  },
  {
    "startMs": 142160,
    "text": "together um this group of people. Um so"
  },
  {
    "startMs": 145920,
    "text": "that one on the back right there, that's"
  },
  {
    "startMs": 148080,
    "text": "John McCarthy, there's Minsky and both"
  },
  {
    "startMs": 151840,
    "text": "most of them look like a real bunch of"
  },
  {
    "startMs": 153680,
    "text": "geeks as you can see. Um but the"
  },
  {
    "startMs": 155760,
    "text": "attractive one on right on the right end"
  },
  {
    "startMs": 158080,
    "text": "is you know my personal hero as more of"
  },
  {
    "startMs": 160160,
    "text": "a language guy um Claude Shannon."
  },
  {
    "startMs": 163840,
    "text": "So this is sort of the start of AI 1956."
  },
  {
    "startMs": 168239,
    "text": "It's where the the term AI came from."
  },
  {
    "startMs": 170879,
    "text": "But really there's other stuff that came"
  },
  {
    "startMs": 173440,
    "text": "earlier than 1956."
  },
  {
    "startMs": 176239,
    "text": "So really starting in the 40s and early"
  },
  {
    "startMs": 178560,
    "text": "in the 50s there was work on"
  },
  {
    "startMs": 180640,
    "text": "cybernetics. So cybernetics sought to"
  },
  {
    "startMs": 183360,
    "text": "tie together communications control and"
  },
  {
    "startMs": 186080,
    "text": "feedback in living things and computers."
  },
  {
    "startMs": 188800,
    "text": "Um so uh you know kubernetes is really"
  },
  {
    "startMs": 192800,
    "text": "the same word as cybernetics just less"
  },
  {
    "startMs": 195519,
    "text": "anglicized um but has very different"
  },
  {
    "startMs": 198000,
    "text": "meaning. Um yeah so this led to the"
  },
  {
    "startMs": 200640,
    "text": "earliest work on neural networks and the"
  },
  {
    "startMs": 203440,
    "text": "perceptron network. So here's Frank"
  },
  {
    "startMs": 205840,
    "text": "Rosenlat's perceptron um that was you"
  },
  {
    "startMs": 209519,
    "text": "know predated the term artificial"
  },
  {
    "startMs": 211840,
    "text": "intelligence and as you can see like in"
  },
  {
    "startMs": 214239,
    "text": "those days you actually used to wire"
  },
  {
    "startMs": 216159,
    "text": "neural networks now this modern matrix"
  },
  {
    "startMs": 218959,
    "text": "multiplication but another thing that"
  },
  {
    "startMs": 221599,
    "text": "actually started before the term"
  },
  {
    "startMs": 224480,
    "text": "artificial intelligence and has been my"
  },
  {
    "startMs": 227760,
    "text": "long-term area of interest is doing"
  },
  {
    "startMs": 230560,
    "text": "things in natural language processing"
  },
  {
    "startMs": 232640,
    "text": "and natural language processing"
  },
  {
    "startMs": 234720,
    "text": "began as machine translation. And so"
  },
  {
    "startMs": 237920,
    "text": "here we are again in 1954,"
  },
  {
    "startMs": 241360,
    "text": "two years before the term artificial"
  },
  {
    "startMs": 243200,
    "text": "intelligence was coined. And here's the"
  },
  {
    "startMs": 246400,
    "text": "front page of the New York Times. Now,"
  },
  {
    "startMs": 249200,
    "text": "there's a way in which the front page of"
  },
  {
    "startMs": 250959,
    "text": "the New York Times in 1954 is"
  },
  {
    "startMs": 254000,
    "text": "surprisingly reminiscent of the issues"
  },
  {
    "startMs": 256799,
    "text": "that you see in the front page of the"
  },
  {
    "startMs": 258560,
    "text": "New York Times um in 2026 because it's"
  },
  {
    "startMs": 263120,
    "text": "exactly the same issue. President"
  },
  {
    "startMs": 265600,
    "text": "proposing ending citizenship for people."
  },
  {
    "startMs": 268560,
    "text": "Um not much has changed in the um"
  },
  {
    "startMs": 271120,
    "text": "intervening 75 years. Um allegations of"
  },
  {
    "startMs": 275280,
    "text": "various kinds of behavior akin to"
  },
  {
    "startMs": 277199,
    "text": "treason. All sounds very familiar. Um,"
  },
  {
    "startMs": 280000,
    "text": "but this was the top half of the page."
  },
  {
    "startMs": 281840,
    "text": "And if you went down to the bottom half"
  },
  {
    "startMs": 283680,
    "text": "of the page, um, you found this article."
  },
  {
    "startMs": 286880,
    "text": "Russian is turned into English by a fast"
  },
  {
    "startMs": 289280,
    "text": "electronic, um, translator. A public"
  },
  {
    "startMs": 292080,
    "text": "demonstration of what's believed to be"
  },
  {
    "startMs": 294000,
    "text": "the first successful use of a machine to"
  },
  {
    "startMs": 296320,
    "text": "translate meaningful text from one"
  },
  {
    "startMs": 298080,
    "text": "language to another took place here"
  },
  {
    "startMs": 300639,
    "text": "yesterday afternoon. Um, and here's a"
  },
  {
    "startMs": 303520,
    "text": "little bit of video of showing that"
  },
  {
    "startMs": 305520,
    "text": "system."
  },
  {
    "startMs": 311919,
    "text": "into"
  },
  {
    "startMs": 316479,
    "text": "one of the first nonmerica"
  },
  {
    "startMs": 326000,
    "text": "were made that the computer would"
  },
  {
    "startMs": 327280,
    "text": "replace most human translators."
  },
  {
    "startMs": 332560,
    "text": ">> Okay. Um, and we'll come back to that"
  },
  {
    "startMs": 335520,
    "text": "again in a little bit. Um, but coming"
  },
  {
    "startMs": 338400,
    "text": "off of the founding of AI, um, by John"
  },
  {
    "startMs": 342880,
    "text": "McCarthy, fairly soon after that, John"
  },
  {
    "startMs": 345680,
    "text": "McCarthy moved to Stanford and founded"
  },
  {
    "startMs": 348960,
    "text": "Stanford Artificial Intelligence, the"
  },
  {
    "startMs": 350960,
    "text": "Stanford Artificial Intelligence Lab,"
  },
  {
    "startMs": 353680,
    "text": "starting from 1963."
  },
  {
    "startMs": 356000,
    "text": "And the original Stanford AI lab was in"
  },
  {
    "startMs": 359199,
    "text": "this um, building up in the foothills."
  },
  {
    "startMs": 361840,
    "text": "Um if you down if you know the South Bay"
  },
  {
    "startMs": 364639,
    "text": "well um if you've ever been to a"
  },
  {
    "startMs": 366880,
    "text": "restraero and you look over next to a"
  },
  {
    "startMs": 369520,
    "text": "restraero um where there's the Portola"
  },
  {
    "startMs": 372160,
    "text": "pastures um horse area or if you ride a"
  },
  {
    "startMs": 375759,
    "text": "horse um that's where this AI lab used"
  },
  {
    "startMs": 378400,
    "text": "to be and that was the site of many of"
  },
  {
    "startMs": 380720,
    "text": "the um founding um work in artificial"
  },
  {
    "startMs": 384400,
    "text": "intelligence and also a lot of other"
  },
  {
    "startMs": 386319,
    "text": "stuff. Um the Stanford AI lab was the"
  },
  {
    "startMs": 389199,
    "text": "where the very first video game"
  },
  {
    "startMs": 391520,
    "text": "tournament was played in 1972."
  },
  {
    "startMs": 395520,
    "text": "Um now McCarthy himself um was um"
  },
  {
    "startMs": 400720,
    "text": "mathematical logician and nearly all of"
  },
  {
    "startMs": 403520,
    "text": "his work was sort of building out these"
  },
  {
    "startMs": 405440,
    "text": "not notions of mathematical logic but in"
  },
  {
    "startMs": 408960,
    "text": "his thinking he was very wide ranging"
  },
  {
    "startMs": 412240,
    "text": "and so he"
  },
  {
    "startMs": 415840,
    "text": "he liked the idea of how could we build"
  },
  {
    "startMs": 418800,
    "text": "an embodied artificial general"
  },
  {
    "startMs": 420880,
    "text": "intelligence and was very happy to sort"
  },
  {
    "startMs": 423520,
    "text": "of provide the facilities and the money"
  },
  {
    "startMs": 426400,
    "text": "for other people to start to explore"
  },
  {
    "startMs": 428479,
    "text": "this. Um so this was the Stanford AI"
  },
  {
    "startMs": 431759,
    "text": "Labs um very first robot um the Stanford"
  },
  {
    "startMs": 435759,
    "text": "cart not very fancy um the old um sale"
  },
  {
    "startMs": 439759,
    "text": "building um but more relevant to what"
  },
  {
    "startMs": 442560,
    "text": "we're going to talk about today for"
  },
  {
    "startMs": 445039,
    "text": "simulation infrastructures for AI um was"
  },
  {
    "startMs": 448800,
    "text": "another robot um Stanford SRRI shaky"
  },
  {
    "startMs": 452080,
    "text": "robot"
  },
  {
    "startMs": 453520,
    "text": ">> mach"
  },
  {
    "startMs": 457599,
    "text": "audio Um, so this was a robot that could"
  },
  {
    "startMs": 461919,
    "text": "um perceive the environment, move"
  },
  {
    "startMs": 464800,
    "text": "around, move boxes and so on in the"
  },
  {
    "startMs": 468400,
    "text": "environment and do things like that."
  },
  {
    "startMs": 472880,
    "text": "Okay. And so this sort of started to"
  },
  {
    "startMs": 476319,
    "text": "explore this idea of having an embodied"
  },
  {
    "startMs": 480319,
    "text": "in artificial intelligence that had an"
  },
  {
    "startMs": 483280,
    "text": "understanding of a world and its"
  },
  {
    "startMs": 485520,
    "text": "environment."
  },
  {
    "startMs": 487599,
    "text": "Um, just one little distraction that's"
  },
  {
    "startMs": 490240,
    "text": "not really about AI. Um, who here has a"
  },
  {
    "startMs": 493680,
    "text": "connection to Stanford? Stanford"
  },
  {
    "startMs": 496720,
    "text": "connections. Who here has a connection"
  },
  {
    "startMs": 499199,
    "text": "to Berkeley? Berkeley connections. Less"
  },
  {
    "startMs": 502400,
    "text": "Berkeley people. Anybody with a"
  },
  {
    "startMs": 503840,
    "text": "connection to both places? There are"
  },
  {
    "startMs": 505440,
    "text": "some people who've been to both of them."
