The Godmother of AI on jobs, robots & why world models are next | Dr. Fei-Fei Li (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) A lot of people call you the godmother (00:00:01) of AI. The work you did actually was the (00:00:04) spark that brought us out of AI winter (00:00:06) >> in the middle of 2015, middle of 2016. (00:00:09) Some tech companies avoid using the word (00:00:12) AI because they were not sure if AI was (00:00:15) a dirty word. 2017ish (00:00:19) was the beginning of companies calling (00:00:21) themselves AI companies. (00:00:23) >> There's this line, I think this was when (00:00:24) you were presenting to Congress, there's (00:00:25) nothing artificial about AI. It's (00:00:27) inspired by people. It's created by (00:00:28) people. And most importantly, it impacts (00:00:30) people. (00:00:31) >> It's not like I think AI will have no (00:00:33) impact on jobs or people. In fact, I (00:00:36) believe that whatever AI does currently (00:00:39) or in the future is up to us. It's up to (00:00:42) the people. I do believe technology is a (00:00:45) net positive for humanity. But I think (00:00:48) every technology is a double-edged (00:00:50) sword. If we're not doing the right (00:00:52) thing as a society, as individuals, we (00:00:55) can screw this up as well. you had this (00:00:57) breakthrough insight of just okay we can (00:00:59) train machines to think like humans but (00:01:00) it's just missing the data that humans (00:01:02) have to learn as a child (00:01:03) >> I chose to look at artificial (00:01:05) intelligence through the lens of visual (00:01:07) intelligence because humans are deeply (00:01:10) visual animals we need to train machines (00:01:13) with as much information as possible on (00:01:15) images of objects but objects are very (00:01:19) very difficult to learn a single object (00:01:22) can have infinite possibilities that is (00:01:24) shown on an image in order To train (00:01:27) computers with tens and thousands of (00:01:30) object concepts, you really need to show (00:01:32) it millions of examples. (00:01:36) Today, my guest is Dr. Feay Lee, who's (00:01:39) known as the godmother of AI. Feet has (00:01:42) been responsible for and at the center (00:01:44) of many of the biggest breakthroughs (00:01:45) that sparked the AI revolution that we (00:01:47) are currently living through. She (00:01:49) spearheaded the creation of ImageNet, (00:01:51) which was basically her realizing that (00:01:53) AI needed a ton of clean labelled data (00:01:55) to get smarter. And that data set became (00:01:58) the breakthrough that led to the current (00:01:59) approach to building and scaling AI (00:02:01) models. She was chief AI scientist at (00:02:04) Google Cloud, which is where some of the (00:02:05) biggest early technology breakthroughs (00:02:07) emerged from. She was director at Sale, (00:02:09) Stanford's artificial intelligence lab, (00:02:11) where many of the biggest AI minds came (00:02:13) out of. She's also co-creator of (00:02:15) Stanford's human- centered AI institute, (00:02:17) which is playing a vital role in the (00:02:19) direction that AI is taking. She's also (00:02:21) been on the board of Twitter. She was (00:02:22) named one of Time's 100 most influential (00:02:25) people in AI. She's also on the United (00:02:28) Nations Advisory Board. I could go on. (00:02:30) In our conversation, Fay shares a brief (00:02:32) history of how we got to today in the (00:02:34) world of AI, including this mind-blowing (00:02:37) reminder that 9 to 10 years ago, calling (00:02:39) yourself an AI company was basically a (00:02:41) death nail for your brand. because no (00:02:44) one believed that AI was actually going (00:02:45) to work. Today, it's completely (00:02:47) different. Every company is an AI (00:02:49) company. We also chat about her take on (00:02:51) how she sees AI impacting humanity in (00:02:54) the future, how far current technologies (00:02:56) will take us, why she's so passionate (00:02:58) about building a world model, and what (00:03:00) exactly world models are. And most (00:03:02) exciting of all, the launch of the (00:03:04) world's first large world model, Marble, (00:03:06) which just came out as this podcast (00:03:08) comes out. Anyone can go play with this (00:03:10) at marble.worldlabs.ai. (00:03:12) It's insane. Definitely check it out. (00:03:14) Fei is incredible and way too under the (00:03:17) radar for the impact that she's had on (00:03:18) the world. So, I am really excited to (00:03:20) have her on and to spread her wisdom (00:03:22) with more people. A huge thank you to (00:03:24) Ben Harowitz and Condisa Rice for (00:03:26) suggesting topics for this conversation. (00:03:28) If you enjoy this podcast, don't forget (00:03:29) to subscribe and follow it in your (00:03:30) favorite podcasting app or YouTube. With (00:03:32) that, I bring you Dr. Fay Lee after a (00:03:35) short word from our sponsors. This (00:03:38) episode is brought to you by Figma, (00:03:39) makers of Figma make. When I was a PM at (00:03:42) Airbnb, I still remember when Figma came (00:03:44) out and how much it improved how we (00:03:46) operated as a team. Suddenly, I could (00:03:48) involve my whole team in the design (00:03:50) process, give feedback on design (00:03:52) concepts really quickly, and it just (00:03:54) made the whole product development (00:03:55) process so much more fun. But Figma (00:03:57) never felt like it was for me. It was (00:03:59) great for giving feedback and designs, (00:04:01) but as a builder, I wanted to make (00:04:03) stuff. That's why Figma built Figma (00:04:05) Make. With just a few prompts, you can (00:04:08) make any idea or design into a fully (00:04:11) functional prototype or app that anyone (00:04:13) can iterate on and validate with (00:04:14) customers. Figma make is a different (00:04:16) kind of vibe coding tool. Because it's (00:04:18) all in Figma, you can use your team's (00:04:20) existing design building blocks, making (00:04:22) it easy to create outputs that look good (00:04:25) and feel real and are connected to how (00:04:27) your team builds. Stop spending so much (00:04:29) time telling people about your product (00:04:31) vision and instead show it to them. Make (00:04:34) codeback prototypes and apps fast with (00:04:36) Figma Makeake. Check it out at (00:04:38) figma.com/lenny. (00:04:40) Did you know that I have a whole team (00:04:42) that helps me with my podcast and with (00:04:44) my newsletter? I want everyone on that (00:04:46) team to be super happy and thrive in (00:04:48) their roles. Just Works knows that your (00:04:50) employees are more than just your (00:04:51) employees. They're your people. My team (00:04:53) is spread out across Colorado, (00:04:55) Australia, Nepal, West Africa, and San (00:04:58) Francisco. My life would be so (00:05:00) incredibly complicated to hire people (00:05:02) internationally, to pay people on time (00:05:03) and in their local currencies, and to (00:05:05) answer their HR questions 24/7. But with (00:05:08) Just Works, it's super easy. Whether (00:05:10) you're setting up your own automated (00:05:12) payroll, offering premium benefits, or (00:05:14) hiring internationally, JustWorks offers (00:05:16) simple software and 24/7 human support (00:05:19) from small business experts for you and (00:05:21) your people. They do your human (00:05:23) resources right so that you can do right (00:05:25) by your people. just works for your (00:05:27) people. (00:05:30) [Music] (00:05:31) Fay Fay, thank you so much for being (00:05:33) here and welcome to the podcast. (00:05:35) >> I'm excited to be here, Lenny. (00:05:36) >> I'm even more excited to have you here. (00:05:39) It is such a treat to get to chat with (00:05:40) you. There's so much that I want to talk (00:05:42) about. You've been at the center of this (00:05:44) AI explosion that we're seeing right now (00:05:46) for so long. We're going to talk about a (00:05:48) bunch of the history that I think a lot (00:05:50) of people don't even know about how this (00:05:52) whole thing started. But let me first (00:05:54) read a quote from Wyatt about you just (00:05:55) so people get a sense and in the intro (00:05:57) I'll share all of the other epic things (00:05:58) you've done but I think this is a good (00:06:00) way to just set context. Fay is one of a (00:06:02) tiny group of scientists a group perhaps (00:06:04) small enough to fit around a kitchen (00:06:06) table who are responsible for AI's (00:06:08) recent remarkable advances. A lot of (00:06:11) people call you the godmother of AI. And (00:06:14) unlike a lot of AI leaders, you're an AI (00:06:17) optimist. You don't think AI is going to (00:06:19) replace us. You don't think it's going (00:06:21) to take all our jobs. you don't think (00:06:22) it's going to kill us. So, I thought (00:06:23) it'd be fun to start there. Just what's (00:06:25) your perspective on how AI is going to (00:06:28) impact humanity over time? (00:06:30) >> Yeah. Okay. So, Lenny, let me be very (00:06:32) clear. I'm not a utopian. So, it's not (00:06:36) like I think AI will have no impact on (00:06:38) jobs or people. In fact, I'm a humanist. (00:06:42) I believe that whatever AI does in (00:06:47) currently or in the future is up to us. (00:06:49) It's up to the people. So I do believe (00:06:52) technology is a net positive for (00:06:55) humanity. If you look at the long course (00:06:57) of civilization, I think we are an (00:07:01) fundamentally we're an innovative (00:07:03) species that we you know if you look at (00:07:07) from you know written record thousands (00:07:11) of years ago um to to now humans just (00:07:14) kept innovating ourselves and innovating (00:07:17) our tools and with that we make lives (00:07:20) better. we make work better, we build (00:07:22) civilization, and I do believe AI is (00:07:25) part of that. So, that's where the (00:07:27) optimism comes from. But I think every (00:07:31) technology is uh is um a double-edged (00:07:34) sword. And uh if we're not doing the (00:07:37) right thing as a species, as a society, (00:07:42) as communities, as individuals, we can (00:07:45) screw this up as well. H there's this (00:07:48) line I think this was when you were (00:07:49) presenting to Congress. There's nothing (00:07:51) artificial about AI. It's inspired by (00:07:53) people. It's created by people and most (00:07:54) importantly it impacts people. Uh I (00:07:57) don't have a question there but what a (00:07:58) what a great line. (00:07:59) >> Yeah