DeepSeek: separating fact from hype

Today, we’re bringing you a bonus episode zeroing in on DeepSeek, the Chinese AI lab that’s recently taken over the news and the app stores, beating out OpenAI's ChatGPT. Max Zeff is talking about it all with Ion Stoica, Professor of Computer Science Division at UC Berkeley and the cofounder and executive chairman of software startup Databricks. Listen to the full episode to hear more about: Why Stoica believes the future of AI lies in "doubling down on open source." Microsoft's decision to host DeepSeek on Azure. What the U. S. can do to foster accelerated innovation – with a look back at SB-1047 and a look ahead to 2025. The controversy surrounding claims that DeepSeek used OpenAI’s models to train its own. Equity will be back next week, so stay tuned! Equity is TechCrunch’s flagship podcast, produced by Theresa Loconsolo, and posts every Wednesday and Friday. Subscribe to us on Apple Podcasts, Overcast, Spotify and all the casts. You also can follow Equity on X and Threads, at @EquityPod. For the full episode transcript, for those who prefer reading over listening, check out our full archive of episodes here. Credits: Equity is produced by Theresa Loconsolo with editing by Kell. We’d also like to thank TechCrunch’s audience development team. Thank you so much for listening, and we'll talk to you next time. Learn more about your ad choices. Visit megaphone.fm/adchoices

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Anthony Ha?

  This episode of Equity is brought to you by Schneider Electric. Schneider Electric, your energy technology partner.

• 0:17Speaker 01

  Hello and welcome back to Equity, TechCrunch's flagship podcast about the business of startups. I'm Max Zep, and today we're zeroing in on DeepSeek, the Chinese AI lab that's been going absolutely viral this week. To talk about it all, I'm joined by Jan Stoika, a professor of computer science at UC Berkeley, and also the co-founder and executive chairman of Databricks, as well as AnyScale. He's very active in the AI policy world, and he's perfect to break all this down for us. Jan, welcome to the show.

• 0:44Speaker 02

  Thanks for having me, Max.

• 0:46Speaker 01

  Of course. So I think, right off the bat, I'd love to kind of just quickly separate facts from hype here. And could you kind of break down for us, what to you are the real breakthroughs that DeepSeek has achieved with its recent models, and what is overhyped and has gotten too much play?

• 1:05Speaker 02

  Yeah, so I think I will start with three things which are pretty factual. So first, obviously, it's performing very well on the benchmarks. Not only that, but also on our own chatbot Arena, which, as you probably know, it's a very popular in-the-wild large language model evaluation platform, and what is now a number three, at least the last time I checked. So clearly, it matches or provide roughly similar performance as a top large language models, proprietary models. The number two, what we know, because the model is released, we know that serving queries on DeepSeek R1 and V3 is very efficient.

• 1:51Speaker 02

  And obviously, the number three, it's open source. So I think that's what definitely, in my mind, behind the hype, it's changing a little bit the trajectory of evolution of these models, in the sense that it's accelerated. Now, in terms of fundamentally, DeepSeek is not a new architecture. Of course, it's a same transformer and things like that. Also, most of the techniques they are using are not necessarily new. However, what they are doing is that, and they are doing its superb job, of taking these techniques and really pushing them

• 2:32Speaker 02

  and making them work very well together. Many of these techniques are published in research papers, but they are more brittle and similar to that. And then you get this compounding effect because you are going to get take fifty percent efficient from this technique, fifty percent from this technique, fifty percent from these techniques. And when you are going to multiply all of those, together you are going to get an order of magnitude or more. Like to give an example, you know, everyone knows it's a mixture of experts model, you have multiple experts. So the key here is about, and that's why it's more efficient, right? Because when you train or query the model, it's only a fraction of neurons are activated.

