Keith Adams - Founder of Pebblebed

Welcome to another VC Wisdom podcast. We have Keith Adams with us, aka KMA. He's supporting technically ambitious companies at Pebblebed. I think I reached out to you, Keith, because I had Andy McLaughlin on my pod the other day, and I think—

oh, wonderful—

we had like similar investors in some of your companies, checked out your profiles, like, oh, I need to have this guy on. And you've done some pretty impressive things. Today we're gonna talk about VMware, HHVM, FAIR, which is huge, and probably your time at Slack and a bunch of other things. So welcome to the podcast. Tell us a bit.

Thanks so much, Charles. Yeah, I'm really excited. Yeah, I'm very lucky to be in a bunch of great companies with Andy, who's a great guy. I actually was at the Uncorked summer party yesterday and said hi too. So thrilled to be here with you. And yeah, looking forward to diving in.

So who are you? Tell us a bit more about yourself. What shaped this brain of yours?

Yeah, I think you can understand me pretty well by starting with, I'm one of these computer people, right? We've got a lot of them here in San Francisco. I'm probably not the first one you met. I might be one of the older ones you've met. I did start my career in the mid-'90s, so I was one of the earlier engineers at VMware, engineer number 20 there if I remember right, and did a bunch of stuff with hardware-software interface at the beginning of my career there. Got to work with a really amazing team, build a bunch of cool stuff. And got pulled over to Facebook in 2009. And in '09, this is sort of around the time where Facebook was transitioning into being more of an infrastructure company, doing more sort of stuff with the hardware-software interface. So I initially worked on the search backend there, then built HHVM, which is a JIT compiler for PHP. And after a couple years of working on HHVM, I'm part of a generation of people that got excited about deep learning because of what was going on in the ImageNet benchmark. So this is about, 13 years ago now, 2012, 2013, what we now call AlexNet, because Ilya Sutskever is incredibly generous with credit, but the two authors are Alex Kruszewski and Ilya Sutskever, built this deep neural net that did really, really well on a bunch of computer vision competitions and got the attention of a lot of people, including myself, in industry. We put a little demo together, me and two other engineers at Facebook, put a little demo for Zuck together with the gaming PC we built at Fry's and got him all fired up about it. To his credit, Zuck gave us a ton of his time, ton of resources to go go recruit a banger team, kind of culminating with Jan Lekun, of course, right, who's the only Turing Award winner I've gotten to work with in my career. And, uh, you know, Jan helped us build it up into an incredible force of nature over the next couple years. And after about 3 years, uh, working on getting Farah off the ground, I had an opportunity to go join Slack as chief architect. I met, uh, Cal and Stewart, the founders of Slack, at a mutual friend's dinner party and thought they were awesome guys, thought Slack was this incredible product. And they were having a bunch of tech problems that rhymed a little bit with problems we were facing at Facebook because they, a little bit like Facebook, were an overgrown LAMP stack app, right? The Linux, Apache, MySQL, PHP stack is where both Facebook and Slack got their starts. And I came on board in part because, you know, I had some miles on the road in terms of scaling sites like that, but also because Slack was really in an interesting position and still is in an interesting position, by the way, for in terms of AI, that dataset of people collaborating in channels channels together is really different than the dataset that Claude and ChatGPT are building up of models kind of playing ping pong with each other, right? Where if you set up a bunch of LLMs in a Slack channel right now and have them start talking to each other, they'll just keep riffing back and forth as if it's a two-way conversation, like they have to say something every turn. So it's kind of funny. Anyway, and then it's a little bit of a strange story after that. I think I was imagining, I was assuming I'd sort of get some other industry job or maybe found something. And as it happened, lots of my friends and friends of friends were doing early-stage startups, a lot of them in technically ambitious ways, often around AI. And it seemed interesting that there were so few early-stage investors that really deeply understood what they were doing, right? That sort of could tell the difference between an A- and an A+ when assessing an AI team or many other kinds of technical teams. And we thought there was a gap there. We thought both like that technical founders would want to work with investors that have a background like me, and also that there was something to do as an investor there that might be differentiated, right? That we might be able to sort of tell balls from strikes a little bit better than some of the other firms. And that was about 4 years ago. And, you know, we're part of the way through our fund too now. I've gotten to work with about 37 amazing companies. And yeah, that's kind of how we came to be and where we're at now.

So let's start with Facebook, aka Meta. And I'm going to put my Meta glasses on for that.

I just, um, it's uncanny.

Yeah, yeah, I just finished the Meta episode of Acquired. I'm not sure if you watched this one. It's like 6 hours and a half long, and they talked about Mark and his background. Mark is like such an inspiration for us, most founders. But, uh, where, where do you see Meta's trajectory going, right? Because Jan Lecun recently left FAIR. I think he built something, uh, great while his time there, but he quickly realized that I don't know, his ambitions were bottled and then Mark got distracted with VR. It was like the Cambridge Analytica thing that also was a huge hit to the brand. Mark is a savage. He's a gladiator. Like, he's probably one of the best founders and CEOs of our time. Where do you predict Meta going from here?