I I I feel pretty deeply. I you (00:08:02) know I started um working in AI two and (00:08:06) a half decades ago and I've been having (00:08:09) students for the past two decades and (00:08:11) almost every student who graduates I (00:08:14) remind them you know when they graduates (00:08:17) from my lab that your field is called (00:08:20) artificial intelligence but there's (00:08:22) nothing artificial about it. (00:08:23) >> Coming back to the point you just made (00:08:24) about how it's kind of up to us about (00:08:26) where this all goes. What is it you (00:08:28) think we need to get right? How how do (00:08:29) we set things on a path? I know this is (00:08:31) a a very difficult question to answer (00:08:33) but just what should what what's your (00:08:35) advice? What do you think we should (00:08:36) >> Yeah. (00:08:37) >> How many hours do we have? (00:08:39) >> How do we align AI? There we go. Let's (00:08:41) solve it. (00:08:41) >> Also, I think people should be (00:08:44) responsible individuals no matter what (00:08:47) we do. This is what we teach our (00:08:49) children and this is what we need to do (00:08:51) as grown-ups as well. No matter which (00:08:55) part of the AI development or AI (00:08:59) deployment or or AI application you are (00:09:02) participating in and most likely many of (00:09:05) us especially as technologists were were (00:09:08) in multiple points we should act like (00:09:11) responsible individuals and uh and care (00:09:14) about this actually care a lot about (00:09:16) this. I think everybody today should (00:09:18) care about AI because it is going to (00:09:21) impact your individual life. It is going (00:09:24) to impact your community. It's going to (00:09:26) impact the the society and the future (00:09:29) generation. And caring about it as a (00:09:32) responsible person is the first but also (00:09:36) the most important step. (00:09:37) >> Okay. So, let me let me actually take a (00:09:39) step back and kind of go to the (00:09:41) beginning of AI. Most people started (00:09:44) hearing and caring about AI is what it's (00:09:47) called today. Just like I don't know a (00:09:48) few years ago when JGBT came out. Maybe (00:09:50) it was like three years ago. (00:09:51) >> Three years ago. Almost one more month. (00:09:54) Three years ago. (00:09:55) >> Wow. Okay. That was JT GBT coming out. (00:09:57) Is that the milestone that you have in (00:09:58) mind? Okay. Cool. That's exactly how I (00:10:00) saw it. But very few people know there (00:10:01) was a long long history of people (00:10:03) working on it was called machine (00:10:05) learning back then and there's other (00:10:06) terms and now it's just everything's AI (00:10:08) and there was kind of like a long period (00:10:10) of just a lot of people working on it (00:10:11) and then there's this what people refer (00:10:12) to as the AI winter where people just (00:10:14) gave up almost people did and just okay (00:10:17) this this idea isn't going anywhere and (00:10:19) then the work you did actually was (00:10:21) essentially the spark that brought us (00:10:23) out of AI winter and is directly (00:10:25) responsible for the world we're in now (00:10:27) of just AI is all we talk about as you (00:10:28) just said it's going to impact (00:10:30) everything we do. So, I thought it'd be (00:10:32) really interesting to hear from you just (00:10:33) kind of like the brief history of what (00:10:36) the world was like before imageet, then (00:10:38) just the work you did to create (00:10:41) ImageNet, why that was so important, and (00:10:42) then just what happened after. (00:10:44) >> It is for me hard to keep in mind that (00:10:48) AI is so new for everybody. when I lived (00:10:52) my entire professional life in AI, it's (00:10:56) there's a part of me that is just it's (00:10:59) so satisfying to see a personal (00:11:02) curiosity that I started barely out of (00:11:05) teenagehood and and now has become a (00:11:09) transformative (00:11:11) force of our civilization. It generally (00:11:14) is a civilizational level uh technology. (00:11:17) So, so that journey is about about 30 (00:11:21) years or 20 something 20 plus years and (00:11:25) uh it's it's just very satisfying. So, (00:11:28) where did it all start? Well, I'm not (00:11:30) even the first generation AI researcher. (00:11:33) The first generation really date back to (00:11:36) the 50s and 60s. And you know Alan (00:11:39) Touring was ahead of his time by in the (00:11:42) 40s by asking daring humanity with the (00:11:45) question can we is there thinking (00:11:47) machines right and of course he has a (00:11:50) specific way of uh testing this concept (00:11:54) of thinking machine which is a (00:11:56) conversational chatbot which to his (00:11:59) standard we now have a thinking machine (00:12:02) but uh that was just a more anecdotal (00:12:07) inspir inspiration. The field really (00:12:09) began in the 50s um when computer (00:12:12) scientists came together and look at how (00:12:15) we can use computer programs and (00:12:18) algorithms to uh to build these programs (00:12:23) that can do things that have been only (00:12:27) capable by human cognition. So um and (00:12:31) and that was the beginning and the (00:12:32) founding fathers the Dartmouth workshop (00:12:35) in the 1956 uh you know we have (00:12:38) professor John McCarthy who later came (00:12:40) to uh Stanford who coined the term (00:12:43) artificial intelligence (00:12:45) and between the 50s60s 70s and 80s it (00:12:50) was the early days of AI exploration and (00:12:54) we had logic systems we had uh expert (00:12:58) systems We also had early exploration of (00:13:01) neuronet network and then it came to (00:13:05) around the late 80s, the 90s and the the (00:13:10) very beginning of the 21st century. That (00:13:13) stretch about 20 years is actually the (00:13:16) beginning of machine learning. It's the (00:13:18) marriage between computer programming (00:13:20) and statistical as uh learning. And that (00:13:25) marriage brought a very very critical (00:13:28) concept into AI which is that (00:13:33) purely rulebased (00:13:35) um uh program is not going to account (00:13:39) for the vast amount of cognitive (00:13:43) capabilities that we imagine computers (00:13:45) can do. So we have to use machines to (00:13:49) learn the patterns. Once the machines (00:13:52) can learn the patterns, it has a hope to (00:13:55) do more things. For example, if you give (00:13:58) it three cats, the hope is not just for (00:14:01) the machines to recognize these three (00:14:03) cats. The hope is the machines can (00:14:06) recognize the fourth cat, the fifth cat, (00:14:08) the sixth cat, and all the other cats. (00:14:10) And that's a learning ability that is (00:14:13) fundamental to humans and many animals. (00:14:17) and uh we we as a field realized we need (00:14:20) machine learning. So that was up till (00:14:23) the beginning of the 21st century. I (00:14:27) entered the field of AI literally in the (00:14:29) year of 2000. That's when my uh PhD (00:14:32) began at Caltech. And so I was one of (00:14:35) the first generation machine learning (00:14:37) researchers and we were already studying (00:14:40) this concept of machine learning (00:14:42) especially neuronet network. I remember (00:14:44) that was one of my first courses in at (00:14:47) Caltech is called neuro network but it (00:14:50) was very painful. It was still smack in (00:14:52) the middle of the so-called AI winter (00:14:54) meaning the public didn't look at this (00:14:57) too much. there wasn't that much funding (00:14:59) but there was also a lot of ideas (00:15:02) flowing around and I think two things (00:15:06) happened to myself that brought my own (00:15:08) career so close to the birth of modern (00:15:11) AI is that um I chose to look at (00:15:16) artificial intelligence through the lens (00:15:18) of visual intelligence because uh humans (00:15:22) are deeply visual animals. We can talk a (00:15:25) little more later, but so much of our (00:15:28) intelligence is built upon visual, (00:15:32) perceptual, spatial understanding, not (00:15:34) just language per se. I think they're (00:15:36) complimentary. So I chose to look at (00:15:38) visual intelligence and um my PhD and my (00:15:41) early uh professor years I um my (00:15:46) students and I are very committed to a (00:15:48) northstar problem which is solving the (00:15:50) problem of object recognition because (00:15:53) it's a building block for the perceptual (00:15:56) world. Right? We go around the world (00:15:58) interpreting, reasoning and interacting (00:16:00) with it more or less at the object (00:16:03) level. We don't interact with the world (00:16:05) at the molecular level. We don't (00:16:08) interact with the world as (00:16:10) um we sometimes do but we rarely for (00:16:13) example if you want to lift a teapot you (00:16:15) don't say okay the teapot is made of a (00:16:18) 100 pieces of porcelain and let me work (00:16:21) on this 100 pieces you look at this as (00:16:23) one object and and interact with it. So (00:16:26) object is really important. So um I was (00:16:30) among the first uh uh researchers to (00:16:33) identify this as a northstar problem. (00:16:36) But I think what happened is that (00:16:39) as a student of AI and then a researcher (00:16:42) of AI, I was working on all kinds of (00:16:46) mathematical models including neuronet (00:16:48) network including Beijian network (00:16:51) including many many models and there was (00:16:54) one singular pain point is that these (00:16:57) models don't have data to be trained on (00:17:00) and uh as a field we were so focusing on (00:17:04) these models but It dawned on me that (00:17:07) human learning (00:17:10) as well as evolution is actually a big (00:17:14) data learning process. Humans learn with (00:17:16) so much experience you know constantly (00:17:19) and evolution if you look at time (00:17:22) animals evolve with just experiencing (00:17:24) the world. So I think my students and (00:17:27) and I conjectured (00:17:30) that a very critically overlooked (00:17:33) ingredient of bringing AI to life is big (00:17:37) data and then we began this image that (00:17:40) project in 2006 2007 we were very (00:17:43) ambitious we want to get the entire (00:17:46) internet's image data on objects now (00:17:50) granted internet was a lot smaller than (00:17:52) today so we I felt like that ambition (00:17:55) was at least not too crazy. Now it's (00:17:58) totally delusional to uh to think a (00:18:02) couple of graduate student and a (00:18:04) professor can do this. But uh and that's (00:18:07) what we did. We curated very carefully (00:18:10) 15 million images on the internet. (00:18:13) Created a taxonomy of 22,000 (00:18:17) concepts borrowing other researchers (00:18:20) work like a linguist work on wordnet and (00:18:24) it's a particular way of dictionarying (00:18:27) uh words and we combine that into image (00:18:31) that and we open source that to the (00:18:34) research community. We held an annual (00:18:37) image net challenge to encourage (00:18:40) everybody to participate in this. We (00:18:42) continue to do our own research. But (00:18:45) 2012 was the moment that many people (00:18:48) think was the beginning of the deep (00:18:50) learning or birth of modern AI because a (00:18:53) group of Toronto researchers led by (00:18:55) professor Jeff Hinton (00:18:58) participated in imageet challenge used (00:19:00) the imageet big data and two GPUs from (00:19:04) Nvidia and created successfully the (00:19:07) first neuronet network algorithm that's (00:19:10) can it didn't fundamental it didn't (00:19:14) totally solved but made a huge progress (00:19:17) towards solving the problem of object (00:19:19) recognition and that combination of the (00:19:22) trio technology (00:19:25) uh big data neuronet network and GPU was (00:19:29) kind of the golden recipe for modern AI (00:19:33) and then fast forward the the the public (00:19:36) moment of AI which is the chat GPT (00:19:41) moment if you look at the ingredients of (00:19:45) what brought Chad GPT to to the to the (00:19:48) uh world technically still use these (00:19:52) three ingredients. Now it's internet (00:19:55) scale data mostly texts is a much more (00:19:59) com complex neuronet network um (00:20:02) architecture than 2012 but it's still (00:20:05) neuronet network and a lot more GPUs but (00:20:08) it's still GPUs. So these three (00:20:11) ingredients are still to at the core of (00:20:14) modern AI. (00:20:16) >> Incredible. I have never heard that full (00:20:19) story before. I love that it was two (00:20:21) GPUs was the f I love (00:20:26) and now it's I don't know hundreds of (00:20:27) thousands right that are orders of (00:20:29) magnitudes more powerful uh and those (00:20:31) two GPUs were they just bought they were (00:20:33) like gaming GPUs they just went to like (00:20:35) the game store right that people use for (00:20:36) playing games (00:20:37) >> as you said this continues to be in a (00:20:40) large way the way models get smarter (00:20:42) some of the fastest growing companies in (00:20:43) the world right now I've had them all (00:20:44) mostly on the podcast Merkore and Surge (00:20:46) and Scale like they do this they (00:20:48) continue to do this for labs just give (00:20:50) them more and more label data of the (00:20:52) things they're most excited about. (00:20:53) >> Yeah, I remember um Alex Wong from scale (00:20:57) very early days. I probably still has (00:20:58) his emails when he was starting scale. (00:21:01) He uh he was very kind. He keeps sending (00:21:03) me emails about how image that inspired (00:21:07) scale. I was very pleased to see that. (00:21:09) >> One of my other favorite takeaways from (00:21:11) what you just shared is just such an (00:21:12) example of high agency and just doing (00:21:15) things. That's kind of a meme on (00:21:16) Twitter. Just you can just do things. (00:21:18) you're just like okay this is probably (00:21:20) necessary to move AI and it was called (00:21:22) machine learning back then right was (00:21:24) that the term most people used (00:21:25) >> I think it was interchangeably it's true (00:21:28) like I do remember the companies the (00:21:31) tech companies I I'm not going to name (00:21:33) names but I was I was uh in a (00:21:36) conversation in one of the early days I (00:21:38) think is in the middle of 2015 middle of (00:21:42) 2016 uh some tech companies avoids using (00:21:46) the word AI I because they were not sure (00:21:49) if AI was a dirty word. And I remember I (00:21:52) was actually (00:21:54) encouraging everybody to use the word AI (00:21:57) because to me that is one of the most (00:22:00) audacious question humanity has ever (00:22:03) asked in our quest for science and (00:22:06) technology and I feel very proud of this (00:22:08) term. But yes, at the beginning some (00:22:11) people were not sure. (00:22:13) >> What year was that roughly when AI was (00:22:14) developed? 2016 I think that was (00:22:17) >> less than 10 years ago (00:22:18) >> that was the changing like um some (00:22:21) people start calling it AI but I think (00:22:24) if you look at the Silicon Valley tech (00:22:27) company companies if you trace their (00:22:30) marketing term I think (00:22:34) 2017ish (00:22:36) was the beginning of companies calling (00:22:38) themselves AI companies (00:22:40) >> that's incredible just how the world has (00:22:43) changed now you Can't not call yourself (00:22:45) an AI company. (00:22:46) >> I know. (00:22:47) >> Just nineish years later. (00:22:49) >> Yeah. (00:22:49) >> Oh man. Okay. Is there anything else (00:22:52) around the history that early history (00:22:54) that you think people don't know that (00:22:55) you think is important before we chat (00:22:57) about where think things are going in (00:22:58) the work that you're doing? (00:23:01) >> I think as all histories, you know, I'm (00:23:04) keenly aware that uh I am recognized for (00:23:08) being part of the history, but there are (00:23:10) so many heroes and so many researchers. (00:23:13) We're talking about generations of (00:23:15) researchers there. You know, in my own (00:23:18) world, there are so many people who have (00:23:20) in inspired me, which I I talked about (00:23:23) in my book. But I do feel our culture, (00:23:27) especially Silicon Valley tends to (00:23:31) assign um achievements to a single (00:23:34) person. Well, while I think it has (00:23:37) value, um but it's it's just to be (00:23:40) remembered. AI is a field of at this (00:23:43) point 70 years old and we have gone (00:23:46) through many generations. Um nobody no (00:23:50) one um could have gotten here by (00:23:53) themselves. (00:23:54) >> Okay. So let me ask you this question. (00:23:56) It feels like we're always on this (00:23:58) precipice of AGI. This kind of vague (00:24:00) term people throw around. AGI is coming. (00:24:02) Is it going to take over everything? How (00:24:04) what's your take on how far you think we (00:24:06) might be from AGI? Do you think we're (00:24:07) going to get there on the current (00:24:09) trajectory we're on? Do you think we (00:24:10) need more breakthroughs? Do you think (00:24:11) the current approach will get us there? (00:24:13) >> Yeah, this is a very interesting term, (00:24:15) Lenny. Um, (00:24:18) I don't know if anyone has ever defined (00:24:21) AGI. (00:24:24) You know, there are many different (00:24:25) definitions including, you know, some (00:24:28) kind of superpower for machines all the (00:24:31) way to can um a machines can become (00:24:35) economically viable agent in in the (00:24:39) society. (00:24:41) In other words, making salaries to live. (00:24:43) Is that the definition of AGI? As a (00:24:46) scientist, I I take science very (00:24:49) seriously and I enter the field because (00:24:52) I was inspired by this audacious (00:24:55) question of can machines think and do (00:24:58) things in the way that human humans can (00:25:02) do. For me, that's always the northstar (00:25:05) of AI. And from that point of view, I (00:25:08) don't know what's the difference between (00:25:09) AI and AGI. I think we've done very well (00:25:14) in achieving parts of the goal, (00:25:16) including conversational AI, but I don't (00:25:19) think we have completely conquered all (00:25:21) the goals uh of of AI. And I think our (00:25:24) founding fathers that Alan Turing, I (00:25:28) wonder if Alan Turing is around today (00:25:31) and you ask him to contrast AI versus (00:25:33) AGI, he might just shrug and said, well, (00:25:37) I asked the same question back in 1940s. (00:25:40) So, so I don't want to get get onto a (00:25:44) rabbit hole of defining AI versus AGI. I (00:25:48) feel AGI is more a marketing term than a (00:25:52) scientific term. As a scientist and (00:25:54) technologist, (00:25:56) AI is my northstar is my field's (00:25:59) northstar and I'm happy people call it (00:26:01) whatever name they want to call it. (00:26:05) >> So let me ask you maybe maybe this way (00:26:07) like you described there's kind of these (00:26:09) components that from ImageNet and (00:26:11) AlexNet kind of took us to where we're (00:26:13) today. GPUs essentially data label data (00:26:17) just like the algorithm of the model. (00:26:20) There's also just the transformer feels (00:26:22) like an important step in that (00:26:24) trajectory. Do you feel like those are (00:26:26) the same components that'll get us to I (00:26:27) don't know 10 times smarter model (00:26:29) something that's like life-changing for (00:26:31) the entire world or do you think we need (00:26:33) more breakthroughs? I know we're we're (00:26:35) going to talk about world models which I (00:26:36) think is a component of this but is (00:26:38) there anything else that you think is (00:26:39) like oh this will plateau or okay this (00:26:42) will take us just need more data more (00:26:43) compute more GPUs. (00:26:44) >> Oh no I definitely think we need more uh (00:26:47) innovations. I I think scaling loss of (00:26:50) more data, more GPUs and bigger current (00:26:53) model architecture is there's still a (00:26:57) lot to be done there. But I absolutely (00:26:59) think we need to innovate more. Um there (00:27:02) is not a single (00:27:04) deeply scientific discipline in human (00:27:07) history that has arrived at a place that (00:27:11) says we're done. We're done innovating. (00:27:13) And AI is one one of the if not the (00:27:17) youngest discipline in in human (00:27:20) civilization in terms of science and (00:27:22) technology. We're still scratching the (00:27:24) surface. Uh for example, um like I said, (00:27:27) we're going to segue into world models (00:27:29) today. You take a a model and and and (00:27:34) run it through a a video of a couple of (00:27:37) office rooms and ask the the model to (00:27:40) count the number of chairs. And this is (00:27:42) something a toddler could do or maybe (00:27:44) maybe