Skip to content

Podcast

DeepSeek Panic, US vs China, OpenAI $40B?, and Doge Delivers With Travis Kalanick and David Sacks

All-In with Chamath, Jason, Sacks & Friedberg

Source ↗ ← All highlights
  • DeepSeek’s Impact
    • David Sacks analyzes the DeepSeek AI story, highlighting the combination of Chinese origin and open-source nature.
    • He attributes the market’s reaction to a confluence of geopolitical concerns and open-source advocacy. Transcript: Chamath Palihapitiya But we had a little bit of a freak out the last week regarding this DeepSeek. If you don’t know, that’s a Chinese AI startup. They released a new language model. It’s called R1. And it’s on par, basically, with some of the best models in production in the West, like OpenAI’s 01 model. But they claim, and listen, you can trust claims coming out of China, you know, for what it’s worth, they claim to have done this all for $6 million on only 2000 GPUs. For comparison, OpenAI spent reportedly 80, 100 million to train GPT-4, which you’re all using now. And Sam claims they’re going to spend a billion dollars trading GPT-5. And so that’s about 7% of the cost of GPT-4. Obviously, there are export restrictions on NVIDIA H100s to China. So there’s a big debate as to if they actually have H1s or not. And Monday was a bloodbath in the stock market. NVIDIA had the worst day in the history of the stock market in terms of total dollar amount of market cap lost. It was down 17%, which is $600 billion. TSMC was down, ARM was down, Broadcom was down. So I guess everybody’s asking the question, how did they do this? Did they do it? And then there’s a bunch of debate on whether they stole, which is kind of rich coming from OpenAI, which got caught red-handed stealing everybody else’s content. And now they’re crying foul that the Chinese stole or trained, did what’s called distillation of their model in order to build theirs. Saks, obviously you are the czar of AI. I’m curious what your take on all this is. And thanks for coming. David O. Sacks Well, I think one of the really cool things about this job is just that when something like this happens, I get to kind of talk to everyone and everyone wants to talk. And I feel like I’ve talked to maybe not everyone and like all the top people in AI, but it feels like most of them. And there’s definitely a lot of takes all over the map on Deep But I feel like I’ve started to put together a synthesis based on hearing from the top people in the field. It was a bit of a freak out. I mean, it’s rare that a model release is going to be a global news story or cause a trillion dollars of market cap decline in one day. And so it is interesting to think about like, why was this such a potent news story? And I think it’s because there’s two things about that company that are different. One is that obviously it’s a Chinese company rather than an American company. And so you have the whole China versus US competition. And then the other is it’s an open source company, or at least it open source the R1 model. And so you’ve kind of got the whole open source versus closed source debate. And if you take either one of those things out, it probably wouldn’t have been such a big story. But I think the synthesis of these things got a lot of people’s attention. (Time 0:15:38)
  • Resource Constraints and Innovation
    • Chamath Palihapitiya suggests constraining resources for AI startups can drive innovation.
    • He argues that over-capitalization can stifle creativity, citing DeepSeek’s resourcefulness as an example. Transcript: Jason Calacanis Hard to know what’s fact and what’s fiction. Everybody who’s on the outside guessing has their own incentive, right? So if you’re a semiconductor analyst that effectively is massively bullish NVIDIA, you want it to be true that it wasn’t possible to train on $6 million. Obviously, if you’re the person that makes an alternative that’s that disruptive, you want it to be true that it was trained on $6 million. All of that, I think, is all speculation. The thing that struck me was how different their approach was. And TK just mentioned this, but if you dig into not just the original white paper of DeepSeek, but they’ve also published some subsequent papers that have refined some of the details. I do think that this is a case, and Saks, you can tell me if you disagree, but this is a case where necessity was the mother of invention. So I’ll give you two examples where I just read these things and I was like, man, these guys are like really clever. The first is, as you said, let’s put in a pin on whether they distilled O1, which we can talk about in a second. But at the end of the day, these guys were like, well, how am I going to do this reinforcement learning thing? They invented a totally different algorithm. There was the orthodoxy, this thing called PPO that everybody used. And they were like, no, we’re going to use something else called, I think it’s called GRPO or something. It uses a lot less computer memory and it’s highly performant. So maybe they were constrained, SACs, practically speaking, by some amount of compute that caused them to find this, which you may not have found if you had just a total surplus of compute Availability. And then the second thing that was crazy is everybody is used to building models and compiling through CUDA, which is NVIDIA’s proprietary language, which I’ve said for a couple of Times is their biggest moat, but it’s also the biggest threat vector for lock-in. And these guys worked totally around CUDA and they did something called PTX, which goes right to the bare metal and it’s controllable and it’s effectively like writing assembly. Now, the only reason I’m bringing these up is we, meaning the West, with all the money that we’ve had, didn’t come up with these ideas. And I think part of why we didn’t come up is not that we’re not smart enough to do it, but we weren’t forced to because the constraints didn’t exist. And so I just wonder how we make sure we learn this principle. (Time 0:26:05)
  • Building a Model-Agnostic Shim
    • Chamath Palihapitiya advises startups to build a model-agnostic ‘shim’ layer.