  },
  {
    "startMs": 507440,
    "text": "No. Okay. Um, so I just have to put in"
  },
  {
    "startMs": 511120,
    "text": "my advertisement for early Stanford. Um"
  },
  {
    "startMs": 514959,
    "text": "if you're an artificial intelligence"
  },
  {
    "startMs": 517039,
    "text": "person you know the story in the early"
  },
  {
    "startMs": 520880,
    "text": "days is all Stanford. I mean there you"
  },
  {
    "startMs": 523839,
    "text": "know there are two different stories I"
  },
  {
    "startMs": 525600,
    "text": "can tell right you know if you're in the"
  },
  {
    "startMs": 527600,
    "text": "history of west coast universities"
  },
  {
    "startMs": 530480,
    "text": "really in the first half of the 20th"
  },
  {
    "startMs": 532880,
    "text": "century all the prominent stuff was"
  },
  {
    "startMs": 535279,
    "text": "Berkeley and that Stanford was um this"
  },
  {
    "startMs": 538720,
    "text": "sort of pokey regional school but once"
  },
  {
    "startMs": 541200,
    "text": "you get to the second half of the 20th"
  },
  {
    "startMs": 544320,
    "text": "century which is the era of computers"
  },
  {
    "startMs": 546880,
    "text": "and AI um it all happened at Stanford um"
  },
  {
    "startMs": 552399,
    "text": "So that you know if you look at things"
  },
  {
    "startMs": 554800,
    "text": "like the beginnings of the internet, the"
  },
  {
    "startMs": 556880,
    "text": "Arpanet, um here it is in 1972"
  },
  {
    "startMs": 560480,
    "text": "and it connects up some UC campuses,"
  },
  {
    "startMs": 563600,
    "text": "Berke um UCLA and Santa Barbara and"
  },
  {
    "startMs": 567279,
    "text": "connects to Stanford, MIT, Harvard, but"
  },
  {
    "startMs": 570720,
    "text": "no Berkeley. Go ahead to 1980. Um"
  },
  {
    "startMs": 574480,
    "text": "Berkeley still isn't part of the"
  },
  {
    "startMs": 576240,
    "text": "internet. you know, it's made it as far"
  },
  {
    "startMs": 578000,
    "text": "as Hawaii and London and Berkeley is"
  },
  {
    "startMs": 580959,
    "text": "still not there. Um, and really in all"
  },
  {
    "startMs": 583680,
    "text": "of this period, there just wasn't AI at"
  },
  {
    "startMs": 587440,
    "text": "Berkeley. I mean, there was other stuff"
  },
  {
    "startMs": 589440,
    "text": "that went on at Berkeley, I should be"
  },
  {
    "startMs": 591120,
    "text": "fair. There was important systems work."
  },
  {
    "startMs": 593600,
    "text": "There was ingress database and BSD Unix,"
  },
  {
    "startMs": 596560,
    "text": "which some of you probably remember if"
  },
  {
    "startMs": 598480,
    "text": "you have gray hair like me. Um, but you"
  },
  {
    "startMs": 601680,
    "text": "know, essentially through the 60s and"
  },
  {
    "startMs": 604720,
    "text": "70s and the first half of the 80s, there"
  },
  {
    "startMs": 607360,
    "text": "was just no AI at Berkeley. And so AI at"
  },
  {
    "startMs": 610240,
    "text": "Berkeley really only got underway in"
  },
  {
    "startMs": 612640,
    "text": "1986"
  },
  {
    "startMs": 614399,
    "text": "um when two fresh Stanford PhD grads um"
  },
  {
    "startMs": 617600,
    "text": "Jendra Malik and Stuart Russell um moved"
  },
  {
    "startMs": 621440,
    "text": "from Stanford um to Berkeley to take up"
  },
  {
    "startMs": 624240,
    "text": "professorships. Now of course that's 40"
  },
  {
    "startMs": 626560,
    "text": "years ago now. So, they've had they've"
  },
  {
    "startMs": 628240,
    "text": "had um AI for a while at Berkeley now,"
  },
  {
    "startMs": 630880,
    "text": "but not in the old days. Um, if you'd"
  },
  {
    "startMs": 633920,
    "text": "like to know more about this history of"
  },
  {
    "startMs": 635760,
    "text": "the old days, um, a couple of years ago,"
  },
  {
    "startMs": 638240,
    "text": "we actually put together, um, a story of"
  },
  {
    "startMs": 641040,
    "text": "the first 60 years of, um, AI at"
  },
  {
    "startMs": 643440,
    "text": "Stanford. Conveniently, um, the 60 years"
  },
  {
    "startMs": 646959,
    "text": "stopped just before the arrival of chat"
  },
  {
    "startMs": 649279,
    "text": "GPT and large language models. Um, and"
  },
  {
    "startMs": 653279,
    "text": "you can, um, find it on YouTube."
  },
  {
    "startMs": 656640,
    "text": "Um, okay. So, that was my very brief"
  },
  {
    "startMs": 659200,
    "text": "potted history of AI. Um, what did I do"
  },
  {
    "startMs": 663839,
    "text": "for my life? Well, what I did for my"
  },
  {
    "startMs": 666560,
    "text": "life was to be a natural language"
  },
  {
    "startMs": 669200,
    "text": "processing person. And so, a natural"
  },
  {
    "startMs": 671680,
    "text": "language processing that was where we"
  },
  {
    "startMs": 674880,
    "text": "developed language models. And language"
  },
  {
    "startMs": 678320,
    "text": "models actually in some form go back a"
  },
  {
    "startMs": 683120,
    "text": "very very long way. So the first"
  },
  {
    "startMs": 686079,
    "text": "language model was proposed by Andre"
  },
  {
    "startMs": 689680,
    "text": "Marov who invented Markoff models right."
  },
  {
    "startMs": 692560,
    "text": "So in developing the idea of Markoff"
  },
  {
    "startMs": 695360,
    "text": "models he actually did it with language"
  },
  {
    "startMs": 698399,
    "text": "um taking a novel of Pushkin's Eugene"
  },
  {
    "startMs": 701440,
    "text": "Anagen and developed a character level"
  },
  {
    "startMs": 704720,
    "text": "language model over that. Um but the"
  },
  {
    "startMs": 707279,
    "text": "famous thing is then back to Claude"
  },
  {
    "startMs": 709200,
    "text": "Shannon who we saw earlier who sort of"
  },
  {
    "startMs": 711600,
    "text": "formalized information theory and"
  },
  {
    "startMs": 714160,
    "text": "started to build um word and character"
  },
  {
    "startMs": 717760,
    "text": "engram language models in the late 1940s"
  },
  {
    "startMs": 721120,
    "text": "which is the dominant stuff that we went"
  },
  {
    "startMs": 723360,
    "text": "on and using and in particular um then"
  },
  {
    "startMs": 726800,
    "text": "in 1975"
  },
  {
    "startMs": 728880,
    "text": "again so back in the sort of early part"
  },
  {
    "startMs": 731200,
    "text": "of AI um a famous group at IBM Fred"
  },
  {
    "startMs": 735120,
    "text": "Gelan's group at IBM sort of defined"
  },
  {
    "startMs": 738480,
    "text": "well they came up with the term language"
  },
  {
    "startMs": 740880,
    "text": "model that we still use today and they"
  },
  {
    "startMs": 743839,
    "text": "defined this idea of probabilistic"
  },
  {
    "startMs": 746000,
    "text": "models of text that could be used as a"
  },
  {
    "startMs": 749120,
    "text": "basis for all kinds of speech and"
  },
  {
    "startMs": 751680,
    "text": "natural language processing"
  },
  {
    "startMs": 753040,
    "text": "applications. So really from very early"
  },
  {
    "startMs": 756720,
    "text": "on in the speech and NLP tradition,"
  },
  {
    "startMs": 759440,
    "text": "language models were seen as a central"
  },
  {
    "startMs": 761680,
    "text": "technology that other people in other"
  },
  {
    "startMs": 763920,
    "text": "areas of AI and um machine learning just"
  },
  {
    "startMs": 768399,
    "text": "didn't really know about. And it was"
  },
  {
    "startMs": 770560,
    "text": "what enabled"
  },
  {
    "startMs": 773040,
    "text": "um interesting good things to happen"
  },
  {
    "startMs": 775839,
    "text": "early in speech and various areas of"
  },
  {
    "startMs": 779519,
    "text": "NLP. So both early speech recognition"
  },
  {
    "startMs": 782639,
    "text": "systems but also the kind of um machine"
  },
  {
    "startMs": 786720,
    "text": "translation that you got at Google"
  },
  {
    "startMs": 789040,
    "text": "starting about 2007. It was considerably"
  },
  {
    "startMs": 792800,
    "text": "powered by language models. And so I"
  },
  {
    "startMs": 796079,
    "text": "think there's actually a kind of an"
  },
  {
    "startMs": 797680,
    "text": "interesting story here of the"
  },
  {
    "startMs": 799839,
    "text": "development of our modern AI"
  },
  {
    "startMs": 803279,
    "text": "um sorry which is um slightly different"
  },
  {
    "startMs": 806639,
    "text": "to the story most people remember which"
  },
  {
    "startMs": 810560,
    "text": "um is a story that's dominated by the"
  },
  {
    "startMs": 813600,
    "text": "vision story because the vision sort of"
  },
  {
    "startMs": 815920,
    "text": "came in most people's head to be seen as"
  },
  {
    "startMs": 819120,
    "text": "the sort of entry place of neural"
  },
  {
    "startMs": 821440,
    "text": "networks in a large scale. Um but"
  },
  {
    "startMs": 824320,
    "text": "nevertheless I'll point out the"
  },
  {
    "startMs": 825839,
    "text": "limitation of this early work in"
  },
  {
    "startMs": 827519,
    "text": "language models. It was used for"
  },
  {
    "startMs": 829120,
    "text": "spelling correction, machine"
  },
  {
    "startMs": 830800,
    "text": "translation, all of these things. But at"
  },
  {
    "startMs": 833519,
    "text": "this time, nobody thought of language"
  },
  {
    "startMs": 836000,
    "text": "models as these were going to solve"
  },
  {
    "startMs": 837600,
    "text": "artificial intelligence. We still all"
  },
  {
    "startMs": 840000,
    "text": "bought the old AI story of that we're"
  },
  {
    "startMs": 842560,
    "text": "going to need memories, knowledge"
  },
  {
    "startMs": 844320,
    "text": "representations, planning systems,"
  },
  {
    "startMs": 846480,
    "text": "reasoning systems, all of these classic"
  },
  {
    "startMs": 848959,
    "text": "AI things."
  },
  {
    "startMs": 851199,
    "text": "Yeah. So here's the sort of history of"
  },
  {
    "startMs": 853760,
    "text": "large language models from an NLP"
  },
  {
    "startMs": 856399,
    "text": "suspect perspective. So the very first"
  },
  {
    "startMs": 858880,
    "text": "mention of the term large language"
  },
  {
    "startMs": 860800,
    "text": "models was in 1998 as far as I can tell."
