Things Ben Says on Podcasts

But there's a reason those rockets never ran up. By the way, all things considered, they are underrated. We're a great team. They ran up against one of the greatest teams of all time. And this is why it was so devastating. That's a round one. That was the

ultimate problem for Daryl, but it is what it is.

You can get things wrong and you're sort of engineered to be accurate.

So I said in that article, like this is clearly like half of the political divide. And it didn't like that because it's like, you're going to make people upset. Like, but look, that's literally what's happening. And the reason it's happening is because this is unfalsifiable. It's something that can't be proven. The overall precept of actually government needs to do more. There needs to be this, an entire bureaucracy around controlling this. There needs to be people that have jobs to monitor, to do things. Sorry, it's just a reality. There is a one side of the political divide that leans towards more government, that leans towards more bureaucracy. And this is not a commentary about whether that's good or bad. I'm obviously fairly skeptical of that and concerned about its impact on long-term innovation and all those sorts of things. But we can't just, I have to point that out. It's just a reality.

Not only that, it's going to be a reality that informs how policy is crafted here. And I do have some insight because I'm in Washington, D.C. and have a bunch of friends who are in politics. And I remember, I think it was like nine months ago, I met with somebody who is consulting with a potential 2028 presidential candidate. And she was asking me, like, what should our AI policy be? I don't know what we should be pushing. And I think the entire Democratic Party has been sort of casting about looking for what principle

It feels like it might be a situation of better be lucky than good to a certain extent. Apple has their whole ecosystem. At the end of the day, they do own access to customers. So they can get suppliers. This is the classic aggregator play. If you own access to customers, suppliers come to you, not the other way around. So they can get suppliers for their AI as needed. And by the way, to the extent it's true that people don't want to be productive, they just want a sort of a chatbot. Not only can they serve them a chatbot and finally getting a Siri that works, but you can see a future where this absolutely can work on device. And they actually don't even need to pay for inference costs either because they're using the customer's electricity. I don't think we're quite there. There's a reason they're using Google Cloud and NVIDIA chips, but you can certainly imagine a future where that's the case. And they're in physical goods. Actually, making phones is hard. Having retail, having distribution for physical goods, they're more insulated. The smartphone is so perfect. It's small enough in your pocket. It's big enough to watch basically anything on it. You can run your whole life on it. All your entertainment is there. When we talk about customers who just want to be entertained, the TV is now an accessory. It's all on your phone. I don't see anyone taking over the phone. The question is, is the phone always going to be the center? Or is there a bit where, particularly in the home, this is where OpenAI's efforts here are very interesting, where you want sort of an ambient AI where you just talk to the AI and it tells you what you need. Apple is the best position to provide that, but can they provide that without having leading edge models? Can they provide that if they're so phone centric? Or is it like a Microsoft situation? Microsoft didn't miss mobile. They were very early to mobile. The problem is their mobile was a small PC. They assumed the PC would always be the center and their phones were going to be something that was off that. Apple realized, no, we need to reset. The phone is not going to be accessory to the Mac. The phone is going to be the phone. The iPod helped them realize that and going with Windows and all that. But will they fall into a Microsoft-like trap, assuming the phone's so good, it's always going to be the center. And then let's figure out around it. Or is this five the time when actually Ambient, the cloud, just in general, AI being everywhere, it can manifest through your phone, it can manifest through a device, it can manifest on your computer, is actually better and is actually disruptive to them. I think it's possible. I also think it's totally valid for Apple to double down on what they do. The other thing about the AI stuff is on what basis should we expect Apple to be good at this? At the most crude level, AI is this probabilistic endeavor. Apple is the king of deterministic products. A physical product, you ship that iPhone, you ship it once, and it's got to be good. If it's bad, it costs you billions and billions and billions of dollars. Apple's never had an iPhone recall. It's amazing. That care and decision making and diligence and fierceness in terms of your supply chain and making hard decisions is very, very different than everything that goes into like making great AI. I'm generally prefer companies to do what they're good at. So from my perspective, I'm fine with Apple not doing AI. I want them to keep making great devices. Of

the five potential frontier AI winners, so OpenAI Anthropic, Gemini, SpaceX AI, Grok, and Meta, which of those firms do you think has the most interesting setup?

You've written a lot about this recently.

