The Exchanges

Q of like group schooling settings. Like, can you talk a little bit about how you think about that? Because you also started Schoolhouse Schoolhouse.world with our mutual friend Shashir, founder of CODA. And, and, you know, this is like small groups online and believing that there's still the importance of this. Can you, can you talk about sort of how those pieces fit together and what the motivation was there?

A Yeah, absolutely. And, you know, just a little bit more detail on Schoolhouse. We started during the pandemic, you know, that was a moment where like Khan Academy's usage went up three X and it was clear that it was great if people are using Khan Academy, but they, they needed, everyone was isolated and they needed more human support as well. And I'm always thinking about, okay, how can you give something at scale that used to feel really expensive and tutoring, live human tutoring For the most part is expensive, but I was like, well, what if you could leverage volunteerism? And so it was a bit of a, you know, optimistic idea that there would probably be a lot of people who want to do, you know, what I did with Nadia, that a lot of people would actually love to do that. And so we set up schoolhouse.world as a way, you know, it's, it's, it's another nonprofit, a sister nonprofit to Khan Academy. Its mission is to connect the world through learning. And the whole idea is, is that you get free tutoring and who gives the tutoring. It's, it's, Volunteers. And a lot of those volunteers could be near peers. They could be high school students, college students, but some of them also are teachers who teach at a fancy private school and want to give back. They might be a retired professor, et cetera, et cetera. It's, it's gone well. I mean, it's definitely at a smaller scale than Khan Ac…

AI assessment note: “they needed, everyone was isolated and they needed more human support as well.”

Q Well, what are some of the major forward-looking aspects of this sort of project and research?

A I think there's a ton to do in the context of video generation. Like if you look at make a video, it was very exciting. It was sort of first of its kind, um, capabilities at the time, but it's still, it's a four second video. It's essentially an animated image, right? It's kind of the same scene, the same set of objects that are moving around in reasonable ways, but you're not seeing objects appear, objects disappear. You're not seeing objects reappear. You're not seeing scene transitions. Um, None of this is, is, is in there. And so if you look at, if you think about the complexity of videos that you just regularly come across on various surfaces, this is far from that. And so there is a ton to be done in terms of making these videos longer, more complex, having memory so that if an object reappears, it's actually consistent. It doesn't now look entirely different. Um, things of that sort, sort of being able to tell more complex stories through videos, all of this is, uh, entirely open.

AI assessment note: “there is a ton to be done in terms of making these videos longer, more complex”

Q you have these fundamental breakthroughs and sometimes it's like an, it's an architecture like transformer based models versus traditional NLP. And sometimes it's, um, you know, iterating on a lot of other things that already exist in the preexisting approaches and just sort of solving specific engineering or technical challenges. If you were to sort of list out the bottlenecks to this, what do you think they're likely to be?

A Yeah, I think there's a few different things. One is videos are just sort of from an infrastructure perspective, harder to work with, right? They're just sort of larger, more storage and sort of more expensive to process and more expensive to generate and all of that. So there's just that iteration cycle that is much slower with video than it would be with other modalities. So that is, is one. Um, the second is I don't, I don't think we've still figured out the right representations for video. Right. There is a lot of redundancy in video one frame to the next frame. There's not a whole lot that changes. Um, we still kind of approach them fairly independently as sort of independent images, even if you're generating it as kind of one after the other, or even if you're generating in parallel and then making it finer grain. Um, so I think maybe that could be something that helps with a breakthrough that if we really figured out how to represent videos efficiently. Um, and the third is this, Hierarchical architecture that if you want longer videos, they're just so many pixels that you're trying to generate, right? It's a very, very high dimensional signal, um, compared to anything else, uh, that we're doing. And so just thinking through how do we even approach that, what sort of hierarchical representation, um, makes sense, especially if you want these scene transitions, if you want…

AI assessment note: “I think there's a few different things. One is videos are just sort of from”

Q You made another acquisition of a company called Streamlet that I, I think we're also both familiar with. Can you talk about, um, the rationale for that?

A Streamlit is, is a company that does visualization animation, um, you know, for, for Python applications, but specifically in the world of machine learning. The problem with machine learning is, um, if you're not a programmer, it's pretty damn hard to consume, you know, what, what it is and how it works. Um, but Streamlit is almost reflexively reached for by Python programmers to basically make a machine learning model Consumable by a general business, uh, user, you know, you can manipulate the variables and it just redraws everything. Visualization animation. Uh, and, and that's, we, the reason that we acquired Streamlit is a, you know, that's certainly, we have to have visualization and animation. And by the way, this also touches the world of BI because a lot of people use Streamlit, you know, for the same reason that they would use BI type of products, but this is just much more Um, you know, specific to all kinds of reporting and use cases and, and, and, and dashboarding. Um, so what we wanted to do with Streamlit is to bring it inside Snowflake. We call it Streamlit in Snowflake. And the reason is you need to have that hardcore, trusted, sanctioned governance perimeter, um, because otherwise people, people will not allow the business to use these kinds of applications. Governance is a really big deal because the data needs to be sanctioned and trusted. And the business sh…

AI assessment note: “the reason that we acquired Streamlit is a, you know, that's certainly, we have”

Q I guess related to that, it feels like one of the next mountains that you all are conquering has to do with the incorporation of AI into your product. And obviously you care more about how the product impacts the users, but I'm a little bit curious, you know, when did Instacart start focusing on AI and ML and how do you feel that this wave of generative AI differs?

