Our next session is with two guys who founded one of the iconic companies of our era. Ion Stoica is a professor at UC Berkeley. He founded Databricks with five of his students and visiting professor Ali, and Patrick was one of the students. So I'm sure you're the smartest one of the lot, right?
But, but maybe Ion, I'll start with you. We've been talking about pilots to production and you've been very, very bold about saying that's the biggest issue. What are the problems with it in your perspective? Because you're coming from academia, but you've got a bunch of companies. What are the big issues you're seeing here?
I think the problem with AI is a problem we already have even with software systems before: it's easy to build prototypes, it's hard to productize them. And productization is about reliability, making sure that what you build, that kind of feature as a prototype, is going to run, you know, predictably and reliably in production, right? It's doing what you expect and so forth. And I think that this is hard, even harder with AI today, right? So it's about how you are ensuring reliability. Some people say, you know, alignment, you know, how is it going to be aligned with the user intent, right? That's kind of one of the hardest problems. Of course, there are harder problems in terms of the cost, which you all know is quantitatively you can measure that. So we know that it's expensive, but yeah.
Yeah, what about like security and trust and those pieces?
Security and trust basically go side by side with reliability, right? Because you don't want, you know, like I've said that it's aligned with the user intent, you do not want someone to... certainly the user intent hopefully is not to attack or, you know, to breach some other company, so to speak. And I think, but there it's, and it's again, it's like in many of these cases, it's like we've done that at least for some instances like for decades, right? In terms of security, kind of reliability. So what you do, kind of testing for kind of the particular workloads you think that you are going to have for particular use cases, and you are going to, when it comes to security and everything, you are going to put the best practices, you are going to use the best practices, right? Including securing the sandboxes. So this is what you are going to do. So I think that what we see right now is that a lot of people who are kind of deploying AI, it's almost maybe they are not familiar or something like that, but kind of forget some of these core best practices which have been, you know, people have been doing for years. Yeah.
Can I add on? I know that wasn't my question, but...
By the way, I noticed that I'm an uppercase CO-FOUNDER and Ion is a camelCase co-founder.
I think that means I'm more of a co...
That's above my pay grade. I think I'm more important.
You're also on the left. You're the first, right?
I'm first.
But I just wanted to add on to what Ion was saying. So one thing I do at Databricks personally is we have about 10,000 employees and I'm the one who's kind of in charge of our own internal rollouts of AI in the company, and you know, about 3,000 of them are software engineers. So it's quite interesting, the technology is really powerful, but the safety and governance stuff is actually what is the biggest bottleneck in our ability to consume more at this point. Some of it is alignment like Ion said, and I think alignment is a difficult technical problem for the people making the models, but it's even more difficult when you have many organizations using models because what constitutes aligned behavior for Databricks is specific to Databricks. So for instance, like we had an issue where one of our software developers said, 'Hey, I'm trying to test this. Can you help me test this feature?' And I think it was Claude or Codex, I don't remember. It decided that one objective sub-goal in its objective of testing this feature was to take all the source code and upload it to this location on the public internet because it needed to upload an image artifact that it was going to reference from some other place. And you know, it's not like that violated any obvious problem. If it was like an individual at home using Claude Code, maybe that would be completely fine. But at Databricks, having a copy of our source code on the internet is like a major security breach. Now we have monitoring, we caught it immediately, it was fine. But I think even in principle, it's difficult to do perfect alignment because what constitutes aligned is different for different parties. So a major thing we focus on is like guardrails and our own network sandboxing and all of this stuff. But I think the bottleneck is in many ways the bottleneck towards broader value is about security, safety, cost and these other things, that the raw intelligence capabilities are very good.
Absolutely. Yeah. It's been... this is the thing, it's easy to build and you are going to see all the time fantastic demos, right? The capabilities are there. The question is how you are going to apply these capabilities in production safely, reliably, securely.
And we were joking, even the AI companies themselves keep having security breakouts, you know, and it's funny because they market it as saying, 'Oh, our AIs are so intelligent that they managed to escape from our...' You ask any security person, like they just had basic security mechanisms not in place.
But we saw some really cool demos up there for robotic stuff. It was even though they were demos, they were really cool from the skilled guys.
And I want to piggyback on something which I think is very important, what Patrick said. It's actually in most of the cases when you go back, you know, always you are going to get, you know, you get this kind of surprising behaviors, whether it's kind of escaping the sandbox or whatever, and or, you know, uploading to the internet the source codes. But when you go and look and you try to understand, it's actually the AI as a tool did what it was supposed to do given its context, and the context is coming from the instructions you give it and from the training data it was trained on. This is the context. Now if something is missing from that context which you believe it should have been there, you don't even think about it because it's common sense, then AI will just exploit that gap. They are very effective at exploiting the gaps because they are relentless, and if there's a context gap, that's a gap.
Patrick, you talked about model intelligence is commoditized, right? But lots of money is going to the AI models, but the data infrastructure below that, I know you're Databricks, so I know your answer to this, but how's the split between the models and the data and clean data? How should the audience be thinking about it?
