AI 实验室的护城河只有资本｜对话 Databricks 联合创始人、伯克利教授 Ion Stoica（英文访谈）

Ion, it’s very nice to have you. Do you mind recapping the message you delivered today?

1. AI Infrastructure Grows More Complex

Yeah, sure. Thank you. I’m very glad to be here. I think the message I delivered today had several points. Number one, AI infrastructure is becoming increasingly complex for many reasons: the hardware is becoming much more complex, applications are becoming much more complex, and the speed of innovation at both the hardware and application levels has increased dramatically. Because people want ultimate efficiency, given the very high cost of running these AI workloads, everything is tightly coupled and integrated. You need to develop everything together as one big system.

2. AI Remains a Tool

Number two, there’s been a lot of discussion about AI breaching different systems and about cybersecurity. The question is how dangerous AI is for our society in general, for humanity. My position is that we should take a step back and remind ourselves that, at the end of the day, AI is a system, a tool. It’s very complex, and it’s going to behave in ways we don’t expect. This isn’t the first time: we’ve had complex systems before, like the internet and large distributed systems, which behave in unexpected ways. We should focus on understanding these attacks as much as possible and use that understanding to improve safety and guide systems toward the goals we want. But we should never forget that, at the end of the day, these are tools to improve our productivity and, in general, our lives.

To address your first point, we’ve seen so many vertically integrated companies optimizing every layer of the AI stack. Where do you think the opportunities are for AI infrastructure startups?

3. The Next Infrastructure Opportunities

I think opportunities for infrastructure startups are always about identifying core problems, solving them, and excelling at that. You have bigger opportunities when you focus on emerging problems and workloads. One is clearly post-training. We’ve seen many companies enter that space over the past year or so, but post-training is going to continue to grow because it’s so effective at improving the capabilities of already capable AI models, helping them excel at particular tasks. That’s also driven by the fact that there are now a lot of competitive open-weight models, many coming from China, that you can use for post-training.

Mm-hmm.

You can significantly improve their capabilities and exceed the capabilities of the best proprietary models on certain tasks. Another area is data, especially multimodal data. We’re going to have more and more multimodal models and applications, and finding better ways to process video, images, text, and audio together is going to be an increasing challenge.

A third area, which connects the first two, is robotics. Robotics is still early, and many of these companies are still building their models. Having an open-source stack, or a stack that works well for robotic applications, could be very important. You’re not building trillion-parameter models; the problem is a little more tractable, and I think that’s an opportunity.

Another related challenge is moving data around. Even when you’re training, because of the scarcity of GPUs and AI chips in general, even a big company may get resources and GPUs from different cloud providers. Maybe you’ll build your own data centers, too.

Right.

Then you need to put all of that together and make it easy to use and to train across different data centers, clusters, and so forth. There’s the question of how you train and how you provide a little more portability. That’s a little bit of what SkyPilot is doing, along with moving data very efficiently between these locations. You’re dealing with huge amounts of data: you train on tens of trillions of tokens or even more. Those are some of the aspects.

We’ve seen the AI labs competing head-to-head and accelerating the pace at which they release new models. What do you think of the evolving landscape of these models, like frontier models versus open-source models? Do you think open-source models will eventually close the gap? What’s the moat or defensibility of an AI lab?

4. Capital Is the Core Moat

First of all, as you know, I’m a strong believer in open source. A lot of the things I’ve done are open source, so my opinion will be colored by that. To answer your question, I think the moat of the biggest labs is capital. As you can see, if others can match their capital, or even have less capital but enough, they can create competitive models. That’s why, to this day, you don’t have a lab that’s absolutely dominating. It can dominate for a few months, maybe, but then things equalize. Everyone even forgot that at the end of last year, just a few months back, arguably the best model was Gemini.

Yeah.

Right?

I still remember that.

So, with proprietary models, you can see how fast things are changing. In February and for several months after that, it was obviously Claude. Then there was Fabul [?], and then Astra [?], a pretty good model for coding right now. It may be a little bit better now. You also have Opus 5.5, which is very good. I’m talking about coding.