a a a elementary school kid could (00:27:47) do and AI could not do that, right? So (00:27:51) um there's just so much AI today could (00:27:53) not do then let alone thinking about how (00:27:57) did you know um someone like Isaac (00:28:00) Newton look at the movements of the (00:28:03) celestial bodies and and and derive an (00:28:07) equation or or a set of equations that (00:28:11) governs the movement of all bodies that (00:28:14) level of creativity extrapolation (00:28:17) abstraction we have no way of enabling (00:28:21) AI to do that today. And then let's look (00:28:24) at emotional intelligence. If you look (00:28:26) at a student coming into a teacher's (00:28:30) office and have a conversation about (00:28:33) motivation, passion, what to learn, (00:28:35) what's the problem that's that's you (00:28:38) know really uh bothering you. that (00:28:41) conversation as powerful as as today's (00:28:45) conversational bots are, you don't get (00:28:48) that level of emotional cognitive (00:28:51) intelligence uh from today's AI. So (00:28:54) there's a lot we can do better. Um and I (00:28:58) do not believe we're done innovating. (00:29:00) >> Uh Demis had this really interesting (00:29:02) interview recently from deep mind Google (00:29:04) where someone asked him just like what (00:29:05) do you think uh how far are we from AGI? (00:29:08) What does it look like when it's through (00:29:09) there? He had a really interesting way (00:29:10) of approaching it is if we were to give (00:29:12) a the most cutting edge model all the (00:29:14) information until the end of the 20th (00:29:17) century see if it could come up with all (00:29:19) the breakthroughs Einstein had and so (00:29:21) far we're never near that but they can (00:29:22) >> no we're not in fact it's even worse (00:29:26) let's give AI all the data including (00:29:30) modern instruments data of celestial (00:29:33) bodies which Newton did not have and (00:29:36) give it to that and just ask AI to (00:29:39) create the six 17th century set of (00:29:42) equations on the laws of bodily (00:29:45) movements. Today's AI cannot do that. (00:29:49) >> All right, we're a ways away is what I'm (00:29:50) hearing. (00:29:51) >> Yeah. (00:29:51) >> Okay, so let's talk about world models. (00:29:53) This is uh to me this is just another (00:29:55) really amazing example of you being (00:29:58) ahead of where people end up. So you (00:30:01) were way ahead on okay, we just need a (00:30:03) lot of clean data for AI and neural (00:30:06) networks to learn. uh you've been (00:30:07) talking about this idea of world models (00:30:09) for a long time. You started a company (00:30:10) to build uh essentially there's language (00:30:13) models. This is a different thing. This (00:30:14) is a world model. We'll talk about what (00:30:15) that is. And now uh as I was preparing (00:30:18) for this, Elon's like talking about (00:30:19) world models. Jensen's talking about (00:30:21) world models. I know Google's working on (00:30:22) this stuff. You've been at this for a (00:30:24) long time. And you're actually just (00:30:25) launched something that's going to we're (00:30:27) going to talk about uh right before this (00:30:29) podcast airs. Um talk about what is a (00:30:32) world model? Why is it so important? I'm (00:30:34) very excited to see that more and more (00:30:36) people are talking about role models (00:30:39) like Elon, like Jensen. Um, (00:30:43) I have been thinking about (00:30:46) really how to push AI forward all my (00:30:50) life, right? and the large language (00:30:53) models uh that came out of uh the (00:30:57) research world and then open AI and and (00:31:00) all this for the past few years were (00:31:03) extremely inspiring even for a (00:31:06) researcher like me. I remembered when (00:31:09) GPT2 came out and that was in I think (00:31:13) late 2020. (00:31:16) I was um co-director um I still am but I (00:31:20) was at that time uh full-time (00:31:22) co-director of Stanford's uh human (00:31:24) center AI institute and I I remember it (00:31:27) was you know the public was not aware of (00:31:30) the power of the large language model (00:31:32) yet but as researchers we were seeing it (00:31:35) we're seeing the future and I had pretty (00:31:38) long conversations with my natural (00:31:41) language processing colleagues like (00:31:44) Percy Leang and Chris Batting, we were (00:31:46) talking about how critical this (00:31:48) technology is going to be and Stanford (00:31:52) uh AI institute, human center AI (00:31:53) institute, hi was the first one to (00:31:56) establish a full research center um (00:31:59) foundation model. We were Percy Le Young (00:32:01) and and many researchers led the first (00:32:04) uh academic paper um foundation model. (00:32:07) So so it was just very inspiring for me. (00:32:10) So, of course, I come from the world of (00:32:13) visual intelligence and I was just (00:32:16) thinking there's so much we can um push (00:32:18) forward on beyond language because (00:32:22) humans um humans have used our sense of (00:32:29) spatial intelligence and world (00:32:31) understanding to do so many things and (00:32:34) they are beyond language. Think about a (00:32:37) very chaotic (00:32:39) first responder scene, whether it's fire (00:32:42) or some traffic accident or or some (00:32:46) natural disaster. And it's if you (00:32:51) immerse yourself in those scene and (00:32:53) think about how people organize (00:32:55) themselves to to rescue people, to stop (00:32:58) further disasters, to put down fires, to (00:33:02) to a lot of that is movements, is is (00:33:07) spontaneous understanding of objects, (00:33:10) worlds, hum (00:33:13) situational awareness. Language is part (00:33:16) of that. But a lot of those situations (00:33:19) language cannot get you to put down the (00:33:21) fire. So that is what is that? I I was (00:33:25) thinking a lot and in the meantime I was (00:33:27) doing a lot of robotics research and I (00:33:30) it ca it dawned on me that the lynch pin (00:33:34) of connecting (00:33:37) the additional intelligence in addition (00:33:40) to language and connecting embodied AI (00:33:44) which are robotics. connecting visual (00:33:47) intelligence is this sense of spatial (00:33:50) intelligence about understanding the (00:33:53) world and that's when um I think I um it (00:33:57) was 2024 I gave a TED talk about spatial (00:34:01) intelligence and world models and uh I (00:34:05) start formulating this idea uh back in (00:34:09) 2022 (00:34:11) um based on my robotics and computer (00:34:13) vision research and then one thing that (00:34:16) is really clear to me is that I really (00:34:20) want to work with the brightest uh (00:34:22) technologist and and move as fast as (00:34:26) possible to bring this technology to (00:34:28) life. And that's when we founded this (00:34:30) company called World Labs. And you can (00:34:33) see the the the word world is in the (00:34:36) title of our company because we believe (00:34:38) so much in world modeling and spatial (00:34:40) intelligence. (00:34:42) >> People are so used to just chat bots and (00:34:43) that's a large language model. So the (00:34:45) simple way to understand a world model (00:34:46) is you basically describe a scene and it (00:34:49) generates an infinitely (00:34:51) explorable world. We'll link to a the (00:34:53) thing you launch which we'll talk about (00:34:55) but just is that a simple way to (00:34:56) understand it? (00:34:56) >> That's part of it Lenny. I think a (00:34:58) simple way to understand a world model (00:35:01) uh is that this model can allow anyone (00:35:05) to create (00:35:08) any worlds in their mind's eye by (00:35:11) prompting whether it's an image or a (00:35:14) sentence (00:35:15) and also be able to interact in this (00:35:18) world. whether you're browsing and (00:35:21) walking or or picking objects up or or (00:35:24) or changing changing things as well as (00:35:29) to reason within this world. For (00:35:31) example, if if the person consuming if (00:35:35) the agent consuming this output of the (00:35:38) world model is a robot, it should be (00:35:40) able to plan its path and and help to (00:35:44) you know tidy the kitchen for example. (00:35:48) So, so world model is a (00:35:52) a foundation that that you can use to (00:35:56) reason, to interact, and to create (00:35:59) worlds. (00:36:00) >> Great. Yeah. So, robots feels like (00:36:02) that's potentially the next big focus (00:36:06) for AI researchers and just like the (00:36:08) impact on the world. And what you're (00:36:10) saying here is uh this is a key missing (00:36:13) piece of making robots actually work in (00:36:16) the real world. Understanding how the (00:36:17) world works. (00:36:18) >> Yeah. Well, first of all, I do think (00:36:20) there's more than robots that's (00:36:21) exciting. Um so, but I agree with (00:36:24) everything you just said. I think uh (00:36:26) world modeling and spatial intelligence (00:36:29) is a key missing piece of uh uh embody (00:36:33) AI. I also think let's not underestimate (00:36:36) that humans are embodied agents and (00:36:39) humans can be augmented by AI's uh (00:36:43) intelligence just like today humans are (00:36:46) language animals but we're very much (00:36:48) augmented by AI when helping us to you (00:36:52) know do language tasks including (00:36:54) software engineering. I I think that uh (00:36:57) we shouldn't underestimate or maybe it's (00:37:00) it's um we tend not to talk about how (00:37:04) humans as an embodied agents can (00:37:07) actually benefit so much from world (00:37:10) models and spatial intelligent u models (00:37:13) as well as robots can. So the big (00:37:16) unlocks here, robots, which uh a huge (00:37:19) deal. If this works out, imagine each of (00:37:21) us has robots doing a bunch of stuff for (00:37:22) us. Goes into, you know, they help us (00:37:24) with disasters, things like that. Uh (00:37:26) games obviously is a really cool (00:37:27) example. Just like infinitely playable (00:37:30) games that you just invent out of your (00:37:31) head. And then creativity feels like (00:37:34) just like being fun, having fun, being (00:37:35) creative, thinking of m wild new worlds (00:37:37) and and environments. (00:37:39) >> And also design. humans design from (00:37:42) machines to buildings to homes and also (00:37:46) scientific discovery right there is so (00:37:48) much u I I like to use the example of (00:37:52) the discovery of the structure of