Yeah, well, by the way, I would say I don't think I used the word commoditized in the last section. But I do think that the pure intelligence has almost saturated to the point where for typical enterprise work, so for what we all do when we go into the office in terms of like raw intellect, raw reasoning ability, I actually think, you know, I hate to say this, but I think many of the models are smarter than I am. If it's given exactly the same amount of information, I think it will reason in a better way than I do. But then the question begs the question, okay, why do we not have then 100,000 additional employees that are very productive at Databricks, you know, or infinite additional employees? And the issue is kind of like what Ion was talking about, they don't have sufficient context. And actually it makes me appreciate how much implicit context there is for anyone working in a company. Some of it is just like the background diffusion of information into our heads that we're just slowly, tacitly picking up during our day-to-day lives. But then at Databricks, what we focus on therefore is like that bridge between your company's context, what is happening in your company, and that could be big picture things: who works where, what's the org chart, what's in our structured data, you know, that's sort of where we started, Databricks, what's in our unstructured data. There needs to be a bridge between that and then this sort of disembodied reasoning thing that's very, very intelligent. And so that's certainly where we're focusing a lot of our time and effort is on bridging those two. And then the question is how do you do that without giving up your own IP in some sense, that the collective intelligence of every employee in your company is the IP of your company in most cases, like very in the US and advanced economies where we're knowledge and service industries. It's not like the IP of most companies is a patent that they have about how to make some widget. It's mostly the collective knowledge of the people in the building. And so what we focus on at Databricks is like how can you take that collective knowledge, infuse, you know, leverage it with your AI tools while also keeping it private and part of your IP basically.
So let me go off script because a lot of the stuff you're talking about is context. Patrick, you guys have a pretty in-office culture if I understand correctly from what I know, right? Is there some value to that because a lot of people are wrestling with kind of in-office, remote, but I think we're back, right? But what's your culture out there? Because I'm thinking about context in a human sense almost in some ways.
By the way, it does give you appreciation for how richly humans communicate because you see these AIs trying to communicate with each other. It only can, it's teletype, it's back to like the oldest computer interface, and there's so much nuance missing in the communication. And it gives you appreciation of how much humans are conveying to each other beyond just the words that we're saying. Like for instance, I move my hands a lot. But at Databricks, we were very focused on the big thing was during the pandemic everyone decided, let's be, we all went home because we had to, like in the whole world everyone worked at home. And then I believe what happened is people developed this false belief that like, oh, it seems almost equivalent to what we were doing before, like I'm able to talk with the same people and we have such great VoIP technology now. But we at Databricks kind of felt like we're not sure we're ready to become a super distributed remote company just because of this. And even during the pandemic we kept hiring in our locations, so like you had to join Databricks ready to go to an office even though there was a period where we're not in the office because for us pretty quickly it became clear that there's a lot missing.
And there's so much quick exchange of ideas that can happen when people are physically together, and we just felt like we didn't want to let go of that. So we were atypical in that, and now I think it's kind of funny because the trend has moved back in the other direction where people are suddenly re-remembering that humans like to be around other humans.
Yeah.
Ion, so the winning architecture is going to be, you know, a distributed model to some extent, right? It's not going to be one giant model, it's going to be composed almost like software, lots of little models, lots of agents. How do you think, what's your view of the future? You're a professor, you look 10 years ahead, 5 years ahead of us, right? And so how do you think about this? You start great companies, you just did AnyScale, so it's like, what's your view of the future here?
I don't know about 5, 10 years, you can kind of predict at this stage given what happened in the past few years. But I think I would say a few things. So first of all, if you are talking about the enterprise and actually what its customers like Databricks customers and many of your customers, it's like the tasks they have to do, many of them are repetitive and at the end of the day are not super complex. It's not about solving the Navier-Stokes problems, right? And for those, you can actually, as people know for now for a long time, you can actually fine-tune a model, post-train a model which is kind of reasonably a few billion parameters, and you can make it work well at that task better than one of these frontier models. And the added benefit because it's so small, it's not that creative, so it doesn't do things you do not want it to do, right? So clearly there is no question about that. So when you are going to have a task, fine-tuning, post-training a model is going to give you a lot of benefits and it's going to be better and cheaper than bigger models. And then if you are going to have a lot of tasks, then what you are going to do, you are going to put a router or you are going to combine maybe, you know, multiple tasks that can be a pipeline, right? You can have more agentic, more kind of complex agentic applications, right? Or you are going to distribute the different tasks to different, like a routing, right, gateway, like also you have at Databricks. So I think that's kind of for the enterprise, this is what you are going to have and you have to do, right? So it's going to be more reliable, cheaper, and you can, if a post-trained open source model is going, you know, it basically also can access your data, right, while maintaining the privacy, everything that Patrick mentioned about that, right? You can get that context, rich context, enterprise context, ontology, data ontology and things like that, while otherwise you wouldn't send that data to some OpenAI or Anthropic, right? So I think definitely in the enterprise that's kind of will be the future. This doesn't mean that you are always going to have, kind of at the edge, at the boundaries you are going to try to solve harder and harder problems. So these kind of frontier models are going to be very valuable. That's why still people spend a lot of money on them, right? But yeah.
Yeah, I'll also just add one to concur with Ion. I think the fastest growing product right now that we have at Databricks is this AI gateway that Ion was mentioning. And it's because if you want to take, if there's a Pareto curve of sort of cost on one side and quality, and you know, how many tasks can this, for this level of dollars per task, you know, how good of tasks can this thing do, and if you want, and so there's not just going to be one model even just because of that fact, like, you know, you may have a relatively low-cost model that can get a lot of simple tasks done, and you may have a very, you see how much they spent to solve Navier-Stokes, they spent like, I don't know, hundreds of thousands of dollars at least. So it solves a very complex task at extremely high cost. And so the way to take advantage of this, and by the way, new models keep coming out like every 10 minutes now that somehow push this, some of them don't, some of them are actually within the Pareto frontier. So those don't get used because they're just not better, they're not cheaper or better than any other models, but many of the newer ones push the frontier out in some way. So now I can, you know, I can for this fixed amount of cost I can do more, or for this fixed difficulty of task I spend less. But the only way to take advantage of that is you need some middleware in your company that can manage this like huge explosion of models and help you have policies and decide who can use which models and test new ones. And we use that thing called, we built a product for that called a gateway, and it's probably our fastest growing product in several years I think.