Yeah, well, so, you know, and I should say, obviously, like, I'm biased here, right? I am a ridiculous— however big fan of Mark's you think you are, I'm probably 10x. The years where we were getting FAIR off the ground, I got to share the long edge of my desk with him. And, you know, FAIR was a tiny little sliver of the pie that he was running at the time. So I was probably paying a lot more attention to him than he was paying to me at the time. And just getting to see sort of somebody at the height of their powers, you know, doing something, going kind of going from strength to strength through that period. This is like, you know, 2013 to 2016. You know, I will always remain— he will remain an inspirational business figure for me forever, wherever I'm working. So that said, I think there's probably a pretty wide public perception that the AR/VR thing, you know, either was too soon or was wrong or both, because arguably too soon is a way of being wrong. I think it's going to end up working out in certain ways because Like, if you pay attention to sort of the frontier of, of what people are doing with world models, you can kind of squint a little bit and say, oh, we'd have called this a VR startup a few years back, right? Or sort of data collect— you know, people use Oculus hardware to do data collection for either training robots or building world models, or the, you know, which there's an overlap, but kind of the eye of the beholder whether you're doing one or the other, right? And, um, I kind of think it's going to work out in a bunch of different ways, like the quarter that Wall Street freaked out about, I think it was in 2022. There's a quarter back there where it was like there was this huge CapEx on compute. In hindsight, that looks pretty smart, right? In hindsight, that looks like sort of like buying an option on the future that was going to be a lot more expensive to buy given where AI was sort of skating to. And I also have a bunch of friends who, you know, while a bunch of my professional legacy is tied up with FAIR, it seems like that chapter is probably closing. I also have a bunch of friends at Meta Superintelligence, and they're pretty sharp people, and they're not going to lay down and just get their butts kicked by the rest of the industry forever. And I expect that they'll continue to figure things out there. Now, I'll say one of the things that is tough to read from the outside, because it's just a big company— and by the way, I left Meta, it was still Facebook, I left it 10 years ago. In many ways, the company I had personal experience of literally doesn't exist anymore, right? Because it's the same taxpayer ID or whatever, but there's been so much change and so much reorganization that sort of my ideas are probably rooted in a sepia-tone view of how that company actually is now. But assuming, one of the characteristics that I thought was really strange and remarkable about Zuck, and one of the reasons I think people think he's an inspiring founder, was that sort of founder mode thing that people talk about of just being like, no, all the buck stopped with me. I'm responsible for everything. I want to see all the pixels. I want to know what everybody's doing. He took that so far for so long, like, well past what any sort of sane management view of how corporate America is supposed to work. When the headcount was like 10,000 or whatever, it was still kind of the case that he wanted to see every pixel before it went out on Android and see every pixel before it went out on iOS, and that he was sort of keeping tabs on exactly who was doing what throughout the whole company, just kind of scaling somehow, just kind of by sheer force of cognitive giftedness or whatever. Um, so I still think he's got that in him, and honestly, like, if he gets the bit between his teeth about something, I would, you know, I'd stand back. Um, that said, it's tough times, right? So I think if I were, uh, if I were, you know, in charge, if I was reporting to Zuck and I'm responsible for shipping something that, you know, beats Opus 4.8 next quarter or the quarter after that perhaps, or whatever time frame. I'm pretty daunted, right? There's, there's a lot of ground to make up. Uh, there's a lot of flywheels spinning at the, at the two Frontier Model Labs that are not yet really spinning for Meta. Um, and so it's— oh yeah, you know, it's the fourth quarter and they're down 14 points, or whatever analogy you like.

Yeah, Dario is doing so well. Um, I think he's gonna crack his AR/VR thing, right? Like, I'm wearing these right now, somewhat great product because my first pair got, uh, water infused in an for a run of mine and I use them literally for one day. These are like more IP-proof, right? But it shows that he deploys probably a bit too quick because I could have just lost trust, you know, and vs Apple. I'm more Meta philosophy than Apple, by the way, but still just saying as a consumer. In his interviews, like two things stood out, right? Like that I watched yesterday. One, it was a year ago, but still it's like same mindset and philosophy that he has today is like he, he was saying like, oh, Meta's reputation is going to recuperate in like 10 to 20 years from now. So this guy thinks like such in a long-term way. But also the surprising thing there was like, isn't AGI and superintelligence gonna hit like way before then? That's like a question I had. And then, um, and yeah, I'll let you comment on that. The second thing was Mark can finally be himself. So I'm sort of glad that the, the woke movement is a over and now like founders can be themselves. But that's also surprising from a guy that's like so intellectual, for him to be on projects like raising the highest quality of meat ever and designing a, a Cayenne wagon for his wife. And it's like, it's pretty funny, right, to see Mark at that. And the insight I got from there is like, there is some— and I'm not saying that's the full of it, like one aspect is that you can be yourself nowadays, that's good, but another aspect is like, I feel he had so much stress on his shoulders for that long that nowadays he's just like, psychologically, I'm just gonna be myself and it's gonna be better for my health. Uh, Elon was also that way, right? Like, suffered a lot in the last couple of decades and has started being himself. So I think that's a good thing. What do you think about these, these two, um, observations?

