You know we talked a lot about innovation, technology innovation, organizational innovation. All innovations have obstacles, right? There are challenges and our entrepreneurs face them daily and overcome them. But I think one of the most difficult challenges if you live in the Bay Area is to cross the Bay Bridge to come here. And if you live here, you know what I'm talking about. Both of our distinguished guests have done that challenge, overcome this challenge this morning, so thank you for coming. Really, this topic of AI, digital AI, physical AI, there could not be two better stalwarts, industry titans than Shankar and Yan. Just briefly, Shankar Sastry was the dean of engineering at Berkeley for over a decade and has been the professor leading robotics innovation for multiple decades. He's been my mentor for 30 years. And Yan has built some of the most iconic companies in the age of AI and started this discipline before AI was AI. So, we're really grateful to have you here.
And Yan, I'll start with you. When you started Databricks with some of your colleagues more than a decade ago, what was the opportunity that you saw then in AI and what is the opportunity that's still left today?
Yeah. So you know it was a great journey. So I think that when we started Databricks, what we focused on in terms of persona was the data scientist at that time, and basically focused on data scientists who need to process and do analytics and get insights from data at scale. So we focused on cloud deployment, but that was our persona. And then it was a journey, right? Because initially we did a lot of, and in the first slides we talk about AI at that time was machine learning, and it was kind of the old machine learning, logistic regression and collaborative filtering, things like that. But we are showing how to scale it. And actually, we started to sell this, you know, our product, and we had it after some time, you know, we release it and we see that it's a hosted product, so we see that people use it, we're happy. And then after a few months, you know, we hire a marketing person and say, well, you know, you need to go and get these people to have testimonials, right? To add to our website. So he goes there and start talking with people and say, well, no one is really doing machine learning. And I'm saying, why? Because this is what I thought you bought it for. Well, you know, we didn't have the data either, like we didn't have enough data, the data quality was not there, we need to clean it up and so forth. Now the good thing was Spark was very good also for data preprocessing and so forth, so that's why we are seeing a lot of people using it, but they were using it for something else and we sold it for machine learning. Of course, eventually they clean up the data and so forth and they started to do more machine learning. But the lesson here obviously, like everyone knows, the data is extremely important.
So how do you see that today? How has that data problem been solved and is that the fundamental reason why we see the AI adoption increasing today? What's really been the unlock?
I think it's everything. Obviously, data plays a very important role, right? I mean, it's like that's why people are spending a lot of money to get the data, to get new data. They spend a lot of money to get the data being labeled by experts. And they spend a lot of money creating new data through agentic systems, right? You try to accomplish a task and then you create a lot of data in the process and you take the traces with whether the task was successful or not, are fed back to train or pre-train, post-train the models. So there is no question, I mean, it's like data is extremely important. Of course, architecture and compute are equally important. But you cannot do it without data because at the end of the day all these models are still prediction engines, right? At the core they predict the next token. Of course, they have a lot more semantics baked in implicitly, but at the end of the day they are still prediction engines.
If you were starting a company today, and I know your lab is a factory of startups that have done really well, what is the next big thing that you would want to focus on given the current state and where we're going over the next decade?