• 3:16Speaker 02

  Now, this is well known from G P T four, it's also a mixture of experts, but in general, it's the ratio of number of the neurons which are active versus the total number of neurons. Typically, it's between, you know, twenty five percent or maybe in the best case scenario, twelve point five percent. And the reason is that the lower the fraction, the learning is less stable. But DeepSeek is five point five percent. Another example, they used eight bits when they are training instead of sixteen bits. They optimized communication and a few other things. I'm going to stop here, but you'll get the idea.

• 3:54Speaker 01

  Right, so you touched on a lot of the kind of efficiency gains that they were capable of achieving. But I'm curious kind of how you think that this plays out for chip makers here. I mean, you work with lots of chip companies and there was some concern that these efficiency gains mean that we don't need these enormous clusters anymore. Can you kind of break down that argument?

• 4:18Speaker 02

  I think that's a good question. I am in the camp which believes that on the contrary, the lower cost will spur the demand in terms of, you know, you are going to accelerate. The growth and the adoption of the AI for North enterprises and that growth in volume is going to offset the reduction in the cost. And in fact, you know, I am always wondering because we already, why people say that or take that for granted? Because we already have some data about it and the data I'm referring to is just think about OpenAI.

• 4:55Speaker 02

  The cost of serving that requires the prompts decreased by a factor of what? Almost one hundred in two years. And what happened with the revenue is OpenAI, you know, just exploded. Obviously, the reason they are able to provide lower costs because they have much more efficient models. So, you know, this is always happens and but I think that at the beginning of an industry, if you reduce a cost, then you just spur the demand. It's very different from an industry which is mature and it's already penetrated. You know, the new technology penetrated, so to speak, all the industries or the industries which is a good fit for.

• 5:38Speaker 02

  And now you reduce a cost, right? Because then the volume is not going to increase as much because already everyone who should use that technology is using it. But with AI, it's very different.

• 5:50Speaker 01

  Right. And I'm curious if you could kind of help me break down how this really changes the AI landscape in Silicon Valley? Because I think that there's a lot of implications here. One I want to start with when we're talking about this is the kind of battle, the push and pull between open source and closed source. I mean, this is an open model and what do you kind of make of this open model being so competitive?

• 6:14Speaker 02

  This is a subject dear to my heart. Thank you for asking. And just to provide a context, I've been involved in open source projects for quite a bit at Berkeley. And Berkeley itself has been the cradle of the open source all the way from the Unix BSD and the third database systems like Ingress or Open Source and many others. And the project I in talking about, I've been involved in for like Apache Spark, which has started data brings it's all based on that project or any scale you mentioned is based on another project Ray open source to scale machine learning workloads. So I'm really a believer in open source and I think that from my perspective and why open

• 6:59Speaker 02

  source, obviously the advantage of the open source is that it accelerates the progress because you have a common artifact, everyone can work on it, share it and improve it. So the speed of iteration and progress can accelerate. And I think that unfortunately what happened in the US last year is that and for the past two years or so, three years, the development and the research and the progress in this large language model space came in kind of silos, right? You have Anthropic, you have OpenAI, you have Google, now you have XAI, which is one notable

• 7:43Speaker 02

  exception, which is Meta, right? And then the academia is not really part of the game because the lack of resources, right? So you have this kind of very insular approach and obviously from the point of view of maximizing the progress is not optimal because if you want to optimize the progress, you want to have all these people, your best people working on the problem. So I do think that this is a failure of not realizing that and encouraging or even discouraging the open source because if you remember, it was a lot of discussion last year about why we shouldn't build powerful open source model in the US, right?

• 8:27Speaker 02

  And there are a lot of regulatory frameworks which are pushed through the states, maybe SB 1047, if you remember, probably is going to come back, almost every state is pushing it. It was also government level, national level with executive order. So the theory was, well, you know, if you build very powerful open source, then the competitors will take advantage, you'll have it for free. But the alternative was, and I remember we discussed that even last year, that if you do not do it. Then Azure will do it. Now, what is better from the United States perspective?

