[ { "start": 0.743, "end": 1.185, "text": "Thank you, chris." }, { "start": 1.742, "end": 3.633, "text": "Then nothing goes wrong." }, { "start": 3.714, "end": 9.732, "text": "No, hopefully App is not focused." }, { "start": 10.183, "end": 10.605, "text": "Show it to me." }, { "start": 13.924, "end": 16.786, "text": "Ah, now she's back, now she's the focus." }, { "start": 17.608, "end": 19.442, "text": "Let's do it Alright." }, { "start": 19.843, "end": 21.649, "text": "Recording, So recording." }, { "start": 22.041, "end": 22.784, "text": "Let's get it started." }, { "start": 26.692, "end": 28.7, "text": "How do you pronounce your name, Richard Zocker?" }, { "start": 29.643, "end": 37.138, "text": "Richard Zocker, if it's English and social, if you want, you can do it properly." }, { "start": 37.279, "end": 40.168, "text": "Richard Zocker, we say it with O and O and Tote." }, { "start": 41.501, "end": 46.992, "text": "Maybe I use the German version, as I'm also German and my listeners most likely know that." }, { "start": 48.022, "end": 49.739, "text": "So, welcome to the Alpha List podcast." }, { "start": 50.602, "end": 56.513, "text": "I am your host, tobi, and today I'm here with pronounced in German, richard Zocker." }, { "start": 56.86, "end": 64.8, "text": "And Richard Zocker is the founder and CEO of Ucom, the AI search engine that puts you in control." }, { "start": 65.221, "end": 89.488, "text": "That's why he has this great domain, and they raised 25 million, i think, in series A funding last year, and Richard actually previously served as chief scientist at Salesforce and sold his old company, MetaMind, to Salesforce, and I think you personally reported to Mark." }, { "start": 89.628, "end": 90.37, "text": "Is that correct?" }, { "start": 91.523, "end": 92.366, "text": "I worked with him directly." }, { "start": 92.82, "end": 96.91, "text": "There's one person between us officially who runs all of technology at Salesforce Srinni." }, { "start": 98.021, "end": 100.649, "text": "Okay, reporting lines are not the most important thing to touch." }, { "start": 101.523, "end": 102.527, "text": "So what was that all correct?" }, { "start": 102.86, "end": 107.552, "text": "And I think you were a PhD, a computer science PhD, from Stanford." }, { "start": 108.042, "end": 109.572, "text": "So you are indeed a nerd." }, { "start": 109.592, "end": 110.719, "text": "That's right." }, { "start": 111.682, "end": 115.171, "text": "I used to teach at Stanford for four years when she has an adjunct professor." }, { "start": 116.682, "end": 122.416, "text": "And just one addition we raised 25 an hour A and 20 an hour C ground earlier." }, { "start": 122.737, "end": 132.286, "text": "Okay, great, and you're right now sitting in your childhood room at home in Dresden, so you're also German." }, { "start": 133.903, "end": 136.19, "text": "And when did you move over?" }, { "start": 136.27, "end": 136.711, "text": "Like, why?" }, { "start": 137.262, "end": 138.346, "text": "Like, maybe we start there." }, { "start": 138.84, "end": 139.522, "text": "Yeah, yeah." }, { "start": 139.562, "end": 151.494, "text": "Originally I just wanted to work with the absolute best AI people When I was a PhD during my undergrad, already starting 2003, i studied linguistic, computer science and lipstick." }, { "start": 151.8, "end": 156.266, "text": "I love languages and math and programming And so I got into that." }, { "start": 156.98, "end": 175.671, "text": "And then I switched in my masters to computer vision but discovered what at the time was called statistical machine learning and pattern recognition And fell in love with it and felt like, wow, man, if you can really get good at statistical machine learning, it seems like such an amazing tool that could do so many different things." }, { "start": 175.9, "end": 188.8, "text": "I wanted to nerd out a lot about statistics and probability and linear algebra And then during my PhD in the early days I essentially I wanted to work with the top people." }, { "start": 189.281, "end": 202.8, "text": "So I went to the US and eventually got into Stanford after a few rejection rounds earlier And just yeah, discovered at Stanford neural nets, rediscovered them." }, { "start": 202.94, "end": 230.651, "text": "You know they're kind of one of those methods that no one really loved for a long time And yeah, anyway, without going into too many details, i found that a lot of AI at the time was kind of graduate student descent of defining and designing features for your AI problem, and then I thought wouldn't be cooler if you could really learn everything from scratch, from raw data, and that's kind of what we ended up doing with neural nets for natural language processing." }, { "start": 232.488, "end": 233.916, "text": "Did you meet Sebastian there as well?" }, { "start": 233.956, "end": 234.8, "text": "He's also Stanford right." }, { "start": 235.564, "end": 236.046, "text": "I did yeah." }, { "start": 236.066, "end": 237.071, "text": "Sebastian Thrun." }, { "start": 237.172, "end": 238.8, "text": "Thrun, that's right." }, { "start": 240.207, "end": 242.76, "text": "Yeah, we actually communicate quite a lot these days, okay cool." }, { "start": 243.301, "end": 244.964, "text": "And yeah, he's a good friend of ours." }, { "start": 245.444, "end": 267.552, "text": "We hang out together at our ranch in the Bay Area often now And, yeah, we actually did overlap during the heyday of the AI, the Stanford AI lab, when they're still and during full time there, and Daphne Koehler and Sebastian and Chris Manning, you know some of those folks are still there and some of them have mostly left or don't spend as much time at the university anymore." }, { "start": 268.88, "end": 269.361, "text": "Okay, cool." }, { "start": 269.742, "end": 276.239, "text": "Is there like a detail in your personal life when you actually got attracted to computers and why?" }, { "start": 277.06, "end": 284.802, "text": "Like, i mean, you started off with the machine learning bit and statistics, but I mean, why Yeah?" }, { "start": 285.783, "end": 287.446, "text": "I guess in the very early days." }, { "start": 288.788, "end": 297.771, "text": "You know, like the first one was probably just computer games, like with many kids You know, you have like we had a very early days of Commodore, commodore 64." }, { "start": 298.98, "end": 300.602, "text": "And now sort of my first computer." }, { "start": 300.642, "end": 314.226, "text": "And then I really only got into programming in the early days through web programming, because I all of a sudden felt like oh, wow, wait, if I can program something here, everyone can see it, my friends can immediately try it out." }, { "start": 314.94, "end": 328.8, "text": "So I got into some some more sort of nifty JavaScript programming, like things that can move around in the browser and you can link tags, you know, yeah, php and getting like databases going, and it just seemed very powerful." }, { "start": 329.181, "end": 348.509, "text": "I worked for my mom's company and it's a real estate company, so I helped her kind of set up a database so she can have different, you know, listings of apartments and houses and stuff that she was selling in her real estate company And implemented that website and I got more into programming." }, { "start": 348.57, "end": 373.508, "text": "Actually, only once I could more easily share it as a kid, during high school, and then, yeah, and then during, you know, obviously, once I started studying, i got more and more into it, but I wasn't, you know, to be honest, i felt like, in comparison to some people, i was a late bloomer, like there's some people who were like I started coding when I was 10 years old or something, and I really only got into it at like 16." }, { "start": 375.382, "end": 377.443, "text": "Okay, 16 or so, okay, cool." }, { "start": 379.141, "end": 381.108, "text": "And what did she do with Metamind then?" }, { "start": 382.201, "end": 386.047, "text": "So Metamind basically started at the end of my PhD." }, { "start": 386.407, "end": 386.828, "text": "You know I was." }, { "start": 387.389, "end": 396.826, "text": "I finally got an EB1A visa in the US It's kind of called Person of Extraordinary Ability And before that I was on an F1 student visa." }, { "start": 396.866, "end": 397.508, "text": "So I couldn't work." }, { "start": 398.86, "end": 403.8, "text": "And once I got that visa I basically had a bunch of folks asking me hey, can you join us full time?" }, { "start": 404.061, "end": 405.547, "text": "I'm like, no, i kind of want to be a professor." }, { "start": 406.623, "end": 410.857, "text": "Then they said, well, can you like consult for us or just do a project?" }, { "start": 410.917, "end": 415.384, "text": "And so I did some projects And actually one of those projects was kind of funny." }, { "start": 415.424, "end": 429.611, "text": "The very first project ever at the end of my PhD that I got paid for was this guy who wanted me to build him a very fancy