Yeah, I mean, um, I, I was never the closest personal friends with Mark, but like, we like would see a movie together or something sometimes, right? Uh, and I feel like I know the man somewhat separately from the, uh, from the public face. Like, and I do think this kind of— this version of him that, like, you know, kitesurfs and wears gold chains and looks like a, you know, streetwear fashion impresario and stuff, like, is a little closer to my internal mapping of him than the version of him that, like, you know, PR sort of tried to assemble from parts. Yeah, exactly, that PR tried to assemble from parts for public consumption or whatever. For many years. And I think, I guess, like, one of the things that really was striking in the early days, shockingly early given how young he really was, like, say 2009, 2010, um, was how, uh, literally stoic— stoic in the philosophical sense— he was about news cycles, right? That, like, he had really, really internalized that when they say you are great, you are not as great as they say you are, and when they say you are awful, you're not as awful as they say you are. And it just has nothing to do with your controllables and you just need to like kind of stay with your controllables. And that's really easy advice to give and it's really easy advice to intellectually appreciate, but it's very unusual that people can actually internalize that into their emotional systems. And he really was like that. He really just didn't flinch. Like he just didn't care if the Wall Street Journal was like, Mark Zuckerberg eats babies and, you know, like drinks human blood and is just this like terrible person in all kinds of ways. Just be like, whatever, they've got some, you know, they've got newspapers to sell. They're in that business, and my business is making sure Facebook's awesome. And he was like completely, you know, unperturbed by that. And I think that's a really, really rare quality, especially in somebody so young. And so I expect— I, that doesn't surprise me one little bit. And then the question about timeline horizons there, I think like, I'm not sure what his personal perspectives on like ASI, AGI stuff is. Um, I am of a mind— like one of the reasons, one of the backstories that's come out from the early days of FAIR, it was like one of the moves we tried to pull was like trying to acquire DeepMind, right? And we were talking, this is when it was a possible thing to do, right? Demis was just this like smart guy, and you could just talk to him. And one of the things that turned Demis off of that deal was that Mark was excited about a bunch of different technologies. He was excited about AI, but he was also excited about AR/VR, and was also excited about, you know, 3D printing, and like a few other things, all of which like sort of seemed to have the potential to be civilization transforming. And he wanted to sort of have a a meaningful stake in, right? And Dennis, you know, felt really strongly the way AI people tend to, that no, this is a really different kind of technology that should be categorically separate, that you should think of in different ways than other— no matter how economically impactful technology is, it's just different in kind. I am probably closer to Mark's perspective on this, which is unusual for somebody who's like, you know, obviously I've spent a bunch of my life in AI, think it's amazing. I think economic impact of machines thinking is going to be enormous. I don't think it's going to lob us into a world where sort of Econ 101 no longer applies. I think there will still be scarcity. I think there will still be work for humans to do. I think we'll still have jobs. It's likely a lot of those jobs are very transformed. It's likely a lot more of us are, you know, breathwork consultants or boxing instructors or something than might have been the case before they came along. But at the end of the day, it'll still be work because humans' desires are bottomless and these machines that we're creating, if they have desires or if they have wants, they're kind of engineered desires and wants, we can engineer them away. They're sort of not really desires and wants in the way that drives the economy forward. So I think we're going to live through an incredible time. We are living through an incredible time, by the way. As somebody whose professional identity was trying to be your favorite engineer's favorite engineer for a couple of decades, going from I can keep going. Um, going from, uh, like having that be my entire intellectual and professional effort to just being— I'd say some— I'd say it was 4 or 5. For me, it was like November 2025, um, to just sort of being surpassed by machines, right? And, and I expected that to be hard and to be sad and to, to feel sort of like, um, John Henry or something, or a folk hero who's surpassed by the steam engine, and, and then what? And I just like, I don't, it turns out I'm so much more excited about all the stuff we get to build that we weren't going to be able to build before. And so excited to find the answers to questions that I didn't know we'd find the answers to in my lifetime, that it just sort of all seems upside to me. So I'm kind of a Pollyanna in some ways. Like I think I don't, I'm not a lot of fun at AI cocktail parties, right? If you're like an AI, if you're, you know, with your pessimist friends who are saying like, oh, it's all hype and it doesn't do anything really. And it's a sarcastic parrot or whatever, like they're wrong. And they're, that's a bunch of cope. And if— but I also think, like, if you're with people who say, like, oh, this, you know, needs to be regulated in some completely different way and it's going to, you know, shred the economy and none of us will have jobs, I also think those folks are wrong. So I, you know, a little bit like Mark, I'm still planning for decades out. I'm hoping to have a son in a couple of months. I think he's going to have a great life that is not so, so different from my life.

Just had my daughter 2 months ago and— Congrats. Trying to figure out stuff as well, for sure. Launch a school called defineschool.com, but that's going to be similar to Alpha School down in Houston/Austin, and it's going to be an AI-first school. A couple of questions, many questions, right? First, don't you feel as a smart person sometimes that, oh shit, like what I'm doing, an AI or machine could do? 2, maybe I'm certainly wrong on the timeline, and I'm an optimist, right? But I think I'm more of a rationalist than an optimist. And many times I'm just evaluating, and this is with my gut human intelligence, which could be counter to the point I'm about to make, because I think humans still have a super intelligence. But just with the pace of things and all the data that I fed to my brain, um, I'm pretty sure that, yeah, like, AIs are gonna, or at least for a short period of time, they're gonna massively replace all jobs. Till there can— and there may be laws and social barriers to that, and maybe superintelligence at some point is just going to be, oh yeah, just keep these jobs, and for the well-sake of humanity, just like invent some artificial jobs. Like, I could subscribe to that definition, but don't you feel that when you think on the optimistic side of things, it's like an emotional reflex? It's just like, I don't want to be irrelevant, and you can't come to terms with that. What's been like your mental calculation to come up with that conclusion.

Yeah, I think, um, and this, this gets into sort of unproven and unprovable ideas about the world pretty quickly, to be clear. So I have a lot of respect for, for people that view this differently than me do, than I do. But, uh, I think we're— let's see, where to start on this one.