DNA if (00:37:55) you look at one of the most important (00:37:58) piece in DNA's discovery history is the (00:38:03) X-ray defraction photo that was captured (00:38:06) by Rosalyn Franklin and it was a flat 2D (00:38:10) photo of a structure that looks like it (00:38:13) looks like a cross with defractions. You (00:38:16) can you can uh Google those photos. But (00:38:19) with that 2D flat photo, (00:38:24) humans, especially two important humans, (00:38:27) James Watson and Francis Crick, in (00:38:30) addition to their other uh information, (00:38:33) was able to reason in 3D space and (00:38:38) deduce a highly three-dimensional double (00:38:41) helix structure of the DNA. And that (00:38:44) structure cannot possibly be 2D. You (00:38:48) cannot think in 2D and deduce that (00:38:52) structure. You have to think in 3D (00:38:55) spatial um use the the human spatial (00:38:58) intelligence. So I think even in (00:39:01) scientific discovery um spatial (00:39:03) intelligence or AI assisted spatial (00:39:06) intelligence is critical. (00:39:08) >> This is such an example of I think it (00:39:10) was Chris Dixon that had this line that (00:39:12) the next big thing is going to start off (00:39:14) feeling like a toy. When Chad GBT just (00:39:17) came out, if like I remember Salman just (00:39:19) tweeted as like here's a cool thing (00:39:20) we're playing with. Check it out. Now (00:39:21) it's the fastest growing product all of (00:39:23) history changed the world. (00:39:24) >> Yeah. (00:39:24) >> Uh and it's oftentimes the things that (00:39:26) just look like okay this is cool. Uh (00:39:29) that it's fun to play with and end up (00:39:30) changing the world most. (00:39:32) >> Yeah. (00:39:33) >> This episode is brought to you by Cinch, (00:39:35) the customer communications cloud. (00:39:38) Here's the thing about digital customer (00:39:39) communications. Whether you're sending (00:39:41) marketing campaigns, verification codes, (00:39:44) or account alerts, you need them to (00:39:45) reach users reliably. That's where Cinch (00:39:48) comes in. Over 150,000 businesses, (00:39:51) including eight of the top 10 largest (00:39:53) tech companies globally, use Cinch's API (00:39:55) to build messaging, email, and calling (00:39:58) into their products. And there's (00:39:59) something big happening in messaging (00:40:01) that product teams need to know about. (00:40:03) Rich Communication Services, or RCS. (00:40:06) Think of RCS as SMS 2.0. Instead of (00:40:09) getting text from a random number, your (00:40:11) users will see your verified company (00:40:13) name and logo without needing to (00:40:15) download anything new. It's a more (00:40:17) secure and branded experience. Plus, you (00:40:19) get features like interactive carousels (00:40:21) and suggested replies. And here's why (00:40:23) this matters. US carriers are starting (00:40:25) to adopt RCS. Cinch is already helping (00:40:28) major brands send RCS messages around (00:40:31) the world, and they're helping Lenny's (00:40:32) podcast listeners get registered first (00:40:34) before the rush hits the US market. (00:40:37) Learn more at get started at (00:40:38) cinch.com/lenny. (00:40:41) That's s i nch.com/lenny. (00:40:45) >> I reached out to Ben Horowitz who loves (00:40:47) what you're doing. A big fan of yours. (00:40:49) Uh they're investors I believe. And (00:40:51) >> yeah, we we've known each other for for (00:40:54) many years, but yes, right now they are (00:40:56) investors of uh Warlaps. (00:40:58) >> Amazing. Okay. So I asked him what I (00:40:59) should ask you about and he suggested (00:41:01) ask you why is the bitter why is the (00:41:03) bitter lesson alone not likely to work (00:41:07) for robots. So first of all just explain (00:41:10) what the bitter lesson was in the (00:41:12) history of AI and then just why that (00:41:14) won't get us to where we want to be with (00:41:15) robots. (00:41:16) So well first of all there are many (00:41:18) bitter lessons but (00:41:21) but the bitter lessons everybody refers (00:41:23) to is a u is a paper written by Richard (00:41:26) Sutton who won the touring award (00:41:29) recently and he does a lot of (00:41:31) reinforcement learning and Richard has (00:41:33) said right if you look at the the (00:41:35) history especially the algorithmic (00:41:37) development of AI it turns out simpler (00:41:41) model with a ton of data always win at (00:41:45) the end of the day instead of the the um (00:41:49) the you know more complex model with (00:41:52) less data. I mean that was actually this (00:41:55) paper came years after imageet that to (00:41:58) me was not bitter it was a sweet lesson (00:42:02) that's why I built uh image net because (00:42:04) I believe that uh big data plays that (00:42:07) role so why can bitter lesson work in (00:42:11) robotics alone well first of all um I (00:42:15) think we need to give credit to where we (00:42:18) are today robotics is very much in the (00:42:21) early days of (00:42:23) experimentation. It's not the the (00:42:26) research is not nearly as mature as say (00:42:29) language models. So many people are (00:42:33) still um experimenting with different (00:42:36) algorithms and some of those algorithms (00:42:38) are driven by big data. So I do think (00:42:42) big data will continue to play a role in (00:42:46) robotics and um but what is hard for (00:42:51) robotics there are a couple of things (00:42:53) one is that (00:42:55) it's harder to get data it's a lot (00:42:58) harder to get data you can say well (00:43:00) there is web data this is where the (00:43:02) latest robotics research is using web (00:43:05) videos and I think web videos do do play (00:43:09) a role but if you Think about what made (00:43:11) language model work. A very as someone (00:43:15) who does computer vision and and spatial (00:43:18) intelligence and robotics, I'm very (00:43:19) jealous of my colleagues in um in (00:43:22) language because they had this perfect (00:43:26) setup where their training data are in (00:43:29) words eventually tokens and then they (00:43:33) produce a model that outputs words. So (00:43:37) you have this perfect alignment between (00:43:40) what you hope to get which we call (00:43:42) objective function and what your (00:43:45) training data looks like. But robotics (00:43:48) is different. Even spatial intelligence (00:43:50) is different. You hope to get actions (00:43:54) out of robots. (00:43:56) But your training data lacks (00:43:59) actions in 3D worlds. And that's what (00:44:03) robots have to do, right? actions in 3D (00:44:06) worlds. So, you have to um find (00:44:09) different ways to fit a uh what do they (00:44:14) call a a a a square in a round hole that (00:44:19) what we have is tons of web videos. (00:44:23) So then we have to start talking about (00:44:26) uh adding supplementing (00:44:29) data such as teleaoperation data or (00:44:33) synthetic data so that the robots are (00:44:36) trained with this hypothesis of bitter (00:44:39) lesson which is large amount of data. I (00:44:42) think there's still hope because even (00:44:45) what we are doing um in world modeling (00:44:48) will really unlock a lot of this uh (00:44:52) information for robots but I think we (00:44:54) have to be careful because we're at the (00:44:56) early days of this and bitter lesson is (00:44:59) still to be tested uh because we haven't (00:45:04) fully figured out the data for another (00:45:08) part of the bitter lesson of robotics I (00:45:10) think we should be so (00:45:12) so realistic about is again compared to (00:45:16) language models or even spatial models, (00:45:19) robots are physical systems. So robots (00:45:24) are closer to self-driving cars than a (00:45:27) large language model. And that's very (00:45:29) important to recognize. That means that (00:45:33) in order for robots to work, we not only (00:45:37) need brains, we also need the physical (00:45:40) body, we also need application (00:45:43) scenarios. And if you look at the the (00:45:45) the the the (00:45:47) history of self-driving car, um my (00:45:50) colleague Sebastian Thrum uh uh took (00:45:53) Stanford's car uh to win the first DARPA (00:45:58) challenge in 2006 or 2005. It's 20 years (00:46:02) since that prototype of a self-driving (00:46:06) car being able to drive 130 miles in the (00:46:10) Nevada desert to today's Whimo and um on (00:46:15) the street of San Francisco and we're (00:46:18) not even done yet. There's still a lot. (00:46:20) So that's a 20 year journey. And (00:46:23) self-driving cars are much simpler (00:46:25) robots. They're just metal boxes running (00:46:27) on 2D surfaces. And the goal is not to (00:46:31) touch anything. Robot is 3D things (00:46:36) running in 3D world and the goal is to (00:46:39) touch things. So the journey is going to (00:46:42) be you know there's many aspects (00:46:45) elements and of course one could say (00:46:48) well the self-driving car early (00:46:50) algorithm were pre-deep learning era. So (00:46:54) deep learning is accelerating uh the (00:46:56) brains and I think that's true. That's (00:46:58) why I'm in robotics. That's why I'm in (00:47:01) spatial intelligence and I'm excited by (00:47:03) it. But in the meantime, the car (00:47:05) industry is very mature and productizing (00:47:10) also involves the mature (00:47:13) use cases, supply chains, the hardware. (00:47:16) So I think it's a very interesting time (00:47:18) to work in these problems. But it's true (00:47:21) Ben is right. we might still be subject (00:47:26) to a number of bitter lessons (00:47:29) >> doing this work. Do you ever just feel (00:47:31) awe for the way the brain works and is (00:47:33) able to do all of this for us? Just the (00:47:36) complexity just to get a a machine to (00:47:39) just walk around and not hit things and (00:47:41) fall. Does it just give you more spec (00:47:43) for what we've already got? (00:47:44) >> Totally. We we operate on about 20 (00:47:48) watts. (00:47:50) That's dimmer than any light bulb in in (00:47:52) the room. I'm in right now. And yet we (00:47:56) can do so much. So I think actually the (00:47:59) more I work in AI, the more I respect (00:48:02) humans. (00:48:03) >> Let's talk about this uh product you (00:48:06) just launched. It's called Marble. A (00:48:07) very cute name. Talk about what this is, (00:48:09) why this important. I've been playing (00:48:10) with it. It's incredible. We'll link to (00:48:11) it and for folks to check it out. What (00:48:13) is Marble? (00:48:15) >> Yeah, I'm