History shows that no company so far has proven it can build a moat other than capital, which is very important. Then you have capital and revenue, which is a huge moat. But technologically, even if you look at Chinese models, it was DeepSeek in February last year. Everyone was talking about it; it amazed everyone. But right now, you have GLM 5.3 and Qwen 3, which are arguably better than DeepSeek, at least for some tasks.

This also makes me think that, because the labs haven’t been able to accumulate a technical moat, or anything like that that’s visible, open source still has a chance. I hope that open source will become more and more popular and close the gap. But it doesn’t necessarily need to close the gap completely. For every task, there’s a threshold: if a model is good enough, it’s going to solve that task. Being 10 times better, or 10 times more capable, at that task won’t matter. It’s like a car: if you need a car to go safely at 100 miles per hour, that’s what you need. The fact that a car can go 300 miles per hour doesn’t really matter if your task isn’t racing, but driving between Santa Clara and San Francisco.

Because of that, you’ll have more and more models, and many more open models that are perfectly good enough, or more than good enough, for more and more tasks. Some companies, like Vercel, say that in the past few months, their AI routers have seen a transition. I don’t remember the exact numbers, but a few months back, 70% were proprietary models, and now 70% are open-source models.

Steve Jobs says a lot about how, ultimately, it all comes down to taste. Do you think taste is also a moat?

5. Taste Shapes the AI Experience

I absolutely think so. Taste is a human trait. For instance, you can have two people trying to give you the same information and explain the same difficult concept. You might find one person who can explain it in a way you understand in 10 or 15 minutes, while another person gives you an explanation of the same thing that you still can’t understand after struggling with it for 3 hours.

Or think about what people look at when they use a device.

It’s UX, right? It’s what Steve Jobs thought. UX is very important. The same device with better UX is much easier to use. You can use more of its capabilities.

Mm-hmm.

It’s the same thing with models: the way they convey answers, what they tell us back, and the way they may ask for clarification about the task we give them. It’s extremely important, like communication between humans. Most problems between humans come down to miscommunication. That’s part of human nature. As long as we’re building systems for humans, we need to take their needs into account and meet them where they are. To some extent, this is also what we call taste.

Yeah. Can we also say that one advantage, or the moat, of those different AI labs is that they have talent with better taste?

If that were true, you wouldn’t have so many models switching rankings every 2 months or so. When it was Google, Google dominated. When it’s Microsoft, Microsoft starts dominating. You don’t see it happening that way. It seems that we have multiple labs with similar capabilities, at least in terms of people. I don’t think that’s the case. Of course, people are very talented, but I do think capital plays a bigger role. If you have capital, you can attract talent. Just look at how people move between all these labs.

6. Open Source Spreads Innovation

That’s why I think open source is so important: it ensures the diffusion of innovation. It allows everyone to contribute. That’s a problem right now: in each of these labs, only people within the lab can contribute.

Mm-hmm.

That’s the strength of open source: you can have people in all corners of the world contributing. I gave the example of Linux. It’s built like Linux. This kind of thing happened in the U.S. So that’s the kind of future I hope for: more and more open AI. You do have quite a bit of that in China, but I hope for even more. Not just open weights, but as open as possible: open data, open software—all the software used to train—and open evaluation.

Mm-hmm.

Then everyone can contribute, and everyone will benefit.

With so many models released, it’s harder and harder to tell their quality, and benchmark evaluation and reliability have become much more important. This is probably what you’ve been focused on. What is your vision for Arena?

7. Real People Evaluate AI

For me, the ultimate goal of Arena is to help people build better AI. By better AI, I mean more accurate, more reliable, more predictable, and, of course, more cost-effective. To do that, fundamentally, you need evaluation. How else can I do it? How can I know which model is best for my task without evaluation?

Arena’s belief is that the best evaluations are done using real people who run their real tasks, try to solve real problems, and express preferences, implicit or explicit, about how well models or agents have done at solving those tasks. So it’s real; it’s not an artificial benchmark. I’m not saying artificial benchmarks or static benchmarks aren’t important. There’s no question about that. But we believe that, as long as we’re building systems for humans, the ultimate benchmark is humans.

Qingyun Cao

Yeah. I think this has been a very helpful conversation today. Thank you very much.

Thank you.

Qingyun Cao

Yeah, thank you.

Thank you so much.