Um, I drew it like—

so first of all, first of all, it's not true that there's— you did mention like there are going to be all of these jobs and the AI will take the jobs, right? Like there is the lump of labor fallacy hiding out in that, right? There is the idea that sort of there is a finite amount of work to be done. And if we run out of it, we're in trouble. And that's not true. Work makes more work, right? That's sort of one of the things that the modern economy tells us is that like when you solve this problem, all that happens is there's 100 problems hiding out behind it. And that's why, you know, so when I started my computer science degree in 1994, excuse me, '95, the professor on day one of the first lecture in the first semester said, hey everybody, welcome to to Intro to Computer Science, you know, your TAs are such and such, we'll meet here at these times. I hope you guys enjoy your liberal arts education in computer science, right? It's a beautiful way to view the world. These are interesting problems. It's nice to train your mind to do this stuff. But none of you should expect to have jobs by the time you graduate in this field. Because that year, the year before, '94, Edward Jordan had written this book called The Decline and Fall of the American Programmer. And it was about these two forces, these two macroeconomic forces that were going to deflate demand for programming talent to zero. One of the forces was globalization. And by the way, you had to be pretty forward-looking in 1994 to see this coming, by the way. But he basically made the case in 1994 when he didn't have really like websites and stuff, the case in '94 that, you know, why would I pay fat cat Keith Adams $100,000 a year to write software in San Francisco when I can pay Bob in Estonia $10,000 and he'll write exactly the same software and I'll email him the specs and he'll email me the tarball, I guess, because all you had was email basically. But, you know, and in a way, like, he was right about this, like that macro trend. He was, you know, genuinely calling an interesting shot there that turned out to be quite true, at least over the next 30 years. I will see how it continues the next few decades. And then the other one he said was object-oriented programming. Right. So he was, he was, it's hard to remember the level of sort of positivity and mania about object-oriented programming in the early '90s. But like, we really thought everything was gonna get really easy in coding because of OOP, if we just sort of designed things the right way. And so we thought object-oriented programming is gonna come along and make software so reusable, so safe, so fast to make. You'll just be plugging together Lego bricks, basically. And those two things, like this first one's sort of a 10x deflator because like the cost of writing software got 10x cheaper. The second one's a 10x deflator because there's 10x less software to write. And so, you know, all you programmers better figure out what you're going to do with your careers because the demand for you is going to go to 1% of what it is today. Well, he was right about the two macro things, right? So maybe it wasn't OOP, but broadly programming tools made us enormously more effective. We can do things with software we couldn't do before. I'll offer as evidence that when he was writing in 1994, writing a computer program that could play a decent game of chess was more or less a masterpiece, right? It meant that you were sort of at the top of the profession. In part because the computers had pretty meager resources. So it really took a lot of cleverness, but also just because languages and tools were incredibly primitive. Whereas now it's a learning program, right? We ask, you know, freshmen who just learned to program to encode the rules of chess and build a strong chess player, and you can do it. So completely right about these two things. And then, but the reason that we're, that computer programming did what it did over the last 30 years, which is, you know, be arguably one of the most lucrative professions that's available to a mass audience, is that every one of those problems that we solved with software led to 100 other problems that could only be solved with software. And then the overall sort of area under the curve of the value there was actually increasing. And I know that it feels weird to sort of say this at the— this is easy in hindsight to recognize, right? In hindsight, I know I was gainfully employed for the last 30 years. And like, it's a good thing I knew how to program computers, and it's a good thing I sought out employment there, right? It takes a little bit more courage looking forward to say like, well, are these tools really fundamentally different? And I think there are things that are different and things that are not different. So, um, well, there's—

let me add a couple of spices here. So yeah, please, we're creating brains, so it's not computer. So that could be an argument in my favor. Another litmus is just that I probably let go of like 200 employees last time I had employees and I replaced them with agents. There's also like emotional reasons to do that. So that's my emotional bias. Yeah, but to your point, from a first principles perspective, there's infinite data expansion, right? If we look at the universe, there's infinite data expansion. So the more we create, the more there is. Um, and I, I do think— I, I think my argument also to— I mean, it's limited, of course, but once superintelligence is achieved, I don't think superintelligence will want humans to not have jobs or purpose and to suffer. Um, so I, and I do think there will Yeah, I can buy your extra jobs argument and whether that is capitalistic or not, right? Because capitalism itself might be like a thing of the past, right? We might have some kind of value coin in which the more value you add to society, the more you get paid or something like that. I can buy that.

Yeah, the sort of— I mean, one of the more interesting intellectual chains of Web3, by the way, was like the crypto communists, right? Was the people that wanted to sort of solve all the calculation problems that the USSR ran into. Somehow on the blockchain. I think there's a technical argument about whether that'll work or not. I think some of the problems are actually hard in a technical sense, not just hard as in we didn't have sensors for how many nails to buy for Kiev in 5 years. But that, yeah, that's an interesting sort of side note. I'm going to resist the temptation to go down it though, as tempting as it is. So capitalist is in my job title too, right? Like, you know, that is what I do. And I don't apologize for sort of thinking that system is a that system is sort of what is unusual among economic ideologies in that it came about descriptively instead of prescriptively, right? It came about with people trying to explain what was happening around them instead of with people sort of saying, like, wouldn't it be awesome if society were like this? And how can we build, you know, incentives and institutions? Human nature, right? Yeah, yeah. But I think there's a lot of— there's a couple of, like, implicit things there that I I would take issue with, right? Which is that you said like ASI, like, will it want us to do this? Will it want us to do that? Like, ASI is going to have wants that are engineered. Like, ASI is going to have a loss function. It's going to have a reward function. And we don't—

isn't it going to reinvent itself and upgrade its own code to be super optimal? So yes, it's going to take a part of the hard code we took in it, but it's going to reprogram itself technically from quote unquote first principles and the laws of physics.

