What's up, guys? How you guys doing?
Great. Thank you everyone for staying so late in Presidio. It's so fun.
Yeah, thanks for not leaving, except those two people.
Yeah, they're going to be back.
Well, cool. Yeah, thanks for doing this. I love the backdrop of the foghorn to remind you that it's fairly late in the evening. I wanted to start off with something fairly serious to kind of kick this off, right? Which is, what are your OpenClaw agent names? And please don't say Claudia because that's taken. I feel like you have really good ads because your Twitter handle is @LadyAI, so you have the best name for a lot of your agents probably. But mine's my dog's name Enzo, and then he's basically the engineering manager, and we have a bunch of different dog names managing different dogs. And that's been working out really well. It's easy to just come out. I go to these Italian Greyhound meetups, and I just pick up names and just name them.
Oh wow, that's really cute.
That's cool. I have a couple, but the main one I use is named Bombbot, which is a long story, so I'm not going to go into it. Doesn't matter. But yeah, I have one with a Telegram chat in my family that we use to adjudicate fights in the family chat.
So actually let's just dive into that. Like what are you using? Are you using this system? Because we were on some group chats and you were like, "I'm not really sure what this is useful for," but have you found usage for this thing?
So candidly, I have not. I think there's a lot of productivity porn out there right now about this idea that you can automate your life using OpenClaw and it'll run a fucking business for you. It'll do all this crazy shit. It's going to call people to handle customer support and just—I think the reality is that the number of people who are getting this kind of productivity out of OpenClaw is like 20, and the vast majority of other people have tried it once and they're like, "Ah, this is kind of finicky," or they're using it for very limited uses. So for me, I think it's kind of like a chatbot on steroids. There's some automation in there, but really if you want to build some kind of automated system, you're probably better off vibe coding it yourself and building it in exactly to the spec that you want as opposed to obeying the pretty rigid mechanisms that OpenClaw prescribes. But if you want something out of the box that will talk to you on WhatsApp and generate GIFs and do all this other fun stuff, yeah, OpenClaw can do that really quickly. But if you try to give it access to email and say, "Great, take over my life and don't make any mistakes," you're going to have a really bad day.
Yeah, I think the myth is that this was supposed to be a productivity tool, but we're actually spending more time refining this thing, trying to squeeze some juice out of it. I was in a board meeting last week and the CEO was just heavily distracted trying to set up OpenClaw. It was a disaster.
Wait, during the board meeting?
During the board meeting.
Is this company not doing well?
Deflect. Tell us, are you using this thing or is it working for you?
Yeah. So many people assume that VCs are really, really busy. But then, you know—
Wait, I'm really busy. What are you talking about?
We're busy, but we already kind of optimize a lot of the productivity as an investor—research and a lot of things, portfolio. And I was like, "OpenClaw is great, so what can I do?" I tried everything, but then I realized that my life is kind of simple in a way in terms of work, life, and all these things. So we actually co-hosted an OpenClaw meetup, like the very first one in SF, and I went to ClawCon and was demoing my own thing. I get so much inspiration like, "Oh, I can do that, do that." But then I come back home after the drive, and do I need that? Do I really need it?
Okay. So you were at this ClawCon conference. What did you see? Tell me about the stuff you were seeing that people were actually doing.
A lot of AI girlfriends, a lot of kind of like a concierge. But I think the inspiration that I got was the removal of an interface because you've been in tech companies, I've been in tech companies, and we learned a lot about UX/UI, but the way that this community was thinking, the interface was changing—how it's contractual, autonomous. It actually brought back a lot of flashbacks that I had to how I joined the crypto industry in like 2017. It was my first hackathon in SF for crypto, and how a lot of things are early and we kind of build for smart contracts and DAOs, not for humans. Crypto is just brutally bad for UX, and people always say, "Oh, we need better on-ramps, better UX," and all these things. But now the idea about interfaces is changing. That was the biggest takeaway for me, rather than single different ideas that are inspiration. This is like a small hint that we're changing the whole way that we interact with models, apps, or any kind of services.
Actually, the analogy that very much comes to mind—because what struck me about your answer was that there still wasn't really any examples of anyone doing anything. It's like, "Oh, AI girlfriend." Did we already have that? Isn't that the first thing people did when they had AIs? But what strikes me is it's actually very reminiscent of early Ethereum, right? If you read the Ethereum white paper, there's all this stuff in there about, "We're going to create tokens and finance and prediction markets and blah blah blah." But if you went on early Ethereum, there was kind of nothing to do, right? In the early days of Ethereum, people made some tokens, they made some scams, they made some kind of gambling games. It was all kind of stupid. If you went there, you were kind of like, "Well, no one's really doing anything. There's these big stories that people are telling, but there's not a lot of juice there, and maybe it's shit. Maybe I should go home and ignore this whole Ethereum thing."