  },
  {
    "startMs": 864639,
    "text": "But you know really that was sort of a"
  },
  {
    "startMs": 866560,
    "text": "how are we going to store all this text"
  },
  {
    "startMs": 869040,
    "text": "um to build a model. So the first"
  },
  {
    "startMs": 871199,
    "text": "interesting um connection of large"
  },
  {
    "startMs": 873440,
    "text": "language models was in 2000 when Joshua"
  },
  {
    "startMs": 876240,
    "text": "Benjio and colleagues defined new"
  },
  {
    "startMs": 879120,
    "text": "probabilistic language models. the first"
  },
  {
    "startMs": 881680,
    "text": "neural language model. Um, but as you"
  },
  {
    "startMs": 884160,
    "text": "can see from those stats of 32 million"
  },
  {
    "startMs": 887040,
    "text": "um token corpus, 31,000word vocabulary,"
  },
  {
    "startMs": 890720,
    "text": "this was a teeny model because at that"
  },
  {
    "startMs": 893199,
    "text": "point they just didn't have enough"
  },
  {
    "startMs": 895040,
    "text": "compute to do anything interesting. Um,"
  },
  {
    "startMs": 898240,
    "text": "but interestingly as early as 2007"
  },
  {
    "startMs": 902320,
    "text": "at Google they were able to solve the"
  },
  {
    "startMs": 905760,
    "text": "compute problem and the data problem."
  },
  {
    "startMs": 908639,
    "text": "Um, so most people forget this now, but"
  },
  {
    "startMs": 911680,
    "text": "in 2007,"
  },
  {
    "startMs": 913680,
    "text": "um, Google had a language model that was"
  },
  {
    "startMs": 916800,
    "text": "built on two trillion tokens of text."
  },
  {
    "startMs": 919839,
    "text": "So, you know, that's a bit smaller than"
  },
  {
    "startMs": 922000,
    "text": "the state-of-the-art models now that"
  },
  {
    "startMs": 923680,
    "text": "might be trained on 15 trillions of"
  },
  {
    "startMs": 925680,
    "text": "text, but it's actually the same order"
  },
  {
    "startMs": 927600,
    "text": "of magnitude, right? So we all already"
  },
  {
    "startMs": 930800,
    "text": "in sort of two decades ago there was the"
  },
  {
    "startMs": 934000,
    "text": "scale of data and compute to build a"
  },
  {
    "startMs": 937279,
    "text": "form of language model. The big problem"
  },
  {
    "startMs": 939760,
    "text": "was they didn't have enough model"
  },
  {
    "startMs": 942079,
    "text": "flexibility that the kind of modern"
  },
  {
    "startMs": 944160,
    "text": "powerful neural networks that we use"
  },
  {
    "startMs": 946000,
    "text": "today hadn't yet been invented."
  },
  {
    "startMs": 948880,
    "text": "Um and so it was then only in 2018 that"
  },
  {
    "startMs": 952959,
    "text": "modern large language models started to"
  },
  {
    "startMs": 955759,
    "text": "appear using transformers. But the early"
  },
  {
    "startMs": 958959,
    "text": "versions of those um were back to small"
  },
  {
    "startMs": 962079,
    "text": "amounts of data. So the early GPT model"
  },
  {
    "startMs": 965360,
    "text": "only used 3.3 billion tokens. So down"
  },
  {
    "startMs": 968639,
    "text": "three orders of magnitude in size again."
  },
  {
    "startMs": 971199,
    "text": "So we reverted to not enough data. And"
  },
  {
    "startMs": 974000,
    "text": "then it was eventually only in 2020 with"
  },
  {
    "startMs": 977199,
    "text": "GPT3 forward that all of those things"
  },
  {
    "startMs": 980160,
    "text": "came together and language models really"
  },
  {
    "startMs": 982560,
    "text": "took off as we know about it now."
  },
  {
    "startMs": 986399,
    "text": "Um and so that led to this surprising"
  },
  {
    "startMs": 989360,
    "text": "victory of natural language processing."
  },
  {
    "startMs": 991920,
    "text": "I mean it actually gives me a bit of a"
  },
  {
    "startMs": 993519,
    "text": "laugh um that if in the sort of mass"
  },
  {
    "startMs": 996480,
    "text": "media if you see a reference to AI these"
  },
  {
    "startMs": 999600,
    "text": "days with pretty high probability"
  },
  {
    "startMs": 1001440,
    "text": "they're actually going to be talking"
  },
  {
    "startMs": 1002880,
    "text": "about a large language model which"
  },
  {
    "startMs": 1005120,
    "text": "wasn't the way it used to be where NLP"
  },
  {
    "startMs": 1007120,
    "text": "used to be a fairly marginal area of"
  },
  {
    "startMs": 1009199,
    "text": "artificial intelligence. And as we all"
  },
  {
    "startMs": 1011680,
    "text": "know, these large language models have"
  },
  {
    "startMs": 1014399,
    "text": "allowed us to do amazing amazing things."
  },
  {
    "startMs": 1017759,
    "text": "And so the ability of these systems to"
  },
  {
    "startMs": 1021600,
    "text": "reason and solve complex problems like"
  },
  {
    "startMs": 1024319,
    "text": "math problems is just completely"
  },
  {
    "startMs": 1026640,
    "text": "stunning. I mean, you know, even for"
  },
  {
    "startMs": 1029280,
    "text": "someone like me who's spent my 30 years"
  },
  {
    "startMs": 1031439,
    "text": "working in natural language processing,"
  },
  {
    "startMs": 1033760,
    "text": "I kind of find it hard to believe that"
  },
  {
    "startMs": 1036400,
    "text": "you can pick really difficult math"
  },
  {
    "startMs": 1039120,
    "text": "problems of a kind I certainly could not"
  },
  {
    "startMs": 1041679,
    "text": "solve myself and just feed it into a"
  },
  {
    "startMs": 1044240,
    "text": "large language model and somehow it can"
  },
  {
    "startMs": 1047199,
    "text": "um chunk along doing its test time"
  },
  {
    "startMs": 1049280,
    "text": "thinking and come up with right answers."
  },
  {
    "startMs": 1051760,
    "text": "I mean it's just been a sort of a"
  },
  {
    "startMs": 1053840,
    "text": "stunning breakthrough from an unexpected"
  },
  {
    "startMs": 1056559,
    "text": "direction. Um so that's been most of my"
  },
  {
    "startMs": 1059760,
    "text": "world but then here we are um now"
  },
  {
    "startMs": 1063600,
    "text": "getting back into the main topic of the"
  },
  {
    "startMs": 1065760,
    "text": "talk of well where are we um why is"
  },
  {
    "startMs": 1070160,
    "text": "embodied artificial general intelligence"
  },
  {
    "startMs": 1073360,
    "text": "a north star that we should be looking"
  },
  {
    "startMs": 1075360,
    "text": "at and the reason for that is even"
  },
  {
    "startMs": 1078720,
    "text": "though it's been so amazing what can be"
  },
  {
    "startMs": 1082720,
    "text": "done with large language models they're"
  },
  {
    "startMs": 1086480,
    "text": "still this textbased description of the"
  },
  {
    "startMs": 1089760,
    "text": "world. And it turns out that you can do"
  },
  {
    "startMs": 1093120,
    "text": "a lot of stuff with a textbased"
  },
  {
    "startMs": 1095280,
    "text": "description of the world. I think it's"
  },
  {
    "startMs": 1097440,
    "text": "true that we can do just way more than"
  },
  {
    "startMs": 1101440,
    "text": "almost anybody believed possible with a"
  },
  {
    "startMs": 1103840,
    "text": "textbased description of the world."
  },
  {
    "startMs": 1106160,
    "text": "Certainly a lot of prominent people in"
  },
  {
    "startMs": 1108640,
    "text": "robotics and computer vision spent a"
  },
  {
    "startMs": 1110880,
    "text": "decade saying, \"Oh, you'll never be able"
  },
  {
    "startMs": 1112559,
    "text": "to do that with a large language model.\""
  },
  {
    "startMs": 1115120,
    "text": "where actually we've been able to do a"
  },
  {
    "startMs": 1116960,
    "text": "lot of that with a large language model."
  },
  {
    "startMs": 1119039,
    "text": "But still at the end of the day um we do"
  },
  {
    "startMs": 1122720,
    "text": "actually want to deal with the world"
  },
  {
    "startMs": 1124240,
    "text": "around us and have artificial"
  },
  {
    "startMs": 1126400,
    "text": "intelligence that can operate in our"
  },
  {
    "startMs": 1128880,
    "text": "world. And so the question is then how"
  },
  {
    "startMs": 1131120,
    "text": "can we build these embodied artificial"
  },
  {
    "startMs": 1133440,
    "text": "general intelligences?"
  },
  {
    "startMs": 1135440,
    "text": "And so gradually I started to um get a"
  },
  {
    "startMs": 1138799,
    "text": "bit more interested in well how can we"
  },
  {
    "startMs": 1141840,
    "text": "start to incorporate the visual world"
  },
  {
    "startMs": 1144240,
    "text": "and I started to look at things like"
  },
  {
    "startMs": 1146480,
    "text": "visual question answering and how you"
  },
  {
    "startMs": 1149120,
    "text": "could connect between um text and then"
  },
  {
    "startMs": 1152960,
    "text": "generative models of visual worlds and"
  },
  {
    "startMs": 1155760,
    "text": "reason about that. And so today I'm"
  },
  {
    "startMs": 1158080,
    "text": "going to be telling you more about how"
  },
  {
    "startMs": 1160480,
    "text": "you can build out that line of work and"
  },
  {
    "startMs": 1163440,
    "text": "be using simulation as a way to start to"
  },
  {
    "startMs": 1166960,
    "text": "approach um embodied artificial general"
  },
  {
    "startMs": 1169840,
    "text": "intelligence. So why do we want"
  },
  {
    "startMs": 1172960,
    "text": "simulation? Um there are sort of two"
  },
  {
    "startMs": 1175760,
    "text": "ways that you can go about um starting"
  },
  {
    "startMs": 1178960,
    "text": "to build intelligent models of the"
  },
  {
    "startMs": 1181360,
    "text": "physical world. One way of doing it is"
  },
  {
    "startMs": 1184880,
    "text": "that you're actually going to directly"
  },
  {
    "startMs": 1186960,
    "text": "learn in the real world. So you can um"
  },
  {
    "startMs": 1190960,
    "text": "set up hardware or collect your YouTube"
  },
  {
    "startMs": 1194240,
    "text": "videos in the real world and start"
  },
  {
    "startMs": 1197440,
    "text": "learning intelligent policies for how to"
  },
  {
    "startMs": 1200320,
    "text": "act. So this was the kind of approach"
  },
  {
    "startMs": 1202720,
    "text": "that was used in Google X's QOP um work"
  },
  {
    "startMs": 1205919,
    "text": "that you've got your um rows of robot"
  },
  {
    "startMs": 1208400,
    "text": "arms that are doing things. you're"
  },
  {
    "startMs": 1210240,
    "text": "recording what they're doing and you're"
  },
  {
    "startMs": 1212320,
    "text": "in the physical world starting to learn"
  },
  {
    "startMs": 1215280,
    "text": "a policy. Um, that's a really unpleasant"
  },
  {
    "startMs": 1220160,
    "text": "way to try and make progress. So, when"
  },
  {
    "startMs": 1222880,
    "text": "people are do making progress in this"
  },
  {
    "startMs": 1225520,
    "text": "kind of world, you're getting about"
  },
  {
    "startMs": 1228080,
    "text": "10,000 hours of teleyop, that means a"
  },
  {
    "startMs": 1231840,
    "text": "human is moving the robot around um data"
  },
  {
    "startMs": 1235200,
    "text": "to start to train a model. um it's not a"
  },
  {
    "startMs": 1238720,
    "text": "very appealing picture. If we want to"
  },
  {
    "startMs": 1240960,
    "text": "start having generally good robotics,"
  },
  {
    "startMs": 1243679,
    "text": "we're just not going to get very far if"
  },
  {
    "startMs": 1246400,
    "text": "we're sort of um chugging along at the"
  },
  {
    "startMs": 1249120,
    "text": "speed of these robots um moving."