A Yeah. You know, it's interesting because obviously AI is having a big moment, but Instacart has been focusing on AI since the very beginning. Like if you think about it, We are fundamentally a search engine. We do a billion searches a month, a quarter, sorry. And you know, it's all powered by AI. Every time a shopper goes to the store and a product is not on the shelves, we suggest replacement. We have suggested a billion replacement at this point. And like, that's all powered by AI. The way in which we match shoppers to go to specific stores, all powered by AI. So the list goes on. Um, but I think the thing that's Changing now with generative AI is that we are finally at a point where we can answer people's questions using natural language and the way they've been asking themselves these questions before. If you take a step back and you think about commerce online, it is very bizarre that we have forced human to train themselves to express their commerce needs in the form of keywords. I mean, people don't think about like, you know, how to put dinner on the table for the week, uh, in terms of like cilantro, eggplants, sour cream, which is exactly what they're currently typing on Instacart right now. Instead, what they think about is like, I have a budget of 300 bucks. I have to put, you know, X meals on the table. My daughter is lactose intolerant. We Traditionally eat Mediter…

AI assessment note: “Instacart has been focusing on AI since the very beginning”

Q I have to ask because you have multiple different experiences that are really deep and interesting in advertising from Facebook To what you're doing with retailers and CPG brands today. Um, if you zoom out, what predictions do you have for the impact of AI, uh, on advertising?

A Oh, I think it's going to be fantastic for ads. Um, I think it's really about much more personalization and really like this kind of ideal that, you know, you're going to get the perfect ad at the perfect moment for you. I think it's also going to be an explosion in creative. Uh, when you think about all of the variation of creative that can be created now, so that again, we match the right person to the right user, like that's going to be just Incredible. Uh, and that's gonna, again, make advertising a lot more relevant. And then the other thing is that I think advertising fundamentally follows the consumer products and the consumer use cases. I mean, that was one of our guiding principle at Facebook. And because we're gonna see such an explosion of new consumer experiences, I think as a result, we're gonna see new placements for ads that are gonna be interesting. So for example, We talked about, you know, ChatGPT plugins. We talked about Ask Instacart. Like, it is a very obvious place to start putting some sponsored content so that if you're asking, you know, what are the, like, best dairy-free snacks for my kids, you know, we can show you organic results, but we can also show you some sponsored results there. And these are placements that are, you know, gonna be really interesting for advertisers to tap into because, again, there's a much Clearer expression of intense and ju…

AI assessment note: “I think it's really about much more personalization and really like this kind of ideal”

Q different. It feels like there was always some good talent pools there. I know Google had a giant office there and Meta set up one and, you know, it was really sort of flourishing over the last decade. And now it definitely feels like a very strong standalone ecosystem. But are there other aspects of either recruiting or other things that, that you really benefited from being in New York?

A Yeah, I would say there's really two things. The first one is from a customer perspective, we're sort of out of the echo chamber of the Bay Area, which makes it easier. I would say to latch on to what really matters to customers and what's not just, you know, a fantastic idea. You told three people and then repeated three others, and then it came back to you and it sounds even better now. And there's a lot of companies of basically all many non-tech companies in New York, and you can sell to, and you can get a good idea of what they need basically. The second aspect I think that benefited us is that So it's a bit more difficult to recruit in New York. Uh, there's less pure tech talent. There's less deep tech talent in New York than there is in the Bay Area, but the retention is a lot higher. So, you know, if you give people, you know, great responsibilities, interesting work, treat them well, they're going to stay with you for three, four, five years more, you know, um, which I think in the Bay Area is pretty much on the, you know, very high end of what you can expect. What we see from looking at data, we have data from a, Most of our customers are engineers and most of our users are engineers. Um, and so, and we, we see basically when, when their individual accounts churn at our customers, uh, organizations, and we see that it's not rare for companies in the Bay Area to have, …

AI assessment note: “Yeah, I would say there's really two things. The first one is from a customer perspective”

Q Right. It's not totally intuitive to somebody who comes from outside of defense. Why higher volume of systems is suddenly a need, or is it more the idea that you're placing something that would have been people before? Um, so can you sort of explain that as a thesis?