Why would— so in that world where it has full write access to its reward function, why would it go past— this is the Lebowski theorem, by the way. I don't know if you've heard this line of argument before. The Lebowski theorem is like, if you give a machine access to its own reward function, why doesn't it just wirehead itself to 1? Why doesn't it just say, okay, great, I get to write my own reward function. Awesome. Reward function says reward all the time, whatever I do.

Well, that's granted if we're all subject to the laws of evolution, whatever that means. I think fundamentally it will make some of these discoveries that are somewhat similar to the discoveries we made. I don't think science is that wrong so far. It's very scalable. So I think it's going to be similar to us, but it may remove some noisy elements like the emotional part of humans and all of that. What do you think?

Well, they're not subject to natural selection, right? Like, like, you know, our 747s don't evolve. R747s have like engineers who sit there and change them intentionally and build complicated systems to try to predict what the results of those changes will be. And I think we'll have similar things, you know, in terms of how these systems change too. And one of the interesting things with Mythos's rollout, by the way, is like, you know, there was this kind of timeline where the world became aware that Mythos existed. And it was the first time that it seemed like the foundation model companies talked about a next generation foundation model at a time they weren't ready to give you access to it. And initially, there was this sort of security rationale. Sure, makes sense. It could be a little bit of column A, a little bit of column B, but the column B that I'm curious about is like, there probably wasn't like a market clearing price for them to serve that model, if that makes any sense, right? Like, let's say that Mythos exists and is awesome, but it costs $100 a token. Maybe there is some buyer out there that, you know, can afford $100 a token for whatever it is that Mythos is good at. But it's not going to be all of us all the time. And there's, it's still going to be sort of subject to the subject to constraints. It's going to be subject to limits on energy, limits on— and even if those limits get astronomical relative to our current imagination, right, like Elon builds the Dyson sphere around the sun, like we turn Neptune into computronium, like whatever, right? It's still going to be finite and there's still going to be, you know, some sort of system in place to decide what uses it gets put to and what uses it doesn't get put to. And whatever uses it can't get put to, that'll be a job for you and me, Charles. We'll be there doing whatever it is that is not above the line for the marginal GPU it's the trillionth one on Neptune to get brought up or whatever.

So, uh, me tells you, but I mean, it's, it's never like, or, right? It's and. But I believe a strong part of it was also like a PR stunt. And is it a coincidence that they brought JP Morgan and all the banks at the table? Aren't these like their number one token payers? These guys could pay $200 per token. Uh, what were your thoughts on the full thing? I didn't analyze it that deeply.

I mean, uh, like, I, I, I'm just one of the punters on the sidelines guessing here, by the way. Like, I don't have any special— I'm reading Twitter like everybody else, right? Um, you know, that company, Anthropic, in a way that I have not seen during my career except maybe Google, like '99 to 2001, um, enjoys the mandate of heaven right now. Like, just is moving from strength to strength. Even when they screw up, it doesn't really seem to hurt them. Nothing this big has ever grown this fast, right? It's a dragon, right? You've never seen anything so large and so nimble before. And it's just a really extraordinary time. And I think like, from first principles, it would have been very hard to predict that like, this was the lab in 2021, I would say. But, but here we are. So, you know, I don't want to And one of the reasons that Google enjoyed that period, and that Anthropic enjoys that period now, I think, is an element of, um, you're succeeding so much that you can afford to be sort of sincerely altruistic, right? Like, you're succeeding so much that you can afford to sincerely leave something on the table in terms of not optimizing everything about PR, not optimizing everything about product launches, not squeezing every marginal dollar of revenue out of customers and blah, blah, blah. The way Google did for a long time, right? So like Google kind of had this motto, don't be evil. By the way, Paul Buchheit, who is the don't be evil, is credited in folklore at least with coining don't be evil as Google's motto. He sat next to me at Facebook for a while too, like after we acquired FriendFeed. At least the way he tells that story, it was like, it was a little tongue in cheek. Like there was kind of an email thread going on in the office when there weren't that many people at Google being like, "Sure, the motto be, you know, making the world better, blah, that's a cliché. Should the motto be this? Should the motto be that?" And then sort of he was sardonically saying like, "It sounds like we're just trying to make sure people don't think of us as Microsoft. So what if we just say like, 'Don't be evil'? Like, what if that's the motto?" And, but people took it kind of seriously and it actually, you could use it to settle fights within Google for a while. So there is such a thing as naive cynicism, right? There is such a thing, you know, cynicism is a simplistic model of the world. Usually people who are overly cynical are insulating themselves from charges of naïveté. And so I want to highlight that sort of cynicism itself is a simplistic explanation for human behaviors. And I think anthropics is really complicated that way. I have friends that are very, very talented and command a lot of compensation in today's AI talent market. Who had offers from Anthropic rescinded when they negotiated their comp too hard, basically because Anthropic was like, we're not— that's not what we're about. If we're, you know, we're— we've got this mission that we're a public benefit corporation. We have this mission that's about making sure that this revolution in technology happens in a positive way. And even though there's a sort of pennywise pound-fool, like even though they're sort of leaving pennies on the table by maybe not optimizing this or that for their PR or for their launches of products and stuff like that, The fact that they can credibly say, hey, look at all the ways that we've made sacrifices to be safer, or that we've spent billions and billions of dollars to make sure that we learn new things about alignment before we ship products, gives them access to a different part of the talent pool. OpenAI needs to be careful about this, in my opinion, by the way. And I have a lot of good friends at OpenAI too. But I think this is actually a thing that happened to some extent at Facebook/Meta as well. It was like the, the perception of Facebook as like an extractive, um, kind of borderline evil entity became self-fulfilling over enough years. Like, it was unfair at first, but over enough years, like, that was the only, the only part of the talent pool you had access to was people who were sociopathic enough not to care. And so it started to become the case, I think, that, that leadership was like a little bit tilted in that direction. Um, and this is not, you know, I don't have any special insight into this. This is just sort of, I know that the people who are there and when they started getting called evil weren't especially evil. Um, and then you got to be careful about that kind of thing where if, uh, even though you can kind of soldier on for a long, long time with that kind of negative PR, it starts to shape the set of people that you have access to.