autocomplete That could autocomplete a lot more code, for instance." }, { "start": 430.361, "end": 446.348, "text": "And so I basically took a recurrent neural net and I trained it a little bit and I said here, this is probably like your solution and you're going to want to just get a ton more training data into this and make it a lot bigger, but this is kind of what you're going to want to use." }, { "start": 447.321, "end": 449.407, "text": "And it's interesting because that is basically like GitHub." }, { "start": 449.428, "end": 453.76, "text": "Co-pilot now is large neural nets and autocomplete code." }, { "start": 456.665, "end": 461.594, "text": "But here's this little code snippet with some first version that does something." }, { "start": 462.661, "end": 463.282, "text": "It turns out." }, { "start": 463.742, "end": 471.572, "text": "I guess it was another couple of years and a couple billion more parameters than the model that I had that I sent over." }, { "start": 472.934, "end": 474.356, "text": "That really got it to work." }, { "start": 474.88, "end": 500.596, "text": "But yeah, so it felt like, well, you know, you're only at the time a handful of companies that could attract some of these top people that build neural nets at the time still from scratch, you know, like C++, python was barely coming up, and I felt like this is technology that could be useful for a lot more use cases and it felt like I didn't just want to go one, i still want to go into academia." }, { "start": 500.676, "end": 516.228, "text": "But then even said well, i'll postpone my faculty job for a year and work essentially on this company, metamind, that would give neural network technology to a lot of other organizations and make it very, very easy to train." }, { "start": 516.34, "end": 534.942, "text": "So we basically allowed you to just drag and drop some images into the web browser or text documents into the web browser and then immediately back out comes a neural network with like three lines of Python code or some curl command and you can just run and have a functioning system." }, { "start": 535.263, "end": 545.026, "text": "And that worked out pretty well And eventually we actually got acquired after less than two years by Salesforce, where I found that man." }, { "start": 545.046, "end": 555.948, "text": "It would have taken us a long time to scale that technology in terms of sales not in terms of technology, but in terms of really getting it into the hands of a lot of people took a lot longer." }, { "start": 556.561, "end": 558.387, "text": "And then during that time I also decided, maybe." }, { "start": 559.02, "end": 560.867, "text": "I started teaching on the side at Stanford." }, { "start": 561.66, "end": 567.727, "text": "We did a little bit of research in a startup and I realized, oh, teaching is fun, but maybe three months of teaching a year is enough." }, { "start": 568.78, "end": 569.905, "text": "It does get a little repetitive." }, { "start": 570.16, "end": 572.988, "text": "And so I ended up not becoming a full-time faculty." }, { "start": 574.02, "end": 577.023, "text": "And then at Salesforce we scaled the research team massively there." }, { "start": 577.043, "end": 580.266, "text": "That was super fun and we had a ton of impact on real product." }, { "start": 580.326, "end": 586.372, "text": "So Right, right, like Salesforce has quite a sales engine, i guess That's right, that's right, no pun intended." }, { "start": 586.58, "end": 587.906, "text": "Yeah, there is a force behind that." }, { "start": 588.26, "end": 602.386, "text": "Yeah, I can imagine How much like, if you look back, i guess, when you started off, infrastructure was like still a key problem right For machine learning." }, { "start": 602.88, "end": 615.99, "text": "I mean, these days it's kind of easy, but back then, like the term MLOps wasn't really invented I guess, and it was hard to bring it live I guess." }, { "start": 617.021, "end": 617.443, "text": "That's right." }, { "start": 617.483, "end": 624.968, "text": "Yeah, i mean, in the early days of MetaMind, there's literally no PyTorch or TensorFlow yet, and so it took a long time to like." }, { "start": 625.08, "end": 633.142, "text": "You had to in some cases, do some of the math yourself to like figure out your gradients and back propagation and so on." }, { "start": 634.08, "end": 647.168, "text": "So that's where it started And we build our own frameworks, try to make them fast and be deep down in CUDA kernels and C++ with the team and it was like not, it's not very easy, and of course, this has gotten so easy." }, { "start": 647.24, "end": 654.846, "text": "In fact, we call it the software 2.0 stack now, a little bit after Andre Kapathy's blog post and this new software 2.0 stack." }, { "start": 655.36, "end": 695.166, "text": "Things of software vary differently, right, there's data set collection, then data cleaning, data labeling, then choosing the right AI model, implementing it possibly from papers or taking an open source one from Huggingface, potentially fine tuning it of things like weights and biases and full disclosure I'm an investor in both of those companies And then running it, watching out for distributional shift and changes over time, which a company like Kulina would do, and just like keeping your data sets clean and easily extractable." }, { "start": 695.76, "end": 700.271, "text": "So having nowadays LM ops is like another level of ML ops." }, { "start": 702.344, "end": 719.908, "text": "And there's great companies like vector databases, like Chroma or data sets, companies like Active Loops and ML tools, and so yeah, so lots of interesting like Xenom L is another one of these ML ops companies that is open source and, again, full disclosure." }, { "start": 720.482, "end": 722.207, "text": "I have an investor in all of those companies." }, { "start": 723.14, "end": 725.849, "text": "And so now through AIX, our venture fund." }, { "start": 726.02, "end": 732.967, "text": "But long story short, there's so many tools and in some ways, metamind was almost a little too early, which, as an academic, is great." }, { "start": 733.1, "end": 735.428, "text": "Just any like oh wow, it's so visionary, you know what's happening." }, { "start": 735.6, "end": 738.266, "text": "But as a startup founder, it's not that helpful to be too early." }, { "start": 738.42, "end": 740.468, "text": "You wanna be at the right time, right, right, right." }, { "start": 741.081, "end": 742.426, "text": "If you're too early, you can also die." }, { "start": 742.94, "end": 746.763, "text": "The startups need to be just at the right time with the right technology." }, { "start": 746.864, "end": 748.189, "text": "I'm like academia." }, { "start": 748.4, "end": 751.25, "text": "And so, long story short, you're 100% right." }, { "start": 751.34, "end": 759.768, "text": "There's so many tools now and each of these different aspects of the new software 2.0 stack has often a unicorn associated with it." }, { "start": 759.808, "end": 761.664, "text": "At this point, Yeah, and right now." }, { "start": 761.704, "end": 765.327, "text": "for me, it's always a question like is now the time to get into that stuff?" }, { "start": 767.362, "end": 777.026, "text": "or even like, stand on the sideline for a bit longer and just wait until it improves at 10 axis again, right, that's so weird." }, { "start": 779.445, "end": 789.169, "text": "But before we know it off more there, maybe you can tell us a bit more about what makes youcom special and why did you choose that direction?" }, { "start": 790.161, "end": 793.943, "text": "You wanted to compete with Google, or, yeah, you want?" }, { "start": 793.983, "end": 794.345, "text": "to still." }, { "start": 795.2, "end": 795.823, "text": "Do we want to?" }, { "start": 795.863, "end": 797.945, "text": "yeah, do we want to or do we have to?" }, { "start": 798.02, "end": 799.545, "text": "So at a very high level." }, { "start": 800.86, "end": 811.751, "text": "It started by seeing and being part of making natural language processing so much better And you see, wow, we can now have a single model in 2018." }, { "start": 812.58, "end": 837.085, "text": "My co-authors, brian McCann, sending in the tish we invented prompt engineering, where you could, as input to the model, ask any kind of question like what is the sentiment, what is the summary, what is the translation, and a single model would respond to you, and so that made it very clear to us that you could do a lot better than here's a list of blue links." }, { "start": 838.963, "end": 843.086, "text": "Yet you see that and then no one's doing it, and so that's clearly an opportunity." }, { "start": 843.56, "end": 879.088, "text": "So to tell you to answer your question, ucom is a chat first search engine that actually leverages all the latest AI And, as a user, if you're interested in oh, i want to be able to generate images, or I want to have AI write my essays, or I want AI write my code you see those features in