very excited. So first of all, (00:48:17) Marbo is uh one of the first product (00:48:19) that World Labs uh has rolled out. (00:48:22) Worldlabs is a foundation frontier model (00:48:25) company. We are founded by four (00:48:28) co-founders who have deep technical (00:48:31) history. My co-founders Justin Johnson (00:48:34) uh Kristoff uh Lassner and Ben (00:48:37) Mildenhal. We all come from the research (00:48:40) field of AI, computer graphics, computer (00:48:42) vision. And uh we believe that spatial (00:48:45) intelligence and world modeling is as (00:48:49) important if not more to uh language (00:48:51) models and uh complementaryary to to (00:48:54) language models. So we wanted to seize (00:48:57) this opportunity to create deep uh tech (00:49:02) research lab that can connect the dots (00:49:05) between um frontier models with (00:49:08) products. So, Marvel is an app that's (00:49:13) built upon our frontier models. We've (00:49:16) spent a year and plus building the (00:49:19) world's first uh generative model that (00:49:22) can output genuinely 3D worlds. That's a (00:49:27) very very hard problem. Um and uh and I (00:49:32) it it was a very hard process. Uh we uh (00:49:36) we have a team of incredible founding (00:49:38) team of incredible technologists from (00:49:41) you know incredible uh teams. And then (00:49:47) around um just a month or two ago, we (00:49:51) saw the first time that we we can just (00:49:55) prompt with a sentence and an image and (00:49:58) multiple images and create worlds that (00:50:01) we can just navigate in. If you put it (00:50:04) on goggle, which we have an option to (00:50:06) let you do that, you can even walk (00:50:08) around, right? So it was even though (00:50:11) we've been building this for for for (00:50:13) quite a while, it was still just all (00:50:15) inspiring and we wanted to get into the (00:50:18) hands of uh people who need it. And then (00:50:21) we know that so many creators, (00:50:24) designers, people who are thinking about (00:50:28) uh robotic simulation, people who are (00:50:30) thinking about uh different use cases of (00:50:34) uh navigable interactable (00:50:37) um uh immersive worlds, game developers (00:50:40) will find this useful. So we uh develop (00:50:43) developed Marble as a first step. It's (00:50:46) it's again still very early uh but it's (00:50:50) the world's first uh model doing this (00:50:52) and it's the world's first uh product (00:50:55) that allows people to just uh prompt we (00:50:59) call it prompt to worlds. (00:51:01) >> Well, I've been playing around with it. (00:51:02) It is insane. Like you could just have a (00:51:04) little sh world where you just (00:51:05) infinitely walk around Middle Earth (00:51:07) basically and there's no there's no one (00:51:09) there yet but uh it's insane. You just (00:51:11) go anywhere. There's like dystopian (00:51:12) world. I'm just looking at all these (00:51:13) examples. (00:51:14) >> Yes. Uh, and my favorite part actually, (00:51:16) I don't know, I don't know if this is a (00:51:17) feature or bug, you can see like the (00:51:19) dots of the world before it actually (00:51:21) renders with all the textures. And I (00:51:23) just love to like you get a glimpse into (00:51:25) what is going on with this model. (00:51:26) basically create. (00:51:27) >> That's so cool to hear because this is (00:51:30) where as a researcher I I I'm learning (00:51:34) because the the the the dots that lead (00:51:37) you into the world was a an intentional (00:51:42) feature uh visualization. It is not part (00:51:46) of the model. It's uh the model actually (00:51:48) just generates the world. We we were (00:51:51) trying to find a way to guide people (00:51:53) into the world and a number of engineers (00:51:56) uh worked on different versions but we (00:51:59) converged on the dot and so many people (00:52:02) you're not the only one told us how (00:52:04) delightful that experience is and it it (00:52:07) was really satisfying for us to hear (00:52:10) that this intentional visualization (00:52:13) feature that's not just the big hardcore (00:52:16) model actually has delighted our users. (00:52:19) >> Wow. So, you add that to make it more uh (00:52:22) like to have humans understand what's (00:52:24) going on more, get more delightful. Wow, (00:52:26) that is hilarious. It makes me think (00:52:28) about LM and the way they it's not the (00:52:30) same thing, but they talk about what (00:52:31) they're thinking and what they're doing. (00:52:33) >> Yes, it is. It is. (00:52:35) >> It also makes me think about just the (00:52:36) Matrix. Like, it's exactly the Matrix (00:52:39) experience. I don't know if that was (00:52:40) your inspiration. (00:52:42) >> Um, well, like I said, a number of (00:52:43) engineers worked on that. It could be (00:52:45) their inspiration. It's in their It's in (00:52:48) their uh It's in their subconscious. (00:52:50) >> Yeah. (00:52:51) >> Okay. So, just for folks that may want (00:52:52) to play around with this, maybe use it. (00:52:54) What's like what are some applications (00:52:55) today that folks can start using today? (00:52:57) What's what's your goal with this (00:52:59) launch? (00:53:00) >> Yeah. So, um we do believe that world (00:53:03) modeling is very horizontal, but we're (00:53:05) already seeing some really exciting uh (00:53:08) use cases. virtual production for movies (00:53:11) because what they need are 3D uh worlds (00:53:16) that they can align with the camera so (00:53:18) when the actors are acting on it uh they (00:53:22) can you know they can uh position the (00:53:24) camera and shoot the the segments really (00:53:27) well and uh we're already seeing um (00:53:30) incredible use in fact I don't know if (00:53:34) you have seen our launch video showing (00:53:36) marble it was produced by a virtual uh (00:53:40) production company. We we collaborated (00:53:42) with Sony and they use marble things to (00:53:45) shoot those videos. So our we were (00:53:48) collaborating with those uh uh technical (00:53:50) artists and directors and they were (00:53:52) saying this has cut our uh production (00:53:55) time by uh 40x. (00:53:58) In fact it has tox. (00:54:00) >> Yes. In fact, I had to because we only (00:54:02) had one month to work on this project (00:54:05) and and there were so many things they (00:54:08) were trying to shoot. So, so using (00:54:10) marble really really significantly (00:54:13) accelerated the production of virtual (00:54:16) virtual production for VFX and movies. (00:54:19) That's one use cases. We are already (00:54:22) seeing our users putting uh taking our (00:54:25) marble scene and taking the mesh export (00:54:28) and putting games you know whether it's (00:54:30) games on VR or games uh just just just (00:54:33) fun games that they they have developed (00:54:36) we have had um we were showing uh an (00:54:40) example of uh robotic simulation because (00:54:44) uh when I was I mean I'm still am a (00:54:48) researcher doing robotic uh training. (00:54:52) One of the biggest pain point is to (00:54:54) create synthetic data for training (00:54:56) robots. And these synthetic data needs (00:54:58) to be very diverse. They need to come (00:55:00) from different environments with (00:55:02) different objects to manipulate. And uh (00:55:05) and one path to it is is to ask uh (00:55:09) computers to simulate. Otherwise, humans (00:55:12) have to, you know, (00:55:14) build every single asset for robots. (00:55:17) That that's just going to take a lot (00:55:19) longer. So we already have researchers (00:55:22) reaching out and wanting to use marble (00:55:24) to create those synthetic environments. (00:55:26) We also have unexpected um user uh (00:55:31) outreach in terms of uh how they want to (00:55:35) use marble. For example, a psychologist (00:55:39) team called us to use marble to do (00:55:42) psychology research. It turned out some (00:55:45) of the psychiatric patients they study, (00:55:48) they need to understand how their brain (00:55:51) respond to different immersive scenes of (00:55:55) different features. Uh, for example, (00:55:57) messy scenes or clean scenes or or (00:56:00) whatever you name it. And it's very hard (00:56:03) for researchers to get their hands on um (00:56:06) these kind of immersive scenes. and it (00:56:08) will take them too long and too much (00:56:11) budget to uh to to create. And Marble is (00:56:16) a really almost instantaneous way of (00:56:20) getting so many of these um experimental (00:56:23) uh environments into their hands. So, (00:56:26) we're seeing um uh we're seeing multiple (00:56:29) use cases at this point, but the the (00:56:32) VFX, the game developers, the simulation (00:56:35) uh uh developers as well as designers (00:56:38) are very excited. (00:56:39) >> This is very much the way things work in (00:56:41) AI. I've had other AI leaders on the (00:56:43) podcast and it's always like put things (00:56:45) out there early as soon as you can to (00:56:47) discover where the big use cases are. (00:56:49) the head of CHAJBT told me how when they (00:56:51) first put out ChatJBT, he was just (00:56:53) scanning TikTok to see how people were (00:56:55) using it and all the things they were (00:56:56) talking about and that's what convinced (00:56:58) them where to lean in and and help them (00:57:00) see how people actually want to use it. (00:57:02) I love this last use case of like for (00:57:04) therapy. I'm just imagining like like (00:57:06) heights, people seeing dealing with (00:57:09) heights or snakes or spiders, which (00:57:12) >> it's amazing. A friend of mine last (00:57:14) night literally called me and talked (00:57:16) about his height scare and asked me if (00:57:19) marble should be used. That's amazing. (00:57:22) You went straight there. (00:57:23) >> That's, you know, cuz I'm imagining all (00:57:25) the like the exposure therapy uh stuff (00:57:28) like this could be so good for that. Uh (00:57:30) that is so cool. Okay, so let me I (00:57:32) should have asked you this before, but I (00:57:33) think there's a qu there's going to be a (00:57:35) question of just how does this differ (00:57:36) from things like V3 and other video (00:57:39) generation models? It's pretty clear to (00:57:41) me, but I think it might be helpful just (00:57:43) to explain how this different from all (00:57:44) the video AI tools people have seen. (00:57:47) >> Wordlab's thesis is that spatial (00:57:49) intelligence is fundamentally very (00:57:51) important