A Yeah. So when you look at how the U S has operated, we have very expensive, very exquisite systems that are largely manned. So, you know, you can think aircraft carriers, fighter jets, bombers. These are outrageously expensive platforms, uh, that cost, you know, just huge amounts of money to build, to maintain, to train pilots on. On top of that, we have a pilot shortage, right? We don't have enough pilots to be able to fly all these things. We have recruiting shortfalls consistently. Uh, and, uh, and even more so, we have an industrial capacity issue. Our ability to make, uh, you know, nuclear submarines is going down. We are producing less nuclear submarines per year, year over year than we have in the past. So we're like, like, 1.7 per year we can produce these submarines. It's not getting better, and the workforce to make these is retiring. It's a very artisanal process. It's hard to train up people. We have very few suppliers in the industrial base to do it. So we're not going to outbuild China. That's not an option. On these large, you know, very expensive platforms. It's just, it's not going to work. It's a bad strategy. Um, the other part of this is you look at what happened in Ukraine, where, you know, a lot of how we thought warfare would progress is kind of playing out. It's almost like surprising to the degree to which you don't see even the Russians deploying fight…

AI assessment note: “we have very expensive, very exquisite systems that are largely manned.”

Q those are on the scale of hundreds of millions or billions of dollars. And many of these contracts end up being public, right? Or in the, in the public record. Are there any contracts that you can talk about or just so people get a sense of the scale at which, you know, some of the, the contracts that you've closed or programs or records that you've won have been?

A Yeah. So that, that's, I think that's kind of like how we view success is this sort of, can we get to scalable programs? And we sort of measure our success by, are we getting things to field In volume, uh, in a way that this is actually getting out. And so the, you know, the first one we had was with Customs and Border Protection doing border security. I think it was like two and a half years into the company. We had, uh, I think it was like a, Two hundred fifty million dollar kind of contract to start to deploy that. Um, and then I think maybe about a year or two later, we had landed a counter drone contract with SOCOM for a billion dollars, uh, to, you know, kind of scale out and get these technologies deployed. So, you know, these are pretty meaty contracts. Uh, you know, the, the, the nature of the defense business is, uh, you have these sort of very concentrated, large, uh, captures you're going after, which is, Uh, a bit high stress at times, but it's sort of the nature of what it is. Um, and, um, but that, that's what you have to do to succeed. Now, to, to actually be able to, to do those, to convince the government you are a good partner and trustworthy to do this, we've established a lobbying group day one of the company, right? We were on the Hill talking about what we do, why it's beneficial to national security and why it's beneficial to the taxpayer, why this is a …

AI assessment note: “landed a counter drone contract with SOCOM for a billion dollars”

Q history of these things being intertwined. Um, and more recently, a number of new companies in AI, like Applied Intuition, Scale, obviously, Anduril, Are strongly engaged in defense work. And again, I think that's in sharp contrast to even five, six years ago where it was a quite controversial or at least unpopular thing to do. How do you think this Renaissance came about? Like what changed or what shifted?

A Well, I, I think Vladimir Putin has a way of changing people's minds about the necessity of defense. And I, I think the, uh, invasion of Ukraine was, was a pretty large sea change in people's view of this, where prior to that, I think there was a belief that State on state conflict wasn't real. This wasn't gonna happen. Nobody was crazy enough to do it. But I think we've seen that if a dictator thinks that force can be successful as a means of getting to their political ends, that is on the table. They will try to use it, right? And so I, I think the recognition that we need to provide the best technology to the U.S. and to U.S. allies, that we need to be able to, you know, kind of solve these problems in a more ethical way. Um, be able to, you know, kind of deter this sort of aggression. That has been a massive shift from what we've seen, where this has become, uh, a very clear cut issue. I think the other side of it is, you know, you kind of look at US policy, uh, and views towards China, and I think we're in, still in the middle of the shift, but, you know, G has kind of made clear in a lot of ways, his intentions, like just listen to what he says, right? It is a very aggressive posture. Uh, he does not view the U S as like, we're gonna have some, you know, highly, uh, friendly relationship. Like, you know, the, it is a very tricky situation. And he has said repeatedly that …

AI assessment note: “invasion of Ukraine was, was a pretty large sea change in people's view”

Q What was that vision for you besides like, you know, obsessed with computers, wanted to work on them?

A Yeah, I, I more or less had two of them. So the first vision I had when I was a teenager was I wanted to be a computer science professor. So I just looked at what computer scientists did and thought this is the most amazing stuff I've ever seen. And I went to a science and technology high school and the way that it worked where I lived is like a really rural area. And so the science and technology It was a governor's school, so it was centrally located in each high school in these four or five counties that surrounded the governor's school got to send two students each, and so I was one of the two students that got selected from my high school to go to this thing, and my computer science professor there was this guy, Dr. Tom Morgan, and I just sort of felt like he'd opened up this entire new world to me, like it was just Thrilling to learn all of this stuff. And I was like, yeah, I want to be like Dr. Morgan. Uh, and a lot of this stuff for me is about, you know, like who those influential role models have been, uh, in your life. And so as soon as I like met Dr. Morgan, I was like, oh, I should just go be a computer science professor. And that was the path I was on until I was about 30 years old. Um, when I, you know, I was a compiler optimization and computer architecture programming languages person, and I got Pretty disillusioned with what being a computer science professor …

AI assessment note: “the first vision I had when I was a teenager was I wanted to be a computer science professor.”