a search engine context, first on ucom, and then maybe three, six months later, you see Google and Bing and others copy those capabilities, and by those times, we're usually already ahead again." }, { "start": 880.08, "end": 898.547, "text": "But it is a little bit frustrating because now the big guys are copying more and more of those features And their marketing of course says we're the first And we're like actually we've already launched a chat model for a search engine with citations last year in December And you're really just copying what we've done." }, { "start": 899.901, "end": 903.371, "text": "And what were your thoughts when chat GPT was released?" }, { "start": 903.48, "end": 907.367, "text": "Like, was it, ooh, this is good, or were you already there?" }, { "start": 908.824, "end": 909.667, "text": "Or like, what were you thinking?" }, { "start": 910.36, "end": 926.85, "text": "Yeah, of course we saw that earlier than most and, like new chat, actually our sort of search chat bot with citations and stuff came out, i think just two or three weeks after chat GPT And I think overall it's been a positive." }, { "start": 927.28, "end": 946.306, "text": "You know, it would have been even better if we had launched the first one and the only one, maybe, but I can tell you that before chat GPT, whenever we innovated too much in search, we had the majority of users come to us and say it's a little too different to Google." }, { "start": 946.38, "end": 947.704, "text": "And I'm so used to Google." }, { "start": 948.226, "end": 949.289, "text": "You can't be too different." }, { "start": 949.96, "end": 958.249, "text": "You know, we were the first to bring a Reddit app into search engines, like early last year So over a year ago now and that worked." }, { "start": 958.38, "end": 965.349, "text": "But when we innovated too far away from a list of blue links with some add-ons, users weren't ready for it." }, { "start": 965.86, "end": 969.19, "text": "But that, thanks to chat GPT, has massively changed." }, { "start": 969.74, "end": 985.211, "text": "People are now much more open to new ways to find information online, new ways to get things done, and if you think about a Venn diagram of chat and search, there's a large overlap, but it's also even bigger." }, { "start": 985.26, "end": 1001.583, "text": "So chat is even bigger and can be even bigger than search, because the truth is you wouldn't ask Google, write me an HTML website with a commenting function in JavaScript, or write me this front-end code, or train in a neural net to do name entity recognition or any of those things." }, { "start": 1002.06, "end": 1006.145, "text": "You wouldn't ask that to a search engine because all you get back is a bunch of blue links and then you'd have to find the code." }, { "start": 1006.981, "end": 1011.362, "text": "But on ucom the answer is literally hundreds of lines of code." }, { "start": 1011.643, "end": 1016.305, "text": "That will just be the answer, and you just copy and paste it with a copy and paste button right there and you're done." }, { "start": 1017.943, "end": 1024.529, "text": "And so the Venn diagram search is maybe not entirely a subset, but certainly there's a large overlap." }, { "start": 1025.0, "end": 1025.984, "text": "But chat is even bigger." }, { "start": 1027.1, "end": 1039.607, "text": "OK, before we talk more about those facts like what is just briefly, as a technical podcast, what is, briefly, your infrastructure?" }, { "start": 1040.12, "end": 1047.085, "text": "What do you need to operate a modern search engine which is more powerful than Google?" }, { "start": 1047.501, "end": 1048.404, "text": "What do you need for that?" }, { "start": 1049.72, "end": 1049.961, "text": "Oh boy." }, { "start": 1051.58, "end": 1067.805, "text": "So I could nerd out a little bit on that, but part of me doesn't want to share too much of our stack because we have a lot of competition, big and small, that are literally trying to copy what we imagine every week, and it is a small copy." }, { "start": 1067.94, "end": 1070.79, "text": "Whatever you can disclose, ok, yeah, exactly." }, { "start": 1070.9, "end": 1074.189, "text": "So I guess there's some obvious things like data breaks." }, { "start": 1075.38, "end": 1081.671, "text": "There are big vector databases like Pinecone and Chroma." }, { "start": 1083.54, "end": 1088.35, "text": "In the back end there are some very large LMs sitting on GPUs." }, { "start": 1090.18, "end": 1094.952, "text": "There are both internally as well as APIs for those." }, { "start": 1098.006, "end": 1103.225, "text": "Yeah, those are some of the parts of the stack that are pretty crucial that come up a bunch." }, { "start": 1105.92, "end": 1118.23, "text": "And I guess you run on Google Cloud and use Kubernetes, then We don't usually share which cloud we're on, but we're not on." }, { "start": 1118.391, "end": 1118.633, "text": "Google." }, { "start": 1120.183, "end": 1122.255, "text": "I was just joking, ok, i mean, yeah, it's interesting." }, { "start": 1123.282, "end": 1128.61, "text": "You think, oh, they're the enemy, but then Google can also be potentially a partner in some ways." }, { "start": 1130.645, "end": 1141.547, "text": "And in technology you would think that companies like Samsung and Apple, who are producing phones that are basically competitive, would hate each other and never talk." }, { "start": 1141.7, "end": 1161.307, "text": "But turns out, samsung also provides a lot of the hardware pieces that go into an iPhone with their semiconductor area, and then they compete later on on the actual phones that are being fully put together." }, { "start": 1162.34, "end": 1167.145, "text": "And so, yeah, we try to be pragmatic, but yeah." }, { "start": 1171.404, "end": 1175.261, "text": "And what is the business model of a search engine of the future Like?" }, { "start": 1175.341, "end": 1178.267, "text": "is it subscription-based Or will it be subscription-based?" }, { "start": 1180.124, "end": 1191.593, "text": "Or I mean ads feels a bit dead right if you look at Google And there won't be so many impressions anymore in so many clicks." }, { "start": 1191.82, "end": 1195.03, "text": "So it's a really good question." }, { "start": 1195.16, "end": 1198.85, "text": "Yeah, we're actually currently very focused on that." }, { "start": 1199.56, "end": 1208.529, "text": "To be honest, until we had launched UChat, we sort of had 30% month-over-month growth." }, { "start": 1208.68, "end": 1209.223, "text": "It was all right." }, { "start": 1209.78, "end": 1225.104, "text": "But once we launched UChat, we all of a sudden had, in January, 300% month-over-month growth And then went from a couple of 100,000 users to several millions of users, and that's when we started to really have to care about our cost to serve." }, { "start": 1226.281, "end": 1237.709, "text": "So we focused on that and got that under control in March, and now we're again been very nicely, organically growing, and now we're focused solely on monetization." }, { "start": 1237.86, "end": 1249.85, "text": "So, indeed, on the one hand, people pay for chat models, but people don't pay for search, and we're a chat-based search engine." }, { "start": 1250.0, "end": 1255.846, "text": "So that puts us in an interesting new category that really doesn't exist like that." }, { "start": 1255.92, "end": 1261.768, "text": "I think all the existing search engines will add this sort of as an add-on, but they're not chat-first the way we are." }, { "start": 1261.92, "end": 1272.147, "text": "And so, long story short, we will explore some private advertisements similar to Duck Go that is not user-dependent but it's only a query, or I guess, nowadays conversation-dependent." }, { "start": 1274.863, "end": 1287.845, "text": "And then we will also explore subscription models for various expert or add-on features, that even bigger models, even more complex things that cost even more money to serve." }, { "start": 1288.94, "end": 1293.51, "text": "And then we have also now had a lot of interest in APIs." }, { "start": 1294.8, "end": 1300.327, "text": "So I feel like that's a little bit distracting for a consumer search engine." }, { "start": 1300.52, "end": 1334.75, "text": "But with all this interest and we have a lot of developers on our platform and those developers are saying I love the search engine, but I want to use those same models for my own stuff, so we're thinking about offering uh api's as well to people great, uh, and don't you sometimes think, okay, this could be open, uh like, maybe it shouldn't be, maybe it should be belonging to a to, to like a for-profit entity, uh like, like wikipedia or something, uh, or what are your thoughts about that?" }, { "start": 1335.506, "end": 1355.567, "text": "yeah, i i actually thought about this a lot and i, i i bring it up from time to time when i get excited, but the truth is that, uh, it open, having an open source index, is only one of the many necessary ingredients for running a search