The frontier labs are in a neck-and-neck race that burns talent and barely holds an edge for weeks. Pedro argues the open research ecosystem still dominates breakthrough odds — and why locking up publishing kills the golden goose.

Big labs chase edges they can’t keep; talent, publishing, and why open academia still beats closed labs; IP, secrets, and culture.

Transcript

Pedro: What did they do when the chat GPT mortal threat came out? They just shoehorned everybody into doing this thing. And they complained and went and did it. But privately, they were oh, man, I used to do this in research and I did this. And in the short term, maybe that’s what they had to do. In the long term, they’re killing the golden goose. And here’s the thing.

Again, Google, Microsoft, Bell Labs famously, they weren’t publishing their research out of the kindness of their heart. It was a net benefit for them in several ways. One is that the AT T reason was part of why they were allowed to be a monopoly, officially, if these other companies are near monopolies, is by saying, oh, but we’re going to contribute by having this public research. Another one is that the research is an enormously great calling card.

The talent, not the money, is the most important thing. And the talent is attracted to where it is that the great stuff is happening. And most important, the talent wants to publish. If you say to the talent you can’t publish, you lose the talent. But different people will fall on different sides of that. But at the end of the day.

And then also there’s another very important one, which we’ll see how it plays out in the AI area, is that by making stuff public, and you’re usually very strategic about what you make public and what you don’t. There’s things that they will tell. There’s things that are in the paper, and then there’s what makes it irreproducible. They’ve perfected the art of irreproducible research. It’s like, we ace the benchmarks, and I’ve had this conversation over and again, but we’re not telling. This isn’t research, it’s PR. But to the extent that you do this, you help develop an ecosystem of stuff. Some companies do this completely explicitly, Meta is a great example.

And not by accident, Because that’s not what their business is But you build an ecosystem of stuff that either builds on top of what you make money with or just having that ecosystem means you can draw on it, and it’s a multiplier. Academic research, to take the most clear-cut example of this, is actually great, I’m an academic and you’re Google. Collaborating with me is fantastic, because you get all this stuff for free.

And I’m training my students and stuff that is relevant to you for free. And then they go to work for you. What could be better, right? Now, they’ve forgotten all this recently, because the need to keep it secret, overrode that, And they understand, but here’s the thing. My bet is that in the longer term, those big breakthroughs are surprises, and everybody says oh, all the big breakthroughs come from the labs now. And the funders ask me, and others, should we still be funding AI because all these companies are putting, 100 times more. Basic research is more important than ever.

Because the open system, the open research system where everybody does their research, And they get credit for it. But then every year they publish it, Or every X, whatever these days. And then everybody starts from the database. This is a driver of for example, I worked in finance for a while, And there they’re super secretive. Renaissance, D Shaw, etc. They make money, but the system as a whole makes very little progress because it’s all atomized. All these labs, it doesn’t matter how many researchers they have. And as long as there’s a large, open, academic research ecosystem, its probability of generating the next breakthrough dwarfs the lab’s probability.

Pablos: Great. That’s a really well put. There’s a kind of I had a long career doing invention work, filing patents on these things. And mostly in deep tech stuff, but Silicon Valley had such a bad attitude about patents that it basically destroyed the ability to patent software inventions anyway. And we weren’t, we kind of steered off of that, but it also looks to me we’re in this moment now where it’s coming back to bite them because they can’t patent anything. They can’t protect anything. Trade secrets aren’t working because their employees are constantly quitting and going to the next guy and back.

In some sense, those secrets aren’t staying secret anyway. And so they’re not getting a major advantage out of them. And it feels like it’s a for the purposes you describe of trying to have an open scientific dialogue about this stuff, it’s kind of happening in a de facto way anyhow, there’s just no tools. I meet these founders who have incredible AIs. You could make an API, run it in a black box, charge people to use it, tell them, never tell them how you do it.

That’s probably the best you can do. But by the time you scale past, I don’t know, 13 people, you’re going to end up with somebody walking off with the crown jewels on a USB key. There’s just no, there’s, I don’t really see how we could protect these things well.

Pedro: I agree with what you’re saying. There’s a few other important aspects. It’s true Silicon Valley has been very cavalier about patents, but software patents, but the underlying reason for that, not the whole story, but the underlying reason is that by nature, software is hard to patent because there’s always a million variations. In hardware, in drug design, here’s the drug, If you tweak doesn’t work, In software, if I invent something, then tomorrow you invent independently or knowing about my thing, or, even deliberately, Something that is different enough that it’s not covered by my patent, but it still works.