I don't, you know, maybe it's hard probably to say very precise, but if you take a step back, you also try to figure out what are the problems out there, right? Because eventually whatever you start it has to solve a problem. If you don't solve a problem, if your product doesn't solve a problem, people are not going to buy it. And if you look about all AI, what is AI doing today? In general, when you build a product or you solve a problem, there are fundamentally three stages. One, you come up with the idea, with specification, whatever what you want to do, what you want to solve, what product you want to build. Then you build it and then you evaluate it, right? The problem is solved, whatever the product is going to sell and so forth. And then there is iteration until hopefully you get to the desired outcome. So in many cases what these large language models are doing is, in some sense not in the physical world but in the digital world, they commoditized the second part, coming up with solutions, ideas and so forth. But this means that the other parts are going to become the bottleneck. You know, coming up with clear specification what you want to build and so forth, and then evaluating that what you build really works as intended, right? It's aligned with your intentions. That's kind of, it's very hard. So you have to, you know, I would try either to try to alleviate the first problem, how you are going to come up with much more clear specification, formulation of the problem you are going to solve, which is relatively the last one because if you don't have kind of clear specification, how do you know that what you build, what AI built, is going to do what you want it to do? And then kind of evaluation. And the evaluation there are many, many aspects. One, it's about accuracy, but the other one it's also about performance, how long it takes to do evaluation. Because that's why, as long as you are remaining in the digital, you can almost have arbitrary speed up. The moment you are going to hit the physical, or including humans, right, things slow down. Right? For instance, you do drug development, I can give you, you know, I can come up with 100 kind of things and proposal of new drugs, but you still need to go through clinical trial, you need to wait fundamentally years in order to see or there are side effects and things like that. So that's kind of this, you know, once you remove the bottleneck about how to come up with a solution, then it's about how to verify. The same way with programming, right? It's like you can generate very easy right now impressive programs. Now the question are they going to work and are going to deliver, particularly in use cases which are high value and so forth? Because at the end of the day, not Anthropic, it's liable, if something goes wrong, it's you who use Anthropic are going to be liable. So you better make sure that that is going to work. And the more complex the artifact generated, the harder it is to say this is really working as you intended.
So let's stick with that. So the next step of LLM, if I'm hearing you right, I think what you're saying is for coding, because code can be formally specified and verified, we had fast adoption. But if it's drug development or physical AI, autonomous trucking, the AI adoption may not be as fast.
Yeah, it won't be, you know, it's going to be fast, but things will not move as fast as people believe because you eliminate this kind of middle, you know, that maybe was a bottleneck before, I don't know, in some cases. And okay, that you commoditized that, it disappeared, but the other bottlenecks about like we said, verification, evaluation and so forth, it's not going to change. It can become even more complex, right? And by the way, with the programming, it's not as simple. It's, you know, of course these tools are super impressive, but it's not a panacea. I can tell you, I've been, you know, everyone who use it, how many people use Claude or Codex for doing really kind of more complex, say, infrastructure software?
It's hard. Like, I mean, look, it's like Anthropic, you know, someone forwarded to me, you know, they hire software engineers and pay them a lot, right? You know, it was just something like base salary 570k. So why do they still hire them? Why do they use Claude Code, right? But because, you know, it's very hard, you know, formally specified programs is damn hard. It's very hard. Actually, we kind of have some kind of loose specification and then verifying them is actually, even before that, if you really build infrastructure software or something like that, most of the time you don't spend in writing the code. Most of the time you spend in debugging, improving, you know, maintaining the code. This is where you spend the time even before that. So it's, again, you see that I'm just trying to say that you see these kind of things even in programming and software engineering, which is true is much easier to verify and much quicker to verify than other artifacts. But even there, you see these challenges.
What I find so fascinating about this conversation is that these are two of the world's foremost scientists and innovators in this discipline of AI. And I think from here, I don't want to put words in your mouth. What I'm hearing is that I don't think you're saying that we're going to get AGI or ASI anytime soon.
Oh, yeah. Where, which is what we hear from, you know, this is, if you're writing the S1 for Anthropic, you would say it's coming soon, right? Sure. I mean, it's again, it's what is AGI and how much it matters, right? It's like because it's like in time, and this is going for decades, there are tasks for which computers are better than humans. And time goes on, more and more, and the computers will be better than humans at more and more tasks. There is no question about that, starting with the calculators, right? In at the end of the 60s, 70s, right?