engine, turns out." }, { "start": 1355.607, "end": 1366.251, "text": "You have to have all of that technology also, be extremely fast and available all around the world, and different data centers, and those just cost a lot of money." }, { "start": 1366.706, "end": 1380.191, "text": "Uh, it's kind of like open source traditionally felt like well, if you have the source code, now i can just run that program on my own machine, and so having a source code was all you need." }, { "start": 1381.045, "end": 1386.548, "text": "But when you think about a i, having the source code is only one thing." }, { "start": 1386.628, "end": 1404.454, "text": "You need to also have the weights of the model, and then you also need to have the training data to have to be able to train the model, and then you also need to actually have massively large beefy machines to be able to even run the model at all, let alone train it." }, { "start": 1405.386, "end": 1424.235, "text": "Uh, and so having just the source code is is not, you know, it's like one quarter or one third of the equation, and so what that means is that, especially for search engine, uh, it's very hard to make it work because you just can't avoid the cost as a pure, simple, on-prem open source project." }, { "start": 1424.605, "end": 1434.257, "text": "You get the code and then you run it yourself, um, and that just doesn't work for a search engine, like people can't run a search engine for the entire web on on their laptop." }, { "start": 1434.425, "end": 1455.911, "text": "It's just laptops aren't big enough, uh, and so so that's one problem, uh, and then of course you know it's all not not very cheap to run, um, so i'm still kind of open, i like, i like the idea, um, but i have not yet seen how it would work financially and just project wise, like and when you double click into it, even things like firefox, for instance." }, { "start": 1456.011, "end": 1458.289, "text": "In theory it's an open source project." }, { "start": 1458.45, "end": 1464.667, "text": "Yeah, sure, nonprofit, yeah, act as it gets ninety five percent plus of the entire motel f foundations." }, { "start": 1464.767, "end": 1466.674, "text": "Money comes from google from google." }, { "start": 1466.714, "end": 1472.359, "text": "Yeah, and so if you come in and chat with them and like, hey, can we add our search engine to the list of defaults?" }, { "start": 1472.419, "end": 1479.323, "text": "not even make it the default, because we know you get three hundred million dollars or something from google for that to be the default, but just add us to the list." }, { "start": 1479.364, "end": 1484.69, "text": "They're like, well, how much money are we going to get for your advertisements when we add you to the list?" }, { "start": 1484.851, "end": 1487.607, "text": "not even the default option, just one of the choices." }, { "start": 1488.229, "end": 1511.874, "text": "And if you don't make revenue, um, because we're currently, you know, we haven't had ads um, we're just now starting to explore them it's just that the conversation goes nowhere and even if you have the revenue there, it's just a very um, it's a very sort of touchy subject for them because, uh, if something goes wrong on that front, the entire organization loses its budget from google." }, { "start": 1511.974, "end": 1524.777, "text": "And so, uh, long story short, if you were like some open source project, right, and you wouldn't give them any money, no one will actually get to choose it, even in another open source quote unquote project that is a non-profit." }, { "start": 1524.817, "end": 1526.048, "text": "So, yeah, i don't know." }, { "start": 1526.108, "end": 1527.994, "text": "There's a lot of uh intricacies." }, { "start": 1528.948, "end": 1538.203, "text": "We're continuing to learn, uh, running a search engine on the financial side, the partnership side and so on, where you realize, oh man, like it's not always the best technology that will win, it will also." }, { "start": 1538.223, "end": 1550.665, "text": "You know, you have to have good partnerships, you have to have strong sales and like strong financial model that makes it all work yeah, and I mean even looking at wikipedia, like you, you brought the firefox example." }, { "start": 1552.349, "end": 1562.571, "text": "Um, wikipedia is also uh backed um and uh potentially also influenced uh through through backing right." }, { "start": 1562.671, "end": 1566.854, "text": "I mean that's always a question like how open can information really be?" }, { "start": 1567.646, "end": 1569.271, "text": "um, what do you think?" }, { "start": 1569.331, "end": 1570.114, "text": "like who will?" }, { "start": 1570.986, "end": 1579.628, "text": "you just mentioned, training data has to be paid as well and generated as well like who will generate tomorrow's training data?" }, { "start": 1581.732, "end": 1596.453, "text": "I mean, i think it'll be mostly you know the world and people, um, it's kind of interesting, right like just this age old saying uh, when you're uh, when the product is free, you are the product, um, and so the internet has been mostly free." }, { "start": 1597.465, "end": 1608.394, "text": "So now the AI kind of takes all the data from the internet, um, that is publicly available, uh, and tries to learn from it, um, and so I think I think that will continue to happen." }, { "start": 1608.434, "end": 1623.727, "text": "But then I think where companies will be able to differentiate and build a little bit of a mode is, uh, to collect interesting new, unique datasets, um and uh, and so companies will continue to collect data." }, { "start": 1623.767, "end": 1625.652, "text": "And then, of course, there's more complexity." }, { "start": 1625.713, "end": 1640.31, "text": "You can look into biomedical data and, you know, there it's even more unique to have your own dataset very interesting, and how do you think online marketing can function in that world, like in the age of, of search via AI?" }, { "start": 1640.41, "end": 1643.246, "text": "so conversational search, um." }, { "start": 1643.727, "end": 1652.21, "text": "How will I be able, will I be able, to positively or negatively influence my, my ranking or your answers?" }, { "start": 1652.25, "end": 1654.476, "text": "like, um, how will that look like?" }, { "start": 1655.245, "end": 1657.215, "text": "yeah, yeah, that that's a whole new world." }, { "start": 1657.296, "end": 1666.216, "text": "We already have a couple of SEO folks who are like reaching out and say, hey, how can I you know we're seeing traffic from, from ucom like how can I influence that?" }, { "start": 1666.845, "end": 1668.029, "text": "the truth is, right now you can't." }, { "start": 1668.45, "end": 1672.409, "text": "Um, it's very, it's even harder than it is with SEO." }, { "start": 1673.191, "end": 1698.95, "text": "Uh, and, and maybe that's okay, maybe maybe that is part of the point um, i think there's also non-zero chance that in the future, you may have advertising inside the answers, so the answers will be influenced in some way or another through ads, um, and, and then you have to pay to have an unbi, like even more unbiased kind of answer than you would expect otherwise." }, { "start": 1699.251, "end": 1705.346, "text": "And and again, if it's free, you know, then these llms have to earn something somehow, and maybe it's just at the end." }, { "start": 1705.386, "end": 1708.228, "text": "It's like you might be interested in this thing, and then there's a link to an ad." }, { "start": 1708.269, "end": 1708.429, "text": "that's." }, { "start": 1708.509, "end": 1709.713, "text": "That's what we're exploring right now." }, { "start": 1710.325, "end": 1712.271, "text": "But, uh, who knows?" }, { "start": 1712.451, "end": 1715.913, "text": "yeah, so I, to be honest, we haven't thought about it too much." }, { "start": 1716.013, "end": 1716.616, "text": "I think we're." }, { "start": 1716.797, "end": 1721.133, "text": "We're currently thinking a lot about how to monetize this, just to stay alive as a company." }, { "start": 1722.947, "end": 1736.553, "text": "I think, like really thinking this through uh can drive you really crazy, right um, like thinking about truth and and how how truth might look like tomorrow, right like, are you afraid of that as well?" }, { "start": 1736.613, "end": 1748.202, "text": "a bit, or um, i generally, you know, i think, as a startup founder, um, the default, uh, mo, shouldn't be that you're afraid, but more that you're." }, { "start": 1748.283, "end": 1755.288, "text": "You know you're optimistic, otherwise it's very hard to um stay alive, uh and and motivate it." }, { "start": 1755.65, "end": 1761.538, "text": "Okay, you have to uh really remind yourself that every crisis is an opportunity." }, { "start": 1762.04, "end": 1769.127, "text": "Every crisis is an opportunity, which is easier said than uh felt uh in the middle of a crisis." }, { "start": 1769.689, "end": 1786.97, "text": "Um, but yeah, long story short, um, you know, for