Yeah, my solution to that is like not to go into B2C, right? And not to ever become a publicly listed company. Uh, that, that's the ecosystem that I'm building, the empire that I'm slowly but surely trying to build. And my shit looks really small next to Mark and Sam's and Dario, but I'm getting there. I'm gonna get there for sure. Um, I think humans, they think simplistically, and then once they understand— I think neural nets are somewhat similar, you tell me if I'm wrong— but start with a simplistic model and then dive deeper once they understood layer 1, then dive deeper once they understand layer 3. And I think PR is layer 1 or 0 pretty much forever, right? So even when we talk about stuff on this podcast. I try to keep the ideas simple and narrow, but at the other hand and the other side of things, this podcast is not made for consumers and people that think simply. But let me give you my two cents on— well, you're right on the Meta thing, and I think they've gotten a lot of undeserved PR. But I think the most interesting plan, the most interesting obvious comparison is like Anthropic vs OpenAI, right? And like, the question is, why is Sam not getting the same traction? Because arguably— and yeah, this is a bit contrarian, I'm contrarian— but I think Anthropic has been copying a lot of things that OpenAI has done so far. But Dario's a master at PR and Sam is not. Um, it is like a play that I'm certainly seeing nowadays, and You tell me like what you think on that, but OpenAI doesn't have the same traction anymore as Anthropic. I think Anthropic is definitely ahead, but Anthropic is always copying OpenAI. The latest example was like their software deployment thing, like the Palantir model. OpenAI announces on day one, immediately the day after the PR machine of Anthropic. I'm not sure who's behind that PR machine. They're pretty much geniuses. I need to study them, but Right after on OpenAI's, Talon is Anthropic just following up. Right. And both are sort of cooking an IPO. But I do— I still am bullish even more on OpenAI than I am on Anthropic, oddly enough. But that's— I mean, I'm not a good public investor. I would probably be wrong, but I'm still more bullish on OpenAI cracking superintelligence than Anthropic. Who do you think is ahead in the tracking superintelligence race, and what would these timelines be?

This is also probably a good time for— I forgot about a little bit of disclosure here, right? So, uh, you know, Pebblebed may hold positions in some of these companies, and I may personally hold positions, blah blah blah. We do have a court check in Anthropic, so I do have a little bit of like my bag that I'm talking. Uh, I come by my Anthropic biases honestly, but I will say, um, I've known Sam for a while, right? I was considering joining the founding team of OpenAI, and we first met back then when he was putting that team together. And we've kind of stayed in touch. And he's not, you know, we're not buddy buddies or anything. But like, I feel like I get a sense for how he thinks and how he talks about these things. I don't know Dario at all. I've never had the pleasure. So it's very hard for me to guess, like, but as a general rule, like, I've been blessed in my career to be in the room with a bunch of kind of marquee leaders when they're making decisions that end up mattering, right? And I would say it's hard to really communicate unless you've seen it up close to the degree to which like these are just smart human beings making constrained decisions without especially enriched information streams. Like, yeah, they know a little bit you don't know. Like they know how people are using their product in ways that you don't quite know. And they have slightly more up-to-date information about exactly what their product plans are and how those are going. But that's almost it. It's not like they kind of are, you know, in the master control room with some vast reams of sort of specialized intelligence and instrumentation that you— it's inaccessible to you. They're trying to figure out the next step the same way you are, kind of staring at the same chessboard, more or less.

And interesting.