Software, because software is so multifarious, There’s a million ways to do anything, Which in a way is the beauty of it, but from the point of view of patents, there’s a reason why software patents are naturally very hard to enforce, right? Now, having said that, you have to remember there’s more than one way to protect your IP. Again what the quant funds do is trade secrets. You can keep it all very, secret. And I’m sure these, at least the older companies all have some of this, but now, and actually let me give you a very salient example. Probably the most salient in the world right now. ASML.

Pablos: Sure.

Pedro: Knows how to do EUV lithography. It took them 20 years to crack that nut. Nobody else has. One day, somebody else will, of course, a lot are trying, or they will find something else that is as good, some of the other companies are saying, But in the meantime, these guys have had a fantastic run.

Out of just keeping that secret. Now, the problem is that in AI, I, as far as I know, I don’t see any signs of that having happened to anybody. Nobody has a secret sauce, which does not mean that, no one will ever have a secret sauce. It’s possible, If I have this black box that you’re talking about, and then there’s also the problem of distillation, Which is but setting that aside for now, I can imagine, for example, I’m just going to construct a scenario where I have a black box that has two layers inside it. There’s the layer that maybe you could distill. But it’s not the high value.

That is directly talking to people. This may be a fun way to put it. Your brain, You and I talk with each other, And figuring out how language works and whatnot is that hard. You’re like, even for human beings, because babies do it all the time, talking to their parents. But. But all of that is being run on the hardware that we don’t know how the heck.

Pablos: It works.

Pedro: If you have a trade secret on how the wetware works, boy, you’d be making trillions of dollars.

Pablos: Let me run a couple of examples. First of all, I think just to dig into those things real quick, finance, hedge funds, quant trading firms, they’re all back East. Back East, they’re good at secrets. Back East, they’re good at, prosecuting people for violating secrets.

Pedro: The laws are different there.

Pablos: They’re different there. The culture is different there. And get a stick to work with. Now, I think you’re right. In a lot of ways, wasn’t as successful overall. They also are in a position to not, just use the stick, but use the carrot because they make enough money. They can pay people to keep their fucking mouth shut. I get worked. And you could say the AI companies can pay people, to keep their mouth shut, but it’s not working.

And I think that’s a cultural gap. I think there’s, some nuance there. ASML has the most complicated machine in all of human history. They have thousands of suppliers, each of which do their tiny bit and have no idea how the rest of the machine works. There is IP behind every single bit of it. And so there’s just, they’re in the ultimate IP protection scenario. It’s not about software, maybe I think the way this has to play out is I like that the foundation model companies are so successful at showing what’s possible, raising a shit ton of capital, doing the data center build out.

Cause regardless of what happens, they could all go to zero. I don’t give a shit. I’m not invested in them. I don’t care, but we’re going to have that compute available for whatever we want, whether it’s open source model or a Chinese model or any other model or whatever comes next, we at least will have the compute capacity. And I’m a computational maximalist. I think the compute matters.

We could probably take, you could tell me I’m wrong, but probably some shitty AI algorithm from the eighties. And if you throw this much compute at it, you’re still going to get amazing results. So think it’s less about the breakthroughs in AI and deep learning in the last decade or something than it is about just throw compute.

Pedro: No, I apologize, but I’m just going to agree with you there. In the following way. First of all, I don’t want anybody to have that moat. Also as an AI researcher, I don’t think ASML is the world where we’re going to wind up in AI.

Pablos: No.

Pedro: No, the paradox, which I, as an academic researcher find very enticing and very delicious is that I think from the point of view of value creation. Discovering those AI fundamentals that haven’t been discovered yet is going to make, by far the biggest contribution. It’s like, Newton, respect for the rocket scientists, Newton was more important. But in terms of capturing value, that’s going to shrink to zero.

Precisely because you’re only going to be keeping, it secret. And if, I had my way, it would never be secret to begin with. It’s published by an academic researcher. However, There’s a playbook, which a lot of these companies and their VCs are furiously following, and they may be right: this is a race.

The race moves very fast. It’s a bunch of hypotheses, but let’s roll with that. And see what that scenario is. Because a lot of money is at stake for a lot of people in this Let’s say that we are competitors and my stable of researchers that I keep captive really does come up with let’s suppose there’s a big breakthrough, Just to simplify that nobody has. Boom. The one, we call the master algorithm, suppose somebody invents the master algorithm, And then, and then I know, and the Googles have known this for a long time. I’m not, I’m only going to be able to keep it secret for so long.

But I just keep it secret for a few years. And then the other people, even after they have it, it’ll take them a while to figure it out. In the meantime, let’s say I have a five year advantage, In that five year advantage, I start doing what we were talking about before of having all the specialized models, the continual learning, And that’s how I get the lock in. What being the first one to discover buys me is the advantage that is short lived, the lock in that comes from that lives forever. This is not, this is a perfectly plausible scenario.