and spatial intelligence is (00:57:53) not just uh uh it's not just about (00:57:58) videos. In fact, the world is not (00:58:00) passively watching videos passing by, (00:58:04) right? Um I I love uh Plato has the (00:58:08) allegory of the cave analogy uh to (00:58:12) describe vision. He said that imagine a (00:58:15) prisoner tied on his chair uh not not (00:58:19) very uh humane but um uh in in a cave uh (00:58:24) watching a full life theater uh on the (00:58:29) in front of him. But but the actual live (00:58:32) theater that actors are acting is behind (00:58:35) his back. It was just lit so that the (00:58:39) projection of the the uh the action is (00:58:42) on a on a wall of the cave and and then (00:58:46) the goal the the task of this prisoner (00:58:49) is to figure out what's going on. It's a (00:58:51) pretty extreme example, but it really (00:58:54) shows it describes what vision is about. (00:59:00) is that to make sense of the 3D world or (00:59:03) 4D world out of 2D. So spatial (00:59:07) intelligence to me is deeper than owning (00:59:11) creating that flat 2D world. Spatial (00:59:15) intelligence to me is the ability to (00:59:20) create, reason, interact, make sense of (00:59:25) deeply spatial world, whether it's 2D or (00:59:29) 3D or 4D, including dynamics and all (00:59:32) that. So, so World Lab is focusing on (00:59:35) that. And of course, um the ability to (00:59:38) create videos per se, could be part of (00:59:41) this. And in fact uh just a couple of (00:59:44) weeks ago we rolled out the world's (00:59:46) first uh realtime (00:59:48) demoable realtime video generation on a (00:59:52) single uh H100 GPU. So we we we part of (00:59:56) our technology includes that. But I (00:59:59) think Marvel is very different because (01:00:01) we really want creators, designers, (01:00:06) developers to have in their hands a (01:00:10) model that can give them uh worlds with (01:00:14) 3D structure so they can use it for for (01:00:17) their work. And that's where that's why (01:00:20) Marble is so different. (01:00:22) >> The way I see it is it's a it's a (01:00:23) platform for a ton of opportunity to do (01:00:26) stuff. uh as you described videos are (01:00:29) just like here's a oneoff video that's (01:00:30) very fun and cool and you could and (01:00:32) that's it and that's it and you move on. (01:00:33) >> By the way, we could in Marble we could (01:00:36) allow people to export in video form. So (01:00:39) you could actually, like you said, you (01:00:41) go into a world. So So let's say it's a (01:00:44) hobbit uh cave, you can actually, (01:00:47) especially as a creator, you have such a (01:00:50) uh specific way of uh uh moving the (01:00:54) camera in a trajectory in the director's (01:00:57) mind, right? And then you can export (01:00:59) that uh from Marble into a video. (01:01:02) >> What does it take to create something (01:01:03) like this? Just like how big is the (01:01:05) team? How many how many GPUs you (01:01:07) working? Like anything you can share (01:01:08) there? I don't know how much of this is (01:01:09) private information, but just what does (01:01:10) it take to create something like this (01:01:12) that you've launched here? (01:01:13) >> It takes a lot of brain power. (01:01:16) So, we just talk about 20 watts per (01:01:20) brain. It's uh so from that point of (01:01:22) view, it's it's a small number, but but (01:01:25) it's actually an incredible, you know, (01:01:27) it's a half billion years of evolution (01:01:30) to get give us those power. Um we have a (01:01:34) team of 30ish people now and uh we are (01:01:39) predominantly (01:01:40) uh researchers and research engineers (01:01:44) and uh but we also have designers and (01:01:47) and product. We we actually really (01:01:50) believe that we want to create a company (01:01:52) that's anchored in the deep tech of (01:01:56) spatial intelligence but uh we we we are (01:02:00) actually building serious products. Um (01:02:04) so so we have we have this uh (01:02:07) integration of R&D and productization (01:02:11) and of course we use you know a ton of (01:02:14) GPUs. That's a that's the technical (01:02:17) >> I'm so happy to hear. (01:02:20) >> Well, congrats on the launch. I know (01:02:21) this is a huge milestone. I know this (01:02:23) took a ton of work. So, I just want to (01:02:24) say congrats to you and your team. (01:02:26) >> Let me talk about your founder journey (01:02:28) for a moment. So, you're a founder of (01:02:30) this company. You started how many years (01:02:32) ago? Couple years ago, two, three years (01:02:33) ago. (01:02:33) >> Oh, a year ago. A year ago. (01:02:36) >> A year. Okay. (01:02:37) >> 18 month. Yeah. (01:02:39) >> Okay. What's something you wish you knew (01:02:41) before you started this that you wish (01:02:42) you could like whisper into the ear of (01:02:44) Fay of 18 months ago? (01:02:46) >> Well, I continue to wish I know (01:02:51) the future of technology. I think (01:02:53) actually that's one of our founding (01:02:55) advantage is that we see the future (01:02:59) earlier in general than than most (01:03:01) people. But still, man, this is so (01:03:03) exciting and so uh amazing that that (01:03:07) what's unknown and what's coming. But I (01:03:10) know the reason you're asking me this (01:03:12) question is not about the future of (01:03:14) technology. You're probably more, you (01:03:16) know, look, I I did not start a company (01:03:20) of this scale (01:03:23) at 20 year old. So, you know, I started (01:03:26) a dry cleaner when I was 19, but that's (01:03:29) a little smaller scale. we got to talk (01:03:31) about that (01:03:31) >> and and then I you know um founded (01:03:34) Google Cloud AI and then I founded an (01:03:37) institute at Stanford but those are (01:03:39) different beasts. I did feel I was a (01:03:43) little more prepared as a a founder of (01:03:47) the the grinding journey that um that I (01:03:51) um compared to maybe um maybe the the (01:03:55) the 20 year old founders. But I still (01:04:00) I'm surprised and and and uh it puts me (01:04:04) into paranoia sometimes that how (01:04:08) intensely competitive uh AI landscape is (01:04:13) from (01:04:15) from the model the technology itself as (01:04:18) well as talents. And you know when I (01:04:21) founded the company um we did not have (01:04:25) these incredible stories of how much (01:04:28) certain talents would cost you know um (01:04:32) so these are things that continue to (01:04:34) surprise me and uh and I have to be very (01:04:38) alert about. (01:04:40) >> So the competition you're talking about (01:04:41) is yeah the competition for talent the (01:04:44) speed at which how things are moving. (01:04:46) >> Yeah. (01:04:47) >> Yeah. you mentioned this point that I (01:04:49) want to come back to that you if you (01:04:51) just look over the course of your (01:04:53) career. You were like at all of the (01:04:55) major uh collections of humans that led (01:04:59) to so many of the breakthroughs that are (01:05:01) happening today. Obviously we talked (01:05:02) about Imageet also just sale at Stanford (01:05:04) is where a lot of the work happened at (01:05:07) Google cloud which a lot of the (01:05:08) breakthroughs happened. What brought you (01:05:10) to those places? uh like for people (01:05:13) looking for how to advance in their (01:05:16) career, be at the center of the future, (01:05:18) just like is there a throughine there of (01:05:19) just what pulled you from place to place (01:05:22) and pulled you into those groups that (01:05:24) might be helpful for people to hear? (01:05:26) >> Yeah, this is actually a great question, (01:05:28) Lenny, because I do think about it and (01:05:30) uh (01:05:32) obviously we talked about it curiosity (01:05:35) and passion that brought me to AI. That (01:05:37) is more a scientific northstar, right? I (01:05:40) did not care if AI was a thing or not. (01:05:44) So, so that was one part. But how did I (01:05:47) end up choosing (01:05:49) um in the particular places I work in (01:05:52) including starting world labs is (01:05:57) I think I'm very grateful (01:06:00) to myself or maybe to my parents' jeans. (01:06:05) I'm I'm an intellectually very fearless (01:06:08) person and I have to say when I hire (01:06:11) young people I look for that because I (01:06:15) um (01:06:16) I think that's a very important quality (01:06:19) if one wants to make a difference is (01:06:22) that when you want to make a difference (01:06:25) you have to accept that you're creating (01:06:29) something new or you're diving into (01:06:31) something new. people haven't done that. (01:06:33) And if you have that self-awareness, you (01:06:37) almost have to allow yourself to be (01:06:40) fearless and to be courageous. So when I (01:06:44) uh for example um came to Stanford, you (01:06:49) know, in the world of academia, (01:06:52) I was very close to this thing called (01:06:55) tenure um which is, you know, have the (01:06:58) job forever in in at Princeton. But I (01:07:03) I choose to chose to come to Stanford (01:07:06) because I love Princeton. It's my alma (01:07:08) mater. It's just at that moment there (01:07:12) are people who are so amazing at (01:07:14) Stanford and the Silicon Valley (01:07:16) ecosystem was so amazing that I was okay (01:07:21) to take a risk of restarting my tenure (01:07:24) clock. (01:07:25) um going to um becoming the first uh (01:07:30) female director of sale. I was actually (01:07:34) relatively speaking a very young faculty (01:07:36) at that time and I wanted to do that (01:07:40) because I care about that community. I (01:07:42) didn't spend too much time thinking (01:07:44) about all the failure cases. Obviously, (01:07:47) I was very lucky that the more senior (01:07:50) faculty supported me, but I just wanted (01:07:52) to make a difference. And then going to (01:07:55) Google was similar. I wanted to work (01:07:58) with people like Jeff Dean, Jeff Hinton, (01:08:02) and um all these incredible Dennis, the (01:08:06) the incredible people. Um (01:08:10) I you know, so so the same with World (01:08:13) Labs. I I I have this passion and I also (01:08:18) believe that people with the same (01:08:21) mission can do incredible things. So (01:08:23) that's how it guided my through through (01:08:26) life. I don't overink (01:08:29) of all possible things that can go wrong (01:08:32) because that's too many. (01:08:34) >> I feel like that's an important element (01:08:35) of this is not focusing on the