• 9:08Speaker 02

  Is it better to have everyone using an open source model developed in US? Or is it better to have everyone using an open source model developed by someone else? And by the way, people should understand, it's not about only DeepSeek, it's open source. Alibaba's models are open source. A few days back, it was this Azure model, Kimi, which is open source. So fundamentally, the open source creates a create an artifact on which more people can use to build on, iterate and improve it.

• 9:40Speaker 01

  Right.

• 9:40Speaker 02

  It's fundamental.

• 9:41Speaker 01

  Yeah, I remember, you know, in terms of the Biden AI executive order, you signed a letter along with some other open source researchers and venture capital firms, saying that you thought the Biden AI executive order could restrict open source efforts. And with with SB 1047, you were a very staunch opponent of the bill because of how it would impact open source. So now we're in this space where, you know, you've been fighting these efforts that you say would restrict open source for the last year or two. And now a Chinese open model has, you know, taken the lead. So do you feel emboldened? Do you feel like we need to do more as a country to prop up open source?

• 10:26Speaker 02

  Absolutely. So so look, I think that we got into this situation, right? And it's very important to understand what got us in this situation, right? Because if you don't understand it and you cannot change that structure of the ecosystem which got us in this situation, what you should expect that the trends will continue. And if you look at the trends on the chatbot arena of the progress of DeepSeek versus the best models, if you just extrapolate to six months, you know, it's not like this model is going to match. It's going to exceed significantly the best models developed in US. You know, nothing changes. You should expect the same trends.

• 11:11Speaker 02

  Now, there are voices, actually, online, and I heard today, it's like, well, you know, it's like, actually, they cheated somehow. They used maybe OpenAI data, they used the models from Llama, so it's bad format as they release these models, which is absolutely false. It's like, it's like, Llama, so far, is not a M O E model, a mixture of experts. It's not, it's, it's a different architecture. So, with OpenAI, if you do the math about, they, they train on fourteen trillion tokens, how much it will take? It would be tens of millions of dollars to just use OpenAI.

• 11:50Speaker 02

  But even if you can use it, how you are going to use it to get all the diversity you, you need in, in, in the training dataset? How would you do that without OpenAI noticing? It's just crazy. Maybe, maybe, you know, they use some of the data, like everyone is using. But I, I hear this and say, okay, we should even more should clamp down on the open source models. Even more, we should, the restriction of the export, right, export control and things like that. I think that it's a losing proposition. I think that it's just trying to slow down someone you don't have control, instead of trying to accelerate, you know, here where you have more control.

• 12:31Speaker 02

  For me, the only, I think that the natural, again, I am biased, right, and I all recognize that. For me, the natural way forward is to double down on open source. But even more, to open source the entire recipe, the entire pipeline of training. Because right now, when you are talking about open source, in the case of large language models, we are only talking about open source weights. You have the model weights, the numbers, but you don't have the data. Only just train, you don't have the algorithms, the code, you don't have the evaluation, the use while training. So I think you need to open source all of those. It will be great, ideally.

• 13:13Speaker 02

  Open source data, open source code, open source weights, obviously, and open source evaluation. Because you want to maximize the talent which is going to work on solving this problem, because only then you can accelerate the innovation. And by the way, if you have the open source, you know, open data, open code, and open evaluation, you can build so much more safer models.

• 13:43Anthony Ha?

  Yeah. What moves our world forward? Energy. History has been shaped by it, revolutions won with it, and progress dependent on it. For 190 years, Schneider Electric has been a leading force in this story, creating the technologies that harness energy and transform the competitive landscape. And today, in the dawning era of energy and industrial intelligence, Schneider is helping customers build more efficient, resilient, and sustainable operations across every sector, from data centers to homes, from buildings to industry. Schneider Electric, your energy technology partner. Discover more at se.com.

• 14:21Speaker 01

  I'm curious what you make of Microsoft. They just announced that they're going to host DeepSeek's models on Azure. And I'm curious what you make of that development?