Pablos: For sure. And I think you’re right. That is a lot of what people are trading on right now with their, AI companies, AI investments and that. I’m trying to figure out if I have the master algorithm, is that the best I can do? Is there some, and part of it is I think that the I’m much less academically oriented. I’m looking at how do you take these inventions, bring them to life in the world, scale them up fast.

And I think, and I think in some sense, AI took too long. We should have done it sooner. And we didn’t because it was stuck in an academic context, the commercial investment in it was too low. So we didn’t get these benefits and we didn’t get on this track until much later than we needed to. And I, and that’s, so I think there’s, I don’t have the same sensibility about commercialization being evil that a lot of people do. To me, this is how you attract the resources to bring these things to life. Everyone can benefit.

Pedro: Just to be clear, I’m not saying you have that.

Pablos: Oh no, I get it. But I, so, and by extension, that’s, really where my motivation for looking for pragmatic business models and IP protection for these things comes from. How do I, how do I invest enough? How do I have enough protection that I can afford to invest in this thing? To scale it up and make it meaningful and show that it’s possible. But it’s, a wild, moment in time. All right, let’s find another tweet. I like this one. It says, these will be the good old days of AI. That’s great. It’s true.

Pedro: They are.

Pablos: I was talking to an 18 year old today who just graduated from Yale and math, smart kid, tons of potential. And he’s thinking about going to grad school. And again, I’m, the pirates. I’m grad school would be a complete waste of your time. Now is the wild West. Everything is happening. And I said, you should just get on the next bus to San Francisco, hang out at coffee shops, talk to randos.

They’ll introduce you to their friends who are the smartest people in math, doing AI. You’ll just, that’s how San Francisco works. You’ll find your way around. You’ll find something in super enticing and interesting to work on. It’ll probably crash and burn. I said, do that 10 times over the next decade. And you’ll be the most interesting person on earth. Just do that. And if you want to go get a PhD, then it’ll be there. And lots of people could argue with me about that advice, but I think you would.

Pedro: Oh, for sure.

Pablos: Good. Let’s, you should do that. But I, but I think the point is like, this is an exciting time. A lot is happening and you want to be in the thick of it because these will be the good old days of AI. Now you get to give that kid advice.

Pedro: Listen, someone asked Warren Buffett you live in Omaha.

And he said, I want to stay as far away from Wall Street as possible. And having been on Wall Street, I know exactly what he’s referring to. You get captured into the very, deep echo chamber and local optimum and San Francisco in AI right now is the worst and best of this that you’ve ever seen. It’s go back slightly earlier, the scenario that you just described, go to San Francisco, hang out at coffee shops, meet amazing people. The problem with that is that hear this from people who are there. Including people who moved there. For this, 20 somethings. And it’s like, almost everyone I meet is LARPing.

Pablos: Yes, that’s right.

Pedro: And I, and need to wade through the LARPing to try to find the real people, the real people. There’s some real people there for sure.

Pablos: I was there in the nineties when, I was there in the nineties when the LARPing started in late nineties. And that there is that absolute problem. You do have to weed through the noise. You have to rise above that noise floor somehow. But the difference is like people will try to help. And I think that, I don’t think San Francisco is a great place to live or to be or to build a company.

I prefer founders who left San Francisco to start their company somewhere else. To me, that’s great. I think the vibe there, especially if you’re 18, you’re going to get exposed to so much more. You’re going to learn so much more. You’re going to get so many more different kinds of input and opportunities. You’re just going to see possibilities. And I would characterize grad school, the way you characterized finance in New York. It is old and slow and staid. And it’s a it’s not where you’re going to see the most potential. That’s what I think. But you’re, I’m not a professor.

Pedro: Let me.

Pablos: Straighten me out.

Pedro: Let me say a couple of things. And all fairness, I do have conflicts of interest or biases because I’m a professor and I’m in Seattle.

Pablos: I love it.

Pedro: But let me say the pros and cons of both of these things. Actually Aaron Levy of box, He grew up here.

I asked him, so is it better to do a company in and this was pre the current boom already when AI was growing quickly. I don’t know, maybe 10 years ago or so. It’s an early 2000s company. Reference, this, but and I asked him to start a company in Seattle or the Bay Area. And the thing about, for example, you, I think, we could also go into that, but you don’t want to be too in the echo chamber, but you want to be too disconnected from it either, because as you said, it’s a fabulous generator of ideas.

And again, for the people, the gold rush, Even the guys who didn’t find gold, they probably look back on those years as that was my great adventure. You want that. And go to San Francisco and have time of.

