I want you to get to know a Romanian who left the country at the age of 29. It happened in 1994. He was, as I said, a young and promising computer scientist. Today, he is one of the most important specialists in the world in the race to develop artificial intelligence. As specialized magazines have estimated, billions. This is Mr. Ion Stoica, a research professor at Berkeley University, but also the one who owns the company. He founded the company Databricks, which was again valued at over 60 billion. Olari, I say good evening or good morning. I would like to start the discussion with your permission with that year 1994. A leap in time. And if at the moment you left Romania, you would have imagined yourself in your mind's eye today at the head of a company that is, as I have read, reflecting the power of technological innovation and entrepreneurial vision. Please tell me.
When I left, I left to do my doctorate in the United States. Obviously, you can't imagine what the future holds for you. There have been many developments in technology that were hard to predict. But I've always been in a privileged position to, as they say, have a front row seat to all the innovations that have happened in the last 30 years, from the internet to artificial intelligence.
Yes, I would like you to explain to us at this moment what your company Databricks does. It's true. We read about artificial intelligence, but I would like your explanation to be at a level of understanding that those who are following us at this moment can understand.
So Databricks was created in 2013 and it was founded based on an open source project that we developed at Berkeley. In fact, a student here was the main developer who is of Romanian origin and he is Matei Zaharia now and he is a professor at Berkeley. And what the company does is help other companies to make the most of the data they have. So now, any company collects a lot of data about users, about how different products perform and so on. Now with this data, the problem is once you have it, how do you use it to improve the business, to grow the business, to improve the products, to improve the user experience and so on.
Can you give us a concrete explanation of using Databricks so that people understand what they could use the data extracted from a company for?
Before I say that, I would like to say that the value of data is very, very high. So as has been practically demonstrated by Google and Facebook, now Meta, and many other companies because these are data companies. So they use all the data about users to improve their business, which is to effectively show the best ads, advertisements. But the cases where the data is used and that bring a lot of value are very different and practically any industry can benefit. For example, in pharmaceuticals, in the pharmaceutical industry, there are a lot of companies. So in fact, almost all companies now use the data that they have and generate a lot of other data to develop new drugs, new treatments and so on. Or for example, a car industry uses data to improve and reduce costs, to reduce downtime, to produce and innovate parts so that they build cars from as few parts as possible. And obviously here we are talking about self-driving cars and so on. Why? Because there is still a problem where data and data processing is very important. And obviously one way in which you can be able to process data is to use artificial intelligence. Artificial intelligence in a way is a tool in which you send it some data and it gives you an answer. And this is the service that Databricks offers, provides a platform that you can run more of these products that Databricks offers, in which you can enjoy the value of the data that you have.
Yes, you know that I, for example, would be very curious if someone could help me understand exactly what the viewers want from this show, for example, what information they want. By the way, there are many of them who are delighted to listen to you, you know, in Romania because it is truly a joy to know that our Romanians have succeeded in the world at large, leaving the country. And it is true that at this moment you are at the center of a very important global race, namely the development of artificial intelligence. I want you to tell me at this moment who is the leader. Is the leader the United States? China?
So now, from the point of view of the best models, they are still developed in the United States. So at Berkeley we actually have a platform on which we evaluate these models. It is called Chatbot Arena. Anyone can go there, it is a leaderboard ranking and can see how different models perform. And who is in first place? Still OpenAI? Or? Well, I could tell you right now. Yes, so now actually the first are Gemini from Google, after that is ChatGPT from OpenAI, and after that is DeepSeek from China.
Well, this means that two superpowers are fighting in artificial intelligence: the United States and China. You can say that. And the leader seems to be the United States for now.
Yes, why do you say for now? Once again, it's very hard to predict the future. And at least China has grown quite a lot. And it's not surprising in a way. They have a lot of specialists, big resources, and we talked earlier about data. So data is very, very important to develop very good models. It's like a person: when you have more data, when you have more information to learn from, you can become better. And so it is with those models. And China has a lot of data, has more people, has maybe to use more data, there are less regulations on what data you can use, and they have a lot of specialists. Yes, so we'll see. And on this, another thing is open source. So DeepSeek, one thing that is impressive is that it's open source. So these models have basically, you imagine there are very many numbers that are used when you send a question to generate an answer. And once the numbers to these, everyone can run that model and use it. And open source means that these numbers, called weights, are available and anyone can use them.
Because you're talking about DeepSeek. Is it an intellectual theft, as we've seen some voices accuse, or is it an amazing achievement of Chinese computer scientists?
It's a model, an impressive product. So no one knows exactly what data it used to learn on. Here, some say that maybe they used certain data from the models generated from the OpenAI models. But that's not what's important. What's important is that the model they created is very performant and very efficient. And as I told you, this model is open source. So everyone can take it to look at it. So you can't have any argument against that. That means the model is very good regardless of what data it used.
I understand that. Yes, the United States is still the leader in the field of artificial intelligence. What is the explanation for the very large investments?
So for a very long time, there is this solution to train models that are much more efficient. And the solution is to learn on more data and to have larger models. And to have larger models and to learn on more data is increasingly expensive. It costs a lot. For example, GPT-1, which was the first model, it seems to me it was a few hundred thousand dollars. GPT-2 was 4 million to learn on all the data. GPT-3 was 100 million, and so on. So the cost increases exponentially to learn these models, to create these models. And so investments in the United States have created the opportunity to create these models. Now, once you have these models and a certain performance, people try to make them as efficient as possible. And that's what happened with DeepSeek in a way. So they adopted the most avant-garde techniques and they made them work, and they put them together, and they made a model that is impressive.