• 14:31Speaker 02

  Yeah, people are voting with their money, right? It's a lot of demand. And now one of the concerns, I think it's from Microsoft, makes perfect business sense. Because what is a worry now? Oh, is a worry is that these models are hosted in China. So the traffic is going to China, can be recorded and things like that. So Microsoft, I didn't know about this, but it's a great move, right? It's like now you have in Azure.

• 15:02Speaker 02

  And all the, you know, the data is not going at least to China, so to speak. I think the other thing it's about, I wanted to say is that makes no mistake. I mean, when, when you say, oh, these guys, they cheated and things like that. In some sense, it's a little bit of arrogance, right, which is not good. China has first class AI experts and is a sheer number for sure more than we have in US. We should have, you know, all the respect in the world for what they achieved, right? Because only then we can do better, right?

• 15:35Speaker 01

  Yeah, I'm, I'm curious. I mean, I'm a little surprised to hear you kind of defend them because I feel like we're in a moment right now where there's a lot of people who are saying, you know, there, there's chip analysts who are saying, no, they have hundreds of millions of dollars of GPUs. They didn't use this small amount. And then there's, you know, OpenAI just recently said, we have evidence that they, you know, we used our models, they distilled from us. So, I mean, what do you, what's your response?

• 16:02Speaker 02

  Look, what I'm trying to say, I think that, like I said, they may have used some data, the style from OpenAI is very hard to believe that they used huge amount of data. So that's one. But then, I, again, I am focusing on what you can verify. Everyone can verify their performance. But they have also, if you think about reasoning, you know, they have DeepSeek R1.0, right? Which shows reinforcement learning is quite successful. At least for reasoning task where you have verifiers, whether the output is correct or not. So the performance is beyond doubt.

• 16:44Speaker 02

  But the other thing is like the cost of serving, the efficiency of the model. Where did they copy? There are only things which are in public, right? Because certainly, OpenAI presumably doesn't have as efficient models because probably the cost will be lower. And the third thing, it's open source, we should embrace it because it's good for the community. These companies, they can actually benefit from these techniques they employed. It's, it's a factor of what, 10, 20, they are cheaper than for all, right? I, I, I think that's the one I look. So these are facts, not hypotheses about how such. Now about the chip controls, the export controls. There are cases in which they are effective, you should do it.

• 17:29Speaker 02

  There are cases in which maybe they are more counterproductive. Like I'll give you examples, which are the encryption algorithms, long time ago. So it's each case, it's different. But in this particular case, I don't know, it's like the fact that there are experts' control, you can clamp down on that. But let's not forget the scarcity and constraints also fuel innovation. For me, you know, these are things which I cannot control. And things, you know, some people believe they can control. I think it's very hard to control now. It's like China has a very large footprint in the world. The point is, the main question, it has to be how you are going to have accelerated innovation in the U.S. That's the main question.

• 18:13Speaker 02

  Because it's very hard to stop them, okay? And how much energy are going to spend in slowing down, stopping them? And how much energy and efforts you are going to spend in increasing the velocity of innovation in the U.S.? That's the question. I prefer to think about, hey, if I cannot stop or even slow down significantly, how I make sure that I have a chance still to win? And the only chance you have to, you can win in this, under this assumption, is to move faster, to make faster progress.

• 18:47Speaker 01

  And Yom, what kind of talks are you having with, you know, policymakers, maybe people in the Trump administration, if any, about efforts to compete with China on AI and invigorate America's open source development?

• 19:04Speaker 02

  Yeah, look, I'm telling everyone I'm meeting. And again, I am a researcher at heart. I'm not a Washington kind of guy or something like that. I'm telling the same thing. I think that when I look at the situation, I think the only way I see to compete is this, go one step forward than they went in open source. Open source, the entire pipeline, and get everyone...

