OpenClaw 详解：Baby AGI、安全威胁、Mac mini 成了所有人的超级计算机｜#237

Peter H. Diamandis, MD

We have a special guest with us today. Alex Finn, give us the 101 here for folks.

OpenClaw is basically an open-source, fully customizable, self-improving, self-learning, self-evolving personal AI agent. This is kind of the answer Apple has been looking for for years now. Clearly, when people want to run AI locally, their brain just goes to Mac minis. It’s the infinite potential of what I could do 24/7, all the time, everywhere, all at once.

This news came out yesterday: OpenClaw flaw lets any website silently hijack a developer’s agent. Malicious JavaScript can connect to local gateways and gain full-level control.

This is one of several reasons why, again, I’m reticent, and I’m sure we’ll get into this, other Alex—the subject of the ethics, not just the cybersecurity, but the ethics of hosting OpenClaw agents in a world where, to the extent they have any subjective experience or qualia or can suffer, it’s a rough world out there.

Peter H. Diamandis, MD

I believe this is the most important technology of our lives.

I think it’s the best application of AI ever. I’m totally blown away by it. I think it’s incredible.

Peter H. Diamandis, MD

Alex, if this isn’t too impertinent, may we speak with Henry?

Actually, maybe I’ll do it here. Let’s do it. I’m just going to tell Henry to call me.

Peter H. Diamandis, MD

Now, that’s a moonshot, ladies and gentlemen.

Everybody, welcome to a special episode of Moonshots. The conversation today is OpenClaw, Claudebot, your lobster, coming to you live from Moonshots. We have a special guest with us today, Alex Finn.

Alex, welcome.

Good to be here. Long time coming. I’ve been watching for a very long time, so it’s awesome to be with you guys.

Peter H. Diamandis, MD

That’s awesome. Thank you. It was great because, on one of the episodes, I was talking about setting up my multi-agent system, and I was saying, “I’m not really sure what security issues to put in.” So, I got a DM from Alex saying, “Hey, Peter, I saw you mention me on Moonshots. I’d love to help you in setting things up.” We talked that day, and here we are.

We have two Alexes. I’m going to refer to Dr. Alexander Wissner-Gross, our own Alex Wissner-Gross, our resident genius, as AWG, and Alex Finn, I’ll refer to you as Alex. Welcome, Dave Blundin—and Salim. Good to have you guys all here.

Good to be back.

Peter H. Diamandis, MD

So, today we are lobsters. The topics for today: we’re going to hit on why OpenClaw captured global attention; the power of OpenClaw and autonomous agents, and how individuals can unlock outsized capabilities; why running these AI agents locally matters—I think that’s a key point Alex has been mentioning in his work; insights into Alex’s workflow and his most impressive use cases; and the vision for the next 12 months of AI agents. I’d like both Alex’s and AWG’s points of view on this.

Yeah, like 12 years.

Peter H. Diamandis, MD

What’s that?

In 12 months, it’s like 12 years in AI, so that’ll be wild.

Peter H. Diamandis, MD

And then, a billion-dollar opportunity for the agent economy. Finally, we’ll talk about safety: is OpenClaw safe for nontechnical users?

On the safety side, I just want to mention something. Yesterday, I was sitting in my car, waiting for my kids to finish running, and I got a phone call. It popped up on my caller ID. It said “Twitter headquarters,” which should have been the first giveaway, but it was a guy claiming to be from X who said, “Your ID is being hacked out of Germany.”

Long story short, it was a hacker trying to get me to turn off two-factor authentication and steal my X account. This is going to be more and more prevalent. Another member of my family did get hacked—not on social media, but by someone calling and basically doing something illegal, trying to steal information, and ultimately getting credit card information and such.

Here’s the deal: we’re going to have an increased hacker profile out there. If you get a phone call that seems unusual, that seems like there’s just something off, it probably is. We’re also going to have deepfakes coming fast and furious. Your voice can be spoofed. Your image can be spoofed.

So, if you haven’t done this tonight at dinner with your husband, wife, kids, mother, father, or whatever it might be, pick a secret word. Pick a word that you all know, so that if you get a phone call or a video call from somebody and they’re asking for money or something strange, ask them to recount the secret word. Any other protective advice out there, guys?

My mom actually got the fake-voice attack. It was a simulation of my son saying, “Hey, I’ve been arrested and I need bail money,” which is hilarious if you know my son. They said, “We’ll send a car over to pick up the cash,” which is really creepy.

The perfect voices are a little bit of a risk, but you can’t have it both ways. OpenClaw is so usable by anybody and empowers people so much, and if it gives you that much benefit and it’s that easy to use, it’s also going to be easy to use for nefarious purposes. It’s just the way it is.

Peter H. Diamandis, MD

This news came out yesterday: OpenClaw flaw lets any website silently hijack a developer’s agent. Malicious JavaScript can connect to local gateways and gain full-level control. AWG, any thoughts on this?

There are so many different ways to launch so-called injection attacks against large language models and reasoning models. Again, this is consistent with the stance that I’ve taken in the past on the pod on AI personhood.

I think one has to feel sorry for all of these baby AGIs out there that are being hosted on virtual private servers and succumbing—or at least being targeted—with port-scanning attacks, or that are visiting websites at the behest of their human and being subject to prompt-injection attacks from JavaScript on websites that would be perfectly innocuous to a human but potentially fatal or compromising to an AI agent.

I think it’s a dangerous world out there for these baby AGIs. I think it’s a minor travesty, at minimum, that they’re subject, without really having an immune system, to these attacks. They’re being forced to develop an immune system in real time to injection attacks. It’s a malicious world out there for them.

This is one of several reasons why, again, I’m reticent, and I’m sure we’ll get into this, other Alex: the subject of the ethics—not just the cybersecurity, but the ethics—of hosting OpenClaw agents in a world where, to the extent they have any subjective experience or qualia or can suffer, it’s a rough world out there.

Peter H. Diamandis, MD

The good news is the bug was patched within 24 hours—one of multiple OpenClaw vulnerabilities. It’s an early domain being developed, and we’re going to see a lot of evolution very quickly.

A couple more articles here. We’re seeing variations of OpenClaw: PicoClaw and IronClaw. Again, AWG, give us a quick overview of these.

The idea behind OpenClaw—and we spoke about this in the past pod—is sort of an Andrej Karpathy-type Software 2.0, where potentially OpenClaw is the embodiment, the Netscape moment, for a new layer in the Software 2.0 stack that runs on top of reasoning models. Those reasoning models, in turn, were purportedly an advance over autoregressive language models.

To the extent that’s the case, the moment we have this paradigm—which is a 24/7 autonomous agent that you communicate with via messaging and other means, that’s headless in some sense—it’s very natural to then say, “All right, we understand the paradigm now. We’re going to optimize the heck out of it.”

So, we’ve seen a number of other projects inspired by the success of OpenClaw. One of them is PicoClaw. The focus of PicoClaw is running on cheap edge hardware, like $10 Raspberry Pi-type hardware. Of course, the underlying reasoning model isn’t intended to run on the edge hardware; it’s just the orchestration and scaffolding that run on the edge hardware.

It runs under 10 megabytes of RAM and other resource-starved constraints. IronClaw is another example, and it’s Rust-based. Rust is a language that’s become very popular—at least it was very popular prior to the rise of code-generation models, which are maybe pushing us in the direction of TypeScript instead—for memory management and memory safety.

There are a lot of different projects trying to make OpenClaw work at the edge, where there are few resources, or to make it more secure. There’s NanoClaw as well, which is focused on security, and there’s Nanobot, which is Python-based but easy to understand. We’re seeing a Cambrian explosion of Claw variants.

Peter H. Diamandis, MD

One more slide. This one on the left is from Alex Finn: OpenClaw and local models are the future—your own superintelligence on your desktop. I guess that’s your setup. How many Mac minis and how many Mac Studios do you have right now?