  },
  {
    "startMs": 1252880,
    "text": "But interestingly, if we go back to"
  },
  {
    "startMs": 1255039,
    "text": "Shaky in the 1970s,"
  },
  {
    "startMs": 1257760,
    "text": "Shaky had a different answer. Shaky"
  },
  {
    "startMs": 1260880,
    "text": "said, \"Well, we shouldn't just have the"
  },
  {
    "startMs": 1265039,
    "text": "real world that's surrounding us. We"
  },
  {
    "startMs": 1268320,
    "text": "should also have in the robot's head a"
  },
  {
    "startMs": 1272400,
    "text": "world model that has an internal"
  },
  {
    "startMs": 1276240,
    "text": "representation of what the world is"
  },
  {
    "startMs": 1278799,
    "text": "like. So this idea of a world model as"
  },
  {
    "startMs": 1282320,
    "text": "an abstracted internal representation of"
  },
  {
    "startMs": 1284720,
    "text": "the world which can be used for"
  },
  {
    "startMs": 1286880,
    "text": "simulation and planning that's an idea"
  },
  {
    "startMs": 1289440,
    "text": "that goes back to cognitive science. So"
  },
  {
    "startMs": 1291520,
    "text": "it was first proposed by Kenneth Craig"
  },
  {
    "startMs": 1294640,
    "text": "um in the 1940s. So he argued that"
  },
  {
    "startMs": 1298559,
    "text": "humans and other creatures have world"
  },
  {
    "startMs": 1301520,
    "text": "models inside their own heads. Um so um"
  },
  {
    "startMs": 1305840,
    "text": "that they can think about alternatives"
  },
  {
    "startMs": 1308240,
    "text": "and how they're likely to play to play"
  },
  {
    "startMs": 1310799,
    "text": "out and therefore they can think of good"
  },
  {
    "startMs": 1313360,
    "text": "plans um before acting. And so that's"
  },
  {
    "startMs": 1317760,
    "text": "also what we'd like our robots to do is"
  },
  {
    "startMs": 1321120,
    "text": "to have the same kind of abilities. Um"
  },
  {
    "startMs": 1324159,
    "text": "so this was Shaky's world model. So,"
  },
  {
    "startMs": 1326400,
    "text": "Shiki had a kind of a blocks world where"
  },
  {
    "startMs": 1328960,
    "text": "it was actually a grid world as you"
  },
  {
    "startMs": 1330720,
    "text": "might remember from your early AI"
  },
  {
    "startMs": 1332880,
    "text": "textbooks, but it was representing um"
  },
  {
    "startMs": 1335919,
    "text": "what was in different places in the"
  },
  {
    "startMs": 1337760,
    "text": "world and what were the attributes of"
  },
  {
    "startMs": 1340240,
    "text": "different things and it was all stored"
  },
  {
    "startMs": 1342320,
    "text": "in this kind of logical representation"
  },
  {
    "startMs": 1345039,
    "text": "in those days. So how now can we start"
  },
  {
    "startMs": 1350000,
    "text": "um building a 21st century version of an"
  },
  {
    "startMs": 1354159,
    "text": "embodied artificial general"
  },
  {
    "startMs": 1356080,
    "text": "intelligence? And I think the right way"
  },
  {
    "startMs": 1358320,
    "text": "to do it is to work out how to build"
  },
  {
    "startMs": 1361360,
    "text": "good action condition world models. And"
  },
  {
    "startMs": 1364640,
    "text": "today I want to talk about a practical"
  },
  {
    "startMs": 1368559,
    "text": "effective way to build a simulation"
  },
  {
    "startMs": 1371120,
    "text": "infrastructure that can be used um for"
  },
  {
    "startMs": 1374559,
    "text": "as a basis from embodied AGI."
  },
  {
    "startMs": 1378159,
    "text": "So the idea of a world model is now"
  },
  {
    "startMs": 1380400,
    "text": "normally formalized in terms of"
  },
  {
    "startMs": 1382640,
    "text": "reinforcement learning ideas that what"
  },
  {
    "startMs": 1385360,
    "text": "we do is we have observations of a world"
  },
  {
    "startMs": 1388960,
    "text": "but we assume that underlying those we"
  },
  {
    "startMs": 1391679,
    "text": "have a semantic abstracted"
  },
  {
    "startMs": 1394159,
    "text": "representation of the state of the world"
  },
  {
    "startMs": 1396880,
    "text": "in our head. And what the world model"
  },
  {
    "startMs": 1399520,
    "text": "does is gives us an ability to try and"
  },
  {
    "startMs": 1403200,
    "text": "predict um when an action is taken in"
  },
  {
    "startMs": 1406080,
    "text": "one state, what new state is going to"
  },
  {
    "startMs": 1409200,
    "text": "emerge. And so the crucial thing there"
  },
  {
    "startMs": 1412640,
    "text": "is that the world model is abstracted"
  },
  {
    "startMs": 1416799,
    "text": "and has more semantics. Quite a lot of"
  },
  {
    "startMs": 1419840,
    "text": "the time when people have talked about"
  },
  {
    "startMs": 1422159,
    "text": "world models, they haven't really been"
  },
  {
    "startMs": 1425039,
    "text": "thinking about the abstraction and the"
  },
  {
    "startMs": 1427520,
    "text": "semantics, they've just been talking"
  },
  {
    "startMs": 1429919,
    "text": "about can you produce beautiful"
  },
  {
    "startMs": 1432400,
    "text": "generative AI video. Um so here's um"
  },
  {
    "startMs": 1436559,
    "text": "Genie 3. Um this is kind of a cute"
  },
  {
    "startMs": 1439360,
    "text": "example by Riley Goodside who was the"
  },
  {
    "startMs": 1441760,
    "text": "same person who got a lot of fame in the"
  },
  {
    "startMs": 1444240,
    "text": "early LLM days by being a good prompt"
  },
  {
    "startMs": 1447360,
    "text": "engineer. um and he's now playing around"
  },
  {
    "startMs": 1450240,
    "text": "um here with Genie 3. And you know the"
  },
  {
    "startMs": 1453120,
    "text": "the kind of things you can do with Genie"
  },
  {
    "startMs": 1455440,
    "text": "3, you know, it looks beautiful, it's"
  },
  {
    "startMs": 1458799,
    "text": "wonderful, it's pretty fluid, it seems"
  },
  {
    "startMs": 1463120,
    "text": "great, but you know, most of visual AI"
  },
  {
    "startMs": 1466720,
    "text": "for this period has been just judged by"
  },
  {
    "startMs": 1469760,
    "text": "the pixels. If you've got beautiful"
  },
  {
    "startMs": 1472159,
    "text": "pixels, you've got a beautiful piece of"
  },
  {
    "startMs": 1474559,
    "text": "software, but these pixels are trying to"
  },
  {
    "startMs": 1477840,
    "text": "simulate observations. They don't"
  },
  {
    "startMs": 1480480,
    "text": "actually have or represent any of the"
  },
  {
    "startMs": 1483039,
    "text": "semantics behind the world. And"
  },
  {
    "startMs": 1485440,
    "text": "therefore, they aren't good at realism."
  },
  {
    "startMs": 1488640,
    "text": "They aren't good for giving a basis to"
  },
  {
    "startMs": 1490799,
    "text": "plan. They aren't good for actually"
  },
  {
    "startMs": 1492960,
    "text": "having a robust simulation"
  },
  {
    "startMs": 1494880,
    "text": "infrastructure of the real world. And so"
  },
  {
    "startMs": 1497600,
    "text": "we want this kind of simulation"
  },
  {
    "startMs": 1499360,
    "text": "infrastructure because with a good"
  },
  {
    "startMs": 1501600,
    "text": "simulator we actually have causal"
  },
  {
    "startMs": 1503919,
    "text": "knowledge of how the world works which"
  },
  {
    "startMs": 1506799,
    "text": "allows us to predict and plan how things"
  },
  {
    "startMs": 1509840,
    "text": "will work in any situation. And so"
  },
  {
    "startMs": 1513200,
    "text": "that's the kind of world um that we're"
  },
  {
    "startMs": 1516080,
    "text": "wanting to have and make available at"
  },
  {
    "startMs": 1519679,
    "text": "Moon Lake. So the starting point is a"
  },
  {
    "startMs": 1523120,
    "text": "real world observation. Um, so given an"
  },
  {
    "startMs": 1527360,
    "text": "image or a bit of video, we want to be"
  },
  {
    "startMs": 1530320,
    "text": "able to interpret this image as an"
  },
  {
    "startMs": 1533200,
    "text": "observation which is always partial of"
  },
  {
    "startMs": 1536240,
    "text": "what's actually in the underlying world."
  },
  {
    "startMs": 1539279,
    "text": "And then what we want to do is"
  },
  {
    "startMs": 1541679,
    "text": "reconstruct a model of this world which"
  },
  {
    "startMs": 1545200,
    "text": "actually allows us to do stuff in it and"
  },
  {
    "startMs": 1548559,
    "text": "see how it reacts. that this will give"
  },
  {
    "startMs": 1550640,
    "text": "us a basis of being able to work out"
  },
  {
    "startMs": 1554159,
    "text": "causality, work out how to plan and"
  },
  {
    "startMs": 1557600,
    "text": "reason in a repres representation"
  },
  {
    "startMs": 1560000,
    "text": "condition way which gives us the basis"
  },
  {
    "startMs": 1562880,
    "text": "for intelligent robot actions inside"
  },
  {
    "startMs": 1566480,
    "text": "this world."
  },
  {
    "startMs": 1568320,
    "text": "So what might one do here? Well, the"
  },
  {
    "startMs": 1571360,
    "text": "first thing you can do is take that um"
  },
  {
    "startMs": 1574720,
    "text": "picture and feed it into one of our"
  },
  {
    "startMs": 1577279,
    "text": "well-known um other companies products"
  },
  {
    "startMs": 1580000,
    "text": "and say, \"Okay, make a simulation of"
  },
  {
    "startMs": 1582720,
    "text": "this world.\" Um and so this is what you"
  },
  {
    "startMs": 1586159,
    "text": "get from Marble. And it's a pretty good"
  },
  {
    "startMs": 1588559,
    "text": "simulation of the world. And if what you"
  },
  {
    "startMs": 1591200,
    "text": "want to do is just uh walk around in"
  },
  {
    "startMs": 1594480,
    "text": "this world and see it from different"
  },
  {
    "startMs": 1596880,
    "text": "angles, this works pretty well. Um the"
  },
  {
    "startMs": 1602080,
    "text": "question is is this a good simulation"
  },
  {
    "startMs": 1604880,
    "text": "and whether it's a good simulation"
  },
  {
    "startMs": 1607600,
    "text": "depends on"
  },
  {
    "startMs": 1610559,
    "text": "what problem you want to solve in the"
  },
  {
    "startMs": 1612559,
    "text": "real world. And then is does this give"
  },
  {
    "startMs": 1615679,
    "text": "you sufficient information to solve the"
  },
  {
    "startMs": 1618720,
    "text": "problem in the real world? And if all"
  },
  {
    "startMs": 1621200,
    "text": "you want to do in the real world is to"
  },
  {
    "startMs": 1623360,
    "text": "be able to wander around and see the"
  },
  {
    "startMs": 1625520,
    "text": "view from different angles, then this is"
  },
  {
    "startMs": 1628080,
    "text": "great. We're done. But a lot of the time"
  },
  {
    "startMs": 1631919,
    "text": "what you'd like to do in the real world"
  },
  {
    "startMs": 1635120,
    "text": "is understand the objects are here and"
  },
  {
    "startMs": 1638640,
    "text": "to be able to do things with them."