• 19:34Speaker 02

  To work on it, we need to find some funding from the universities to run these models and to pre-train new architecture, new models, like new models, right? It's like open source, end-to-end, and funding for universities because, you know, there is not a lot of money to start this, to do some serious work and maybe tens of millions to pre-train new models, even if you are, you know, very efficient. That's what I, that's what I say to everyone, what I'm telling you, what I'm saying here. If you look at my testimony last year and SB 1047, always was the first point about the

• 20:13Speaker 02

  open source and my point was that if we slow down and if we add friction to the development, discourage the development of the open source, the logical conclusion is that within one year, we would use, at least in academia, open source models developed in China or elsewhere. It's a logical conclusion, you know, in academia, you can, you have to use, you are going to use open source because you need to do some research, right? The only thing is that happens earlier than I thought. So I'm always asking everyone, like, what is the other solution? I'm, I'm just curious because it's like people, if, if you say that, oh, you know, they are cheating or I'm going to control them and so forth. Actually, the cheating argument is a little bit of denial, right?

• 20:58Speaker 02

  Because even if something like that happened, there are, like we discussed, clear evidence there is they made significant advancements when it comes to model efficiency, stability, and also in using RL successfully for reasoning. And the other thing is good, is good for the community, is good for everyone, they publish, it's open source.

• 21:21Speaker 01

  Yeah, and just to kind of, you know, wrap up here, I'm curious to get your thoughts on, just a couple of months ago, we were talking about how AI progress had hit a wall and I want you to kind of remove the fact that DeepSeek is from China, remove the fact that it's open source. I mean, just the sheer technological breakthroughs that they came through, what does that tell you about, you know, the pace of scaling laws and just AI progress generally today?

• 21:48Speaker 02

  I think that obviously that is. Some truth to the fact that hitting the wall in terms of large language models and pre-training, right, in particular that data being limitations, in particular, you know, factual data, right, factual data, you cannot generate synthetic factual data. Right, that's oxymoronic, I think. But then it was quite obvious and what what happens is that, and it was very much expected that there will be a lot of progress in some for some of the tasks where it's easy to verify the outcome. Because if you can verify the outcome,

• 22:30Speaker 02

  then I can make faster progress because I generate, these are generative models, that's our generative AI, that was kind of the name. They are very good at generating solutions. If I can verify the solutions, eventually I can select the good solutions. So I think that the reason we see this progress in reasoning, because in reasoning you have this kind of task. If you are talking about math problems, you talk about coding. You do have much easier to verify them. The problem, you know, did you get the solution or not with a coding? You have unit tests or something like that, right? So whenever you have, you can verify it, you can make much, much faster progress. Like in life, you have a clear goal, right? You can verify, you are going to make much progress.

• 23:14Speaker 02

  You have an exam, you need to pass the exam, you know, with a certain score. But you can, you are going to have a very clear target, you are going to learn for that, right? And you are going to make rapid progress. And and I expect that this will happen more and more. And I think there will be a lot of efforts for adding trustful or loving trustful verifiers, they are calling reward models, for more and more kind of tasks. So I think this will drive the research and the progress quite a bit. So I expect a lot of. Progress for the task for which you can verify the outcome.

• 23:53Speaker 02

  I expect a lot of work on learning or creating high-quality verifiers for more and more categories of tasks. I expect that RL will play an increasingly important role. Of course, it's not everything, you know, it's great about RL. It's also very sampling efficient and so forth. By the way, people are worried about, oh, this is how we are going to use all these chips. If you do RL, you can use the chips a lot because it's very sampling efficient. You need to go, you need to generate a lot of trajectories. You never, you know, a lot of chain of cells, right? It's almost unbounded. That's what I would expect.

• 24:32Speaker 01

  Yeah, they'll find a use for those chips somehow. So we're just about out of time. So, but thank you so much, Jan, for your time. We really appreciate it and hope to have you on again soon.

• 24:42Speaker 02

  Thank you, Max.

• 24:49Rebecca Bellan?

  Equity is produced by Teresa Loconsolo with editing by Kel. We'd also like to thank TechCrunch's audience development team. Thank you so much for listening and we'll talk to you next time.