And obviously that was the wrong answer for Ethereum, right? It was just very early, and building things on Ethereum with the level of robustness you need to actually create financial applications just took way more time than actually just building raw Ethereum, right? So it took until like 2017, 2018 until we started to see DeFi. It was like two years of people trying random stuff that just didn't work until DeFi really started to become a thing.
In the same way, I suspect for OpenClaw—OpenClaw has had such a moment, that's so much mindshare. And to be clear, these Twitter boosters who are making up these ridiculous stories about how they're automating their lives and their businesses or whatever, they're also doing important work because they're getting people excited about a vision. And people really want to believe this. They really want to believe someone out there is making $5 million ARR while they sleep because their OpenClaw built some amazing business for them. It's not true, but people are waiting for it to be true. And the moment that they feel like it's true, they're going to go try it again. So I think this ecosystem is just going to keep growing like crazy because everybody is going to keep adding and improving and solidifying this particular ecosystem until it becomes the thing everyone wants it to be.
Yeah. The best way to really learn what we can do and figure it out eventually is just doing it, right? Ever since OpenClaw and Claude Opus, we've been using AI heavily internally at HashKey, and we also came to the conclusion, "Okay, so what should we build?" We couldn't answer it ourselves. We could do a lot of different things, but let's get inspiration from others just like a typical VC, right? So we actually launched the VibeLabs incubator program as very much an experiment. Going back to your point on early Ethereum, we actually launched HashKey Studio, like an accelerator, in 2018–2019, and that's where we were trying to figure out, "What can we do with Ethereum? Let's go gaming, entertainment, all these." That's why we incubated Axie Infinity early on, and Sandbox. So maybe let's do that again.
So we launched VibeLabs, and we just closed applications. We're going through the list and we got like 400 applications, all solo founders, global, and I think like 10% of applicants are at $1 million ARR.
How quickly did that happen?
Within like five weeks, I think, and the whole program is eight weeks. So what we're realizing is that, as a builder, we were like, "Let's just try it and see what people build," and we're realizing that these things are actually working. To be honest, as a crypto investor, these are faster ARR growth than a lot of our crypto portfolios, too, right? So it's not really real yet, but then it's also kind of real in that things are moving and the numbers are coming.
So we're still really open-book iterating: "What can we do here? What's going to be our playbook? Where do we invest? What's the moat?" There's not much answer yet, but we're trying to survive as a VC to figure out how we can add value here, how we can help them. Because these are young 20s builders, solo, building million-dollar businesses. What can we do as a VC? These are connections, right? For example, one guy is building a tourism-related app, and it's making money, and we happen to have an Abu Dhabi office, so we put him in front of the Etihad chairman. That's the distribution we can make for these kind of AI builders, right? That's very different from commenting on your token model, consensus model, or how your code is structured. It's been really exciting and also kind of scary for us, but trying to really ride the wave together with the builders.
Thank you. Something you mentioned, which is Ethereum was kind of flat for a while, DeFi became a thing, TVL became the North Star metric that spurred network effects, right? Because the more TVL you have, more builders came in, that kind of spurred something. So I wonder if there's something like that with OpenClaw that's to come. I'm curious from the audience though. How many have an instance of OpenClaw running? Just curious, raise your hands. Okay, that's a good number. Of the people who raised your hands, how many have connected any form of proprietary data to OpenClaw?
Very few. How many regret it? How many have still got it connected?
She's a believer. We got a believer. Beautiful.
Love that. Okay. Humans are toast. Talk to me after. Yeah, we were to begin with. That's interesting. To zoom out, I was thinking a little bit about the scope of this conference or this event, which is around open-source AI, right? So maybe just to get a little bit of insights, what does open-source AI mean to you?