in the beginning of when we launched you chat, a lot of people first search for themselves their own names, and a lot of those early on were probably the single worst thing that the models uh were." }, { "start": 1787.191, "end": 1788.655, "text": "Yeah, uh, trying to do." }, { "start": 1789.025, "end": 1818.034, "text": "Because you know, our model looked up the internet like facts from the internet, and so if you're uh Michael, um Schultz or something, then you were a chancellor and a Formula One driver, and you know, like uh, whatever Harvard professor, and so it's like it put together a CV from all the different Michaels and all the different Schultz's that it found on the world, on the web, uh, and and put it together." }, { "start": 1818.094, "end": 1822.046, "text": "And now you know people are like, oh, this technology doesn't work, it's bad, blah, blah." }, { "start": 1822.106, "end": 1823.55, "text": "And now, if you look up your name." }, { "start": 1824.012, "end": 1836.928, "text": "More and more often it correctly says oh, like there are too many names with this, i can't tell you, or you probably mean this person, and then it's just like tries to stick to one of the people that is most likely the one you mean." }, { "start": 1837.51, "end": 1842.11, "text": "Um, so so there are a lot of issues, um, in the in the limit, of course." }, { "start": 1842.171, "end": 1847.189, "text": "You know truth and uh and bias and so on are extremely hard." }, { "start": 1847.631, "end": 1856.21, "text": "Like if someone says two people were killed in a drone strike in Yemen, how do you like no AI will tell you if that's true or not." }, { "start": 1856.753, "end": 1868.696, "text": "You know it's just like you need to actually be on the ground, interview people and talk, and so you need to trust journalists to be out there who really try to just get at the truth of a situation." }, { "start": 1869.466, "end": 1870.228, "text": "Uh, and you know." }, { "start": 1870.268, "end": 1887.958, "text": "But then of course you, you eventually get to some approximations of that and there's certain uh journalistic outlets and if you trust them on average and you know they will, you know, have uh corrections if they're wrong, and you know you, you build some kind of trust with them." }, { "start": 1888.541, "end": 1890.652, "text": "Uh, and sometimes even those are are struggling." }, { "start": 1890.773, "end": 1898.711, "text": "You know there's some examples of like the lab leak theory, where for a while you were like a conspiracy nutter and then maybe it was true after all." }, { "start": 1898.911, "end": 1927.676, "text": "And and now people, like even smart people, don't have full consensus on the situation and in those cases I think it'll be important and I think actually chat models will be better in the future for uh at describing these kinds of situations and to expose users to different viewpoints and I think, in the limit, that can even help democracy and and finding the truth." }, { "start": 1928.626, "end": 1936.875, "text": "I guess it's also about the way how you express the certainty of the model right, um, am I sure, or not?" }, { "start": 1936.975, "end": 1945.728, "text": "like, if you ask chat gpt for something like it's always very to the point, um, and and, and there's no way to understand." }, { "start": 1945.768, "end": 1950.471, "text": "Like, like if the model was really thinking it was correct or not." }, { "start": 1950.592, "end": 1954.828, "text": "Right, um, right, is this like 99 percent or is it just like 20?" }, { "start": 1955.269, "end": 1971.762, "text": "uh, you don't know right, right, yeah, we're, we're working on those exact kinds of issues and and uh, looking forward in the next next few months to to output more and more of those like features to help you understand the, the veracity and and certainty of it." }, { "start": 1972.164, "end": 1975.373, "text": "But it's not trivial and it's even especially for text, right?" }, { "start": 1975.393, "end": 1980.718, "text": "I mean for texts like understanding, like is this, is this a real text or not?" }, { "start": 1981.019, "end": 1981.983, "text": "like it's, it's, it's." }, { "start": 1982.424, "end": 1984.473, "text": "It's not possible, right, like or not really." }, { "start": 1985.317, "end": 1986.302, "text": "I mean again, in the limit." }, { "start": 1986.342, "end": 1988.433, "text": "It's not possible, but you can do approximations." }, { "start": 1988.513, "end": 1992.795, "text": "But yeah, the biggest thing is, if you say you're certain, then you also better be right." }, { "start": 1993.377, "end": 1998.784, "text": "Yeah, so it's even worse if you make a mistake and you're saying I'm certain, this is true and then it's not true." }, { "start": 1998.804, "end": 2007.946, "text": "That's even worse than just saying it and it's being being slightly wrong, but it is something that of course, we we continue to to work on um and we continue to make it better." }, { "start": 2007.986, "end": 2028.625, "text": "I think, in general, generative AI is sort of the model I, uh, developed is very good when you can, when it will take you a long time to create and answer an artifact, a piece of code, an image and so on, but it's very quick to verify its usefulness or correctness." }, { "start": 2029.356, "end": 2038.584, "text": "And so when you ask an AI model to write you some front end code, for instance, you just copy and paste it, you run it and you're like this looks the way I want it to look and you're done." }, { "start": 2038.796, "end": 2046.208, "text": "And so it could have taken you an hour to write it, but it only takes you five seconds to run it and then see if it works or not." }, { "start": 2047.616, "end": 2049.765, "text": "And so what that means?" }, { "start": 2051.392, "end": 2058.345, "text": "same with images, right, it could take you hours to draw an image, but you look at it and you're like this looks like crap, or this looks beautiful, and this is what I want, and so you're done." }, { "start": 2060.355, "end": 2064.166, "text": "And for facts, i think they fall in between." }, { "start": 2064.456, "end": 2079.635, "text": "Sometimes we give you a nice citation to that fact, and if that is a particularly important fact, you need to quickly verify it in the citation link or on the web links that we still show you on your comm on the right, and then it is there." }, { "start": 2079.755, "end": 2093.506, "text": "Sometimes the facts that it describes are a little bit harder, and then the use cases are less obvious, and then oftentimes, what that means is you probably want to rely on the links which we still show you on the right, just like your old school search engine." }, { "start": 2094.136, "end": 2102.187, "text": "Okay, okay, so you can still verify the sources, essentially, and that becomes maybe part of your journey then, or your job." }, { "start": 2103.436, "end": 2117.527, "text": "Speaking of jobs, i mean yeah, obviously there's lots of discussion, especially after chat, gpt, which made it very accessible, like made AI finally accessible, like everyone was talking before, but now it's accessible." }, { "start": 2119.556, "end": 2122.355, "text": "How do you think this will change the job world?" }, { "start": 2122.455, "end": 2132.435, "text": "Like will it even lead to our brains being partly not unused but having just different content?" }, { "start": 2132.555, "end": 2137.955, "text": "Like I mean, i remember my dad with that huge atlas when we drove on vacation." }, { "start": 2138.117, "end": 2139.795, "text": "Like I don't know if you have the same experience." }, { "start": 2139.795, "end": 2142.154, "text": "Like we always had that huge atlas." }, { "start": 2142.735, "end": 2146.626, "text": "And nowadays, when I visit my dad, he asked me like how did you get here?" }, { "start": 2148.235, "end": 2152.527, "text": "And I mostly don't know most of the times." }, { "start": 2157.399, "end": 2165.05, "text": "Like how will this change our thinking and the usage of our brain from your perspective?" }, { "start": 2165.396, "end": 2167.863, "text": "And yeah, second question job world." }, { "start": 2167.943, "end": 2168.826, "text": "Maybe we'll do that later." }, { "start": 2169.895, "end": 2170.276, "text": "Yeah." }, { "start": 2170.316, "end": 2182.386, "text": "So I think people over-index on the importance of skills that they had needed when they grew up And the truth is like, yeah, i mean, i find it also funny." }, { "start": 2182.936, "end": 2185.765, "text": "Some people are sort of navigationally challenged." }, { "start": 2186.756, "end": 2197.955, "text": "I like to say it sometimes And you know, they couldn't for the life of them navigate without Google Maps or something And part of me is like, oh man, you would be lost." }, { "start": 2198.116, "end": 2207.875, "text": "But the truth is like most people who can navigate with a map would also be lost without a map in a forest because they