  },
  {
    "startMs": 1641120,
    "text": "Right? So there's some objects here."
  },
  {
    "startMs": 1642960,
    "text": "There's a kettle and there's a box and"
  },
  {
    "startMs": 1645360,
    "text": "there's a cup. But in the marble world,"
  },
  {
    "startMs": 1649039,
    "text": "there's nothing you can actually do with"
  },
  {
    "startMs": 1651279,
    "text": "these things. All you can do is sort of"
  },
  {
    "startMs": 1653760,
    "text": "wander around as a disembodied figure."
  },
  {
    "startMs": 1656720,
    "text": "Um, and so for a lot of purposes such as"
  },
  {
    "startMs": 1659840,
    "text": "doing things with robotics or other"
  },
  {
    "startMs": 1662080,
    "text": "physically accurate worlds, um, we need"
  },
  {
    "startMs": 1665039,
    "text": "to be able to have more understanding in"
  },
  {
    "startMs": 1667520,
    "text": "our model of the world, a better, more"
  },
  {
    "startMs": 1670640,
    "text": "detailed simulation. And so at Moon"
  },
  {
    "startMs": 1673200,
    "text": "Lake, we're wanting to work out what"
  },
  {
    "startMs": 1675440,
    "text": "people actually want to do with their"
  },
  {
    "startMs": 1677200,
    "text": "simulation, and then to build the kind"
  },
  {
    "startMs": 1679600,
    "text": "of simulation that will power that. So"
  },
  {
    "startMs": 1682640,
    "text": "you might think, oh, I actually want to"
  },
  {
    "startMs": 1684399,
    "text": "be able to move around the objects in"
  },
  {
    "startMs": 1686480,
    "text": "the world. And so that's the kind of"
  },
  {
    "startMs": 1689120,
    "text": "starting point of the kind of thing that"
  },
  {
    "startMs": 1691919,
    "text": "we're trying to do at Moon Lake. So if"
  },
  {
    "startMs": 1694559,
    "text": "we want to um have a world in which we"
  },
  {
    "startMs": 1698159,
    "text": "can move things around then we're saying"
  },
  {
    "startMs": 1700960,
    "text": "okay kind of like um image blaster a"
  },
  {
    "startMs": 1703919,
    "text": "system like that we actually need to"
  },
  {
    "startMs": 1706240,
    "text": "take this world and understand what's in"
  },
  {
    "startMs": 1709120,
    "text": "it and then have objects that can be"
  },
  {
    "startMs": 1711520,
    "text": "manipulated. So we separate out a"
  },
  {
    "startMs": 1714559,
    "text": "background and then objects inside that"
  },
  {
    "startMs": 1718000,
    "text": "background with then being able to sort"
  },
  {
    "startMs": 1720480,
    "text": "of have the background world in a"
  },
  {
    "startMs": 1723760,
    "text": "representation that's similar to marble."
  },
  {
    "startMs": 1726320,
    "text": "But then in the foreground there are"
  },
  {
    "startMs": 1728159,
    "text": "various kinds of objects that we can"
  },
  {
    "startMs": 1730159,
    "text": "manipulate and move around. Now, well,"
  },
  {
    "startMs": 1733679,
    "text": "that's a start, but um in the previous"
  },
  {
    "startMs": 1737520,
    "text": "picture, um the tea box was always"
  },
  {
    "startMs": 1740559,
    "text": "closed. And we might wonder if you can"
  },
  {
    "startMs": 1743760,
    "text": "open up the tea box and see what's"
  },
  {
    "startMs": 1746320,
    "text": "inside it. Um and well, this is sort of"
  },
  {
    "startMs": 1749039,
    "text": "a part of how observations of a world"
  },
  {
    "startMs": 1752559,
    "text": "are always partial. Um for what we could"
  },
  {
    "startMs": 1755440,
    "text": "see in the actual photo, there was a"
  },
  {
    "startMs": 1757679,
    "text": "closed tea box. We couldn't even see the"
  },
  {
    "startMs": 1759679,
    "text": "tea box very well. How could we possibly"
  },
  {
    "startMs": 1762880,
    "text": "know what's inside it? And well, the"
  },
  {
    "startMs": 1766000,
    "text": "answer to the way we can know what's"
  },
  {
    "startMs": 1768799,
    "text": "inside it in the 21st century is we go"
  },
  {
    "startMs": 1772080,
    "text": "off and do a little bit re of research."
  },
  {
    "startMs": 1774080,
    "text": "We fire up um getting information um off"
  },
  {
    "startMs": 1778559,
    "text": "the web um in a"
  },
  {
    "startMs": 1782080,
    "text": "a rag style fashion um to work out"
  },
  {
    "startMs": 1785919,
    "text": "what's there. And then we can find"
  },
  {
    "startMs": 1788320,
    "text": "images of this um tea um box on the web"
  },
  {
    "startMs": 1792799,
    "text": "which show it in more detail. They show"
  },
  {
    "startMs": 1794799,
    "text": "a picture of what it's like when it's"
  },
  {
    "startMs": 1796480,
    "text": "open. There are descriptions of it."
  },
  {
    "startMs": 1798640,
    "text": "Organic luxury tea bag collection,"
  },
  {
    "startMs": 1800720,
    "text": "leather gift box. It gives its size. We"
  },
  {
    "startMs": 1803279,
    "text": "can find all about it. So therefore, we"
  },
  {
    "startMs": 1805840,
    "text": "can build um in our simulation um a tea"
  },
  {
    "startMs": 1809840,
    "text": "box with an understanding of the"
  },
  {
    "startMs": 1811520,
    "text": "contents of that tea box. Well, that's"
  },
  {
    "startMs": 1814240,
    "text": "really good. Um, so how now we have um"
  },
  {
    "startMs": 1817600,
    "text": "the tea box with tea bags inside it. But"
  },
  {
    "startMs": 1820720,
    "text": "if we do nothing else, um, they're just"
  },
  {
    "startMs": 1824640,
    "text": "sort of sitting there and movable. So"
  },
  {
    "startMs": 1826640,
    "text": "we'd like to realize the fact, well,"
  },
  {
    "startMs": 1828320,
    "text": "wait a minute, tea bags you can lift up"
  },
  {
    "startMs": 1831520,
    "text": "and you can take out of a tea box. So"
  },
  {
    "startMs": 1835120,
    "text": "then we need to start having these"
  },
  {
    "startMs": 1837279,
    "text": "teaags also um be objects that are"
  },
  {
    "startMs": 1840159,
    "text": "modeled in our simulation that they have"
  },
  {
    "startMs": 1842799,
    "text": "a size and an ability to be manipulated"
  },
  {
    "startMs": 1846000,
    "text": "as well. Um so how are we doing all of"
  },
  {
    "startMs": 1850320,
    "text": "this? And so a distinctive part of"
  },
  {
    "startMs": 1852880,
    "text": "what's happening at Moon Lake is"
  },
  {
    "startMs": 1855919,
    "text": "building although part of this is in the"
  },
  {
    "startMs": 1859600,
    "text": "world of vision"
  },
  {
    "startMs": 1861679,
    "text": "a lot of what we're doing is actually"
  },
  {
    "startMs": 1863840,
    "text": "back in the world of code. So this is"
  },
  {
    "startMs": 1866480,
    "text": "picking up on the idea um that what's"
  },
  {
    "startMs": 1870159,
    "text": "normally referred to as language models"
  },
  {
    "startMs": 1872240,
    "text": "but these days are really symbolic"
  },
  {
    "startMs": 1874159,
    "text": "models which work on not only human"
  },
  {
    "startMs": 1877440,
    "text": "languages but also math code and things"
  },
  {
    "startMs": 1880240,
    "text": "like that that they that has been just a"
  },
  {
    "startMs": 1883360,
    "text": "very powerful substrate with which to"
  },
  {
    "startMs": 1886000,
    "text": "make progress. And so we are producing"
  },
  {
    "startMs": 1889840,
    "text": "controllable manipulable world models by"
  },
  {
    "startMs": 1893760,
    "text": "generating the controllable parts of"
  },
  {
    "startMs": 1896480,
    "text": "these worlds by putting code under them."
  },
  {
    "startMs": 1900080,
    "text": "And in particular, um, we can train"
  },
  {
    "startMs": 1903600,
    "text": "these code models using the same kind of"
  },
  {
    "startMs": 1906880,
    "text": "loop engineering that many of you will"
  },
  {
    "startMs": 1908799,
    "text": "have seen, um, with Claude that we're"
  },
  {
    "startMs": 1911760,
    "text": "having a loop where we're writing code"
  },
  {
    "startMs": 1915679,
    "text": "to render objects,"
  },
  {
    "startMs": 1918720,
    "text": "um, and put on that um, diffusionbased"
  },
  {
    "startMs": 1921679,
    "text": "textures, etc. We can then assess how"
  },
  {
    "startMs": 1925840,
    "text": "good our render is against physical"
  },
  {
    "startMs": 1928480,
    "text": "reality and then we can work out the"
  },
  {
    "startMs": 1931519,
    "text": "deviations. We can revise the code and"
  },
  {
    "startMs": 1935279,
    "text": "make better and better um renders of"
  },
  {
    "startMs": 1937840,
    "text": "what goes along. And so that's the way"
  },
  {
    "startMs": 1940799,
    "text": "that so we're building this detailed"
  },
  {
    "startMs": 1944000,
    "text": "action conditional world model so that"
  },
  {
    "startMs": 1946720,
    "text": "we can take actions on the objects in"
  },
  {
    "startMs": 1949279,
    "text": "the world."