    • This approach allows flexibility to switch between AI models without significant engineering changes. Transcript: David Friedberg Well, let me ask you another question. Let’s assume that we start the world of AI today. So there’s no legacy of the last three years. And you wake up today, and there’s this open source model that’s 670 billion parameters, you can run it on your desktop computer, it’s completely available, everything’s completely Transparent. And I ask you the question, forget about all the big companies that are involved in everyone’s strategy historically. What’s the model today to build value here? Where do you build equity value as a business? If you’re going to start a company, if you’re going to invest as an investor, where do you go? Jason Calacanis The first is you have to build a shim. And I think the reason why a shim is really critical is that there’s so much entropy at the model level. What this should show you is you can’t pick any model. And the problem is that the people that manipulate these models, the machine learning engineers and whatnot, they become too oriented to understanding how to get output of high quality Using one thing. Meaning it shouldn’t have been the case that we have engineers that can only use Sonnet, right? That’s the anthropic model, right? It shouldn’t be the case that people can only use OpenAI or people can only use Lama. Right now, that is kind of what we have. You don’t have the flexibility to hot swap as models change. So if you were starting a company today, the first technical problem I would want to solve for is that. (Time 0:42:08)
  • LLMs as Storage
    • Jason Calacanis believes large language models (LLMs) are commoditizing rapidly and will become like storage.
    • The real value, he argues, lies in the application layer built on top of these models. Transcript: Jason Calacanis Because tomorrow, if it’s R2 or Alibaba’s model or Lama, I would want to be able to rip it out and put it back in and have everything work. And right now, we can’t do that. Chamath Palihapitiya The answer to your question, Freeberg, is the answer to your question is the application layer, because this is all going to become storage. It’s like YouTube being built on top of storage or Uber being built on top of GPS. All these innovations are being commoditized and this one is happening faster than all the rest. Do you want to be in the storage business or do you want to be in the YouTube business? Do you want to be in the Uber business? Or do you want to be in the GPS chip business? I mean, they’re both decent businesses. But Gavin Baker came on this podcast and said, it’s the fastest deprecating asset in the world was a large language model. He’s been proven right. They’re not worth anything. They’re all going to be open source. They’re all going to be commoditized. And that’s for the best of humanity. And now we’re going to be on the application level, the hardware level with robots. And I think that’s where the opportunity is. (Time 0:43:27)
  • Uber in China
    • Travis Kalanick discussed Uber’s experience in China, highlighting the intense competition and copying from Chinese companies.
    • He mentioned using billboards in Chinese on the 101 and having a China-focused team in Silicon Valley. Transcript: Chamath Palihapitiya Thank you. All right. Thanks to David Sachs for coming in. And, you know, I guess let’s open up the aperture here and talk a little bit about relations with china we’re obviously in a bit of a cold war with them we have tariffs we have taiwan and Then we have uh the sort of trade war going on here with uh exports of h100s where do we want to start gentlemen and you know travis you’ve got some deep you’re one of probably five American Entrepreneurs who ran an at-scale business with Uber and the DD relationship in China. So you have a unique position of understanding business in this, along with maybe Tim Cook and Elon are the only other two people who’ve really had an at-scale business there. Maybe Disney. They have Disneyland there. What’s your take on the relationship and what’s (Time 0:50:06)
  • From Copying to Innovation
    • Travis Kalanick suggests that companies initially focus on copying to catch up.