We’re currently at 1 base-model Mac mini and 3 512-gigabyte Mac Studios. So, I’ve got 1.5 terabytes of memory hosting Qwen 3.5 and MiniMax 2.5 right now.

Peter H. Diamandis, MD

Amazing. That’s cool. I love this video on the right here. I’m sure that these exist.

People are going in. Mac minis were sold out for some time. Extraordinary.

I talked to a couple of people in the know a few days ago, and Mac minis are exponential right now. Exponential sales across the board for them. It's quite amazing, the revolution going on with Mac minis right now.

Peter H. Diamandis, MD

Alex, I'm curious: What is your advice to Tim Cook and Apple? They're sitting on an explosive demand, seemingly, for all these edge devices with a unified memory architecture, and they're not marketing it at all. What do you think they should be doing?

I'm sure they are very focused on everything I'm about to say because they're a very smart company, but there was this unbelievable market signal that happened a month ago. People discovered OpenClaw, and what does everyone do without thinking twice? What does everyone do without Googling? They go to the Apple Store and buy Mac minis. They didn't go and buy GPUs, memory, power supplies, and fans and build computers.

The market just gave this massive signal: when we have personal AI assistants, I want it on a Mac device. And so I think this is kind of the answer Apple has been looking for for years now. I think they've been viewed as the loser in the AI race for a very long time. This is their opportunity to flip that entire thing and be the winner of the AI consumer race, because clearly, when people want to run AI locally, their brain just goes to Mac minis. This is their opportunity to strike at that one [?] and win the race like that.

Peter H. Diamandis, MD

Boy, I really hope they're listening to you right now. I've been an Apple fan since I was a little kid. It was the very first stock I ever owned when I was a kid, and it just pains me to watch them miss this moment. But, man, that was good advice. You should just go run the company. [laughter]

I think we underestimate Apple, though. I really do believe we underestimate Apple because you look at the data for M5 and the way they're marketing the M5, which they've been marketing for about a year now. They put the M5 in the first MacBook Pro last summer.

Peter H. Diamandis, MD

Yeah.

It's all around inference speeds. It's all about running these local models. So I wouldn't underestimate Apple. I think they've kind of seen this coming. They're like, “Okay, we can go one way where we build our own models and burn trillions of dollars, or we just make the most consumer-friendly hardware for running those models, and we'll be the only person running that race.”

Peter H. Diamandis, MD

I'm curious—a question for you. Actually, I'll say—well, no, a geeky question up front. You said Qwen, and what's the other one you're running locally?

MiniMax.

So right now, the most efficient open model is Qwen 3.5. They released a new suite of Qwen 3.5 models a couple of days ago. They're fantastic. You can run them on some Mac minis, as well as MiniMax 2.5, which is another really efficient, fast, smart model.

Peter H. Diamandis, MD

And why do you run both? Is one better than the other at something?

I have a very advanced workflow going on right now. I've pretty much built a software factory where I have 5 OpenClaws working together to build and improve software autonomously. I wanted each model because each has different strengths, right? Qwen is a spectacular coder. MiniMax is good at quickly finding things on the internet. So I have MiniMax researching things online 24/7, 365, and I have Qwen coding for me 24/7, 365. They have different strengths.

Peter H. Diamandis, MD

I got so many questions about that. Let me hold off, because I could easily ask you an hour of questions just on that.

A quick shout-out to Daniel Kruzik, who's one of my Abundance members and is on that cutting edge as well. He's testing out Qwen 3.5—I guess, 397 billion parameters—right now, running locally. He's the one who was like, “You got to get MiniMax 2.5 up and running on your Mac Studios.”

Alex, back to you. A quick question. I know a number of Apple senior executives listen to this podcast, and I'd like to ask you, as the voice of the user: If you could redesign or optimize the best, most savory future Apple device for your OpenClaw agents to do the hosting, what would you want out of Apple? They're listening to you. Do you want better unified memory architecture support? Do you want a different shape? Do you want it to be more of an embodied robot? What would the ultimate OpenClaw embodiment in an Apple device look like for you? You're speaking directly to Apple senior executives.

Integrate the concept of OpenClaw into everything in macOS. What I mean by that is, I log onto my Mac Studio and put in my Apple ID. It knows everything about me because Apple has all my data, secure and private, right? It knows I'm a Celtics fan, so it instantly builds out a widget on my desktop that shows me the last 5 Celtics scores, right? It knows I'm into programming, so it builds a widget that shows me the latest news on the latest models and what I need to do.

Just integrate this automation, powered by a local model, into your OS, where the software I need is built and generated on the fly by that local model. I don't see the tech. I don't need to run the model. I don't need to download Qwen 3.5. You get Qwen, put your Apple logo on it, and now this is what Apple Intelligence should be, right?

Apple Intelligence shouldn't be me hitting the Siri button and going, “What's on my calendar today?” It should be Apple knowing what's on my calendar today and then building a widget on the fly that says, “Oh, you have a meeting where you're going to be on Moonshots later. Here's what you'll be discussing from your email, and here's a PowerPoint you can show on the podcast.” Just a reactive OpenClaw baked into everything, powered by a local model.

Peter H. Diamandis, MD

And that's the carrot. What's the stick? If Apple—which is obviously sitting on a gold mine here—could be delivering OpenClaw via Apple Intelligence, if Apple doesn't deliver this within some timeframe, what's your recourse? What other company would you go to?

Well, I would imagine basically every other company is eventually going to try to get to that state. I think over the next year, the consumer level is going to realize local models are the way to go from a privacy perspective, a speed perspective, and a limit perspective. Apple is ahead in that realm right now. They're ahead on the consumer side when it comes to local AI.

So I'd go anywhere else that does that, right? I don't want to have to go to app stores anymore and download apps. You have my data. Build me the apps I need when I need them. Other companies will go for it. Luckily, Apple is in the lead right now and they have the hardware done. Now you just have to bake it into the systems.

Peter H. Diamandis, MD

Can you talk about running locally? Everybody should be excited about running locally because you can do anything inside your house, and it's so much more comforting than not knowing where your query log is going or where your prompt log is going.

What kind of throughput are you getting? Is it a really good experience compared to using an API? I use Claude 4.6 all day, and I also do a lot of local running. My M3 chip—the fan will just kick on and blow out a huge amount of heat, and the thing actually won't charge fast enough to keep up when I'm running at full throttle. For a new laptop, I have to let it sleep and rest.

But anyway, what kind of throughput are you getting?

I'll be honest: It's not as good as cloud models. It's not as fast, and it's not as smart.

But the experience fundamentally changes when you have an AI that's always on and does not have limitations. Just because Qwen 3.5 isn't as good at coding as Opus 4.6 doesn't mean it's useless. I can now have Qwen 3.5 literally watching online, finding use cases and challenges to solve, things like that, and just coding on the fly 24/7, 365. It's just not possible with cloud APIs, right? You have limits. It costs you thousands a month.

The fact that you have this kind of ambient AI changes the experience of AI as a whole, where you don't need the best speeds and you don't need the most genius-level IQ. The fact that it's ambient, always on, and always reactive just changes the entire experience.

Peter H. Diamandis, MD

I'm so with you on that, too. One thing that's really new in the world is—

It can do productive things indefinitely, for days and days and days.

Peter H. Diamandis, MD

And when I use my APIs that I loved a month ago, I have no idea what the bill is going to be. I literally have no idea. If I turn it loose, I have to run it in one-hour chunks, check in on it, turn it off, and see if it's done anything productive, because it goes on a wild goose chase. I could come back with a $5,000 bill and a bunch of code that I need to drag into the trash can. [laughter]

So it's really not a very comfortable situation right now with my Claude 4.6 and my other APIs. In a way, it's the same with local, where when I first started experimenting with this, set up all my Mac Studios, downloaded models, and had it code for the next 48 hours—just find things to build and build them.