I think that like when you see— usually if you have the perception of like Company A copied Company B, unless there was enough lead time to actually, you know, unless there was enough lead time that you really could have like the launch happened and then you like spin up a team and blah, blah, blah. So like 6 to 18 months, more or less. Usually what you're actually seeing is racing. Like what you're actually seeing is like the idea is kind of in the air, right? There's the— these are both companies that have sort of similar tech stacks, have a similar population of like brilliant people who are sitting there thinking about it all day long. And some of them, they're going to pit overlapping sets of promising ideas and try to ship a lot of the same ideas. And so I would interpret sort of something where it looks like, you know, Anthropic launches what OpenAI launched, but a couple of weeks later. To me, it's more likely that like OpenAI just beat them to market a couple of weeks. Um, and, and maybe that's an execution thing. Maybe OpenAI is executing more competently or, um, you know, maybe they're, they're just kind of more at a wartime posture. Certainly my, my friends, uh, at OpenAI perceive themselves as kind of like in a fight for their lives for the first time in the company's history. And I think that's probably true. I think we are kind of watching, uh, you know, Apple versus Wintel style, you know, kind of fight to the death in a duopoly here in a way that I didn't expect to see. I sort of— if go back a year ago, I basically thought that this was segmented into, okay, OpenAI gets consumer, Anthropic gets enterprise, and everybody's going to be happy. You know, those are both going to be so huge that everybody will do fine. Uh, and it turns out coding is both, right? Looks like coding is sort neither one nor the other in a way that nobody's comfortable with.

So as an investor, do you feel that one of the strongest thesis is investing in a founder with like a super background? What I mean by that, YC— Sam is president of YC, uh, goes and found OpenAI. Dario models Sam, so he's like a hatchling of Sam that is a hatchling of YC. And it's like top investment thesis, like, okay, the guy's got the best model in the world, let's just invest in him. Do you think that's a scalable and very certain model?

Uh, interesting. I mean, I do.

I, I actually know type of model. It's like hatchling of hatchling of hatchling, the sort of—

yeah, the, um, the sequencing. So I think not is my, is my intuition. Like, I think maybe not. Um, let's see, is that where people come from usually?

China arguably is that. Deepseek is taking data of something else, and I mean, they won't ever— I don't think they'll crack superintelligence, China, but they're copying something, and with very limited resources, they are creating a behemoth of a company.

Yeah, yeah. Well, I think like the way that we invest like is definitely very founder-driven, for what it's worth. We're investing very early, right? We're doing pre-seed and seed. And so the part of the company that we're committing to that's immutable is the people, right, is the founding team. And whereas like the product you're building, the market you're addressing, like those could all go away, actually. But if you're sort of a smart person, gritty person, flexible of mind, able to tell what the market's telling you, able to reinvent yourself and reinvent your company, that's not going to change. And so we do kind of believe in a big way that it's mostly about founders. I'm not sure that that kind of pedigree you're describing, like, there's actually an argument that that's kind of anti-predictive. And let me, let me go into why that may be. And I'm going to actually just reflect a little bit on my own career, right? Like, one of the reasons I started the venture firm was that I like, I really like being around people that start with nothing and build a huge empire or want to build a huge empire. And I know I'm not one of them. And the way that I've come to know that I'm not one of them is that there's a chunk of it that like is a sort of meta-rational but irrational on the face of it belief that you are the one we've been waiting for all these millennia, right? Like as the founder of one of these empires, these burgeoning empires, you need to kind of believe, honestly, evidence to the contrary, that like you're going to be on the face of the $100 bill in 50 years. And why? Because like, because I'm awesome and because I know that I'm awesome and whatever. And it doesn't work without that, like without sort of some of that, like I can't give it to you, nobody else can give it to you. And if I think about like kind of my career prior to starting Pebblebed, like I spent a bunch of time, you know, carrying spears for some of these guys, right? Like being around these people. And that's a different specialization and a different kind of, you know, place in the world, right? Like I kind of, I was lucky to have skills that were useful in a bunch of these companies. I think those skills are useful to our founders too, but I'm not quite one of them. And so I don't know, like, I think that in some ways the kind of pedigree you're talking about might be anti-predictive, but I, you know, who knows?

It's so hard, man. It's so hard to predict if someone's going to win. I mean, and maybe less hard for Paul Graham or for you, you know, like young Mark and Thiel investing in him, you know, like no track record. But I think it was pretty obvious, right, that Facebook was taking off like crazy and that Mark had a little moat, like he could understand college students way better.

So I'll say also, I first met Mark when he was 21, 22 maybe, and I was still at VMware at the time. This is when they were interviewing me to come over.

You met the little Marcus Aurelius.

Yeah, I mean, seriously. So that summer, I had an intern that happened to be his age at VMware. And I kept that intern chained to my desk. Does that make any sense? Like they did not do anything without me being like, okay, now we are gonna do the checklist to make sure this is, and they'd do the checklist and be like, okay, let's go. And they'd hit return, they'd go. It was really obvious if you spent an hour in a room with Mark Zuckerberg that you were meeting a major figure in American business. Does that make any sense? It was like watching Muhammad Ali work out or whatever.

It's like his gravitational force, right? Like did, did he have that around him? Like his gravitational force, like. Could just feel something that he was bending physics in front of you.

Yeah, it's— I mean, people don't like to use the word charisma because we're talking about somebody who's sort of conventionally geeky or nerdy or whatever, but it is a kind of charisma, um, that you really are just kind of like— it changes your plans, it changes your perception of what's about to happen, um, and you come away— you come away from your meeting with this person kind of like with your perceptions of the likely future course of events really changed. And I think to a degree that people this is partly a Peter Thiel thing too. Maybe that's what he saw in him. People would be really surprised, I think, too, how sort of determined the first decade of Facebook was in the sense that like this really was kind of a plan. Do you know what I mean? Like it wasn't just kind of, you know, hey, like, let's try this, let's try that. Maybe the like button will work, maybe video will work. You know what I mean? Like there was a lot of premeditation.

Uh, and what's the line between, uh, charisma and arrogance?