Yeah. Can I pick up please? You know, I think from listening to the talks this morning, I was really struck about how people feel that it's so frenzied. And I think for Yan and me, if I can also say this, you know, the future's taken a long time arriving. I think AI, you know, I do AI 30 years and suffered through these winters. And you know, I mean at SRI there were people doing a star 40 years ago. I was at MIT last week and they were talking about Norbert Wiener pushing cybernetics and then John McCarthy who, you know, he was trying to make tenure at MIT and he made up the term AI to distinguish himself from Norbert Wiener. And of course, he decamped to Stanford. But anyhow, you know, I think there really is more of a continuum. I think you're feeling so. And to Yan's point, you know, the future takes a long time to arrive. You know, the initial prototypes are easy, but then to integrate them into the fabric of society, that's what takes time. You know, there was a professor in Munich, German military academy called Bundeswehr, who he was in Munich, but he had a Mercedes drive autonomously to Berlin from Munich. And that was in 1990. And you know, I think it is not yet integrated. So I think that this is undoubtedly a fantastically excited time. I wish I was 40 years younger because I think the next 40 years are going to be just a wild trip. But I think I would not feel frenzy. I mean, just the last thing that Yan said, you know, it's not easy to write specifications. It's not easy to ask for what you want. And somebody said you shouldn't go to universities. And maybe I'm a little conflicted. One of your wonderful founders said this, by the way. I think that, you know, universities need to smell the coffee. So don't get me wrong. But I do think, and somebody else spoke about meta-learning. You know, I do think that we still, there's a big role to be spent in terms of asking for what you want which are integrated into our value systems. Not just in software. It is true that what you call embodied intelligence, Jensen calls it physical AI, embodied intelligence, embodied in the physical world is going to be big. But even in software, you know, about agentic, multi-agentic systems negotiating with each other. I mean those are all things which need to have built into them values that we hold dear and that is going to take time. And that I think is sort of design, you know, it's not, Yan talked about making things and then do people care and do they correctly mirror your value systems. You know, he talked about Databricks in the early days, I remember he used to say we spend all our time doing what he called data proctoring, which is cleaning up the data before the algorithm. Whereas in the university we spent all our time just teaching the algorithms but, you know, operated on junk, you know, till you had to do the proctoring. So I do think that it's a pretty exciting time but it's not so out of control. It doesn't feel out of control to me.
Yeah, yeah, it's not so. So I think, look, the way I think about it, AI is still a tool. It's like a calculator, it's like a computer and so forth, your car. There is one difference though. In general, the tools we are using is pretty clear what they do, what they don't do, what they are good at and they are not good. In the case of AI, this kind of boundaries are amorphous, it's not well defined, it's moving. So therefore, you are going to have all of this, you know, very high expectation because you don't know what it's good for. It's like you generalize. People are, you know, there is this quote which I really like, it's, you know, nothing is as good as you think and nothing is as bad as you think, right? As humans we have emotions and we try to magnify any kind of things. But what I'm, that's what I'm trying to say because, you know, you just use any of these tools and go deep and deep and you'll discover the limits and clearly the limits are going to be lower than you initially hoped. And I'm saying that as someone, we try very aggressively in our group to use these tools to improve our research, you know, all of this kind of AI researcher and so forth. We started, you know, one year ago and I just force my students to use these tools for their research. And you can see, you know, it's like that's what I'm saying, when you put it this way, you kind of see the boundaries at some point and then things are going to change, the boundaries are going to move. But I think that's one of the keys, right? It's like if you can figure out what is good for, what is not good for, I think that's one of the keys to use these tools extremely effectively. And yeah, again, AGI is, yeah, it's a continuum, it's not well defined, which is good because everyone, you know, people can claim, you know, AGI was achieved or has not been achieved, right? Everyone is right.
Everybody's right. Shankar, I want to, can I just...
AGI, you know, I think too much is made of this term AGI. On the other hand, if you sit and think about intelligence, some parts about machine intelligence are retrospective, learning from the past. Other parts of intelligence are about planning for the future. But then there's a third part about reacting to why the future isn't what you think it is. So I think intelligence, human intelligence has within it all these tenses. I think when we say machine learning, generally speaking, you're talking about what you do from the past. But I certainly hope the future isn't going to be the past. So if you just assume that you're going to use machine learning, I think it's going to be a pretty boring future. And I don't think it is. So I do think that artificial, and to Yan's point, as human beings, you know, we don't tell somebody these are your specifications and I'm going to check that you meet them. In fact, you expect humans to evolve. You expect their capabilities. You know, that's what it means to have a team. You know, a human bot team, a human software team has to be one where they co-evolve and you build confidence in somebody. So those are all attributes that, those are essentially human attributes and to lump them all, okay, you could call it artificial general intelligence, whatever you want to call it is fine, but exactly what that is is probably a more nebulous goal. And it does run, and to Yan's point, it does run counter to what we teach in engineering where you say here are the set of specifications, build something to that and we'll certify it and you'll do it. That's I think the premise of AI is to go beyond that and that's exciting to me.