Q When you think about, again, just like the focusing of in-sitro, what domains? Do you decide to work in first? Because this approach should be quite horizontal of, of course, then you have, you know, complexity of what that cellular model can be.

A It for sure is. And again, uh, focusing has always been a challenge in the sense that there's so many opportunities and how do we say no to some of them? So we've, uh, what we've done is tried to go in areas where we think there is both a large unmet need in the sense that the current tools that we're deploying are just not Very effective. And at the same time, where we think that the technologies that we are developing internally, uh, provide us with the unique differentiated advantage. So one of those areas has been, um, in neuroscience, because as we know, the unmet need there is humongous. There are so very few effective therapeutic interventions in neuroscience, and that's partly because the model systems that we've been using, specifically animal models, while one can quibble about in which other therapeutic areas they are more or less Um, relevant. In neuroscience, it is very clear that they're probably not, and that's one of the reasons why things work so well in whatever, curing mice of schizophrenia, whatever the heck that means, and then not having much of an impact in human schizophrenia, because it's not really even the same disease, right? Um, the other, so that's on the unmet need side, and on the opportunity side, we know that induced pluripotent stem cells are actually, um, relatively easily Um, uh, differentiated into neurons so you can actually see cellular p…

AI assessment note: “So one of those areas has been, um, in neuroscience”

Q They just built something and then it kind of took off, um, sort of separately. Do you have any sort of advice to founders who are considering changing direction or thinking of new directions for their company, or how do you How do you keep your eye out for the things that are really interesting or working that may or may not be the core thing of what you're doing?

A Well, I think the best way to do it is to find like the good ratio in your company between like exploitation and exploration. Um, and, and I think that's what a lot of startups are not always getting right. Um, not, not only before product market fit, but also after product market fit, I feel like sometimes companies before product market fit are experimenting kind of like too much changing directions every week. Uh, and, and I don't think you learn a lot from, from that. And then after market feeds, uh, product market feed, they can't like stop experimenting and, and kind of like trying new things and trying to stay away from the local optimum in a way, and looking more for like the global, global optimum. So for us, so what, what we've always done and, and I think we'll always do, um, with, with hugging face is to make sure that, you know, Whenever, always kind of like make sure to spend at least like 30 or 40% of the company's efforts on explorating new things and, and kind of like finding the long-term bets that is going to make you, make you successful. Um, and then give these, uh, you know, experiments and initiatives like a chance, right? Uh, for us, we were lucky that, uh, Thomas, one of our co-founders was leading This kind of like, uh, experiments. Uh, so it made it easier. Uh, but we have examples of other initiatives that started as experiments, uh, from team member…

AI assessment note: “find like the good ratio in your company between like exploitation and exploration.”

Q Um, maybe just to zoom out before we, uh, run out of time here, what are you most excited about in the next year of AI or, you know, expanding into the next five years?

A Um, I think, I think I'm, um, we, we talked a little bit about it in the past, and I'm really excited about, uh, biology and chemistry for, for machine learning, um, because I think, The way I see machine learning is really as kind of like this new paradigm to build old tech, right? It's, it's kind of like this analogy from where software one point was like the first paradigm and now we're in software two point, which is like machine learning power technology building. And so if you look at, you know, like the big sectors and the big kind of like impactful topics, um, that it could, it could change. Um, obviously biology and chemistry are, are kind of like, uh, up there. So, um, and, and we're seeing, uh, kind of like the numbers of models and datasets and demos on, on hugging face increasing. Like a few days ago, there was a release of, uh, bio GPT by, by Microsoft. Meta has been doing a lot of work on, uh, protein generation and prediction. Um, so I think there's gonna be really, really cool stuff coming up, uh, on, on these two topics.

AI assessment note: “I'm really excited about, uh, biology and chemistry for, for machine learning”

Q What, what else has been the most interesting in your, your work at Google so far?

A Yeah, I started working out, um, in representation learning and federated learning. Uh, so this is kind of the technology, representation learning in particular is kind of the technology underlying a lot of the, you know, deep neural networks of today, including GPT-III, GPT-IV, and so on. And so this is largely about, um, learning, uh, representations of text, of images, of other modalities, such that you can efficiently encode them, you can learn from them in the future, you can generalize the new, uh, text and images and so on. And so work for this really started, you know, back in the beginnings of the deep learning era, like in 2013 with convolutional neural networks and scaling those up and work to VEC around 20 15 and glove and all these things. And I think since then, you know, we've been working on technologies around self supervised learning, um, around, you know, doing that in a privacy preserving way. And so, you know, after a couple of years of working on that at Google, I had the opportunity to kind of quickly grow and start to lead a team. Um, I kind of got to the point where I was thinking like, okay, I've upscaled in a lot of ways. I've, I've, I've gotten to the point where I can mentor many other researchers in, in a lot of ways. And now it's a great time to be thinking about my next thing and, you know, going for something ambitious in terms of shaping the tr…

AI assessment note: “I started working out, um, in representation learning and federated learning.”