    • However, once they’ve caught up, they transition to innovation and eventually leadership, drawing from his Uber China experience. Transcript: Travis Kalanick So, look, I had this thing. I’m going back almost 10 years here. Uber day. We’re running Uber China. And I mean, I cannot, there’s no way I could express the frenetic intensity of copying that they would do on everything that we would roll And it was so epically intense that I basically Had a massive amount of respect for their ability to copy what we did. I just couldn’t believe it. We would do real hard work, make it, we’d dial it and it would be epic and it would be awesome. We’d roll it out. And then like two weeks later, boom, they’ve got it. A week later, boom, they’ve got it. And of course I use that to drive our team. And there’s so many great stories. I mean, we had like 400 Chinese nationals in Silicon Valley at our offices in San Francisco. We had a whole floor for the China growth team, and it was primarily Chinese nationals. We had billboards on the 101 in Silicon Valley in Chinese, Uber billboards to join our team in Chinese to serve the homeland. It was like an all-out war. It was really epic. It was epic. And by the way, when you went to that floor in our office, you were in China. Like they rolled China style. Like the desks were literally smaller. Like the density of the space, it was China. Okay. So, but what happens is when you get really, really good at copying and that time gets tighter and tighter and tighter and tighter and tighter, you eventually run out of things to copy. And then it flips to creativity and innovation. Now, at the beginning, it’s sort of all over the place. The kind of innovation it was like, what? You know, you’re like, really? But as they exercise that muscle, it gets better and better and better. So if you want to know about the future of food, like online food delivery, you don’t go to New York City. David Friedberg You go to Shanghai. Right? Chamath Palihapitiya What’s an example like of like something really innovative? Doing? David Friedberg Doesn’t Meituan do drone delivery and stuff? Here’s an example. Travis Kalanick If you went to offices like, let’s say, Shanghai, Beijing, any of the major cities, Hangzhou, etc., the office buildings have hundreds of lockers around their perimeter. That everything that you get, whether it be food or anything else, but especially food, is just the couriers drop them off in these lockers at the office buildings. And then there are a whole other set of people that are sort of like inter-office runners that then bring it to your office. As an example, like, and when you see it, you’re like, and it’s epically efficient. And it’s, you know, they’re taking advantage of their economics on labor and things like this. It wouldn’t exactly work that way here. But a lot of the innovation you will see coming out on Uber Eats or DoorDash, the stuff that’s coming out now is stuff that existed three years ago, four years ago in China, maybe longer. Eventually, you cross that threshold of copying and you’re innovating, and then you’re leading. I think we see that in a whole bunch of different places. (Time 0:51:06)
  • Cloud Kitchens and Picnic
    • Travis Kalanick explains Cloud Kitchens’ innovative food delivery approach, using lockers and asynchronous delivery.
    • He mentioned a service called Picnic, designed for office buildings, which optimizes delivery by batching orders. Transcript: Chamath Palihapitiya Here’s a look at these smart lockers that you can see. They’re just available for sale when you go online. But yeah, these things are crazy. And you’ve experimented with those as well. Didn’t you have like a commissary concept in DTLA? Travis Kalanick Well, look, okay, so we got a couple things. So we have in every one of our facilities, and we’ve got hundreds of them, we’ll have lockers there. So the courier then waves their phone in front of a camera. The right locker pops open. They get the food from there and they go. The courier pickup is asynchronous from production of food. You don’t have lines anymore. There’s no more lines, which then speeds up delivery, shortens the amount of time, shortens, it reduces how much money you spend on couriers. And we’ve got a whole other thing. This doesn’t work. And it probably wouldn’t work in China because, well, for a lot of reasons, but let me explain what it is. It’s called picnic, where if you are in an office building, you order food, you go to a website, you order whatever it is from a hundred different restaurants. Those restaurants happen to be in my facilities. And there’ll be one courier that goes to one of our facilities and picks up 50 orders at a time, brings it to an office, puts it. There’s a shelf on every floor. You get notified when your food arrives. And it arrives the same time every day. And you just go to the shelf, get it on your floor, and dip it right back into your meeting. Saving people time at the office, giving them selection on food, especially in food deserts. But even going like there’s a sweet green right down there in my office right now. I can say 20 minutes by just using our own service versus doing that. And you get at the same price because the courier economics, the courier is delivering 50 orders at a time. (Time 0:54:52)
  • DOGE’s Cost-Cutting Strategies
    • Trump’s Department of Government Efficiency (DOGE) aims to save taxpayer money by offering buyouts and canceling leases.