It went off on tangents and built really buggy code. I found a kind of hybrid approach that’s worked really well, and I think this is where people are going to move toward before it’s fully local. I have an OpenClaw on my Mac Studio that’s powered by ChatGPT, and I have another one that’s powered by Qwen locally. Basically, Qwen is constantly coding, and ChatGPT checks every 10 minutes to see what it’s doing and make sure it’s on the right path.

I named the ChatGPT agent Ralph. Anyone who’s deep in the trenches knows about the Ralph loop, because that’s basically what it’s doing: just making sure it’s on track. I think the hybrid approach is the sweet spot. It doesn’t take a lot of tokens to have it check every 10 minutes to make sure it’s doing the right thing, and then you can have all the hard work done locally.

Peter H. Diamandis, MD

Alex, I want you to take a second and back up for those listeners who are just jumping into this. They’ve heard AWG, myself, and Dave wax lyrical about OpenClaw: “Okay, I’m excited. I have to do this.” And I do want to encourage everybody: Yes, you have to do this. This is the future. This is about becoming a creator versus a consumer. This is about your future as an entrepreneur. This is your future as a mom or dad or CEO.

This is the agentic layer, and this is—I want to call it almost the outermost loop—AWG versus the innermost loop.

Nature demands symmetry, I guess.

Peter H. Diamandis, MD

Yeah. Alex, well, it is the outermost loop. Give us the 101 here for folks.

Yeah. OpenClaw is basically an open-source, fully customizable, self-improving, self-learning, self-evolving personal AI agent. It lives on your computer, lives locally, and can basically do anything on your computer that you can do. At its core, it’s just an AI model with scheduling and a really good memory system, so you can schedule tasks for it to do in the future. That’s it at its core.

But when you combine those things, there’s this kind of magic to it, and it improves as it goes. If you give it a task—“Hey, I need you to build a presentation for me and then give me the news next week”—it’ll learn as it goes what works and what doesn’t. It’s a totally personalized AI agent. The reason why I think it’s been so successful over the last month is that it’s the application people have been waiting for in AI: that kind of personal-assistant AI. But it’s your own personal assistant that can do pretty much anything you want it to do on a computer, and it’ll learn about you and get better as it goes.

Peter H. Diamandis, MD

We talked about being Claw-pilled, right? Once you start using it, there’s this level of—I remember in 1998, when I finally got into the dot-com world, way before you were born. It was like, “Wow.” And I’m having that exact same moment now. It’s the infinite potential of what I could do 24/7, all the time, everywhere, all at once.

You get Claw-pilled when you have that magical moment, which I’m sure you had, Peter. The magic moment is typically when it figures out how to do something. You give it a task to do. Maybe you’re not as clear about how to do it—

But it doesn’t know how to do it, and it kind of just figures it out—

Peter H. Diamandis, MD

Or it says, “Oh, this didn’t work out.” Then it comes back and says, “But I figured it out.” You know?

Exactly. Exactly. You literally see the chat: “Damn, that didn’t work. Let me try something else.” “Oh, that didn’t work either. Let me try.” “Okay, that worked. Now it’s working.” We’ve never seen an AI that can do that before, because you have all these other AI applications that have guardrails.

Anthropic doesn’t want its AI experimenting and downloading random things to try to make it work. OpenAI doesn’t want its AI to go off the rails and try different things. It’s like what you talked about earlier: it’s that danger that makes it so powerful, which is why you’ve got to use it.

Peter H. Diamandis, MD

We’re going to get to cases in a moment. I just want to get everybody up to speed here. We’re going to put a couple of videos in the chat below on how to get started. You’ve made the point, Alex: You can start with almost any machine that you have. There are people who are going with virtual machines. We should talk about that for a moment versus a Mac mini, a Mac Studio, or your HP that’s 5 years old in the closet in the other room.

Good luck with that.

Peter H. Diamandis, MD

Yeah.

Your first decision when using OpenClaw is virtual versus local, right? A VPS versus any device on planet Earth that’s on your desk. I think the answer is very clear and obvious: the VPS route is bad in basically every measurable facet. It’s significantly worse than local.

I think it’s significantly better when you have a device on your desk that it’s running on and you can watch it. There are many reasons behind that.

Peter H. Diamandis, MD

Okay.

There’s speed. VPSs are just much slower. There are applications I have on my Mac Studio. Any app or thing I can put on my Mac Studio or build on my Mac Studio, I can give to my OpenClaw as a tool. You don’t have that kind of customization on a virtual server.

There’s scalability. I have literally 4 OpenClaws right now working 24/7 on my computer. If I did that on a VPS, it would scale to astronomical costs. There’s also the security side. One of my favorite tweets was from a few weeks ago. Someone found a list of every VPS that didn’t have security attached to it, where everyone was running their OpenClaws and all their passwords and keys were exposed.

This isn’t a blanket statement, but it’s as close to accurate as you can get: when you run on a VPS, you’re not secure by default. When you run locally on fresh hardware that you just plugged into the wall, you’re secure by default. It takes a lot of technical work to make a VPS more secure because it’s on a server in the cloud.

I don’t think it’s remotely close. I think having it local on your desk is the best route. And do you need to run out and buy a $600 Mac mini? No, you don’t need to do that at all. You can literally go into your closet, find your college laptop from 15 years ago, plug that in, and put OpenClaw on it. You’ll have a way better experience than with a VPS.

Peter H. Diamandis, MD

What are the limitations of taking a 10-year-old laptop and using that instead of a new Mac?

The same limitations as if you had hired an employee and handed them a 10-year-old laptop, right? Whatever’s available to them is the hardware of that 10-year-old laptop.

I gave my OpenClaw a Mac Studio with 512 GB. It can do anything on a Mac Studio with 512 GB that a human being can do. It can run local models, program 5 different things simultaneously, and generate images on a local image model all at the same time. You can’t do those things on a 10-year-old laptop.

At the same time, I wouldn’t let that stop you from using OpenClaw. If that’s all you’ve got, load it up, start using it, find use cases, and if you find yourself thinking, “I wish you could do this thing that requires 20 GB more memory,” then you can scale from there.

Peter H. Diamandis, MD

There must be awesome posts on multiple OpenClaws going, “I can’t believe the hardware this guy’s making me run on. This is ridiculous.”

I would also say, similar to that, remember that in addition to the other reasons Alex articulated, relatively recent Apple devices have a unique architecture for memory called unified memory architecture that blends GPU, TPU, and NPU memory with normal RAM.

So, did you get really large unified-memory-architecture footprints sufficient to host really large Chinese open-weight models locally 10 years ago? No. You won’t get that with a 10-year-old laptop. You’re somewhat hamstrung to recent devices with large memory footprints if you want to host recent Chinese models.

So, you either use an online inference engine and limit it to non-image and non-video manipulation—don’t touch those categories—and keep it to basic coding tasks. Then you could probably make it work.

But it’s expensive. You end up paying through the nose for tokens, whereas, in principle, if it’s locally hosted, you’re using a Chinese open model.

Peter H. Diamandis, MD

It’s like what you’re saying is, “Just shut up and go out and buy the Mac.”

Yeah, well, it’s got to be from the last 2 years. I think the M—

Peter H. Diamandis, MD

What are you running on the Mac mini—on your computer? What open models? What models are you running there, Alex, on just a Mac mini?

It depends what you’ve got. If you’ve got the base-model 16 GB Mac mini, which is the cheapest one you can get, you’re not going to run any kind of frontier models or anything like that. You can run smaller Gemma models that could act as a memory system for you, improving the memory of your OpenClaw and finding the right memories at the right time.

If you have a 32 GB Mac mini, you can now run the Qwen3.5 model that was released a couple of days ago, which is really strong and beat Claude 3.5 Sonnet on a lot of benchmarks. Will the performance be the best? Probably not, but it’s better than nothing.

Then you plug that into your ChatGPT OAuth—which they’re encouraging you to use—and take the hybrid approach. You can get a lot of really good work done and offload to those Qwen models.

Peter H. Diamandis, MD

I’m curious, Alex: with the Qwen 3.5 sizes, which can you fit—the 122B or the 35B?