don't know how to use the sun and so on." }, { "start": 2208.056, "end": 2211.265, "text": "They just kind of use a map and they'd use street signs and whatnot." }, { "start": 2211.695, "end": 2216.988, "text": "But how important is it to you know, be able to find your way out of a forest these days?" }, { "start": 2217.256, "end": 2228.173, "text": "Like it's not being that important for most people for many decades, and so you know skills that are helpful." }, { "start": 2228.193, "end": 2233.235, "text": "over time change And through industrial revolution and through automation, things change." }, { "start": 2234.422, "end": 2236.675, "text": "Most people would not understand how to replicate a toaster." }, { "start": 2237.781, "end": 2247.215, "text": "It's actually a fun thing I like to do on long drives sometimes, which is like play a game where you get dropped 2000 years in the past And now you have all the knowledge you have." }, { "start": 2248.442, "end": 2250.735, "text": "You just time travel by accident, 2000 years into the past." }, { "start": 2250.979, "end": 2251.834, "text": "Now what do you do?" }, { "start": 2253.201, "end": 2256.115, "text": "How do you try to help humanity speed up its technological progress?" }, { "start": 2257.121, "end": 2268.586, "text": "The truth is, most people have no clue how to make electricity, how to transfer electricity, how to, you know, do a lot of basic things, and so, long story short, there's a long one answer to." }, { "start": 2269.127, "end": 2270.229, "text": "Skills will change." }, { "start": 2271.219, "end": 2277.151, "text": "Certain skills will be less important, new skills will be more important And I do think kids will have like." }, { "start": 2277.231, "end": 2277.955, "text": "Education will have to change." }, { "start": 2279.042, "end": 2280.995, "text": "You have to really update the curricula for kids." }, { "start": 2281.863, "end": 2283.154, "text": "You have to teach them how to program." }, { "start": 2284.324, "end": 2285.715, "text": "I've been telling that to a lot of politicians." }, { "start": 2285.755, "end": 2287.027, "text": "Programming will have." }, { "start": 2287.956, "end": 2288.967, "text": "Programming is still important." }, { "start": 2291.545, "end": 2292.591, "text": "It is kind of still important." }, { "start": 2292.672, "end": 2301.455, "text": "In the same way, as you know, you might say translation isn't important or learning a language isn't important because you can just use a translating algorithm like a translation algorithm now." }, { "start": 2302.522, "end": 2310.542, "text": "But the truth is you still need to have a live conversation And you can just be like oh, he's saying that it's transiting, it's translating it, And then you talk back to it." }, { "start": 2310.815, "end": 2319.884, "text": "Of course, that will get faster and faster and will incorporate speech and everything, but you still need to be able to think And so when you think about concretely generative AI." }, { "start": 2320.355, "end": 2331.692, "text": "It goes back to this model that I just came up with a few weeks ago, which is you might not have to generate as much content, but you still need to have the judgment to know if the content is any good." }, { "start": 2332.265, "end": 2341.094, "text": "You still need to be able to verify that the facts in it are true And you need to still know things in order to be able to creatively recombine them." }, { "start": 2342.387, "end": 2348.213, "text": "And so, yeah, will it be as important to be able to draw nicely when AI will draw for you?" }, { "start": 2348.565, "end": 2354.774, "text": "No, will it be as important to crank out simple sort of re-urgitating, summarizing text?" }, { "start": 2355.445, "end": 2360.57, "text": "Probably not, but it would be important to understand how to structure a good argument to convince someone in a conversation." }, { "start": 2361.092, "end": 2370.335, "text": "Certainly, will it be important to be able to read a text that maybe an AI had written and evaluate whether it's going to convince anyone else of some facts." }, { "start": 2370.485, "end": 2374.493, "text": "I think those are all still going to be important and those are new skills that people will have." }, { "start": 2374.645, "end": 2395.637, "text": "And I guess I don't worry too much about people getting like more and more stupid and like sort of the idiocracy movie that a lot of people found very sort of present or like just innovative and sort of, yeah, predictive." }, { "start": 2396.745, "end": 2404.61, "text": "You know there is a little bit of that concern in the longer term time horizon, but I don't think that will happen in the next generation or two." }, { "start": 2405.845, "end": 2410.977, "text": "So we all don't have to be afraid of that." }, { "start": 2411.245, "end": 2419.066, "text": "Like, if I know Ruby as an example, it will be there for a while, most likely, right, you know?" }, { "start": 2419.146, "end": 2427.293, "text": "like, ai will write some code for you, but you still need to be able to fix it and integrate it, and if it's wrong, you need to find it, and if there's a bug, you need to be able to, like, fix it." }, { "start": 2427.865, "end": 2432.191, "text": "And so, yeah, i think there's a ton of ways that you want to verify." }, { "start": 2432.305, "end": 2452.756, "text": "You know, there's some funny signs where it is a Chinese hairdresser and the new English name for the hairdresser is could not connect to a translation service, and you know, if you don't know anything, a single word in the English language, you will print that big sign and put it on your hairdresser." }, { "start": 2452.938, "end": 2455.671, "text": "You know company, yeah, that's a nice example." }, { "start": 2455.732, "end": 2458.452, "text": "yeah, Yeah, I also see it that way." }, { "start": 2459.053, "end": 2475.633, "text": "I see it that way, but I mean most likely programming, or it could happen that programming rather ends up writing specs, right Like writing stories, writing acceptance criteria, and then you get something generated." }, { "start": 2476.627, "end": 2480.932, "text": "I guess that will be popping up soon, like as a new form of no code, maybe." }, { "start": 2482.045, "end": 2482.407, "text": "That's right." }, { "start": 2483.506, "end": 2484.811, "text": "I think that will also happen." }, { "start": 2484.831, "end": 2491.992, "text": "You know, like in many ways the abstraction layers are going to get bigger, like higher and higher, right, and you'll just be even to be more productive." }, { "start": 2492.485, "end": 2514.433, "text": "You know you don't really have to write in assembler anymore No one has to And so you just don't have to be that low level machine code instructions, and that's a good thing, right, like that allows us to be more productive and build higher, more interesting, complex systems." }, { "start": 2514.525, "end": 2525.309, "text": "Now, there will always be some people who have to be close to the metal right And build new CUDA kernels or new kind of speed algorithms for new kinds of hardware." }, { "start": 2525.33, "end": 2527.99, "text": "That keeps getting developed, but it's fewer and fewer people." }, { "start": 2528.745, "end": 2546.533, "text": "It's a less and less important skill for building things, and you know you have things like JavaScript and Nodejs and Swift and others, so that you know those will allow you to build apps more quickly and in front end web development things, and I think that will continue." }, { "start": 2546.573, "end": 2555.154, "text": "And you're right, i think English will enable you to build and is already able to help you build a couple of simple, quick things." }, { "start": 2555.345, "end": 2559.751, "text": "But if you then want to edit them, you want to say, oh, i want this also and I want it to be slightly different." }, { "start": 2560.305, "end": 2573.556, "text": "You can keep trying to explain it to an algorithm, but at some point you will have diminishing returns and you can actually be faster, if you just knew how to do it, rather than keep sort of iterating, say to the algorithm, oh, but like, make the box just like two pixels to the left." }, { "start": 2573.725, "end": 2579.274, "text": "Oh wait, but not if the screen is this big, then like, then just make the button disappear." }, { "start": 2579.805, "end": 2581.683, "text": "Oh wait, but like for that other phone." }, { "start": 2582.046, "end": 2583.873, "text": "Change the margin, change the margin Yeah exactly." }, { "start": 2584.045, "end": 2602.774, "text": "It's just like at some point you realize, wait, it's probably just better if you had explicitly been able to program that and then at that point you'd be faster, and so, but you know, you get again a quick version of it, a default, and then if you know how, what you're doing, you know how to code, you'll be able to modify it and get it to a really good end result even more