Q We're starting to get really good performance here. And so Can we do something that's in the medical domain specifically? And that was MedPalm. And so how did you do that alignment that you mentioned? Was it some form of RLHF? Was it some other form of fine tuning? Was it how you train the model to begin with? Like, what was the difference in terms of MedPalm versus Palm?

A Yeah, absolutely. I mean, when, when we tried evaluating POM in the medical setting, we noticed that our box on multiple choice questions performing pretty well, and when we took a variation of POM, the FlanPOM model, which was, again, work from Jason Wei and team, um, you know, this is an instruction to a model, a model that's been trained to follow instructions better. Um, you know, again, it was able to perform quite well out of the box, and this was the first model that was able to perform Above the pass mark on the med QA set of USMLE style exam, uh, questions. But then what we noticed is that when we evaluated on long form medical question answering, like actually getting the model to generate response, there was a lot of limitations. And when we compare, compared that to clinician performance, it actually didn't do super well. And so really that was the motivation for that med POM, uh, specific alignment. And so what we did there was really thinking about instruction prompt tuning, which was this technique Which we explored in that, in that MedPom paper, which is kind of a data efficient technique, and a technique that doesn't require too much data to work because, you know, getting labels from doctors is expensive, which took a bunch of expert demonstrations of good behavior from doctors, and then used that to tune the parameters of the model, and do that in a way that'…

AI assessment note: “what we did there was really thinking about instruction prompt tuning”

Q is, you know, more clinical decision-making around who gets a certain pharmaceutical. Do you view this as a technology that's initially a physician's assistant? Do you view it as something that helps with adjudication of medical claims and billing? Like there's so many places where this can sort of insert I'm just sort of curious, like, you know, where do you think you'll, you'll see this technology popping up first?

A Yeah, I think we're already starting to see it in some clinical workflows when it comes to documentation and building. I think there, there are a lot of, um, companies and people thinking about Taking models like GPT-IV and applying them in that setting. And I think that, that is definitely going to be something. I think that is also going to be something where players like Epic are going to be able to partner with existing models and I think potentially deliver real value there. Um, and I think, I think that's, that's, that's very exciting. I think that's something that also, you know, general domain models will be potentially quite good at as well. Um, I think where there might be more of a need for specialized models Is when it comes down to, um, kind of higher stakes workflows, and I think that might look in the short term more like a physician's assistant. And so imagine, for example, an agent that can work with a radiologist, help them interpret a scan, and leverage the benefits of AI to kind of help contextualize, um, you know, what a patient's medical record or any previous scans or different angles of scans that, that a patient has had, uh, to help a radiologist write a more accurate report. I think that's something, that's the kind of thing which I think, you know, is in the sweet spot of, of both feasible today, you know, leverages the benefits of AI in terms of taki…

AI assessment note: “we're already starting to see it in some clinical workflows when it comes to documentation”

Q mission critical AI applications in the past. Maybe on that note, like one last ask for you in terms of encouraging some optimism, you know, you're working on the state of the art in this field and thinking about the, um, the barriers to, uh, the applied use, like five years from now, like how do you hope we are using large language models in the, in the medical field?

A Yeah, I guess, I guess I think about this in, in two broad buckets. I think there, there are two broad types of things that we can do for large language models in the medical field. I think the first is increasing the standard of care very broadly. And so that looks a lot like, you know, increasing access to health information, uh, providing assistance to physicians. So the radiology example I gave earlier, um, potentially clinical decision support, like double checking a doctor's decision or quality assurance for a radiologist report. So if, if, you know, a radiologist is dictating a report, they say no plural effusion scene, but then it's written down as plural effusion scene, then maybe an AI double checks that and just make sure that that's, that's what was intended. I think augmenting telemedicine, I think is, is, is kind of a short-term opportunity that I think in the next five years is, is very achievable. I think the other big bucket of things that is very much achievable, um, is augmenting scientific workflows. And I think this could be a longer term thing than five years, but I think there's also short term things that we can do as well. So thinking about looking at, you know, correlations across modalities and existing data to find novel biomarkers for existing diseases that we know about, or, um, kind of using large language models as research assistants. Um, so I t…

AI assessment note: “I think there are two broad types of things that we can do”

Q the basis for the next generation of these types of models? Their performance? Where does it asymptote? How do you think about scalability? How do you think about the underlying silicon that drives it? Is it a data issue? Is it a compute issue? Like, I'm just really interested in how you think about more broadly these really large scale models since you folks are building many of them now.