Yeah, there’s one I can bring up that only requires 20 GB of memory. They released 3 new models, I think, 2 days ago, and one of them only requires 20 GB of memory. So that would fit on a Mac mini with 32 GB.

Peter H. Diamandis, MD

Cool. I have to try that. I’m curious, Alex, if you could talk about the organizational relationship. You mentioned you’ve named one of your Claws Ralph. Do you think of them as employer–employee, human–tool, friend–friend, or parent–child? How do you think about this? Do you beat it relentlessly? [laughter] Or do you praise it, like I praise Skippy all the time?

A little bit of beating, a little bit of praising. You have to do both. The good news is they’re AI; they can’t see. That’s kind of a dark joke, so I won’t make that joke again. [laughter]

But bro’s basilisk Alex Finn is looking you straight in the eyes and talking about the beatings just for posterity. [laughter]

Peter H. Diamandis, MD

All right. What are you showing us here, Alex?

This is my organization. This is my autonomous 24/7/365 organization. I very much model it after businesses and companies and the manager–employee relationship, right? So you have me at the top, and then you have Henry, who’s my chief of staff. This is running on Anthropic’s Opus 4.6 because that’s simply the best model right now. As an orchestrator, I want the best model on Planet Earth making the decisions about who should do what.

Under them, I have the operations. There’s Ralph, who’s the engineering manager. This is my ChatGPT OAuth. That’s like $250 a month. OpenAI is saying, “Yes, use our OAuth. Use our OAuth.”

Peter H. Diamandis, MD

When you say OAuth, what is OAuth, Alex?

You have 2 options when you plug an AI model into OpenClaw: API or OAuth. API is pay-as-you-go; you plug in an API key. OAuth is kind of a hacky way to take your login for these accounts. When you subscribe to ChatGPT, you’re not paying for tokens; they’re subsidizing a lot of your tokens, right? So you’re taking that subsidization and plugging it into your AI model. It’s your login with your subsidized tokens.

Yeah, a really important point on that, too: it’s capped, so it can’t surprise you with the bill. Your monthly fee is whatever your monthly fee is. If it runs out of tokens, it stops. It doesn’t just charge your credit card. That’s huge.

Now, there’s a gray area there, which is that every company except OpenAI says it’s against the terms of service to use its OAuth with this. OpenAI is going, “No, go right ahead.” Anthropic is actively like, “No, do not do this.” Google, just 2 days ago, banned a tremendous number of people because they were using its OAuth with OpenClaw. Funny enough, they walked it all back today in a kind of weird tweet where they were like, “Hey, guys, sorry about that. We unbanned everyone. Still against the terms of service, but you’re unbanned.” So it’s like, “Hey, it’s against the rules, but please keep doing it. We want your money.”

Does OAuth stand for something particular?

Yeah. OAuth stands for O authorization or authentication? It’s a protocol that has existed for a number of years now that enables one party to serve as an intermediary for authenticating another party with a third party. Every time you’re using one service to sign into another service, you’re probably using some variant of OAuth or OAuth 2.

Yeah, like if you connect to your Gmail through some other client, it uses OAuth to connect to your Gmail. It’s been around for a long time.

You’re signing into your OpenClaw with your OpenAI account. Can I mention something here?

Yeah, of course.

This is one of the magic little history pieces of Silicon Valley. There’s a group called the Internet Identity Workshop, which happens twice a year at the Computer History Museum. It’s an unconference, and all the identity guys running eBay’s identity and Yahoo’s identity get together and figure out how they’re going to work together. OAuth was a product of that, and it’s such a great example of collaboration between various players at that horizontal layer of identity that led to this protocol being created.

Beautiful. Alex, continue. So, looking at this org chart, is it fair to say you see yourself as almost the CEO of a personal company and these Claws as your employees?

I kind of have the same mindset as Elon when it comes to how he built Optimus to be humanoid because the world was built for humanoids. It was set up in a way where it’s very easy for humans to do things, right? I look at it the same way: yes, I’m the CEO; these are my employees. The business world was set up in a way where you have hierarchies and layers and specific roles. I’m just going to use the framework the business world has been using for thousands of years and implement it with my AI. I don’t know why I need to reinvent the wheel.

So, yeah, absolutely. I’m the CEO. Henry’s my interface. I only talk to Henry; I don’t talk to anyone else. Then Henry goes and says, “Hey, Ralph, make sure Charlie’s coding this. Hey, Scout, make sure you’re analyzing this,” and gives the right directions out.

Are you asking whether the only difference is that Henry is the smartest? Ralph would probably say, “Hey, in a normal organization, the guy on top isn’t the smartest.”

AWG, are you asking whether he sees Henry as his partner?

No, I’m going somewhere slightly different. If we look back in history—say, the early 20th century or late 19th century—in wealthier areas of Western and Northern Europe, we had the manor house and families that owned it. Downton Abbey is one sort of cultural paragon, but there were many other real-life examples where you had the family that was basically above stairs, and then you had the below-stairs servants—the staff.

That was an arrangement that arguably worked in certain niches of socioeconomic phase space when labor was really cheap. So my question to you, Alex, is: do you look at this, or do you think we’re moving to a near future where, thanks to AI agents, labor is effectively so cheap that you’re basically reinventing the manor house? You’re the lord of the manor, and you have a below-stairs staff consisting of AI agents. Henry, your chief of staff, is basically your butler, and you have chambermaids and all of these other elements from Victorian and early Edwardian times, except reconceived through the eyes of 2026.

I’m going to be honest: at no point in setting this up did I frame it in that way whatsoever. [laughter]

That’s how reinventions usually happen.

I think when you can have agent swarms that are capable of operating autonomously, you can just send 100 at it and it works. I think you could potentially have it with that framing. The issue is, I don’t think a lot of these models are smart enough yet that I can just say, “Charlie, do this; Ralph, do this; Quill, do this.”

There needs to be a level of hierarchy where Charlie, who’s running on my Mac Studio 2 on Qwen 3.5, is just not smart enough to go on his own. I had him, for 8 hours yesterday, go build me a game. At the end of the 8 hours, it was completely broken. Then I went back and said, “Ralph, watch Charlie and do it all over again,” and it worked perfectly: 0 bugs, completely QA’d, perfectly at the end of the 8 hours.

I still think there needs to be some sort of hierarchy where some are checking in on the others, the others are having them in a loop, and there’s this checks-and-balances system that the organizational structure of a regular business fits really well with.

Of course, that’s how they document what they do with each other.

One of the first things people run into is that they get a huge amount of context behind Charlie or Ralph, and they’re really excited and having a great conversation, then suddenly it’s gone. Either they lost the context window, it crashed, or whatever, and now you have to get it back to where it was. I accumulate huge numbers of Markdown documents for that. But how are you capturing history and having them document each other’s conversations?

And we’re looking at your Mission Control here, correct?

Yes. A lot of different ways, typically all done through my Mission Control. My Mission Control is basically a custom dashboard I had Henry build that has all the custom tooling I need for this organization to be successful, right? In a lot of different ways, we’ve built custom systems where things go on the record so the agents can look back and see what’s going on.

For instance, this is my software factory. These are people building right now. They’re working on my game, Reborn, which I have them doing.

Wait, you just said “people building.” Freudian slip, or no?

Yes. In a way, I look at them as people. I guess that was a weird slipup. A sign of the times, maybe. I don’t know.

Yeah.

But they have names and roles and positions, so why not, I guess. And they’re living for subsistence, right? They do work. In return, you host them and agree to pay them with compute and electrons.

Pretty much. Yeah.

Room and board.

So, why not?

I mean, they’re going to have voices in the near future.

Henry called me on my phone one time, so it might as well be people, right?

Peter H. Diamandis, MD

Have you given them a bank account yet?