I think there's a very good case to make that it is more reckless to not be on the frontier if you're a digital company. The counter to meta is actually Microsoft. Microsoft is not on the frontier. The reason why Microsoft has $20 million of free cash flow last quarter, Microsoft paid a $10 billion dividend last quarter. There's some money, but their play is, oh, we're going to play all these off each other. We're going to provide middleware. We're going to provide the platform that enterprises will build on us and we're going to sort of disintermediate the models. I think it's a rational play. It's the IBM play of the 90s. History echoes. Everyone talks about Google, Google like falling in Microsoft. Microsoft follows IBM. And you can see that to an extent. What

did IBM do? What's the analogy? Well,

For sure. But you have this incredible spread. So you have people, I think the vast majority of people who are using ad today are using it as basically a Google substitute or like a recipe maker or whatever it might be. And my suspicion is that the cost to serve those people is extremely low and low in the basically similar to serving them a web page. I would imagine it's marginally higher, but not that much higher. Then you have on the other extreme, people who are actually leveraging test time scaling. It used to be we just scale by making the models bigger and bigger. Now you can scale as far as time. How long do you think about the answer? Well, you could think about the answer for days or weeks or months. That is directly marginal costs. Every second longer you're thinking is costing more money, which speaks to like we think about AI and inference as this one question. That's what I was pushing back on you. But actually the marginal cost question for the different user, the user using free Chat GPT and the user trying to solve a math theorem, they're not even remotely in the same universe. And I think you see this challenge actually in the enterprise in a very interesting way. Microsoft recently, they are shifting their enterprise plan. So they come out with like an E7 plan, $100 per user per month that includes some amount of usage, but then they also are charging for usage on top of that. I think this is kind of a fraught position for Microsoft to an extent, because the positive way to think about Microsoft is they do everything you need as a business. Every individual component might not be the best, but you get it all for one price and they all mostly work together. And if you're particularly a small or medium-sized business or even a large enterprise, there's real. That's right. It makes life easy. The moment you start having to think about how much you're paying, it's not just that that's a new decision, number one, that is untethered from headcount. Microsoft got the benefit is when you were hiring a new employee, you would think about the cost of that employee and baked in the cost of that employee is $100 a month or $50 a month for their license. It was kind of a thoughtless revenue stream for Microsoft. Now, if you think about usage, you have to think every single month, how much do I want to spend? That introduces two problems. Number one, most companies aren't set up to do this. They make budgets like once a year. This idea we're going to be thinking about through our budgetary allotment on like a monthly basis doesn't compute. There's an aspect where they're used to thinking about CapEx decisions or one-time costs. And there's a bit where when I'm talking about this employee, like the loaded cost of employee, it's not CapEx, but it's kind of like CapEx. It's like you make the decision up front and you don't think about it. Anymore. The decision is sort of already made. But if you're thinking about usage, you have to do it again. The final thing is, if you're every month looking at your Microsoft bill and how much did I use, you start thinking about what am I paying for? How good is each of these products? Should I actually just start thinking about and spraying this out? And I think they had to do it because that extreme of user who uses a ton of tokens and is actually leveraging AI costs way more to Microsoft than $100 a month. They can't support them, but they want to hold on to this set cost for the vast majority of employees who can fit in that because they need to ask their customers to think a little bit for those extreme employees, but they don't want them to think too much because that breaks the model in very surprising ways. Are you

surprised at all that the recipe builder user that is very low cost to serve, that there hasn't been a great business model that's emerged around them just yet? Google and Facebook are sort of business perfected in this prior era. They haven't seemed to figure this out at all.

best engineers don't want to go work on this problem.

And so you literally had OpenAI replaying the Dropbox story, but at like 100 exercise, being like, no, we're going to sell subscriptions to consumers. They did. They sold a lot, but they didn't sell enough. If you're going to be in the consumer market, you have to be doing advertising. They're doing advertising now. It's a little weird. They finally pivoted to doing advertising at the same time. They're like, oh, crap, we need to go off the enterprise because Anthropic is kicking our end. So I'm not quite sure what they're doing there. They have been rolling out ad features very rapidly. Things like capy and the connections with retailers. So you know if a purchase went through, so you can do all the tracking and things like that. I'm very interested to see how that goes. There's a bit where had they leaned into advertising immediately as soon as Chat GPT was a hit, I think they would have a killer ad product right now. I think that Google would be in much bigger trouble. I think Meta would be in much bigger trouble because if you have this flywheel, the thing about advertising with consumers is your ability to monetize the consumer

goes up as the bottom line.