A Well, the, the incredible thing about where we've got to at this point is that all of the progress, in my opinion, is a function of compounding exponentials, right? So over the last decade, the amount of compute that we've used, uh, to train the largest models in the world has, um, increased by an order of magnitude every single year. So I went back and And, and had a look at the Atari DQN paper that we published in. Um, and that used just two petaflops, right? And some of the biggest models that we're training today, uh, at inflection, uh, use ten billion petaflops. So like nine orders of magnitude in nine years is like just insane. So I feel like it's super important to stay humble and acknowledge that. There is this epic wave of exponentials, which is unfolding around us, which is actually shaping the industry. And so when it comes to predictions, you have to just like, look at the exponential. It's pretty clear what's going on. That's just on the amount of compute side, the data side, everyone's super familiar with. We're using vast amounts of data and that's continuing. But I think the other thing that people don't always appreciate is that the models are also getting much more efficient. So, um, you know, one of the big breakthroughs of last year, which got some attention, but probably didn't quite get as much given how many breakthroughs there were was the chinchilla pap…

AI assessment note: “all of the progress, in my opinion, is a function of compounding exponentials”

Q the level of society. You know, we were thinking about it in the context of just like, how do you get more people to use this thing, you know? And so I think it's really interesting that people then later realize the big ramifications of this. In terms of, you know, how that actually cascades in terms of social behaviors and other things. How did that lead to starting DeepMind?

A Well, it was clear to me from that moment on, like I left Copenhagen in 2009 thinking this is not the path to significant positive social change. It still needs to continue. And I support those processes obviously, but I'm just saying it is just not something that I feel I could continue to work on full time. And so my heart was set on technology at that point. So I reached out to Demis who, uh, was the brother of my best friend from when I was a kid. We got together, we had a coffee, we went, and actually, we played poker, um, at, at one of the casinos in London, because we both love games, both super competitive, uh, both good at poker, and on that night, I think we both got knocked out pretty early in the tournament, so we sat around drinking Diet Coke, uh, talking about ways to change the world, and we basically were, you know, having exactly this conversation, like, You know, is it going to be, I mean, obviously at that point I was mostly inspired by platforms and software and social apps and connectivity and so on. Whereas, um, you know, Demis was way more in the kind of robotics land and sci-fi land. I mean, he, he was, he was fully thinking that, you know, the way to manage the economy, the way to make economic decisions was to simulate the entire economy. Right. And, and he thought that he was very much obviously had just come off the back of his games like evil genius…

AI assessment note: “So I reached out to Demis who, uh, was the brother of my best friend”

Q I guess last question on sort of your time with Google and DeepMind and because I think there's a lot of really exciting things to talk about in the context of inflection and sort of the broader field and world. What are some of the things you were most excited to have the team create at DeepMind over the years or some of the breakthroughs that you're most proud of?

A Yeah. Well, I mean, in some ways we, we, we definitely sort of pioneered the deep reinforcement learning effort. And I think, um, you know, in principle, it's a very promising direction. I mean, you clearly want some mechanism by which you can learn from raw perceptual data, and that directly feeds into a reinforcement learning algorithm that can update And essentially iterate on that in real time with respect to some reward function, whether that's online or offline, like directly interacting with the real world in real time, or it's, you know, in, in a kind of batch simulation mode, um, you know, and, and that turned out to be very valuable for a specific type of problem, um, where a game-like environment had a very structured scalar reward, and we could Play that game many millions of times. Um, that's part of the reason why we started the Alpha Fold project, because it was actually my group that was, um, looking around for other applications of DQN-like, AlphaGo-like, uh, tools, and, uh, in a hackathon, um, that we did one week, um, someone stumbled across, across this problem. We'd actually looked at it back in 2013 when it was called Foldit. Which was, uh, a very small scale kind of version of this.

AI assessment note: “we definitely sort of pioneered the deep reinforcement learning effort”

Q Got it. And you got obsessed with robotics. You weren't working on logistics at the beginning. Uh, can you talk a little bit about Remotiv?

A Yeah. I mean, when we, you know, the, the tricky thing, and maybe this is true for more startups than you realize, the tricky thing is like, we didn't know we were starting a company when, you know, all I knew was me and my co-founders didn't have jobs. And, you know, it seemed like a cool thing to do to build some robots. We put something on Kickstarter. We ended up selling a 150,000 dollars worth of robots on Kickstarter, which was a huge amount of money to us at the time. And we ended up building those robots in Uh, the apartment of, it was, it was my apartment at the time, but I mean, it was technically really Tony Hsieh's apartment. You know, Tony was kind of a mentor and inspiration to me. I had just read his book. Um, he lived in the exact same dorm I did in college, just 10 years ahead of me. And so seeing like what he had built and Yeah. I didn't even know building a startup was a thing that you could do, but you know, we, he had all these apartments in Las Vegas. He gave us a bunch of them. We started building robots in those apartments and shipping them to people all over the world. And these were really simple, right? Cause we had no money, no credibility. We weren't that good at building robots at the time. So these were really simple. They were basically laser cut out of acrylic and you could attach your phone to it and it would become a little autonomous roving p…

AI assessment note: “We ended up selling a 150,000 dollars worth of robots on Kickstarter”

Q Yeah. Before we, um, get into that, maybe you could just help our listeners visualize the problem of how a delivery actually works. Can you just walk through if it's, you know, the Ministry of Health or one of those hospitals, like, uh, how one drone delivery happens today?