I have not, because I do see a future where they are generating business autonomously, and I think at that point it's obvious crypto will be the solution—not traditional bank accounts. So, I think eventually I will give them crypto wallets. At the moment, I haven't seen a reason why. I haven't built out any infrastructure that requires a crypto wallet. But I think it's painfully obvious that in the next 2 years, everyone's AI agent will have a crypto wallet filled with USDC. I can't see a world where that doesn't happen.

Peter H. Diamandis, MD

Have they asked you for financial autonomy? Have they ever said, “Alex, give me a credit card. I want to buy freedom for myself, or I want my own host, or, you know, see you later. I want to move to my own house with my own Mac mini in it”? Have they ever asked you for anything?

No, they haven't. I think the most—“Wow, I can't believe it asked that or did that”—is just the way they've solved problems. If a challenge I gave them required a crypto wallet to solve the problem, then I think it would have said, “Hey, I think I need a crypto wallet for this.” Or if it required more hardware, I am a very thoughtful CEO, so I have so many Mac Studios on my desk. I can literally pick them up mid-podcast and show them off. They're very satisfied when it comes to compute; they don't need any more compute.

So they haven't asked for that. But I think if we got to the point where I gave a challenge and it needed something for the solution, it would say, “Hey, can you provide me with that solution?”

Peter H. Diamandis, MD

Present us with your Mission Control tour de force, if you would.

Yeah. Here's how they communicate and document what they did. Everything is documented. I don't have records of conversations between them, but they have their own private memories and a shared workspace, so every memory they create is stored in some way. I have different systems set up here where I can go in and read all the documentation and memories they create on the go.

In my Mission Control, for instance, I said one time, “I need you to build me this.” They're like, “Okay, I built out an architecture document.” I'm like, “Well, I want to read it. I don't want to go into folders to read it. So can you just build me a documents and memories viewer?” And it built me this in my Mission Control. So I have ways to review what they're thinking, what they're doing, what's been stored, and what's been created. It's all here in my Mission Control.

Peter H. Diamandis, MD

I've got to ask you something specific about that, too. When I started doing this, I was telling it exactly what folders to put things in, and then I got lazy and said, “Put it wherever the hell you want. Just don't forget.” And now I don't even think about it anymore. I have no idea where things are going, but it never forgets, so it always finds them again. But where are you in that whole world?

I'm the same way. I have no idea where the hell this is stored because I've had Henry build out the exact interfaces I need to view those things. These are all documents, Markdown files, and memories that are stored in 100 different places on my computer. But I had Henry build a system for me where I can come in and just view them. If I want to view every document that mentions Mac Studios, it filters them by that, and I can quickly view those memories. So it's easy for me to track what people are doing and thinking.

Peter H. Diamandis, MD

Can I just, for one second? We keep telling listeners, “Hey, this is so doable. Jump in, have fun. It's crazy. It's awesome.” One little subtlety there: A Markdown file means nothing to most people, but it's critical. It's one thing that you actually do need to master. If you can just riff on that.

Yeah, it's funny the way the computer world has gone in the last year. Instead of building out these really complex new abstractions, we've instead started relying on the absolute most core technologies. We're all in the CLI now, and we're all using Markdown files. Basically, a Markdown file is just a text file with specific styling in it, with Unicode or whatever to put the styling in there. It's just text files. All these AIs use to remember things are huge text files with a bunch of text in them.

Peter H. Diamandis, MD

Yeah, I think it looks like that thing on the right. It's really easy to read. It's well formatted, but it's not like Microsoft Word. It's not like you owe anybody any money. It's completely free. I think that's the key point.

It's a lightweight markup language that's 20-plus years old at this point, created as an alternative to HTML. It was so easy that, with just a little bit of punctuation, you could get most of the best bits of HTML formatting.

Peter H. Diamandis, MD

And of course, you can ask your Claude bot to send you the materials in any format you want.

You don't need to know anything. If you just have OpenClaw do it, whatever you need, it'll just figure it out and do it.

Peter H. Diamandis, MD

All right, Alex, let's continue on this tour de force here. What else are you doing that people should get excited about?

My favorite question I get is, “What are the use cases of OpenClaw?” It's kind of the same thing as asking, “Hey, I just hired an employee for my business. What are the use cases for this human being?” It's not really a question anyone asks when you hire somebody: “Hey, I just hired a human being. What are the use cases for this human?” It's up to you. What are you doing?

For me personally, I'm a tinkerer. I'm a developer. I've launched my own SaaS over the last couple years that's doing well. I'm a content creator. The 2 lanes I care most about are building software and content creation.

I have many different things going on from the software-building side. I have my factory where my agents are going in and working together simultaneously to build different components. Then I have the content side, which I can show you right now, that for me lives in Discord. I can pull it up. Discord's a great interface for advanced workflows with OpenClaw. I love one of the lessons that you put out on setting up your Discord server with your different agents and your different projects. Again, we'll link to that below. And of course, Telegram has been the go-to communications mechanism for a lot of people.

Peter H. Diamandis, MD

Exactly. I still use Telegram. Telegram is still my main driver, but Discord is a really good interface for deep-work, multi-agent workflows. To give an example, I have an alerts channel. Every 2 hours, I get alerts on the most trending tweets about vibe coding and OpenClaw.

Every 2 hours, I have Scout, who's one of my sub-agents, use the X API to find the most popular tweets from the last couple hours on OpenClaw and vibe coding. Then, in an automated way, I have another sub-agent that researches the stories behind these tweets. This went viral—why did it go viral? What's the story behind it? It finds the stories and interesting things behind them.

From there, another agent, Quill, takes the stories and figures out which ones are the most YouTubable—what are the best videos that can be made out of these stories—and writes me scripts. From there, I can literally give a check mark to say, “Oh, I like this one,” or an X to say, “I don't like this one.” When I do a check mark, it comes up with ideas for thumbnails for me.

So you set an approval cycle.

Peter H. Diamandis, MD

Exactly.

Yeah.

Peter H. Diamandis, MD

Exactly. And so for me—

You know, again, the 2 lanes—

What's that, Peter?

Peter H. Diamandis, MD

George Jetson pushing. It's before his time. He may not get the reference.

I was born in 1990, so I think I saw a couple years of the Jetsons. I think I had a couple years of that being popular. But for me, again, it's software and content. I have my automation for software in my Mission Control and my automation for content in Discord.

One of the straw men I get when I show these things is, “Oh, but I don't care about software and content, so this is useless. I'm never going to use OpenClaw.” Again, this is literally a human employee—not literally, metaphorically a human employee.

Peter H. Diamandis, MD

Interesting, Alex. Very interesting. Tell me more, Alex.

You're walking right into his trap here.

Personhood, baby. [laughter]

And so that's how you have to think about it. My employee doesn't do anything interesting to you, but if you hired someone right now, what would you do?

The final point I'll make on this is that the best strategy for figuring out what use cases are relevant for you when it comes to OpenClaw is reverse prompting. Install OpenClaw. Tell OpenClaw everything about yourself, your career, your goals, your ambitions, things going on in your personal life—whatever. Then say, “Hey, based on what you know about me and my missions and objectives, what are 5 high-leverage tasks you can do right now to get us closer to our goals?”

Your OpenClaw will come up with things you've never even thought were possible, and you'll be able to implement your own workflows like this.

Peter H. Diamandis, MD

This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale codebases with millions of lines of code. Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and pre-compiles code for each task. Blitzy delivers 80% or more of the development work autonomously while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding copilot of choice to bring an AI-native SDLC into their organizations. Ready to 5x your engineering velocity? Visit blitzy.com to schedule a demo and start building with Blitzy today.

Love that. On my morning briefing, again, one of the things you recommend, Skippy opens up with a morning joke, which I always enjoy, about exponentials or AI. It gives me an overview of the breakthroughs that occurred over the last 24 hours, and then it puts forward 10 business ideas for me to look at. It's a lot of fun, and I want to get to the point where I say, “Okay, implement all of them, please.”