quickly." }, { "start": 2603.406, "end": 2603.807, "text": "Mm-hmm." }, { "start": 2604.529, "end": 2626.515, "text": "And right now, if I open up Twitter in the morning, i kind of see this huge AI dust storm that I guess many of us are facing these days And I think, like many CTOs out there, are supposed to have an opinion on that Like is it something we can leverage?" }, { "start": 2626.885, "end": 2627.669, "text": "Are we doomed?" }, { "start": 2628.445, "end": 2629.429, "text": "Is it an opportunity?" }, { "start": 2631.429, "end": 2635.811, "text": "What would you recommend to navigate through that dust storm?" }, { "start": 2639.145, "end": 2647.679, "text": "So I do think Twitter is actually a useful place for staying up to date in AI." }, { "start": 2648.385, "end": 2659.717, "text": "I think there is a little bit of hype, but the way I think about it is, indeed, you are like there's sort of different waves of AI." }, { "start": 2660.245, "end": 2660.947, "text": "You know some people." }, { "start": 2661.509, "end": 2663.947, "text": "It feels like for them it's been." }, { "start": 2664.328, "end": 2671.292, "text": "For some people feels like, oh, this has come out of nowhere, right, you haven't had like AI, just didn't work, and now all of a sudden it's super big." }, { "start": 2671.645, "end": 2681.653, "text": "But the truth is there have been tons of improvements over the last decade where, all of a sudden, speech recognition worked and then computer vision and object recognition started to work." }, { "start": 2682.466, "end": 2684.032, "text": "And then translation started to work." }, { "start": 2684.125, "end": 2688.492, "text": "Like no translator who gets paid per word and not per hour is not using AI." }, { "start": 2689.425, "end": 2692.43, "text": "Now they're all using AI to get a raw translation and they're fixing things up." }, { "start": 2693.066, "end": 2707.811, "text": "And now, with generative AI and general sort of excuse me chat models and generative image models, it's become even more close to people, even closer and visceral and sort of obvious to people how much this can change things." }, { "start": 2710.23, "end": 2714.492, "text": "And so, yeah, i think it's going to be a really, really exciting time." }, { "start": 2714.785, "end": 2715.87, "text": "People should stay up to date." }, { "start": 2716.385, "end": 2721.873, "text": "But as each wave comes, the swell is rising, like the capabilities are getting better." }, { "start": 2722.025, "end": 2734.135, "text": "Behind that big wave, but maybe at the very top of inflated expectations, there will be a bit of a let down because AI cannot yet fully program itself." }, { "start": 2734.285, "end": 2737.074, "text": "Like auto-GPT types of things are a little bit overhyped." }, { "start": 2737.665, "end": 2749.013, "text": "But that technology, i think in the next six to 24 months will do some pretty incredible things and will be able to improve itself more and more, and I'm very excited about the future." }, { "start": 2749.505, "end": 2758.953, "text": "I think humanity is kind of in the middle of an exponential increase in capabilities And I don't think that exponential will last forever." }, { "start": 2759.225, "end": 2763.631, "text": "Most exponentials in technology usually flatten out and become an S-curve at some point." }, { "start": 2763.785, "end": 2771.012, "text": "But we don't know how long that exponential growth is going to continue And I'm very excited for it." }, { "start": 2771.065, "end": 2781.692, "text": "I think long-term no one's going to look back and be like, oh yeah, i wish I could do radiology analysis like everyday, non-stop." }, { "start": 2781.885, "end": 2789.472, "text": "I think it'll be cooler if I just explain to an AI what I'm doing, have it, read the textbook and then do it for me, and this is not going to happen overnight." }, { "start": 2789.846, "end": 2792.81, "text": "We still continue to need radiologists and we probably will always need them." }, { "start": 2793.945, "end": 2806.831, "text": "I think the way I think about it, if there are 100 jobs and 100 people doing something, those 100 people will not get replaced entirely by an AI, but they will often different industries, have different time frames and so on." }, { "start": 2807.325, "end": 2815.837, "text": "But they'll probably get replaced by 10 people that use AI, rather than all like zero people on an only AI." }, { "start": 2816.865, "end": 2822.69, "text": "And so I think that that will be exciting And, just like 150 years ago, over 90% of people work in agriculture." }, { "start": 2824.775, "end": 2832.251, "text": "Yeah, for sure No one will look back and say, oh, you know what, it would be cool if we just worked with our hands all day in the heat, in the cold, in the fields." }, { "start": 2834.228, "end": 2841.935, "text": "No one says that now, but at the time when the big tractors coming, those big tractors are a thousand times stronger than you are." }, { "start": 2842.785, "end": 2846.03, "text": "If you just stood in their way they would crush you You're like, oh man, that's really scary." }, { "start": 2846.905, "end": 2849.533, "text": "Overlord tractor, train, whatever." }, { "start": 2850.305, "end": 2851.108, "text": "Maybe you shouldn't." }, { "start": 2851.148, "end": 2851.811, "text": "People were scared." }, { "start": 2852.405, "end": 2854.09, "text": "Big trains are scary the first time you see it." }, { "start": 2855.645, "end": 2867.475, "text": "But long-term, everyone likes the technology, everyone loves convenience and simplifying lives, and now people are doing, i think, more interesting things than working in a field, and I think we will do even more interesting things 100 years from now." }, { "start": 2868.247, "end": 2868.588, "text": "Obviously." }, { "start": 2868.648, "end": 2875.57, "text": "so I guess a good strategy is to get comfortable with it right, especially as a CTO." }, { "start": 2875.811, "end": 2878.889, "text": "these days, what do you think about?" }, { "start": 2879.325, "end": 2897.895, "text": "there are so many businesses popping up, many startups, like there are many bootstrapers who just these days use, like, open AI APIs and build simple stuff or more complex stuff, and it seems that it's harder to have a mode as a business." }, { "start": 2899.731, "end": 2900.814, "text": "So how do you think about that?" }, { "start": 2902.007, "end": 2910.456, "text": "Like, is everything just commoditized then In a few months, or how do you think about that?" }, { "start": 2911.711, "end": 2911.953, "text": "Yeah." }, { "start": 2912.013, "end": 2926.583, "text": "So I think at some point probably, lms are themselves more of a commodity, and my strong hunch is that LMs are going to be kind of like databases." }, { "start": 2926.603, "end": 2942.636, "text": "You know, you sometimes may need an oracle to have a million queries a second and it needs to be super fast and it needs to be very reliable and it's just massive requirements." }, { "start": 2942.725, "end": 2946.315, "text": "And then you just need a very large, very powerful database." }, { "start": 2947.265, "end": 2967.069, "text": "But a lot of people just download some MySQL database, which is the Alpaca Alamas open source models of the world, and then you just hack up your own website and build it, and so I think the LM might not be the differentiating factor, but just like, which database you use isn't a differentiating factor." }, { "start": 2967.451, "end": 2990.735, "text": "You still need to use one and you can still do a lot of cool stuff with it, and I think that will continue to be right And there will be new kinds of modes of users and communities and training those LMs on interesting tasks, getting more and more data about how to do things, and new kinds of modes will arise in the future." }, { "start": 2991.105, "end": 2995.95, "text": "OK, ok, so I still have three questions for you, so we shortly come to the end." }, { "start": 2998.505, "end": 3013.153, "text": "From your perspective, talking about tomorrow, what's keeping you awake at night, like is there something that you think will lead to the next wave of disruption?" }, { "start": 3013.525, "end": 3022.285, "text": "Is there anything like crazy technological that you right now see, where you tell everyone about That?" }, { "start": 3022.326, "end": 3024.732, "text": "it's like really revolutionary." }, { "start": 3026.205, "end": 3032.036, "text": "I mean, I think you don't have to go too far, You just love LMs and we were in the middle of that revolution." }, { "start": 3032.125, "end": 3047.134, "text": "And then there are a few things that I tell everyone internally and the company about, but I don't really want to tell everyone else outside the company what I tell those folks inside the company, So there's a lot to expect there, Like models will get bigger stuff like that." }, { "start": 3047.745, "end": 3049.492, "text": "Well, yeah, those are