I was also thinking about this because right now the centralized or closed-end AIs are basically on a roll in terms of impacting the market, features—I use those more also. But why open-source AI would do well is probably mainly three things: auditability, that we can actually learn and trace; second is composability, very familiar with the crypto people on permissionlessness, you can build on top; and I think third is really the big thing, which is no kill switch, that you continue to build on, but then they won't really kill it from some kind of institution that can make the decision based on political reasons or business reasons. Those are big ones. And other cases, just putting my crypto perspective on—
Sorry, just something you said—there's no kill switch. With an open weights model though, I just wanted to pick on that for a second because there's so much news about this just in the last few days, it might have changed since we've been sitting in the last few minutes. But with distillation, with to your point the lack of transparency in how some of these open weights models are really built, do you really believe that there isn't such a thing as a kill switch when it comes to open weights? Because most of these are built in an obscure manner, right? We don't really know how it's—
Yeah. That's kind of where I put my crypto perspective on, where these open-source projects kind of start slow and take time, and they are still centralized in a way in terms of that kill switch. But eventually when it gets picked up and grows, it's going to be very hard to have that. I used to be a software engineer before, I used to work at AWS, and uniquely I was actually working on a lot of security open-source stacks that Amazon used. What a lot of people think open-source software is, is very philanthropic, open, goodwill, but it's actually very political because there's a lot of money involved, right? Like AWS pushing that Git repo like, "Hey, do it this way, not that way." It was a really interesting experience that I came in as a young engineer thinking open source is a good thing, but it's actually very dynamic.
So that governance and the dynamic hasn't really formed in open-source AI. That's something I'm very curious and optimistic about. Obviously there is a risk there. And another angle here is also, as Americans we talk a lot about the US-China AI angle, but I think what we are actually talking more about is open-source AI versus Silicon Valley AI right now. So I still have that optimism, but really curious how it's going to evolve.
I mean, look, practically speaking, open-source AI is Chinese AI. All the American open-source AI models suck. GPToss 12B was terrible at the time that it released; now it's even worse. So the reality is that—there was this story you alluded to that just came out a couple days ago that Anthropic discovered that a bunch of the Chinese labs were distilling against US models. And this was pretty widely known, it was pretty obvious, right? If you go to GLM-5 and you ask it who are you, it will say, "I am Claude, trained by Anthropic." So it was pretty clear that they were doing this, right? It was very, very obvious.
Very recently they found there's this network of 25,000 accounts that were illegally set up through all these data brokers that, if you're in China and you want to get access to frontier models—you're in China, you don't use Chinese models, right? If you're in China, you want to get on Claude, you want to get on GPT-5, you want to get on Gemini. So there are all these resellers that basically will proxy requests on these shared accounts to allow you to get access to frontier models, because if you're in China, you're also trying to do the same shit we're all trying to do.
So the reality is that it's an arms race, and it's going to continue to be an arms race for the foreseeable future. The problem with this story about decentralized AI, or we just say open-source AI or open-weight models, is that for AI, unlike crypto, basically nobody is ideological about AI. It's like a rounding error of people who exist who care, who are ideological about AI. People want their AI girlfriends, and they want them to be really smart and to remember their birthdays, and that's kind of it.
Yeah, exactly. It's such an incredibly persuasive technology with its sheer utility that people mostly don't give a fuck, right? There's some performative stuff about, "Oh, they're training on the artists, and what about the authors, and what about the blah blah blah," and the New York Times will write about this, but they all fucking use the same stuff that we do, right? And so I find it to be very unlike crypto in that way. In crypto, there really was this ideological contingency that was actually pretty big, people who really did believe that there's something fundamentally rotten about the core of fiat currencies. There are many, many, many fewer people who believe that about AI or who are willing to act on those beliefs.
Now, that being said, today it's actually not that hard because if you look at the frontier open models, Kimi is actually pretty good at tool calling, GLM is actually pretty good at writing. But now that this distillation door has been closed shut—we don't know how tightly it's been closed, but now that it's been closed, you should expect to see the gap grow.
And now with the power of Claude Code, the power of all these agent coding tools, the reality is that if you're not able to distill on these models anymore to the same extent you were, and if you're not getting these RL loops—in the pre-training era, it's very easy to also train on the internet, especially if you don't really care about paying the people who have the data and you just download it online, whatever, then it's very easy to get the same pre-training corpus. Post-training, which is now where 80% of the money is spent, is much, much harder to get the same data because it's all proprietary. They're generating it themselves. You can't just suck it up through the internet. So the only way you can get to the same quality is either, one, you're spending the same amount of money; two, you have the same size teams, which nobody does compared to what we have in the US; or three, you're distilling on the outputs of a model that's already done all that post-training. Those are your only paths to getting to par. So I think it's actually in a way somewhat dire for open weights to be able to continue to compete going forward over the next five years.
You don't think the cost is a factor here? Because theoretically, per token, you could get any one of these open weights models much cheaper than let's say an Opus token, right?