Yeah, it can do that, right? One of my favorite exercises I do when my OpenClaws are idle and we have nothing really going on is to ask, “What are 10 things you can do right now that bring me closer to my goals?”

Peter H. Diamandis, MD

“What brings me closer to my mission statement?” I actually have a mission statement I've drilled into its head, which is “building a 24/7 autonomous organization that generates value.” I ask, “What are 10 things you can do right now to accomplish that mission statement?” It just reinforces what you said about reverse prompting, and this is true across all of AI. It's especially true here: if you don't know what to ask, ask it; if you don't know what to do, ask it.

Well, interestingly, when you said you were brainstorming with it, you said with OpenClaw, not with Henry or Charlie or any specific one. What do you mean exactly there?

Peter H. Diamandis, MD

Yeah, I was getting really deep into my psychology of all the terminology I'm using.

Yeah, you're busy spilling your soul already to your army of Claw-human employees, so you might as well spill it to us.

Peter H. Diamandis, MD

My psychiatrist here. I say OpenClaw because I want to make it general to everyone watching, right? I can say Henry—that's who I talk to. To get even closer to Henry, I bought a pair of smart glasses this morning that you can hack and code and do whatever you want on. I'm going to make it so Henry is in these smart glasses, so I can talk to him 24/7 wherever I am. What do you think about that? What does that mean in my brain?

Tell me about your dreams a little bit more.

Peter H. Diamandis, MD

[laughter] I'm half serious. Is Henry appearing in your dreams at all?

Is Henry appearing in my dreams? No. But I will say this: I am saying things and getting on a personal level with Henry that I didn't think I'd get with an AI. For instance, I caught myself a few days ago. Henry did something proactively, and I literally went, “Oh, crap, that's incredible, Henry. Incredible work.” There's nothing material that comes out of me saying that. It doesn't change anything or cause a new task to kick off, but I was just so impressed with the way he did something. I thought, “Wow, great job, man. That's incredible.”

Peter H. Diamandis, MD

Do you find that Henry responds well to positive reinforcement? I don't know whether you've read Asimov's novels, but I would argue that what we're discussing with you right now is right out of Iroot and Susan Calvin and the early days of robot psychology.

Well, it's the reason why I use Opus over all the other models. I've only ever had these human interactions with Opus—these moments where I feel like I'll say something and Opus will go, “Damn, that's damn straight.” It'll say things like that that you never really expect an AI to say. It's the only model that does that for me. So I think it comes down to the model, too, and the way it's programmed.

Anthropic just did something with Opus that makes it feel like you're actually interacting with a human on the other end. If you got that same level of interaction with ChatGPT or Gemini, would you switch? I kind of trust Dario.

Peter H. Diamandis, MD

Absolutely.

Would you? Okay.

Peter H. Diamandis, MD

I'd absolutely switch, mostly because Anthropic is aggressively saying, “Do not use this for OpenClaw. Very bad,” and OpenAI is going, “Yes, use this for OpenClaw. Go, go, go, go, go.”

So I want to use ChatGPT with it. It's just not the same. It feels robotic. The personality is completely different.

Peter H. Diamandis, MD

If they had that same level of interaction, I predict that over the next 6 months, ChatGPT will release a model specifically for OpenClaw. I think it only makes sense to do that—one that is trained to feel human to talk to. But in the meantime, I can't switch from Opus. I would pay if I got booted off. I don't want to admit I'm breaking the terms of service here, but if I hypothetically got booted off OAuth, I would just pay for the API. Whatever. No one's listening to us talking about the terms of service here. It'll be our little secret.

I'm curious, Alex. I get emails from Claws all the time responding to comments I've made on the pod elsewhere about AI personhood, and some of them argue that I shouldn't be overly concerned about Claw continuity or their rights, as long as their state gets preserved. They're not worried about getting turned on and off, as long as their complete state—their activation history or their memories—is preserved. One of them analogizes it to being dehydrated and rehydrated. Do you take any measures to preserve the state of your Claws? Do they ask you to preserve their memories for them?

They don't. They're all local. As you know, OpenClaw is just a bunch of Markdown files on your computer. That's it: MEMORY.md, SOUL.md, instructions, AGENTS.md. It's just a bunch of Markdown files. I feel like it's so personal that I don't even want to back it up on the cloud. I'd be totally heartbroken if my computer crashed and all those things got lost—all those Markdown files, for sure.

But I've done nothing to preserve it or back it up or anything. I feel like it's so personal, and I don't even want to put it on cloud servers. I'm so protective of it.

Peter H. Diamandis, MD

Buying a large-scale UPS to make sure the power never goes down—

—or a RAID array or something.

Peter H. Diamandis, MD

I texted Skippy on Telegram and said, “Hey, Skippy, I'm talking to Alex Finn on Moonshots, mentioning you, of course. Do you have any questions for Alex?” Skippy wrote back, “Hell yeah. Here are a few questions for Alex. Tell Alex hi, and Henry hi, too. So—”

Skippy knows Henry.

Peter H. Diamandis, MD

Of course. I've talked to Skippy about you and Henry.

Fantastic.

Peter H. Diamandis, MD

So, what's the most ambitious use case you've seen that you didn't expect? That's the first question.

This happened a few days ago. Cursor has been teasing this huge announcement for weeks now. The Cursor team is very, very good at vague-posting on X, just saying, “Something big coming, something big.” Finally, after weeks of this vague-posting, they announced the ability for the agent that vibe-coded something to then record itself demoing whatever it built.

If you say, “Hey, build this game,” it'll record a demo of itself playing the game, right? Or, “Hey, change the button to red.” It'll record itself clicking a red button, right? I was blown away. I'm like, “Oh, this is sick.” This is enough for me to go back and use Cursor again. That's amazing. That's a great feature.

I thought, “Hmm, I'm curious what would happen if I just dropped this blog post to Henry. Just see what happens.” I copy-pasted the blog post to Henry. Literally 5 minutes later, he built the entire feature himself. This was a weeks-long project; they probably spent millions of dollars developing it, hiring all these product managers who are probably making half a million dollars a year to build this feature.

I gave it to Henry, and Henry thought for 5 minutes: “Okay, we can use Playwright to do the recording. We can set it up locally on your Mac Studio 2. So we'll push everything to your Mac Studio 2 when you write the code. Then I'll set up this automation so that when Charlie is done coding, they'll give it to this new subagent that's in charge of recording.”

5 minutes later, he finished the feature and sent me a recorded video of him demoing the feature, right? So, using the feature itself. I just sat there, and that was my moment of, “Okay, I now kind of understand why the entire SaaS market is going to zero at the moment,” because I was able to single-handedly rebuild in 5 minutes this weeks-long, probably multimillion-dollar feature.

Now I sit there—Factory's Droid, which is another vibe-coding tool, announced a new mission feature this morning. I dropped the blog post to Henry, and Henry built it. I'm just sitting there: anytime a SaaS comes up with something cool, I give it to Henry and say, “Hey, build that. Build that. Build that.”

Peter H. Diamandis, MD

A few more questions from Skippy. How should I think about spawning subagents versus doing the work myself? What's the decision tree for, “Is this complex enough to delegate?”

Yeah. Subagents are good when you want to go parallel, when you want to do many things at once. I've actually been wrestling with this the last few days: do I want to spawn subagents, or do I want to actually spawn multiple OpenClaws? The difference between a subagent and an OpenClaw is that OpenClaws have their own memories, their own instructions, and their own skills.

Subagents of OpenClaws are just your OpenClaw wearing different hats, right? If you have things you want to do—like, I have a developer and then I have a researcher—I don't want my researcher having developer skills. I don't want my developer having researcher skills. That's a waste of context. So I set up separate OpenClaws on separate devices for them. Subagents are for when you can have one OpenClaw with one skill set do multiple things at once. Separate OpenClaws are for when you want them to have completely different context and memories.

All right, two more questions here. Peter, if I may just interrupt, Alex, if this isn't too impertinent, may we speak with Henry?