  },
  {
    "startMs": 1950960,
    "text": "Um so that means that we have a kind of"
  },
  {
    "startMs": 1953279,
    "text": "neuros symbolic representation of the"
  },
  {
    "startMs": 1956080,
    "text": "world here and neuros symbolic"
  },
  {
    "startMs": 1958960,
    "text": "representations have a big advantage"
  },
  {
    "startMs": 1961760,
    "text": "that the symbolic representation can be"
  },
  {
    "startMs": 1964320,
    "text": "easily interfaced controlled edited and"
  },
  {
    "startMs": 1967919,
    "text": "maintained in our human world. Um, and"
  },
  {
    "startMs": 1971919,
    "text": "so the message I'd like to give as a"
  },
  {
    "startMs": 1976080,
    "text": "little delta of a message here, um, is"
  },
  {
    "startMs": 1979919,
    "text": "most of you are probably familiar with"
  },
  {
    "startMs": 1982000,
    "text": "Mark Andre's famous statement 15 years"
  },
  {
    "startMs": 1984559,
    "text": "ago, software will eat the world. And"
  },
  {
    "startMs": 1988640,
    "text": "you know, that was mostly right. Um but"
  },
  {
    "startMs": 1992720,
    "text": "I think it wasn't completely right"
  },
  {
    "startMs": 1995360,
    "text": "because it's just not the case that"
  },
  {
    "startMs": 1998159,
    "text": "software ate the physical world. Um it,"
  },
  {
    "startMs": 2002240,
    "text": "you know, a lot of the world went"
  },
  {
    "startMs": 2004320,
    "text": "virtual and um a lot of the world is"
  },
  {
    "startMs": 2007679,
    "text": "inside computers now and yeah it could"
  },
  {
    "startMs": 2010559,
    "text": "eat all of the processes of managing,"
  },
  {
    "startMs": 2014399,
    "text": "counting, supplying all um records of"
  },
  {
    "startMs": 2017919,
    "text": "suppliers, all the stuff that's in the"
  },
  {
    "startMs": 2020000,
    "text": "virtual world. But we hadn't had the"
  },
  {
    "startMs": 2023039,
    "text": "power for software to eat the physical"
  },
  {
    "startMs": 2025679,
    "text": "world. Whereas the hope is with this new"
  },
  {
    "startMs": 2029039,
    "text": "ability of being able to build"
  },
  {
    "startMs": 2032080,
    "text": "verifiable simulations powered by code"
  },
  {
    "startMs": 2035600,
    "text": "of the sort that I've just sketched for"
  },
  {
    "startMs": 2037760,
    "text": "a moment there that this will allow us"
  },
  {
    "startMs": 2040399,
    "text": "to actually have verifiable simulation"
  },
  {
    "startMs": 2043200,
    "text": "which will also eat the physical world."
  },
  {
    "startMs": 2046640,
    "text": "And so that's the kind of thing we"
  },
  {
    "startMs": 2048240,
    "text": "built. And we can go on from here and um"
  },
  {
    "startMs": 2051839,
    "text": "you know keep on making further steps of"
  },
  {
    "startMs": 2054000,
    "text": "this. Right? So we maybe don't want to"
  },
  {
    "startMs": 2055919,
    "text": "only have tea bags um in foil"
  },
  {
    "startMs": 2059280,
    "text": "containers, but we'd like to be able to"
  },
  {
    "startMs": 2061599,
    "text": "um take them out of the foil container"
  },
  {
    "startMs": 2064159,
    "text": "and we don't we somehow want to actually"
  },
  {
    "startMs": 2067200,
    "text": "get the tea bag inside the cup because"
  },
  {
    "startMs": 2069919,
    "text": "that's a useful step for making tea. And"
  },
  {
    "startMs": 2072720,
    "text": "then of course we also want to have um"
  },
  {
    "startMs": 2075359,
    "text": "the jug of um water which we want to be"
  },
  {
    "startMs": 2078639,
    "text": "able to boil and then be able to pour"
  },
  {
    "startMs": 2080800,
    "text": "that into the tea. Um and once we have a"
  },
  {
    "startMs": 2085280,
    "text": "good world simulator like that, the idea"
  },
  {
    "startMs": 2088399,
    "text": "of this is that we are building in all"
  },
  {
    "startMs": 2092800,
    "text": "of the parts of the simulated world"
  },
  {
    "startMs": 2096079,
    "text": "which will allow effective transfer into"
  },
  {
    "startMs": 2100000,
    "text": "the real world. And it's important to"
  },
  {
    "startMs": 2102480,
    "text": "think about that as to sort of what's"
  },
  {
    "startMs": 2105040,
    "text": "necessary for different kinds of"
  },
  {
    "startMs": 2107359,
    "text": "transfer. And the argument um that I"
  },
  {
    "startMs": 2112320,
    "text": "we'd like to make is that you know"
  },
  {
    "startMs": 2114720,
    "text": "normally uh simulation isn't complete"
  },
  {
    "startMs": 2119440,
    "text": "and accurate in every detail because"
  },
  {
    "startMs": 2122800,
    "text": "there are sort of parts of the world"
  },
  {
    "startMs": 2124960,
    "text": "that are important to you and parts of"
  },
  {
    "startMs": 2127359,
    "text": "the world that aren't important to you."
  },
  {
    "startMs": 2129040,
    "text": "I mean, this is the same with human"
  },
  {
    "startMs": 2131280,
    "text": "world models, right? That a lot of the"
  },
  {
    "startMs": 2133119,
    "text": "world we're not actually modeling at any"
  },
  {
    "startMs": 2135119,
    "text": "time, but we're modeling the bits of it"
  },
  {
    "startMs": 2137599,
    "text": "that are important to us. And it's"
  },
  {
    "startMs": 2139760,
    "text": "having that control of which things you"
  },
  {
    "startMs": 2142400,
    "text": "need to model and use is what needed"
  },
  {
    "startMs": 2145280,
    "text": "then to give you the actual simulations"
  },
  {
    "startMs": 2148560,
    "text": "that will be allowed to be effectively"
  },
  {
    "startMs": 2150400,
    "text": "used um in the real world. So, what kind"
  },
  {
    "startMs": 2153680,
    "text": "of applications"
  },
  {
    "startMs": 2155440,
    "text": "um is this going to allow us to do? So"
  },
  {
    "startMs": 2158320,
    "text": "the hope is that um rather than having"
  },
  {
    "startMs": 2162240,
    "text": "to collect 10,000 hours of data in by"
  },
  {
    "startMs": 2167200,
    "text": "teley op in the real world, we can"
  },
  {
    "startMs": 2169520,
    "text": "instead build accurate simulations which"
  },
  {
    "startMs": 2172720,
    "text": "will allow transfer to real. So let me"
  },
  {
    "startMs": 2175359,
    "text": "show you the kind of progress that we've"
  },
  {
    "startMs": 2177119,
    "text": "been making on this. So this is um not"
  },
  {
    "startMs": 2180079,
    "text": "highuting robotics um but is the kind of"
  },
  {
    "startMs": 2183359,
    "text": "physical AI that you find everywhere in"
  },
  {
    "startMs": 2187280,
    "text": "the real world that there are things"
  },
  {
    "startMs": 2189599,
    "text": "happening in physical processes um which"
  },
  {
    "startMs": 2192880,
    "text": "you'd like to be able to um understand"
  },
  {
    "startMs": 2197680,
    "text": "and automate. And so in particular, um,"
  },
  {
    "startMs": 2200880,
    "text": "if we're going to realize any of the"
  },
  {
    "startMs": 2202640,
    "text": "dreams of bringing back America as a"
  },
  {
    "startMs": 2205280,
    "text": "great manufacturing economy, it seems"
  },
  {
    "startMs": 2207839,
    "text": "like we have to work out how to be able"
  },
  {
    "startMs": 2209760,
    "text": "to automate much more in the physical"
  },
  {
    "startMs": 2212240,
    "text": "world. Well, um, using the Moonlake AI"
  },
  {
    "startMs": 2216560,
    "text": "technology, what we can then do is say,"
  },
  {
    "startMs": 2219839,
    "text": "so from this short video, we can"
  },
  {
    "startMs": 2222640,
    "text": "automatically produce an accurate"
  },
  {
    "startMs": 2225200,
    "text": "simulation of that world, turning it"
  },
  {
    "startMs": 2227680,
    "text": "into a 3D world with enough detail about"
  },
  {
    "startMs": 2231280,
    "text": "how things move and what objects are in"
  },
  {
    "startMs": 2234240,
    "text": "the world that we can start to build on"
  },
  {
    "startMs": 2237520,
    "text": "this and use it as a basis of simulated"
  },
  {
    "startMs": 2240800,
    "text": "data which transfers refers accurately"
  },
  {
    "startMs": 2243200,
    "text": "to the real world. So we then have this"
  },
  {
    "startMs": 2246160,
    "text": "model which allows us to get 10,000"
  },
  {
    "startMs": 2248960,
    "text": "hours of simulation um for free. And so"
  },
  {
    "startMs": 2252240,
    "text": "on the basis of that um we can then"
  },
  {
    "startMs": 2255760,
    "text": "train up a robotic system that can"
  },
  {
    "startMs": 2258960,
    "text": "operate um to actually um explore what"
  },
  {
    "startMs": 2264079,
    "text": "you can do in this world and learn the"
  },
  {
    "startMs": 2267839,
    "text": "way to operate in the world so it works."
  },
  {
    "startMs": 2270800,
    "text": "and then we can have the effective and"
  },
  {
    "startMs": 2273760,
    "text": "cheap training of robotic systems."
  },
  {
    "startMs": 2277839,
    "text": "Okay. Um so that's the story. Um thanks"
  },
  {
    "startMs": 2280800,
    "text": "a lot everyone."
  },
  {
    "startMs": 2282701,
    "text": "[applause]"
  },
  {
    "startMs": 2289760,
    "text": "And I do have time for questions I"
  },
  {
    "startMs": 2291839,
    "text": "believe."
  },
  {
    "startMs": 2296320,
    "text": ">> Thank you very much for the"
  },
  {
    "startMs": 2297440,
    "text": "conversation. Um, super deep. My first"
  },
  {
    "startMs": 2300960,
    "text": "question is, would this be applied to"
  },
  {
    "startMs": 2303359,
    "text": "something like the gaming industry and"
  },
  {
    "startMs": 2306000,
    "text": "how could they use it for we're seeing"
  },
  {
    "startMs": 2308960,
    "text": "GTA 6? Everyone is talking about this"
  },
  {
    "startMs": 2311119,
    "text": "new game. Um, I could see a very"
  },
  {
    "startMs": 2314320,
    "text": "application there where you could have"
  },
  {
    "startMs": 2316480,
    "text": "endless interaction with the world model"
  },
  {
    "startMs": 2319280,
    "text": "that they're building. U, have you been"
  },
  {
    "startMs": 2321280,
    "text": "working with the gaming industry at all"
  },
  {
    "startMs": 2322800,
    "text": "or this is more applied to robotics?"
  },
  {
    "startMs": 2324880,
    "text": ">> Um, so yeah, absolutely. Um, another"
  },
  {
    "startMs": 2329280,
    "text": "hugely good area for this is in the"
  },
  {
    "startMs": 2331440,
    "text": "gaming industry. And you know, the real"
  },
  {
    "startMs": 2335040,
    "text": "world is then kind of a virtual world,"
  },
  {
    "startMs": 2337200,
    "text": "but you can have a simulation of your"
  },
  {
    "startMs": 2339760,
    "text": "gaming world um, and then be building"
  },
  {
    "startMs": 2342960,
    "text": "the same kind of embodied intelligence."