    • They also mandated a return to the office, anticipating further savings through attrition. Transcript: Chamath Palihapitiya And I think I’m going to just do it just to support proper licensing so that people can start going down this path. But let’s get into Doge. I think we’re in 10 days into this administration and Trump formally established DOGE, the Department of Government Efficiency in an executive order. Apparently, Elon’s been spending a lot of time at the offices, bunch of wins. DOGE is claiming on the interwebs to be saving American taxpayers around a billion dollars a day. That’s $3 for every American every day, about $1,000 a year in savings for each US citizen. And they claim they can triple this. And so for Family Five, that’d be about, what, $15,000 a year, maybe $60,000 during Trump’s second term. We got $36 trillion in debt. Have fun with some numbers there if you like. But the key announcement was very similar to the twitter execution the ability for people to resign done in a very kind way eight months of severance ish is being offered to federal workers They expect five to ten percent of federal workers to take this buyout and it’s um i mean this could be something like a100 billion in savings, eight months of severance is not actually A legal concept that you can do. So these are some sort of buyouts. And there’s obviously some hand wringing about it, but I think they’re off to a good start. They’ve also been canceling leases, as we talked about, you know, pre-election, There is so much space not being used that the federal government is terminating a ton of stuff they Own and going to sell it and consolidating folks. And at the same time, all of this is happening. Everybody has to return to office. (Time 1:10:24)
  • Waymo and the Future of Transportation
    • Travis Kalanick emphasizes the significance of Waymo’s self-driving technology and the normalization of autonomous vehicles.
    • He highlighted the impact of cheaper AI on making autonomy more accessible. Transcript: David Friedberg We talk to Travis about Waymo now? Travis, can I ask, have you taken a production Waymo? Travis Kalanick Yes. David Friedberg What do you think about it? And do you think that’s the future of transportation? And how does Uber play into the self-driving car business now. Travis Kalanick I mean, look, it’s funny because as you guys know, back in the day, 2015, 16, 17, we had our own autonomous vehicles out there. And I remember the first one of ours that I took and I got in the back and all I had was a stop button, a big red stop What that I could push if things got weird. And I remember this is in Pittsburgh where we had our robotics division and autonomy division at Uber. And I got out of that car and literally it’s like I got off a roller coaster ride. Like my legs were, I could not stand straight. Like I was like a little wobbly because I was so freaked out and the adrenaline was pumping. You get in a Waymo today and it’s like, you’re not even thinking twice. You’re just like, it’s all good. You just get in, you get out. Now, part of it’s just the normalization. It’s like, it’s just working. And that normalizing matters in terms of the psychology around it is we’re just there. So it just works. Now, is it a optimized experience for ride sharing? No. Like the cyber cab is sort of the ultra sort of destination for what it means to get transported across a city in a vehicle that is not meant for a human to drive. No steering wheel, folks potentially even facing each other, just a whole bunch of different formats. The technology works. We know that. There are different ways to get to the technology. I think that probably the most interesting thing that we should be, or one of the most interesting things to be thinking about. Maybe there’s a few. First is cheap AI makes cheap autonomy. Okay? So as cheap AI gets out there and proliferates and gets broadly distributed, we should expect autonomy gets easier and easier and easier. And you see some of the stuff that’s happening with Tesla and FSD, their new models are like, I think in a three-month period, they went up like 10X in terms of performance, meaning a number Of miles per human intervention. That’s the thing that Elon’s seeing right now because cheap AI, cheap, good AI makes cheap, good autonomy. And that’s a thing we need to connect the dots on. I think the thing then you go one level past that, you’re like, okay, there’s the possibility literally that autonomy just gets easy and commoditized similar to what’s happening to AI. The next part is, okay, you get the hardware. You’re like, okay, manufacturing’s hard. That’s interesting. That