downside, (01:08:38) focusing more on the people, the (01:08:40) mission. What gets you excited? What do (01:08:42) you think? Uh I do yeah I do want to say (01:08:45) one thing to all the young talents in AI (01:08:48) the engineers the researchers out there (01:08:50) because some of you apply to world labs. (01:08:53) I I feel very privileged you considered (01:08:55) world labs. I do find many of the young (01:08:58) people today (01:09:00) think about every single (01:09:05) aspect of a equation when they decide on (01:09:08) jobs at some point. Maybe, you know, (01:09:10) maybe maybe that's the way they want to (01:09:13) do it. But sometimes I do want to (01:09:14) encourage young people to focus on (01:09:17) what's important because I find myself (01:09:21) um constantly in mentoring mode when I (01:09:25) talk to job job candidates. Not (01:09:28) necessarily recruiting or not (01:09:29) recruiting, but just in mentoring mode. (01:09:32) When I see an incredible young talent (01:09:34) who is overfocusing on every minute (01:09:39) dimension and aspect of considering a (01:09:42) job when (01:09:45) when maybe the most important thing is (01:09:49) where's your passion? Do you align with (01:09:51) the mission? Do you believe and have (01:09:54) faith in this team? (01:09:56) and and just just focus on the impact (01:09:59) and and you can make and the kind of (01:10:02) work and team you can you can work with. (01:10:05) >> Yeah, it's tough. It's tough for people (01:10:06) in the AI space now. There's so much so (01:10:09) much at them, so much news, so much (01:10:10) happening, so much FOMO. (01:10:11) >> That's true. (01:10:12) >> I could see the stress. And so, I think (01:10:14) that advice is really important. Just (01:10:15) like what will actually make you feel (01:10:18) fulfilled in what you're doing, not just (01:10:19) where's the fastest growing company? (01:10:21) Where's the who's going to win? I don't (01:10:23) know. I want to make sure I ask you (01:10:25) about the work you're doing today at (01:10:26) Stanford at the HCI. I think it's HAI (01:10:30) human centered AI institute. (01:10:32) >> What are you what are you doing there? I (01:10:34) know this is a thing you do on the site (01:10:35) still. (01:10:37) >> So yes, I HAI human center AI institute (01:10:41) was co-founded by me and a group of (01:10:44) faculty like uh professor John Hendy, (01:10:47) professor James Landy, um professor (01:10:50) Chris Manning back in 2018. I was (01:10:54) actually finishing my last the last (01:10:56) sabbatical at Google. Um and uh it was a (01:11:01) very very important decision for me (01:11:04) because I could have stayed in industry (01:11:07) but my time at Google taught me one (01:11:10) thing is AI is going to be a (01:11:12) civilizational technology and it it's it (01:11:16) dawned on me how important this is to (01:11:18) humanity to the point that I actually (01:11:21) wrote a piece in New York Times that (01:11:23) year 2018 to talk about the need for a (01:11:28) guiding framework to develop and to (01:11:32) to apply AI and that framework has to be (01:11:35) anchored in human benevolence is human (01:11:38) centerness and I felt that Stanford uh (01:11:42) one of the world's top university in the (01:11:46) heart of Silicon Valley that gave birth (01:11:48) to important companies from Nvidia to (01:11:51) Google uh should um be a thought leader (01:11:57) uh to create this human- centered AI (01:12:00) framework and to um to actually embody (01:12:04) that in our research education and (01:12:07) policy and in ecosystem work. So I (01:12:11) founded HAI it uh you know after uh fast (01:12:15) forward after six seven years it has (01:12:18) become the world's largest AI institute (01:12:21) that does human- centered um uh research (01:12:26) education uh ecosystem outreach and (01:12:30) policy uh in uh in uh impact. Uh it (01:12:35) involves hundreds of faculty across all (01:12:39) eight schools at Stanford from medicine (01:12:42) to education to sustainability to (01:12:45) business to engineering to humanities to (01:12:48) uh law and uh we we support researchers (01:12:54) especially at the interdisciplinary area (01:12:57) from digital economy to uh legal studies (01:13:01) to political science to discovery of new (01:13:04) drugs. (01:13:05) uh to to new algorithms to that's beyond (01:13:09) transformers. We also actually put a (01:13:12) very strong focus on um on policy (01:13:16) because when we started HAI I realized (01:13:19) that Silicon Valley did not talk to (01:13:22) Washington DC and or Brussels or other (01:13:27) parts of the world and it's re given how (01:13:30) important this this technology is we (01:13:33) need to bring everybody on board. So we (01:13:36) created multiple programs from (01:13:38) congressional boot camp to um AI index (01:13:43) report to policy briefing and we (01:13:47) especially (01:13:49) uh participated in policym including um (01:13:53) advocating for a u a national AI (01:13:56) research cloud bill that was passed in (01:13:59) the first Trump administration and (01:14:02) participate participating in state level (01:14:05) uh regulatory AI discussions. So there's (01:14:09) a lot we did and and I continue to be um (01:14:13) one of the the leaders even though I'm (01:14:16) much less involved operationally (01:14:19) because I care not only we create this (01:14:22) technology but we use it in the right (01:14:24) way. (01:14:24) >> Wow. I was not aware of all that other (01:14:26) work you were doing. Uh, as you were (01:14:28) talking, I was reminded Charlie Mer had (01:14:31) this quote, take a simple idea and take (01:14:33) it very seriously. I feel like you've (01:14:36) done that in so many different ways and (01:14:38) and stayed with it and it's unbelievable (01:14:41) the impact that you've had in so many (01:14:42) ways over the years. I'm going to skip (01:14:45) the lightning round and I'm just going (01:14:46) to ask you one last question. Is there (01:14:48) anything else that you wanted to share? (01:14:50) Anything else you want to leave (01:14:51) listeners with? (01:14:52) >> I I'm very excited by AI Lenny. Uh I (01:14:56) want to answer one question that I when (01:14:59) I travel around the world everybody asks (01:15:02) me is that if I'm a musician, if I'm a (01:15:07) teacher, middle school teacher, if I'm a (01:15:10) nurse, if I'm an accountant, if I'm a (01:15:14) farmer, do I have a role in AI or is AI (01:15:18) just going to take over my life or my (01:15:20) work? And I think this is the most (01:15:24) important question of AI. And I find (01:15:27) that in Silicon Valley, we tend not to (01:15:31) speak heart-to-heart with people with (01:15:35) people like us and and not like us in (01:15:37) Silicon Valley, but like all of us, we (01:15:40) tend to just toss around words like (01:15:43) infinite productivity or infinite (01:15:47) leisure time or or you know, infinite (01:15:52) power or whatever. But at the end of the (01:15:55) day, AI is about people. And when people (01:15:58) ask me that question, it's a resounding (01:16:00) yes. Everybody has a role in AI. It (01:16:04) depends on what what you do and what you (01:16:07) want. But no technology should take away (01:16:10) human dignity and the human dignity and (01:16:14) agency should be at the heart of the (01:16:17) development, the deployment as well as (01:16:20) the governance of every technology. So (01:16:24) if you are a young artist (01:16:27) and your passion is storytelling, (01:16:31) uh, embrace AI as a tool. In fact, (01:16:34) embrace Marvel. I hope it becomes a tool (01:16:36) for you. Um, because the way you tell (01:16:40) your story is unique and this the world (01:16:43) still needs it. But how you tell your (01:16:46) story, how do you use the most (01:16:49) incredible tool to tell your story in (01:16:52) the most unique way is important and (01:16:55) that that voice needs to be heard. If (01:16:58) you're a farmer near retirement, AI (01:17:02) still matters because you're a citizen. (01:17:06) You can participate in your community. (01:17:08) You should have a voice in how AI is (01:17:11) used, how AI is applied. you you work (01:17:15) with people that you can you know (01:17:18) encourage all of all of you to use AI uh (01:17:22) to make life easier for you. If you're a (01:17:26) nurse, I hope you know that at least in (01:17:29) my uh career, I have worked so much in (01:17:34) healthc care research because I feel our (01:17:36) health care workers should be greatly (01:17:40) augmented and helped by AI technology (01:17:43) whether it's smart cameras to feed more (01:17:47) uh in information or robotic assistance (01:17:50) because our nurses are overworked, over (01:17:54) fatigued And as our society ages, we (01:17:58) need more help for for people to be (01:18:00) taken care of. So AI can play that role. (01:18:03) So I just want to say that it's so (01:18:06) important that um even a technologist (01:18:10) like me um are sincere about that (01:18:15) everybody has a role in AI. (01:18:17) >> What a beautiful way to end it. Such a (01:18:19) tie back to where we started about how (01:18:21) it's up to us and take individual (01:18:24) responsibility for what AI will do in (01:18:26) our lives. Final question, where can (01:18:28) folks find Marble? Where can they go? (01:18:30) Maybe uh try to join uh World Labs if (01:18:32) they want to. What's the website? Where (01:18:34) do people go? (01:18:35) >> Well, World Labs website is (01:18:38) www.worldlabs.ai (01:18:41) and you can find um you can find our (01:18:45) research progress there. We we have (01:18:47) technical blogs. You can find Marble the (01:18:50) product there. You can sign in there. (01:18:52) You can find our job posts uh link (01:18:55) there. You can uh you know, we're in San (01:18:58) Francisco. We love to work with the (01:19:00) world's best talents. (01:19:02) >> Amazing. Fay, thank you so much for (01:19:04) being here. (01:19:05) >> Thank you, Lenny. (01:19:06) >> Bye, everyone. (01:19:10) Thank you so much for listening. If you (01:19:11) found this valuable, you can subscribe (01:19:13) to the show on Apple Podcasts, Spotify, (01:19:15) or your favorite podcast app. Also, (01:19:18) please consider giving us a rating or (01:19:20) leaving a review as that really helps (01:19:22) other listeners find the podcast. You (01:19:24) can find all past episodes or learn more (01:19:26) about the show at lennispodcast.com. (01:19:29) See you in the next episode.