  },
  {
    "startMs": 2346079,
    "text": "And the gaming industry has some really"
  },
  {
    "startMs": 2348640,
    "text": "appealing attributes. Well, you know,"
  },
  {
    "startMs": 2350560,
    "text": "there are millions and millions of game"
  },
  {
    "startMs": 2352000,
    "text": "players, so there's lots of easy data to"
  },
  {
    "startMs": 2354480,
    "text": "collect. Um the people who are gaming"
  },
  {
    "startMs": 2358000,
    "text": "have goals which are fairly clearly"
  },
  {
    "startMs": 2360480,
    "text": "known. So you can sort of learn in a"
  },
  {
    "startMs": 2362720,
    "text": "reinforcement learning loop good ways to"
  },
  {
    "startMs": 2365200,
    "text": "act. And so absolutely one of the"
  },
  {
    "startMs": 2367359,
    "text": "applications that Moon Lake has explored"
  },
  {
    "startMs": 2370240,
    "text": "is in the gaming industry and there's"
  },
  {
    "startMs": 2372160,
    "text": "likely to be more of that. Um but"
  },
  {
    "startMs": 2374800,
    "text": "recently we've been actually"
  },
  {
    "startMs": 2377119,
    "text": "particularly emphasizing physical"
  },
  {
    "startMs": 2379040,
    "text": "infrastructure and doing simulations in"
  },
  {
    "startMs": 2380960,
    "text": "the physical infrastructure world."
  },
  {
    "startMs": 2382880,
    "text": ">> Thank you so much."
  },
  {
    "startMs": 2395680,
    "text": "Uh so it's fascinating to me that that"
  },
  {
    "startMs": 2398320,
    "text": "you're turning back to neurosy symbolic"
  },
  {
    "startMs": 2400640,
    "text": "representations and um the part of the"
  },
  {
    "startMs": 2404800,
    "text": "original kind of old school AI involved"
  },
  {
    "startMs": 2407760,
    "text": "a lot of ontology development and"
  },
  {
    "startMs": 2409599,
    "text": "knowledge representation and logic and"
  },
  {
    "startMs": 2411359,
    "text": "representation of that. Do you see a"
  },
  {
    "startMs": 2413359,
    "text": "role for uh deepening the neuros"
  },
  {
    "startMs": 2416880,
    "text": "symbolic representation using kind of"
  },
  {
    "startMs": 2419200,
    "text": "like what we how ontologies have been"
  },
  {
    "startMs": 2421680,
    "text": "developed previously to model um a"
  },
  {
    "startMs": 2424160,
    "text": "representation of the real world? Um and"
  },
  {
    "startMs": 2427040,
    "text": "do you see that as being a kind of an"
  },
  {
    "startMs": 2428960,
    "text": "area of expansion and development for"
  },
  {
    "startMs": 2431440,
    "text": "the type of tools that you're working"
  },
  {
    "startMs": 2432720,
    "text": "in? Yeah, I mean I don't think we're"
  },
  {
    "startMs": 2436000,
    "text": "quite going to go back to old style"
  },
  {
    "startMs": 2438880,
    "text": "ontologies and knowledge bases, but I"
  },
  {
    "startMs": 2441680,
    "text": "mean effectively"
  },
  {
    "startMs": 2443839,
    "text": "that is what we're doing, you know,"
  },
  {
    "startMs": 2446079,
    "text": "apart from it's in new clothing of um"
  },
  {
    "startMs": 2449920,
    "text": "having it being um code that I mean I I"
  },
  {
    "startMs": 2455359,
    "text": "do you know this is an interesting space"
  },
  {
    "startMs": 2458640,
    "text": "and we can talk about it in more general"
  },
  {
    "startMs": 2460720,
    "text": "right there is you know there's sort a"
  },
  {
    "startMs": 2463040,
    "text": "purely neural approach in which your"
  },
  {
    "startMs": 2465599,
    "text": "latent representation of the world is"
  },
  {
    "startMs": 2467839,
    "text": "purely neural and so that's the kind of"
  },
  {
    "startMs": 2469920,
    "text": "thing that the jeoper architecture is"
  },
  {
    "startMs": 2472160,
    "text": "after I mean in some ways that's a good"
  },
  {
    "startMs": 2475040,
    "text": "pure neural approach but on the other"
  },
  {
    "startMs": 2478240,
    "text": "hand that's very hard for humans to"
  },
  {
    "startMs": 2484079,
    "text": "connect to in any way or for other"
  },
  {
    "startMs": 2486079,
    "text": "applications to connect to in any way"
  },
  {
    "startMs": 2488640,
    "text": "and so our bet is that the sort of"
  },
  {
    "startMs": 2491599,
    "text": "practical ical way to have a physical AI"
  },
  {
    "startMs": 2494319,
    "text": "simulation infrastructure for the"
  },
  {
    "startMs": 2496720,
    "text": "foreseeable future is to base it on"
  },
  {
    "startMs": 2499599,
    "text": "symbolic representations."
  },
  {
    "startMs": 2501839,
    "text": "Taking advantage of sort of the huge"
  },
  {
    "startMs": 2503920,
    "text": "power of code that we see everywhere"
  },
  {
    "startMs": 2507119,
    "text": "around us in the sort of um codeex cla"
  },
  {
    "startMs": 2510319,
    "text": "code era that those kind of symbolic"
  },
  {
    "startMs": 2513200,
    "text": "representations allow um neural eye"
  },
  {
    "startMs": 2516960,
    "text": "systems to reason, plan and do all of"
  },
  {
    "startMs": 2519760,
    "text": "these things excellently well. And so to"
  },
  {
    "startMs": 2521920,
    "text": "some extent yes that will be bringing"
  },
  {
    "startMs": 2523920,
    "text": "back notions like older fashion"
  },
  {
    "startMs": 2526000,
    "text": "knowledge representation."
  },
  {
    "startMs": 2534079,
    "text": "Mic keeps on."
  },
  {
    "startMs": 2541440,
    "text": "Hello. Hello. Oh hi. Um so a lot of our"
  },
  {
    "startMs": 2545839,
    "text": "existing um it seems like industrial"
  },
  {
    "startMs": 2548400,
    "text": "robotics use cases or to sort of"
  },
  {
    "startMs": 2550800,
    "text": "automate existing sort of human manual"
  },
  {
    "startMs": 2554240,
    "text": "intensive processes. Um would investment"
  },
  {
    "startMs": 2558560,
    "text": "in"
  },
  {
    "startMs": 2560079,
    "text": "development of like world models would"
  },
  {
    "startMs": 2562079,
    "text": "that allow us to go explore novel new uh"
  },
  {
    "startMs": 2566560,
    "text": "processes or activities where they're"
  },
  {
    "startMs": 2569440,
    "text": "currently not accessible by human labor"
  },
  {
    "startMs": 2572240,
    "text": "or by even the kind of existing um"
  },
  {
    "startMs": 2575280,
    "text": "industrial manufacturing processes that"
  },
  {
    "startMs": 2577440,
    "text": "we have."
  },
  {
    "startMs": 2578319,
    "text": ">> I mean sure absolutely. Yeah. So for the"
  },
  {
    "startMs": 2581119,
    "text": "example at the end I showed a very old"
  },
  {
    "startMs": 2584720,
    "text": "school um conveyor belt system but I"
  },
  {
    "startMs": 2588720,
    "text": "mean you know we're also in this world"
  },
  {
    "startMs": 2591680,
    "text": "of amazing things happening in humanoid"
  },
  {
    "startMs": 2594400,
    "text": "robotics and for any of the new forms of"
  },
  {
    "startMs": 2600720,
    "text": "robotics automation"
  },
  {
    "startMs": 2603200,
    "text": "other cases you could think about doing"
  },
  {
    "startMs": 2605200,
    "text": "things in space well you'll have exactly"
  },
  {
    "startMs": 2607760,
    "text": "the same problem that You want to train"
  },
  {
    "startMs": 2611520,
    "text": "up AI agents to be able to act in"
  },
  {
    "startMs": 2615119,
    "text": "different scenarios and it's extremely"
  },
  {
    "startMs": 2620400,
    "text": "costly and difficult to do that training"
  },
  {
    "startMs": 2623760,
    "text": "in the real world. Um and it's very hard"
  },
  {
    "startMs": 2627599,
    "text": "when training in the real world to get"
  },
  {
    "startMs": 2629760,
    "text": "out into the sort of tale of rare cases."
  },
  {
    "startMs": 2633200,
    "text": "I mean you've seen that um for"
  },
  {
    "startMs": 2635440,
    "text": "autonomous driving, right? There's a"
  },
  {
    "startMs": 2637280,
    "text": "reason why autonomous driving kind of"
  },
  {
    "startMs": 2639119,
    "text": "took 20 years to arrive. Um from you"
  },
  {
    "startMs": 2642160,
    "text": "know Stanley breakthroughs of yay"
  },
  {
    "startMs": 2645119,
    "text": "autonomous car wins the race to actually"
  },
  {
    "startMs": 2647839,
    "text": "having Whimos um widely deployed is"
  },
  {
    "startMs": 2650800,
    "text": "because you have this enormous tail and"
  },
  {
    "startMs": 2652800,
    "text": "the way to get out to that tail for all"
  },
  {
    "startMs": 2654880,
    "text": "of these new applications with um"
  },
  {
    "startMs": 2657440,
    "text": "humanoid robotics, space robotics, etc."
  },
  {
    "startMs": 2660400,
    "text": "is to have good simulation."
  },
  {
    "startMs": 2666560,
    "text": ">> Hey, uh thank you for your talk. I love"
  },
  {
    "startMs": 2668800,
    "text": "the energy. Uh my question is you showed"
  },
  {
    "startMs": 2672319,
    "text": "a video of water. Uh so is everything"
  },
  {
    "startMs": 2675520,
    "text": "like learned from videos or is like"
  },
  {
    "startMs": 2677359,
    "text": "physics ingrained in anything like for"
  },
  {
    "startMs": 2679200,
    "text": "example does your model know any"
  },
  {
    "startMs": 2681359,
    "text": "physical properties of water like"
  },
  {
    "startMs": 2682880,
    "text": "viscosity, Bernali's laws or like for"
  },
  {
    "startMs": 2685920,
    "text": "example friction. uh for some use cases"
  },
  {
    "startMs": 2688800,
    "text": "like self-driving maybe those things are"
  },
  {
    "startMs": 2690640,
    "text": "like not important right you just need"
  },
  {
    "startMs": 2692480,
    "text": "to know the velocity and stuff but like"
  },
  {
    "startMs": 2694160,
    "text": "for other use cases I would imagine"
  },
  {
    "startMs": 2695680,
    "text": "these physical properties are important"
  },
  {
    "startMs": 2697359,
    "text": "so"
  },
  {
    "startMs": 2697680,
    "text": ">> and we absolutely know physics I mean"
  },
  {
    "startMs": 2699839,
    "text": "this is the sense in which this"
  },
  {
    "startMs": 2702400,
    "text": "neurosyolic approach is you know you"
  },
  {
    "startMs": 2706800,
    "text": "call a conservative approach if you will"
  },
  {
    "startMs": 2709359,
    "text": "you know that this is absolutely using"
  },
  {
    "startMs": 2711760,
    "text": "physics engines and knowledge of physics"
  },
  {
    "startMs": 2715760,
    "text": "um to in its generation and control of"
  },
  {
    "startMs": 2719119,
    "text": "movement, right? That if you're what you"
  },
  {
    "startMs": 2722160,
    "text": "know when you've just started off with"
  },
  {
    "startMs": 2724319,
    "text": "an image of water or and you want to"
  },
  {
    "startMs": 2727760,
    "text": "understand how that's behaves, you're"
  },
  {
    "startMs": 2730160,
    "text": "using a physics model to predict how"
  },
  {
    "startMs": 2732079,
    "text": "it's going to behave."