Pablos: Your life.

Pedro: There you go. Now, having said that, what did he say to this question of where’s the best place to start a company? He said something very interesting based on his own experience with Box. That stuck in my mind. He said when you’re starting a company, when you, when you’re still in stealth mode. You’re much better off in a place like Seattle. Because you actually have time to focus. You don’t have these million distractions.

That keep you from focusing. And people aren’t constantly getting poached. Your employees aren’t constantly looking over the fence and Oh, maybe I could go over there. It’s hard to keep it together when your payoff isn’t months away. In that stage, You’re better off being in a place like Seattle. Once you want to hyperscale, there’s no better place to hyperscale than the Bay Area. Precisely because now you’re on the other side of this ledger. If you want something, you want to hire a thousand or 10, 000 good engineers, go to the Bay Area. But now bring this back to current AI. And what we’re just talking about, if you want to make that big breakthrough.

I don’t know if San Francisco is the best place to make it. And then to go to the academia side of things. I’m an academic but also very much a critic of academia. It’s, there’s this interesting dichotomy that academia as a system, as an institution is incredibly conservative and slow moving, but that actually doesn’t matter because you have to realize someone who’s contemplating graduate school, none of that matters. The university is a bunch of infrastructure that, honestly, we overpay for. And these days you barely need a university. A PhD is an apprenticeship. You’re going to apprentice, pick a good person to be apprenticed with and you will move faster.

Than the guys in the labs. And then how do you get through the LARPs to the people who matter? It’s by having something. That you’ve produced. That people go I need to talk to this. And grad school is the best place to do that.

Pablos: Oh, you think. I think the part, what I told this guy was Hey, the North star should be, get in a place where you’re going to be around smart and interesting people. That you can learn from. And whether that’s grad school or a startup or big tech or whatever, wherever the best opportunity for that what you should go for. That’s what I said. I had other advice that could also probably be thrown out, but I.

Pedro: I think I violently agree.

Pablos: That part, I think you’re right. I don’t, I didn’t do grad school, so I don’t know, maybe it is a great place for that context, but it’s a to me, I, the reason I didn’t go to college because I could get companies to buy me better computers than the university had. And I just wanted to play with computers. I got paid to play with the best computers that money could buy.

I felt that was a better deal than college at the time. And then I was in San Francisco during the gold rush in the nineties. And I loved it. And I loved being surrounded by bright people who were super creative. And I got a lot out of that, but I left, I left when it started to suck. And that was 2001. And I came to Seattle. I spent most of my career in Seattle and you can get work done here.

It was at the time anyway, a great quality life. And I think there was a lot going for it, but I never felt I never felt I could do the things in Seattle that I could do in the Bay Area.

Pedro: I, so let me refine this, partly agree with you, but I think there’s, an important nuance, maybe more than a nuance. And to quote somebody who’s very salient these days, Sam Altman, In the days before he was CEO open AI or early days, heard him say something that stuck in my mind because it’s exactly I think it was something that he did with Reid Hoffman, where Reid was interviewing him and said so how did you get to where you are?

And he said a few things, but one thing that he said, Sam, like everybody, has his shortcomings. Some people can’t stand him, but he’s very smart and very good at what he does. Give him at least that, and he’s very responsible for where OpenAI is now. He started out, I think, somewhere in the Midwest. And one of the things that he said, which I would, really try to impress upon every young person starting out is he said, find your people, find your people. What does find your people mean? If you’re interested in AI generically, find your people means go to San Francisco.

But where the right university has a huge advantage is this, To be fair, there’s more than just the advisor. Again, the institution doesn’t matter, but ideally what you want, this is what I looked for when I went to grad school and boy, did that ever work out for me. I was interested in machine learning. In the eighties. Back when nobody believed in machine learning, not even in AI. You looked at AI conferences and they have almost no machine learning papers. There was a lot of automated reasoning, knowledge representation, blah, And I, this was pre web, and I systematically looked at the conferences and where the papers, where is there a sizable machine learning group?

Not just one guy doing machine learning, but there were only two places in the whole universe. Not Stanford, not Berkeley, not MIT. It was CMU and UC Irvine. And I applied to those and I got into one of them. For example, UCI wasn’t a top-four or even top-10 department, but in machine learning, they had half a dozen people and dozens of students doing this.

And what they had at the time that was super valuable: they had so much folk knowledge about how to do machine learning, the stuff that people are doing now for billions of dollars, the methodology, the benchmarks, everything they have all of this. Wow. Find your people means find an AI direction that you really believe in. And then go to where those people are, which could be at Podunk University.

Links

Recorded on July 24, 2026