A Yeah. The, you know, the idea is really simple. It really is that any doctor, nurse, or, you know, healthcare technician can push a button on a phone and summon, you know, we, we talk about what we do as teleportation. It really means they can basically teleport a product directly to their GPS coordinates. That's, that's the vision. It should, it should, you know, feel like magical, uh, teleportation. It should be nearly instantaneous and it needs to be reliable. And that also means we need to be able to work in any weather. Uh, so what's practically happening behind the scenes to make that work is that, uh, you know, any of our, any of our customer hospitals. So today we serve 3000 hospitals and health facilities across the world. Um, anybody at those facilities can press a button on a phone and place an order. That order is then transmitted to our fulfillment center. So each of our distribution centers, Zipline now operates, uh, I think by the end of January, we'll be operating over Global distribution centers. Each distribution center has a fulfillment, uh, area where we stock all of these different medical products. Blood was the first thing that we stocked. Uh, so we get the order, uh, we will confirm it. We'll pack the products, whatever that is that's needed into a box. And then that box gets handed to flight operations. Flight operations then, uh, will pre-flight an air…

AI assessment note: “anybody at those facilities can press a button on a phone and place an order”

Q Uh, so how did you decompose such a, a hairy problem? Was there like, you obviously discovered more of it as you got into it, but was there an overarching philosophy?

A Uh, you know, there was no plan. We were, we were severely naive when we launched in 2016. There's no doubt about it. I mean, we thought that We had done some testing in Half Moon Bay, California, and we thought we were ready to operate at national scale in Rwanda in any weather. It sounds kind of silly when you describe it in retrospect, but you know, we, we were so confident when we launched and for the first nine months, we only served one hospital, not 21. You know, we were wise enough to say, well, we're going to onboard the first hospital and make sure it's working reliably for that hospital before we try to do more. We thought that was going to take two weeks. It ended up taking nine months. Uh, we were killing ourselves, pulling constant all-nighters, trying to fix all of these problems that suddenly became apparent when we started trying to deliver reliably to this one hospital. And I don't think this was a philosophy. Like, I think we were, you know, it's not like we were wise enough to know how hard this was going to be. Maybe if we knew how hard it was going to be, we wouldn't have done it to begin with. But, uh, I do think you might've thought you needed

AI assessment note: “there was no plan. We were, we were severely naive when we launched”

Q Speaking of that, did you, did you collect your own data for this? Because I assume like, you know, there's not like good location, trajectory, audio at. 80 miles an hour pairs just sitting out there on the internet for you.

A Yep. A hundred percent. And so we had to collect, you know, we had to collect our own data by flying a lot of microphones on airplanes, but we also collected a lot of data by designing this little system that we had at our office where we literally just had a microphone array sitting on the ground, 24, seven recording airplanes flying overhead. Uh, so yeah, we had to build the data set ourselves, but I mean, the amazing thing is that four years in, you know, we rolled the hardware into full production, uh, and then, you know, we, Turned it on in shadow mode across a lot of the countries where we operate. And we've now gotten active permission from regulators and are actually flying using it where it can now control the aircraft if it needs to, to deconflict from another airplane. Uh, and you know, that system, it actively listens. It can identify the make and the model of an aircraft. Is that accurate?

AI assessment note: “Yep. A hundred percent. And so we had to collect, you know, we had to collect our own data”

Q and then how that actually leads to some of these solutions? And, you know, maybe more generally, I know that people Have been focusing on sort of mixture of expert models and things like that, like MOEs, and so it feels like a derivative of some of this work in some ways, and so can you just talk about how Realm was initially set up and then how it's evolved?

A The basic idea, if we were to describe it, is let's say the user provides some sort of input to the model. I'm going to describe the inference time way of using it and then the pre-training that was involved as well. So the user provides some sort of input to the model. The model is then able to Encode, embed that information into a dense vector. So this is a vector that will be situated in a larger vector space where we've also embedded other documents. So any other corpora we have on the web are converted into dense vectors as well, and a nearest neighbor search method is used to find the closest documents to the embedding of the input. And once you have those documents, you can then encode those as well in some form that allows It's the model to then use cross attention over those documents. And based on that, it's then able to make a prediction. So the way in which the model depends on that information is through the cross attention. And there are multiple different ways to train the cross attention to kind of use that information. And in the Realm paper in particular, we tried to learn that from the language modeling task itself. So I guess for our listeners, many of you are probably familiar with the fact that language models predict the next token. So given a sequence of existing tokens, you predict the next one. At the time of the Realm paper, we were doing something ca…

AI assessment note: “in the Realm paper in particular, we tried to learn that from the language modeling”

Q and they'll figure out how to, Right, for iOS and stuff like that. How do you think about that in the, in the ML world? Because there is this sort of dogma right now, I feel, that people believe that only somebody who's trained in LLM before can work on an LLM-centric company or things like that. So I'm just sort of curious, you know, how you think about that.

