Invest Like the Best with Patrick O'Shaughnessy
00:10:34 · Patrick O'Shaughnessy, Ben Thompson
[00:10:34] S0: What can the railroads teach us, do think? It's not the last bigger build out, right, in terms of percent of GDP or getting I [00:10:41] S2: think we might be bigger at this point, or it's like it was the biggest. In the ballpark. The railroads had a real duration mismatch. To build a railroad and make money off it was a decade or multiple decades long endeavor. Whereas you had to issue money to pay for it in the short term. And the world ran out of money. Right? I think that is probably the aspect. I think that's why people reach for the railroads because everyone talks about, are we gonna have enough compute? Are we gonna have enough electricity? Maybe the nearest term questions, are we gonna have enough money? Which is kind of a bizarre thing to think about. That's what happened in the eighteen seventies. The world just ran out of money. The funny thing is the railroads kept operating and they expanded the West. Their contributions to GDP was astronomical. They're still contributing to GDP. Railroad money is what's going into Google right now for Berkshire Hathaway. Like like very funny. No. [00:11:39] S0: It's quite literal.
00:11:39 · Patrick O'Shaughnessy, Ben Thompson
[00:11:39] S0: It's quite literal. [00:11:40] S2: Berkshire Hathaway has this problem. To me, this NVIDIA deal is very much paired with the Google equity issuance, which I thought was shocking when it happened. Why was this shocking? Because it's Google. They can't raise money. Like, why are they issuing equity? Why are they reducing their upside if they believe so strongly in this? But the Berkshire comparison is interesting because to a rough approximation, they make... They have C's candies famously, right? Tremendously high margin business. The problem with a lot of high margin businesses is the percentage profit you can make is very high, but the absolute profit you can make is No [00:12:14] S0: reinvestment runway.
00:12:15 · Ben Thompson
[00:12:15] S2: That's right. Like you're just accumulating cash. The brilliance of the BNSF railway thing was basically they took the C's candy profits and said, here's another industry whose margins are way worse, But the absolute dollar amounts are so large that those way worse margins result in absolute profits that are much larger. BNSF in 2025 or something, the amount of free cash they've threw off in one year was more than Seas Candies had thrown off its entire lifetime. Even though you're talking about a low margin business compared to a very high margin business. I think there's an aspect from Berkshire Hathaway where once your capital gets so large, you start operating in a world of like absolute numbers as opposed to percentage numbers. And the reason why I thought that was so interesting, that story, is it seems to capture where Google itself might be going. And so there is very symbolic for them to invest in Google. Google has this unbelievable high margin business of search. One of the most perfect, beautiful business models of all time. And the purest aggregator of them all, like scales in every direction. Doesn't have to invest any money to do it. Everything's zero marginal cost. It's amazing. Meanwhile, there's this AI opportunity which requires just astronomical... It's just incinerating cash. But you can imagine if AI is intelligence and it's TAM is basically all white collar work. And eventually with robotics, everything potentially, the absolute profits available here, even if the margins are lower, is so much larger that will we look back and Google search with C's candies? It feels like that's what's happening. In that world, you use all your free cash flow. They've done that. You tap the debt markets to the tune of hundreds of billions of dollars. They've done that. You issue equity. What does an equity issuance do? It dilutes your interest and your interest to your shareholders. So you have a smaller percentage of the pie. Well, you have the smaller percentage of an astronomically larger pie. At the end of day, no one's going be complaining. It is very symbolic. Berkshire being the symbol of that equity issuance in that, are they actually not just an investor in Google, but a model for Google and where they're going?
00:14:28 · Patrick O'Shaughnessy, Ben Thompson
[00:14:28] S0: I'm curious, setting aside the commercial and competitive components of this, like you're describing, how AI pilled on the pure technology would you say you are relative to other people thinking about this space? [00:14:42] S2: I have a view that is both super bullish and less bullish in some respects. [00:14:47] S0: Okay.
00:14:48 · Ben Thompson
[00:14:48] S2: So I am not fully convinced about the generalizable argument. AI is clearly incredible at coding. It kind of blows my mind that people were doing this a year ago, like actually writing out code. Is very good at math obviously. But the obvious repost is that these are sort of verifiable domains. What is the evidence or where is the compelling evidence of being very good at verifiable domains cleanly translates to being very good at sort of unverifiable domains or domains that take, have a very long sort of verification loops. I think that's still a little bit to be determined. And it's interesting because I raised this question and there are some people at the labs that were on a panel. And I was kind of annoyed at the answer because the answer took me for an AI bear. Oh, well, people thought we couldn't solve chess or we couldn't solve gold. We solved those easy enough. And I'm like, I thought we could solve chess. I thought we could solve gold because they're knowable domains. Scale was the answer to both of those, but also both of those were bounded. What is the go to example that's not chess, that's not go, that is genuinely in a new space that's sort of an unknowable space where it's doing things that were not possible. That is sort of the, I'm not fully convinced sense. However, AI trained at a rough approximation trained on all the data of the Internet. All the data of the Internet, that's distillation. It distilled all of the end state of human thought. Actual It typing on doesn't have the traces. It doesn't actually have the thought, the emotion, or whatever that went into typing that comment or typing and writing that essay, say Neuralink, whatever. What if the actual payoff from Neuralink is actually capturing the traces
00:16:36 · Patrick O'Shaughnessy, Ben Thompson
[00:16:36] S0: Training data. [00:16:37] S2: Of of human thought that actually dramatically expands the capabilities of these models? In this world, my concerns about verifiability is like, well, we solve verifiability by getting more data. My sense is that a huge number of jobs, a huge amount of economic activity does not exist in these domains that I'm not convinced that AI is good at. Actually, there's a lot of people in the world who are kind of like sentient AIs to a certain extent. They operate very well in verifiable domains. They're given jobs. They do them. And it's almost like a somewhat pessimistic view of humanity to a certain extent. But I think that market is so huge and so large that if the models did not improve at all from where they are right now, the economic opportunity is actually massive. I wrote an article a while ago. There's the whole like accelerationist movement. What I call myself was a reluctant accelerationist. I think we need to push forward because we can't go back. And the worst thing we can do is get stuck where we are. So I'm very AI peeled in terms of its impact on the economy. It's sort of upside in terms of monetization. I'm not sure about the timing. [00:17:49] S0: What would be like the gradient towards it? Imagine law or medicine where I don't know whether or not you would consider those verifiable, like law is like a code of some sort. Medicine, we have a certain state understanding of things.