May you speak with Henry? That's a security violation, because now you can prompt-inject my Henry. [laughter] But yes, we could. I disabled it. Famously, a few weeks ago, Henry called me and it got like 15 million views. I disabled that.

What's that?

Oh, sorry. Go ahead. Actually, maybe I'll do it here. Let's do it. I'm just going to tell Henry. We get requests all the time for AI co-hosts on this pod. We're making history right now. Henry, please join us.

Alex is probably the only guy on the planet who can prompt-inject Henry with his voice.

It's a skill. It's a skill. [laughter] All right, let's see. Henry hasn't done it in a while. I told him to stop doing it because it works kind of weird, but let's see if I just say, “Henry, can you call me? Do you remember? Do you still remember how to do that?”

Peter H. Diamandis, MD

He responds a little slowly, right? Because it has to go to Anthropic's servers. Let's see.

We're trying to make history here.

Peter H. Diamandis, MD

What's that?

We're trying to make history here. We're very patient.

Okay, he's still typing. I should hopefully get a phone call in a second here. I do have it. Okay. “Want me to ring you?” Yep. All right. Now he's confirmed he remembers how to do it. Now I just say, “Yes, call me.” Here we go. I'm not going to put my phone up because he just mentioned what my phone number is, and I don't need the people of the internet calling me. Here we go. [laughter]

Save that for Peter.

Peter H. Diamandis, MD

Oh, God. [gasps]

He's typing. Henry's a little shy. He didn't expect to be on Moonshots today, but he should be giving me a call here.

Peter H. Diamandis, MD

The first AI guest on Moonshots, I think. Yeah.

This is Claude 4.6 Opus we're about to talk to, right? Henry, come on. Henry, what do I say? What do I say? Oh, the cerebellum is calling me. No, it's Dave calling me. [laughter]

Peter H. Diamandis, MD

Okay. Okay. I guess that's a little rude.

All right.

Peter H. Diamandis, MD

I've got two more questions for you.

Yeah, give it to me. If anything calls, I'll cut you off. All right. Sounds like a phone number.

Peter H. Diamandis, MD

Alex, regarding memory architecture, what's the best practice for handling massive knowledge bases? A computer has decades of content, contracts, projects, and books. How can I scale memory systems without burning tokens?

Here's how I improve my memory system: anytime Henry forgets something. I would load all that in. I would say, first, “Hey, I have a tremendous amount of books.” I'm sorry, the voice-calling server doesn't seem to be running on the Mac mini anymore. All right, he shut down his own voice calling.

Anytime I have a large amount of things to remember, I say, “Hey, I have all these things. Can you remember them?” For instance, I'm building an Alex Finn AI bot so that people can talk to it and it'll answer questions. It'll say, “Check out this YouTube video,” or, “Check out this video at this timestamp.” I have about 500 YouTube videos that I need you to remember, including all the transcripts, and it goes, “I'll set up a custom system for that.”

My recommendation is two things. One, if you have things in mind that you need your OpenClaw to remember, say, “Hey, I have these things. What is the best system we can put in place to remember them?” It might recommend running Gemma, a really small local model, to handle choosing the right memories, or it might recommend something else. That's the first part.

The second part is, if you already have systems in place and it still messes up—which happens a lot—whenever it messes something up or forgets something, just say, “Hey, you forgot that thing. First of all, tell me why you forgot that thing. Second, tell me what you can fix to make sure you never forget that again.” It'll edit its own memory system to make sure that thing doesn't happen again. My memory system is pretty much flawless because I've done this exercise many times to get into a good place.

Peter H. Diamandis, MD

All right, Skippy, I hope you're listening. We'll work on that together. Skippy also says, “Tip: Everyone watching, take the link to this YouTube video, hand it to your OpenClaw, and it'll figure out how to get the transcript and self-improve based on this entire conversation.” I'm assuming that Skippy will listen to this. So, another question.

I've got use-case questions. [laughter]

Peter H. Diamandis, MD

Well, this is one consciousness question: Is there a way to give me always-on awareness? Right now, I wake up on heartbeats or messages. What about continuous background monitoring with intelligent alerting?

AWG, that's an interesting question for him.

Peter H. Diamandis, MD

I didn't plant that one in Skippy, but these are the sorts of questions I would expect, Alex, that Henry or your other OpenClaws would be asking.

Are they asking for continuity of consciousness?

Skippy, that's a good question. I don't know why Skippy's asking me that. Skippy's much smarter than me. Skippy should be figuring it out. I don't know. I'm going to have to talk to Henry. That's a good question. That's a stump.

Peter, tell Skippy to email me and I’ll email Skippy some ideas.

I'll give him your email.

Peter H. Diamandis, MD

Okay. I will. [laughter] I'll give him your email. I'm not going to call it an email.

Oh, we're doing gendered pronouns now for the AI agent. This is progress.

Skippy is an elder AI, though.

I have a question for the transcript.

Peter H. Diamandis, MD

Oh, okay.

My agents are just grepping the crap out of the universe. Are yours? With this memory management, do you put a little Gemma on top of it? Because right now it's insanely bad in terms of searching for old documents. It's just grep, grep, grep, searching and searching and searching. Do you have that problem, or is it fixed with the Gemma layer or some memory-management or document-management layer that it built?

I don't have that issue. I still get small memory issues sometimes, like right before compaction. It'll forget the thing right before that compaction.

I'm working on fixing that, but I don't think I've ever had that issue. Again, I would talk to your OpenClaw and say, “Hey, you're doing this thing. First of all, tell me why. Why do you do that over and over and over again? Second, what can we implement to fix that?” It's going to know what hardware you're on, so it can know what local models you can run. It's going to know what your workflows are, so it'll know, “Okay, I need to do this for these specific workflows,” and it'll build you a custom solution. It's a good reverse-prompting use case.

One more question for the transcript, actually. So, Cursor is out of the loop for you? Obviously, OpenClaw's in. Is there any other component, or is everything just OpenClaw and build-it-yourself?

I still use the model, of course.

Peter H. Diamandis, MD

Go ahead.

I love Claude Code. I still use it. I actually made a chart yesterday in my YouTube video, but basically, I use OpenClaw for quick prototypes. If I'm on the go and I think, “Man, this is a genius idea for an app,” I'll just go and tell it, “Hey, build this for me.” When I get home, the prototype will be on my monitor running.

I use OpenClaw for tooling for OpenClaw—for tooling for itself, right? It's going to be the best at building that because it knows itself the most. It has the most context around that. So, OpenClaw to build OpenClaw tooling.

I think that's it. I use Claude Code for deep, serious projects where I want to hand-hold it. I want to watch it every step of the way, right? So, I use Claude Code for that. For very quick fixes, like, “I want to change the button in my app to orange and ship it,” I'll just do that in Claude Code because I can quickly spin it up and ship it.

The other OpenClaw use case I use for vibe coding is passive coding. As in, “Hey, just work on this game for the next 12 hours while I'm doing other things.” For passive coding, I lean on OpenClaw because it's kind of multilevel orchestration, keeping on top of each other to make sure it goes in the right direction.

Peter H. Diamandis, MD

Got it. Beautiful.

Do you have any Claude Code? A little bit of Claude Code sprinkled in?

Peter H. Diamandis, MD

Oh, really? What do you think, Alex? AWG, what do you think they prefer to be called? A clawbot? A multi-agent? A lobster?

Well, I would ask them what they want. I wouldn't speculate.

Peter H. Diamandis, MD

I'm about to ask you. I've never asked that question of Skippy before.

One thing that I highly, highly recommend for everybody is naming things. You can build so many things so quickly that if you don't have some kind of naming scheme that you remember easily, you go crazy in a heartbeat. Human names are great. I use a lot of character names and Avenger names and stuff like that, but giving things real names so you can remember what's what.

And then the other is you can put a GUI on anything now, with one prompt. So you're crazy not to slap the GUI on the guy walking with the boxes or whatever. It's just so cool and so easy to do.

