Invest Like the Best with Patrick O'Shaughnessy
00:27:54 · Eric Vishria
[00:27:54] S2: I am worried about energy. If you think about the models as translating compute into intelligence, quite simply, what do models do? They very effectively translate compute into intelligence. What's the demand for intelligence? Well, it seems like a lot. So then it follows that we will continue to have more and more demand on Compute, I'm saying compute broadly, not chips, not storage, not whatever, but then like what do we need for compute? We need energy, a lot of energy. There's all this topic of distillation and Chinese open source and all these things, but the bigger thing to me is that I think China's bringing on ten times as much energy next year as we are. In the US, if energy is what you need for compute and is the bottleneck, and there's unlimited demand for intelligence, then it stands to reason that if we have a lot less energy Then we will have a lot less intelligence, or a lot less tokens, or a lot more expensive tokens. If we have a lot more expensive tokens, then supply demand, you're gonna end up with less. And that seems very bad. To me, I think the energy bottleneck, however that is, it'll manifest in twenty different ways. Gas turbines go up and down, natural gas goes up and down, and solar, whatever, all these different things, rare earths. That to me is probably more concerning, and if I'm thinking about it from a regulatory perspective or government perspective, and I think the administration is doing some things around this, fostering investment and development of all energy, solar, nuclear, gas, like whatever, do it all, we should do it all, and it'll work itself out. This is one of these things where, yes, one will be relatively better than the other, and I don't know, and I'm not sure enough to predict which one's which, but it'll all work. Speaking of
00:29:39 · Patrick O'Shaughnessy
[00:29:39] S0: compute, I would love to hear the Cere Tell the full version of this. The reason I'm asking about it is, I'm deeply interested in compute, I have big investments in compute, and I'm fascinated by it. It's just like the most magical thing to watch happen. It's mind-boggling when you get close to one of these things, what humans have been able to do on these chips and in these systems. I think you invested in 2016 or thereabouts. I think it was your first foray into like extremely difficult hardware type investments. It's so hard. And now the world is full of opportunities like this Teach me everything you've learned about hardware investing through Storybricks.
00:30:16 · Eric Vishria
[00:30:16] S2: Mostly, it's really hard. It's an amazing example to me about the naivete required. The company came in in twenty sixteen, it was five founders in a deck. I didn't want to go to the pitch, but it's my job, 'cause I'm like, "Why are we gonna go to hardware investment? Like, this is crazy. It's been ten years since we made a semi-investment. " I think basically the team was excellent, and then the kind of first slide was just like, "GPUs actually suck for deep learning, they just happen to be a hundred times better than CPUs. " You have to remember, this is pre-Transformer, OpenAI is this weird research lab at this time, Nvidia was worth like forty billion, not four trillion The TPU hadn't been announced, none of that, so this is early. But the whole idea, as soon as you said it, I was like, "Oh shit, of course! Of course! Like, why? " I had spent the last eighteen months trying to figure out applications of deep learning and looking at the security thing and looking at this medical imaging thing and like all this other stuff, thinking like, "Hey, there must be something here that's gonna be really transformed by this stuff. " Anyways, we go through this whole journey, we end up investing, which was amazing. We In between, just built a lot of conviction that this was a great swing. And I'll tell you what I understood, and I really just understood so little, but basically there are three things that we know how to do to speed up deep learning and hardware. Still till today, increase the number of cores, increase the communication between cores, bring the memory closer to the compute. Those are the three things. That's it, those are the only three dimensions we know in hardware. My articulation of what they said to me, honestly all I understood, was, "Let's just take all three of those things to their logical maximum. You have a wafer scale chip, at that time you would have four hundred fifty thousand cores on it, you'd have something like twenty gig of SRAM on the chip, so you never have to go off chip to get to memory, and because they were all in the same wafer, the communication between cores is maximized. This is the best you could do on that process.
00:32:18 · Eric Vishria
And I think the first chip was seven nanometer or something, you That in software, if you have that logical block diagram of like why it works and everything else, you're kind of eighty percent of the way there, and it's a matter of go-to-market execution. In hardware, you're like two percent of the way there. There's things like physics, entire supply chain of vendors. There's of course TSMC, which everyone knows, but it's not just TSMC, there's thirty other vendors that matter in putting all this stuff together and everything else. I didn't know any of that. Fast forward from twenty sixteen to like twenty nineteen, I think Got their first parts back. You go through this like bring up, and then it's like, "Bring up, oh yes, we got a part back. I mean, it's like, "Bring up, yeah, then you gotta go through like, "Wait, bring up, " and there's like fourteen steps of bring up and everything else. And then by like twenty twenty, we had our like first thing that works. Then it's just this march of actually getting it to work. One of the lessons that I've learned on this stuff is basically, you go through all these sims and everything else in hardware and semis in particular, that basically is your roofline. The best it's ever gonna be is what that is. And then every bit of software and reality and compilers and kernels takes away from that roof line. You might start at ten percent of the roof line, once you bring it up, these guys are grinding for months and years to like get closer and closer and closer to the roof line. It's really different, it's really hard. I'm astonishingly bullish if I kind of rewind, part of the reason we made the investment was if you looked at the four prior generations of compute in my lifetime, so you had CPUs You had graphics, then networking, then mobile. There was a new workload each time, so you had multipurpose compute and it led to the CPU. You had massive parallelism, led to the graphics processor, graphics processor offered massive parallelism, which led to graphics. Then with networking, you needed really low latency chips, and so they had low latency chips. And then with mobile, you needed really power efficient chips, and in each case, we ended up with a new hundred billion dollar company.
00:34:23 · Eric Vishria
The first question, going back to twenty sixteen, was, "Is AI that big a new?" Workload, because there have been many, many other attempts for specialized chips for other things that really just didn't end up mattering. There's some fine outcomes, but they just didn't really end up mattering, et cetera. Right, exactly. And so it's just like, well, okay, well, you need something that's a really, really big workload. Okay, so that's one. We had a lot of conviction on that. And then two, was the nature of the workload, did it introduce a new constraint or problem in it? What I learned was basically the AI workload But GPUs didn't solve the core-to-core communication, basically the layers of the network problem. You're like, okay, wait a minute, there is a new constraint, which is communications, communication bound problem. Then you're like, okay, is AI a new giant workload that is going to have specialized chips? Everything that I just said was the entire, everything I knew at that time. That's obviously played out, and in each prior generation, we got Intel, we got Nvidia, we got Broadcom, Avago, we got Qualcomm and- Arm in each of these generations, and there will be these giant winner, standalone winners. Obviously, the TBU itself is a winner, Trainium's a winner. We've had Grok and Cerebrus, etch. It'll keep getting fought out, but I think that will end up being big. And actually, I think there's a new sixth one that's coming generation, which is, and I'm really excited, we've made an investment that's unannounced in this, but I think that for the first time. In a long time, there's actually room for a new CPU approach. The thing that's happening right now, and you see this reflected in all the semi stocks and everything else, is the LLMs which are running on accelerators and GPUs generate code. The code runs on CPUs. And right now it's running on classic CPUs we've had around forever, but there's a whole bunch of constraints on CPUs that have existed, and CPUs have dragged all this baggage forward that you might not need to anymore. And so, I'm actually really excited about that possibility. The next category,
00:36:23 · Patrick O'Shaughnessy
[00:36:23] S0: the next category. Does the experience with Cerebras make you wanna do a lot more investing in companies? No, but why not? In twenty nineteen, we're