  },
  {
    "startMs": 2736240,
    "text": "Okay."
  },
  {
    "startMs": 2740079,
    "text": "Um I know is there time for Oh, there's"
  },
  {
    "startMs": 2744079,
    "text": "someone else at the mic."
  },
  {
    "startMs": 2747440,
    "text": "Hi."
  },
  {
    "startMs": 2747920,
    "text": ">> Hey. Uh, Professor Manning, thank you"
  },
  {
    "startMs": 2749760,
    "text": "for this presentation. Um, I have a"
  },
  {
    "startMs": 2752160,
    "text": "question about seem to row gap. So, I"
  },
  {
    "startMs": 2755760,
    "text": "mean, simulation is great. It's cheap,"
  },
  {
    "startMs": 2757839,
    "text": "is scalable. Um do you have a suggestion"
  },
  {
    "startMs": 2761200,
    "text": "for um how we close the simulation to"
  },
  {
    "startMs": 2764720,
    "text": "real gaps"
  },
  {
    "startMs": 2766751,
    "text": "[snorts] encounter you know like the"
  },
  {
    "startMs": 2769920,
    "text": "like fix part that we are not able to"
  },
  {
    "startMs": 2772960,
    "text": "accurately simulate and like mechanical"
  },
  {
    "startMs": 2776640,
    "text": "tolerancing back latches like all those"
  },
  {
    "startMs": 2779599,
    "text": "things that we are not able to put into"
  },
  {
    "startMs": 2783200,
    "text": "the simulations."
  },
  {
    "startMs": 2784960,
    "text": "Um yeah so traditionally the problem has"
  },
  {
    "startMs": 2787599,
    "text": "always been the simtoreal gap and the"
  },
  {
    "startMs": 2790640,
    "text": "simtoreal gap seeming too large so that"
  },
  {
    "startMs": 2794319,
    "text": "a lot of the training has had to happen"
  },
  {
    "startMs": 2798160,
    "text": "in the real world"
  },
  {
    "startMs": 2801200,
    "text": "and we believe that the answer to that"
  },
  {
    "startMs": 2804800,
    "text": "is you know effectively claude code loop"
  },
  {
    "startMs": 2809520,
    "text": "we're now in this world in which we can"
  },
  {
    "startMs": 2812400,
    "text": "do neural optimization."
  },
  {
    "startMs": 2815680,
    "text": "So we can have the simulation"
  },
  {
    "startMs": 2819280,
    "text": "compared to behavior in the real world"
  },
  {
    "startMs": 2822640,
    "text": "with video and we can automatically"
  },
  {
    "startMs": 2826400,
    "text": "learn to shrink the sim to real gap in a"
  },
  {
    "startMs": 2830000,
    "text": "way that just wasn't possible when you"
  },
  {
    "startMs": 2832560,
    "text": "had people trying to handr write a"
  },
  {
    "startMs": 2836160,
    "text": "physics simulation of something."
  },
  {
    "startMs": 2842240,
    "text": "This this was a great discussion. Thank"
  },
  {
    "startMs": 2844000,
    "text": "you. Um tangential to the question that"
  },
  {
    "startMs": 2847440,
    "text": "was just asked, how do you see like what"
  },
  {
    "startMs": 2850640,
    "text": "do you see the biggest challenges are to"
  },
  {
    "startMs": 2852800,
    "text": "adding like the sociote techchnical"
  },
  {
    "startMs": 2856319,
    "text": "layer, the pieces, the the processes,"
  },
  {
    "startMs": 2858720,
    "text": "the people, the authorities that make"
  },
  {
    "startMs": 2860240,
    "text": "decisions on the objects that you're"
  },
  {
    "startMs": 2862640,
    "text": "looking to simulate. What are the"
  },
  {
    "startMs": 2864560,
    "text": "biggest challenges to developing a"
  },
  {
    "startMs": 2866560,
    "text": "system that has accurate representation?"
  },
  {
    "startMs": 2869359,
    "text": "So for example, if we want to model uh"
  },
  {
    "startMs": 2873119,
    "text": "EV tolls or um doing you know"
  },
  {
    "startMs": 2876400,
    "text": "forecasting of energy at airports uh you"
  },
  {
    "startMs": 2879119,
    "text": "know things like that not necessarily"
  },
  {
    "startMs": 2880640,
    "text": "extending beyond the robotics into these"
  },
  {
    "startMs": 2884079,
    "text": "other um use cases areas."
  },
  {
    "startMs": 2886400,
    "text": ">> Um yeah that's a great area. Um in all"
  },
  {
    "startMs": 2889200,
    "text": "honesty it's not something we've been um"
  },
  {
    "startMs": 2891839,
    "text": "really dealing with at Moon Lake. Um,"
  },
  {
    "startMs": 2894640,
    "text": "you know, at the there are still some"
  },
  {
    "startMs": 2897760,
    "text": "limitations, but you know, large"
  },
  {
    "startMs": 2900560,
    "text": "language models with their slurping up"
  },
  {
    "startMs": 2903440,
    "text": "of enormous amounts of human behavior"
  },
  {
    "startMs": 2906960,
    "text": "data that they're actually getting"
  },
  {
    "startMs": 2910160,
    "text": "better and better at being able to"
  },
  {
    "startMs": 2912800,
    "text": "simulate how human beings, different"
  },
  {
    "startMs": 2915760,
    "text": "kinds of human beings are going to"
  },
  {
    "startMs": 2918880,
    "text": "behave and react in different"
  },
  {
    "startMs": 2920480,
    "text": "circumstances."
  },
  {
    "startMs": 2922079,
    "text": "So I think we are starting to approach"
  },
  {
    "startMs": 2925839,
    "text": "the point in which we can have fairly"
  },
  {
    "startMs": 2929119,
    "text": "good human behavior simulators that are"
  },
  {
    "startMs": 2933119,
    "text": "being powered by the knowledge of large"
  },
  {
    "startMs": 2935040,
    "text": "language models and there are a couple"
  },
  {
    "startMs": 2937040,
    "text": "of companies that are now starting to"
  },
  {
    "startMs": 2938800,
    "text": "look at that."
  },
  {
    "startMs": 2942800,
    "text": ">> Cool. Last question. Yeah."
  },
  {
    "startMs": 2949440,
    "text": "person over here has one for a long"
  },
  {
    "startMs": 2951119,
    "text": "time. [laughter]"
  },
  {
    "startMs": 2956400,
    "text": ">> Thank you. Um so if you look at um"
  },
  {
    "startMs": 2960000,
    "text": "analog chips for instance right the"
  },
  {
    "startMs": 2962319,
    "text": "physics is understood but as you go"
  },
  {
    "startMs": 2964240,
    "text": "higher up the chemical operations then"
  },
  {
    "startMs": 2967280,
    "text": "the heat and those things are still"
  },
  {
    "startMs": 2969520,
    "text": "people are working on it. So that's not"
  },
  {
    "startMs": 2971359,
    "text": "what you call completely"
  },
  {
    "startMs": 2973680,
    "text": "um understood. Now I'll give you another"
  },
  {
    "startMs": 2976559,
    "text": "example of say I'm looking at kidney"
  },
  {
    "startMs": 2978480,
    "text": "related uh literature and stuff and then"
  },
  {
    "startMs": 2981359,
    "text": "liver related stuff"
  },
  {
    "startMs": 2983760,
    "text": "these things it gives a pretty good"
  },
  {
    "startMs": 2985599,
    "text": "answer the kidney stuff it gives answer"
  },
  {
    "startMs": 2987760,
    "text": "but when you look at the relations"
  },
  {
    "startMs": 2989599,
    "text": "between the two people have somewhere"
  },
  {
    "startMs": 2992480,
    "text": "u doctors have made a link and that's"
  },
  {
    "startMs": 2994720,
    "text": "why you are able to see my take is if we"
  },
  {
    "startMs": 2998400,
    "text": "operated completely in the"
  },
  {
    "startMs": 3002000,
    "text": "latent space and we did not know the"
  },
  {
    "startMs": 3004720,
    "text": "connection connections at a human level."
  },
  {
    "startMs": 3006640,
    "text": "Can we explore in the latent space"
  },
  {
    "startMs": 3008480,
    "text": "completely and be able to find hidden"
  },
  {
    "startMs": 3010880,
    "text": "connections over there that could"
  },
  {
    "startMs": 3012720,
    "text": "translate into the real world or am I"
  },
  {
    "startMs": 3015599,
    "text": "thinking something crazy?"
  },
  {
    "startMs": 3016800,
    "text": ">> Yeah, I mean absolutely. I mean to the"
  },
  {
    "startMs": 3019599,
    "text": "extent that we have a pretty good"
  },
  {
    "startMs": 3021760,
    "text": "simulation, we can hope to find"
  },
  {
    "startMs": 3026319,
    "text": "surprising discoveries that turn out to"
  },
  {
    "startMs": 3028640,
    "text": "be correct in the simulated world. I"
  },
  {
    "startMs": 3031839,
    "text": "mean, you know, things go both ways,"
  },
  {
    "startMs": 3035359,
    "text": "right? There are also likely to be"
  },
  {
    "startMs": 3037920,
    "text": "things that turn out to be true in the"
  },
  {
    "startMs": 3039599,
    "text": "real world that weren't in our"
  },
  {
    "startMs": 3041119,
    "text": "simulation, right? There's this famous"
  },
  {
    "startMs": 3042800,
    "text": "statement about all models are wrong,"
  },
  {
    "startMs": 3044880,
    "text": "but some models are useful, right? And"
  },
  {
    "startMs": 3047920,
    "text": "that so anybody's world model,"
  },
  {
    "startMs": 3051280,
    "text": "regardless of whether it's the one we're"
  },
  {
    "startMs": 3052960,
    "text": "building or the world model in a human"
  },
  {
    "startMs": 3055359,
    "text": "head, right, they're not always right."
  },
  {
    "startMs": 3057520,
    "text": "Sometimes we think a person's going to"
  },
  {
    "startMs": 3059040,
    "text": "react in one way and they react in"
  },
  {
    "startMs": 3060640,
    "text": "another way. But nevertheless, a lot of"
  },
  {
    "startMs": 3063280,
    "text": "the time it can let us explore much more"
  },
  {
    "startMs": 3066400,
    "text": "widely and discover new facts and new"
  },
  {
    "startMs": 3068960,
    "text": "connections that we weren't aware of in"
  },
  {
    "startMs": 3070720,
    "text": "the real world."
  },
  {
    "startMs": 3073680,
    "text": "Okay,"
  },
  {
    "startMs": 3076160,
    "text": ">> awesome. That's a wrap everyone. Thank"
  },
  {
    "startMs": 3077920,
    "text": "you so much, Chris."
  }
]