A Yeah, actually, my thinking about this was already pretty clear because of having seen how OpenAI went about this. If you look at the people who Did GPT-III. Eventually they went on to start Anthropic. All of them are basically from physics background. The CEO Dario's a physics PhD and Jared Kaplan is a physics professor. He wrote the scaling loss paper. Uh, so they basically open AI really succeeded in bringing these extremely talented physics people who wanted into machine learning and, and they came into the sort of language models and scaling. And that paid, paid off well for them. And similarly, their whole engineering team, if you look at an engineering team, it's just Dropbox and Stripe people. All like software engineers, solid people who can build infrastructure front end, and now they're doing AI work. So it's always been clear to me that, uh, in order to work in AI, both research and product, you do not need to already have been in AI. And that's being shown clearly with our, uh, with, with Johnny Ho, our co-founder, he was not in AI. He was a competitor programmer, a trader. He had worked at Quora for a year, but, uh, he's as good as anybody can get in picking up new things. Uh, so the other thing also is that LLMs are sort of in this weird territory where the people who use the LLMs for building stuff, Understand it better than the people who actually did gradient …

AI assessment note: “in order to work in AI, both research and product, you do not need”

Q I'll ask a dangerous question, but what do you, what do you think the future of search looks like? Right? Five years plus out. Do we get, do we still have monolithic horizontal providers if the players change? Um, do we get more embedded apps like contextual search as a feature in different places? Are there agents that do things for you? Like, what do you think it looks like?

A I think there's this phrase that's becoming popular. Uh, I think Karpathy was the first one who tweeted it and Satya Nadella is also using it. It's called an answer engine instead of a search engine that directly tries to answer your question instead of providing you a bunch of links or just snippets from the first link. Uh, so we believe in that, like perplexity is the first conversational answer engine, uh, like truly the first. I think nobody built it before us. I believe answer engines will become its own segment, market segment. Like, just like you have a default search engine, there'll be a default answer engine over time if these things really work. And The burden of like getting ranking and search right will, uh, reduce in the sense answer engines can do more heavy lifting than search engines. As these things get really good and like way fewer hallucinations, um, and even, even if they do hallucinate, people can still go and click on these links. They will eventually prefer this experience over the regular like 10 links or 20 links UI that Google has. So that, that's something I'm pretty confident about. Uh, and, and I think the sort of asking follow up questions will become more of the norm. The number of queries and perplexity that go to at least one follow up has been increasing ever since we released the chat UI. So that, that will keep going up. People will get use…

AI assessment note: “It's called an answer engine instead of a search engine that directly tries to answer”

Q We're gonna be talking, I think, about both, um, your experiences in research and academia, and then we'll also separately be talking about Together, which is a company you're involved with now. Um, could you tell us a little bit more about what the center does and what you're focused on?

A Yeah, so the Center for Research on Foundation Models, uh, started two years ago, is under the Human Centered AI Institute at Stanford, and the main mission of the center is, I would say, to increase transparency and accessibility to foundation models. So foundation models are becoming more and more ubiquitous, but at the same time, one thing We have noticed is the lack of transparency and accessibility of these models. So if you think about the last decade of deep learning, it has profited a lot from having a culture of openness with tools like PyTorch or TensorFlow, data sets that are open, people publishing openly about, uh, paper, about the research, and this has led to a lot of community and, uh, and progress, uh, not just in academia, but also In industry with different startups and hobbyists and whoever just getting involved. And what we're seeing now is sort of a retreat of that open culture where models are now, uh, being only accessible via APIs. We don't really know all the secret sauce that's going behind them, and there's sort of limited access.

AI assessment note: “the main mission of the center is, I would say, to increase transparency and accessibility”

Q What's the relevance of China in all of this?

A I think so. If the last decade was about sort of biologics and these new modalities, I think the next is going to be about speed and cost. Like people want more drugs and they want them cheaper. And China is very good at things related to speed and cost. And so all of a sudden in the last couple of years, you've seen this rise of Chinese biotech companies that are able to create molecules and bring them to patients in clinical trials in China, just early phases of clinical development. Uh, really fast and really cheap, even in some of these new modalities. We've seen this huge uptick in pharma going to China and buying molecules that they typically would have bought from American biotechs.

AI assessment note: “China is very good at things related to speed and cost.”