the obvious things." }, { "start": 3049.525, "end": 3061.245, "text": "They get multimodal, like we had the first LM that will output graphs and tables and things like that inside a large language model And that is now also being copied by the big guys." }, { "start": 3062.067, "end": 3077.788, "text": "So how quickly we can continue to innovate and stay ahead of the big guys so users at some point realize, oh well, yeah, i do see the future on ucom three to six months before I see it with the big guys in some close demo and whatnot or close beta." }, { "start": 3078.946, "end": 3083.111, "text": "We'll continue to be something that will keep me up for the coming future." }, { "start": 3084.305, "end": 3086.471, "text": "And when do you expect AGI?" }, { "start": 3088.565, "end": 3090.572, "text": "I think we just don't know." }, { "start": 3091.445, "end": 3102.337, "text": "I think people have, both in terms of the fear as well as the excitement, inflated expectations." }, { "start": 3103.065, "end": 3105.894, "text": "The truth is, no one is really working on it." }, { "start": 3106.185, "end": 3114.529, "text": "Anyway, i could talk about it for a long time, but I would put it somewhere between 30 to 200 years 30 to 200 years." }, { "start": 3114.67, "end": 3127.116, "text": "OK Yeah, because the release of chat GPT kind of led to the general thinking that it's rather five years than 30 years or 200." }, { "start": 3127.885, "end": 3128.507, "text": "Yeah, i mean." }, { "start": 3128.527, "end": 3135.097, "text": "The interesting thing is a lot of people think that I actually have a bet that I think I've won." }, { "start": 3135.185, "end": 3154.114, "text": "I need to look up when the date was, but about five years ago I had a bet with one of the OpenAI founders And he thought they're going to have AGI already And we have zero of the three requirements fulfilled And all three need to be fulfilled for him to win, and so I think people are a little bit too optimistic." }, { "start": 3155.445, "end": 3159.411, "text": "But it also then kind of you double click on it and it's like what's your definition of AGI?" }, { "start": 3159.845, "end": 3163.615, "text": "And the truth is you don't need AGI to automate 80% of jobs." }, { "start": 3164.845, "end": 3167.434, "text": "You just have a system that predicts the next byte pair encoding." }, { "start": 3167.565, "end": 3205.83, "text": "It has no sense of what a word is, it has no conceptual sense of itself, what it means to be alive, and so on, and you can have a huge amount of impact of all of humanity with just very, very strong AI that isn't fully generalized, self-conscious, super intelligent, can actually then just read all the physics books and then suddenly figure out what particles are actually the reality And then set up and build an entire new large hydrogen collider and run the actual write experiments and then bring physics forward." }, { "start": 3206.172, "end": 3206.754, "text": "Good luck with that." }, { "start": 3206.825, "end": 3209.593, "text": "That will take a lot longer than people think." }, { "start": 3211.205, "end": 3214.053, "text": "So, as a last question, i still have a little surprise for you." }, { "start": 3215.625, "end": 3225.475, "text": "Mark Benny of Salesforce told me about a little Easter egg that you personally built into the Einstein engine, as your leftovers at Salesforce are called." }, { "start": 3225.886, "end": 3251.015, "text": "Your product is being called now, and it's a feature that you have to be able to program Apex to use it, and it's called the time machine the Einstein on your phone, thanks, and it's a special package that actually allows you to really travel back in time just back, not forth." }, { "start": 3252.468, "end": 3285.93, "text": "And I prepared a little experiment fire up my good old JetForcer IDE and just hit Run on the code I prepared before, and now we literally travel back in time to August 2014, the year you graduated from Stanford and kicked off Metamind, and we now observe young Richard coding a little or coding a lot, and you now have the chance to whisper something into young Richard's ears." }, { "start": 3286.191, "end": 3286.873, "text": "What would it be?" }, { "start": 3291.54, "end": 3306.724, "text": "Probably to continue doing what you're doing, because I'm very happy of how life turned out, but probably think more about large language modeling as one of the tasks that you should train an AI with." }, { "start": 3307.065, "end": 3330.992, "text": "So not just a lot of supervised multitask learning, not just a single model for all of NLP, but one that just tells you what you're doing, that just does large language models more, and to have even more conviction that that is right, because I had, from 2014 to 2018, we had the best large language models in the world." }, { "start": 3331.345, "end": 3342.209, "text": "Like we caught the pointer models, which was sort of a precursor to transformer models, and we just didn't scale them to cost $30 million for training one model." }, { "start": 3343.145, "end": 3351.173, "text": "But we should have just kept scaling And I've told the first author of those papers let's just scale them up and see if something interesting happens." }, { "start": 3351.825, "end": 3353.111, "text": "But I didn't push it hard enough." }, { "start": 3353.305, "end": 3354.249, "text": "You didn't push hard enough." }, { "start": 3354.725, "end": 3368.954, "text": "I should have pushed the language modeling objective even more, because, ultimately, full language modeling, where you are perfectly good at predicting the next word, is NLP complete and is AI complete?" }, { "start": 3369.665, "end": 3385.871, "text": "If you say, oh, the next word from, i'm in Dresden right now and I want to get back to San Francisco And I navigate all the streets and navigate all the airports and so on, it means you have a perfect world model, you have a perfect planning model, you have a perfect visual model And you can do a lot of things." }, { "start": 3385.945, "end": 3392.589, "text": "You can communicate with people who might not want to get the ticket anymore Because you're five minutes late to the counter, like all of these things." }, { "start": 3392.825, "end": 3403.612, "text": "And so you realize, wow, predicting the next word is an incredibly interesting and hard task And you can do a lot of things with it, assuming you have the right objective functions for it." }, { "start": 3405.245, "end": 3407.348, "text": "Thanks a lot, Richard Great recording." }, { "start": 3408.365, "end": 3408.888, "text": "Many thanks." }, { "start": 3409.745, "end": 3415.45, "text": "Have a safe trip home And I hope I can visit you on your ranch at a certain point in the next years." }, { "start": 3415.951, "end": 3416.975, "text": "You're very welcome, man." }, { "start": 3417.045, "end": 3417.929, "text": "Sounds quite exciting." }, { "start": 3418.965, "end": 3421.212, "text": "So yeah, thanks a lot for your time." }, { "start": 3421.754, "end": 3422.095, "text": "Thank you." }, { "start": 3422.265, "end": 3423.209, "text": "Thanks for your great questions." }, { "start": 3423.329, "end": 3423.972, "text": "It was a fun one." }, { "start": 3424.386, "end": 3429.169, "text": "You asked some questions that no one has asked before, which is getting harder and harder, so have fun." }, { "start": 3430.365, "end": 3430.726, "text": "Great." }, { "start": 3430.967, "end": 3431.368, "text": "Bye bye." }, { "start": 3431.769, "end": 3434.836, "text": "Thank you very much, it was good." }, { "start": 3435.388, "end": 3438.228, "text": "Yeah, nice, yeah, i mean it Nice podcast." }, { "start": 3438.685, "end": 3442.693, "text": "Yeah, it was fun And it always goes on." }, { "start": 3443.946, "end": 3445.11, "text": "Absolutely, it always goes on." }, { "start": 3445.825, "end": 3446.93, "text": "Yeah, it's nice that you're so optimistic." }, { "start": 3448.146, "end": 3450.99, "text": "So good, that's what we have to do, or we have to be a founder in general." }, { "start": 3453.246, "end": 3454.692, "text": "Every crisis is an opportunity." }, { "start": 3455.146, "end": 3456.291, "text": "Absolutely, absolutely." }, { "start": 3458.731, "end": 3467.15, "text": "I mean sometimes we actually live in a simulation when I look out, but It's a bit of a mess what just happened." }, { "start": 3469.095, "end": 3469.596, "text": "Good, thank you." }, { "start": 3470.006, "end": 3470.789, "text": "I'm doing the intro to Philippe." }, { "start": 3471.225, "end": 3472.569, "text": "Should I just put the Fredders on the right CC?" }, { "start": 3472.589, "end": 3476.623, "text": "Yeah, great, and you probably do it more when you're home, right?" }, { "start": 3477.148, "end": 3478.383, "text": "So if you're still there, at the weekend." }, { "start": 3478.665, "end": 3480.171, "text": "Yeah, the next two days are pretty good." }, { "start": 3481.365, "end": 3483.947, "text": "Okay, cool, have fun, see you Ciao." } ]