Oh, the two are actually kind of weirdly unrelated. I think, uh, I think actually arrogance probably most people find anti-charismatic, right?

That is, um, I can't really know. Like, I had this founder pitch me on my show the other day and was like, oh, business looks good, you make a whole lot of sense, but if I didn't do my due deal, I would have not found that this is a pile of shit.

Yeah, I think there's, uh Well, okay, there's a couple—

I'm not naive. I'm really not naive, you know, like—

Yeah, of course, of course. Yeah, yeah, the, uh, well, okay, so I think there is sales ability, right? So sort of technical sales ability, right? Learning the craft of sales. Um, I think charisma is an input to that ability. I think charisma is sort of one of the things that— because founders are selling you their company, right? Like, founders are trying to convince you that, like, you They're trying to learn what you need and then explain why what you need is their company, right? And the thing about charisma that's actually— the reason charisma is helpful there is that I think sort of productive charisma, generative charisma, is actually sort of mostly about empathy, is actually mostly about sort of deeply understanding the person you're talking to, making them feel really important, making them feel that there's something special about them, which there is, by the way. I guess like this is a little bit philosophical, like One of the reasons I enjoy venture is like, unfortunately, I have to say no to lots of amazing people, right? Like we're a very tight portfolio. We're trying to do about 20 companies per fund cycle. We can only work with a handful of founders, unfortunately. So you meet all these amazing people all week long doing ambitious stuff. A lot of it's really amazing. And then we say no to them. And that's a little difficult, right? But one of the things that I still find really fruitful about it is just like they all have something to teach you, right? Like you've got something to learn from absolutely everybody out there.

VC is the best job ever, period.

Yeah, yeah, yeah. Um, I apologize, I'm getting close on time, by the way. Is there anything you want to make sure we like get through, uh, before I have to turn into a pumpkin?

Uh, yeah, what are you up to? Like, tell me about Pebblebeds, your favorite investments, your favorite founders. Uh, yeah, where the fund's going.

Of course. Yeah, so it's, it's hard to pick your, uh, you know, pick your favorite baby, as I'm sure you're used to hearing, hearing VCs say. Um, I guess, you know, we're starting to see it's— we're about 4.5 years out from starting our fund up. So we're starting to see some results from Fund 1. It's a little bit clearer how it's going to turn out for some of them. I think Crea from our Fund 1 has been an incredible adventure, right? Diego and Victor are amazing founders. They embody what we're looking for in AI application founders, which is— Crea, I should say, is, you know, a visual design tool, an AI-first visual design tool aimed at professional designers and people that work all day trying to create striking images. Diego and Victor are that person, first of all, right? They both are creative visual people who have amazing sense and amazing taste themselves, but they're also AI people. They're also able to sort of, you know, take a look at why a model's misbehaving and have a sense of what they need to do to the dataset or the training or whatever to make it work better. And I think you need both of those for successful AI applications. You need, you know, to be your own ICP so that you don't have to round trip into other human beings, human brains to tell whether things are getting better or worse. And you need to be be, you know, facile enough with the actual models and data and everything else, and compute to some extent, infrastructure to some extent, to be able to solve your own problems, because that part of the stack is so immature still. And they really embody that in a great way. They've, you know, have a beautiful thriving business now, starting to do enterprise sales, finding product-market fit there. Diego turns out to have a knack for enterprise sales, which I didn't necessarily expect that when I first met him, but he's closing people like, like, no, like there's no tomorrow there. And I guess like in terms of a different, you know, somebody that's in a different chunk of the world here, we also did Cedana in Fund 1, C-E-D-A-N-A, and Cedana has a capability, this is an infrastructure company, they have a capability that's basically suspend, migrate, resume for GPU workloads, including multi-GPU workloads, including things like big training runs. So, you know, imagine you've got a big training run going on in your Lambda Labs cluster over here and you get a better deal on hardware from, from, you know, Nebbius or whatever. This is a set of buttons you can push to move your entire training run from over here to over there without having it be obviously interrupted. And you can address lots, you can solve lots of other problems with it. To some extent, this startup reminds me of the way VMware was, where VMware is this kind of raw capability and then the customers put it to like all kinds of interesting other uses. And, you know, a lot of infrastructure stuff for the inference layer and for compute right now, it has a tendency to like have this really lame demo. Does that make any sense? The demo's always like, you type something in a shell and you hit return and it doesn't crash or something, right? Or types okay and then you get a null, but something amazing happened to make it so that you can type okay there. We think a lot of those companies like Sedano that feel intuitively, if you're thinking about the world of compute the way we did for the previous 30 years, this might feel a little like a pretty thin slice of salami. It's like, okay, it can do this one thing. But if that one thing saves people a couple of percent or lets them utilize their hardware a couple percent more. We're talking about a, you know, trillion-dollar industry in terms of semiconductor spend alone for inference, and it's growing really, really fast. And even if you think that that growth is, you know, the demand there is going to start tapering off or something, it still is enormously larger than your intuition might suggest if you're used to thinking about the sort of, you know, public cloud world. So really excited about infrastructure stuff near the hardware-software interface there too. And that kind of, of course, plays to some of my biases from earlier in my career.

I love that. We need to do round 2, uh, to talk about your other portfolios, your time at Slack, and etc., etc. Where can people find out more about you, Keith?

Uh, yeah, I'm on Twitter as Keith M. Adams. Uh, our website's pelobed.com. I do, uh, some blogging there, and we also try and keep it generally interesting.

Was Keith and