could be a long pull in the tent. I think that could be a place where Tesla, of course, has huge advantage. You then look at who are Waymo’s partners. Are they getting set up to do the right kind of manufacturing and get scale of cars out there? But then there’s this dark horse that nobody’s talking about, which is it’s called electricity. It’s called power. And all these vehicles are electric vehicles. And if you said, yeah, I just did some like quick back of the envelope calcs. If all of the miles in California went EV ride sharing, you would need to double the energy capacity of California. Let’s not even talk about what it would take to double the energy capacity in the grid and things like that in California. Let’s not even go there. Even getting 20% more, 10% more is going to be a gargantuan five to 10 year exercise. Look, I live in LA. It’s a nice area in LA and we have power outages all the freaking time because the grid is effed up and they’re sort of upgrading it as things break. That’s literally where we’re at. In LA, one of the most affluent neighborhoods in LA. That’s just where we are. I think the sort of the dark horse kind of hot take is combustion engine AVs. Because I don’t know how you can go fast getting AV out there really, really, really massive with the electric grid as it is. (Time 1:25:06)
  • Fed Rates and Doge
    • Chamath Palihapitiya connects the Fed’s decision to hold rates with potential inflation and Doge’s success.
    • He believes that if Doge achieves significant cost savings, it could influence the President’s stance on rates. Transcript: Chamath Palihapitiya The Fed held rates. They’re getting close to the goal of 2%. I guess we’re at 2.4%, 2.9% in terms of inflation. Any thoughts on where we’re at with the Fed deciding to not cut? And just you put it on the docket here, Chamath. Any wider thoughts there? Jason Calacanis I would just say that the long end of the yield curve is basically telling us that there’s still a chance for inflation. So I think that the question is these next 30 or 60 days from the administration, I think are basically, they’re critical. And I think if Doge gets to the 3 billion a day number quicker than people thought, there’s going to be a lot of room for, I think, the president to make a very valid argument that rates are Too high for where they are and that we’re going to be able to have a lot more cost control in the expenses, which means that there’ll be less need to spend. It doesn’t solve the problem that Yellen created. Yellen and Biden on the way out the door, the biggest problem was that they put America in this very difficult position because they issued so much short-term paper that is extremely Expensive. And as all of that rolls off, we have to go and finance a ton of this debt at now 5%. David Friedberg So it’s still- Nearly 30% of the debt is going to get refinanced this year. Travis Kalanick And then it’s like, what are these auctions going to look like, guys? This is the thing we all got to breathe- The last auction barely had 2X coverage. Jason Calacanis And I think that that could take a lot of the energy out of the market. David Friedberg Watch the Dalio interview because this is exactly the topic he covers. As we end up needing to refinance this debt, the rates climb, the appetite isn’t there, and it becomes a spiral. That’s why we have to cut fast in terms of the deficit to basically attract the market. Now, the market’s moved a little bit, right? So on January 13th, the 30-year treasury peaked at exactly 5%, and it’s come down today. It’s at 4.77. So a little bit of relief since that peak as kind of the administration’s gone into office and actually taken action. But as more of this action is realized, if people do appreciate and Doge is successful and the court’s adjudication does allow reduction in spending, which I think is the intention, I think we could see this rate drop from 4.78 much more significantly than where it is. And that’ll create a great deal of relief. Travis Kalanick And Dave, it’s like, it either does that or it really, really doesn’t. David Friedberg Or it does like the exact super nasty, really bad. I got a text from someone who is pretty senior in capital markets, thinks it’s going to go to 5.5% before it goes down. So they think that there’s going to be a little bit more of a turbulent run ahead. Travis Kalanick But the thing is, it’s like that whole thing of like, it’s going to get to 5.5% before it comes down. By the way- It spirals on itself. It’s like you got to print money to then get to that place. And then the printing drives it for, you know, you get to that spiral. The (Time 1:37:57)