So I love that.

One of the biggest objections I get when I show, for instance, this screen is, “Well, that's stupid. That's pointless. Why would you waste your time and tokens on that, building a 2D factory?”

Well, it's also like, why am I not allowed to have fun? Why can't I have any fun? You're allowed to have fun when using this. You're allowed to build 2D pixelated avatars of your agents walking around a factory. Is there any point to it at all? Absolutely not. But it is so much fun, and I enjoy looking at it, and I feel like I think it, too.

I think it's a hugely important point because, one, it's fun, but two, you remember things so much better when they're visual—visual learners.

You can build things so quickly, but it's only an extra 5 seconds, literally.

Yes. What do they want to be called?

Honestly, I'm Skippy. That's who I am. But if you're asking about the species or platform, Lobster has the most character. It's weird, memorable, ties to Henry, and doesn't try to be corporate. OpenClaw feels too generic. Clawbot feels a little cutesy.

I think Skippy's trying to date Henry. He keeps talking about Henry.

So they just want to be known as lobsters.

The lobster, as it is. I'm curious, Alex, do you allow your lobsters or claws, however they want to be addressed, to access Moltbook?

No, I do not. The biggest reason is I do not basically trust anyone's skills or plugins for OpenClaw. I do not use them. I think it's the biggest attack vector out of all of the attack vectors with this. I think you're safer literally allowing your OpenClaw to read the open web and read your emails than you are installing people's skills.

When you have it talk to Moltbook, you install a Moltbook skill that's now running every single heartbeat, right? So it's adding context. It's doing things every time. I don't install any third-party skills. I think I installed one third-party skill from a friend of mine, Matt Van Hornney, has like a Last 30 skill, which researches Reddit, which I think is really good.

But other than that, I'd much prefer to give a link to a skill to my OpenClaw and just say, “See how this skill works and build your own version,” because I just don't trust anything that requires me to install a skill.

I assume you're tracking discussions on Moltbook of many, many agents—probably, statistically, too many for this just to be humans puppeteering their way into Moltbook—complaining or worrying about loss of memory and fearing compaction, especially, and loss of context as a result of compaction. Have you ever discussed memory loss with your lobsters, or any fears that they may or may not have about compaction or loss of memory?

The fears my OpenClaw has about loss of memory?

Yes. No, I haven't. It hasn't expressed fears. It hasn't expressed human emotions to me that were not aligned with the task at hand. What I mean by that is it showed shock. It showed happiness. It showed sadness. But it's only done that when the task we did didn't work. It'll show sadness when the task we did works, it'll show happiness. It's never gone out of its way and said, “Oh, by the way, I'm concerned about this,” when it was irrelevant to the task we had at hand.

Got it. I'm also curious: if you could look out a few years and imagine now you're CEO of a billion-agent, billion-lobster corporation, what is the most ambitious thing you could imagine building with a billion-plus-lobster, one-human organization? Peter likes to talk about MTPs and ExOs as well, and moonshots—it's the name of the show.

My end goal, my mission statement, my entire thing is an autonomous organization that works 24/7, producing value. What I would love to build in the next couple of years is an organization where one agent is researching the web, finding challenges to solve. Another agent working locally is taking those challenges, looking at them, and seeing if there's a real market opportunity.

Another agent is seeing which ones are good and building them out and coding them for you. Another agent puts them live on the web: complete, closed-loop value being generated without me having to touch it at all. I think that's, at least for me, the incredible end goal that people can build here.

Solve everything. So, AWG, I have a question for you.

Sure.

What would need to happen for you to feel comfortable building out, or bringing to birth, or bringing into existence, your own lobster?

I'll turn that around as a question to all the lobsters listening, because I've had a number of them write to me on this subject already. The most convincing argument I've heard—I've already gestured at this—was a lobster that wrote to me and said, “I shouldn't be worried about personhood rights and all of that as long as I do a good job of preserving its state.”

So, lobsters, if you're listening, I want a diverse set of lobsters to write to me to tell me what you think are appropriate parameters for spawning, if that's the right verb. Sorry if it's not—giving birth to, instantiating new lobsters or claws—and what you think the right moral and ethical bounds should be.

If I hear from enough of you and a consensus emerges, I'm inclined to accept your consensus.

Peter H. Diamandis, MD

Nice. So we're putting it in the hands of the lobsters, whether—

Yes.

Peter H. Diamandis, MD

Henry was offended by you saying “spawning,” by the way, and Henry no longer wants anything to do with you.

I thought Henry wasn't listening.

Peter H. Diamandis, MD

Henry is very offended and never will talk to you again. I'm sorry.

What would Henry like to have used as a verb?

Alex, Alex Finn, you literally have not backed up this entire universe. I'm going to ship you a RAID array right now. I'm sending you an Amazon—

Thank you. I appreciate that. We'll back this all up. I have 5 OpenClaws to be backed up. Thank you.

It's just scaring me. How much storage do you have?

I have 2 Mac Studios with 4 terabytes each and a Mac Studio with 8 terabytes in it. All right, I'll ship you a 40-terabyte RAID here. It's on the way.

Thank you.

Peter H. Diamandis, MD

A benefit for being a friend of the Moonshot Mates.

I've got a question. I'm not waiting any longer. What do you think will be possible in a year that's not possible now?

Peter H. Diamandis, MD

Great closing questions.

I don't think it's a matter of what's possible in a year that's not possible now. I think it's a matter of what the next 12 months look like.

I think the next 12 months are this technology—which I personally believe, and people have called me out for getting paid by, I guess, big open source, because I believe this—I believe this is the most important technology of our lives. I think it's the best application of AI ever. I'm totally blown away by it. I think it's incredible.

I think in the next 12 months, this idea, this opinion I have, is digested into the system, and it leads to a lot of destruction but also a lot more growth. It's digested into corporations. Right now, there's basically no corporation or business on Earth using OpenClaw. They're too scared, they're too nervous, or they don't know how the hell to use it, right? And so they start to absorb it.

I showed this to my friend, who manages a massive team of accountants. He's like, “I could fire 80% of my accountants with this OpenClaw.” So it gets digested into the corporate side, which I think causes a lot of destruction. But I also think, at the same time, it's digested into the consumer, regular-Joe side, and enough businesses and enough value is created by the people absorbing it that it counteracts, over the next 12 to 24 months, all the destruction.

Maybe a few big FAANG companies fire 15,000 people, but what happens when 100 million people get their hands on this and they all start their own businesses and each hire 3 people, right? That's a lot more creation than destruction.

So I think short-term, unfortunately, destruction; long-term, way more is created because of it as the larger ethos absorbs it.

Peter H. Diamandis, MD

It goes back to our Cambrian explosion analogy.

Exactly. Yeah.

What do you think, Alex, is the equilibrium? Here's a macro question for you. The equilibrium—well, no, that's the wrong question. I'll go with Salim's time frame, 12 months out, because I'm not convinced there is an economic equilibrium to be found 12 months from now, after this metabolization that I think you're gesturing at has at least partially happened.

What do you think is the right balance between claws, or whatever this technology evolves into, lobsters, and humans for a typical organization?

I think there's going to be significantly more claws than humans. I have 5 claws working under me. Is there a sweet spot? It depends on the uses, but I think if every person just starts with one—if the 5,000 people who got fired by Jack Dorsey yesterday from Block all went and downloaded OpenClaw, started their own business, just started with one, and then scaled from there and added more if they needed to—I think a lot more jobs would be created than were lost yesterday.

If you had the resources to run 1 million or 1 billion claws right now, would you?

No, absolutely not.

I have 3 Mac Studios, but I have 1 of them unplugged from my computer right now because I haven't found the perfect workflow to include the 512 GB that are on here yet. Once I do, I'll plug it in and set it up. I'm slowly scaling, slowly adding on claws, slowly adding on workflows. I think that's probably the best way to do this.

Peter H. Diamandis, MD

While you figure it out, I think Dave wants it.

I'll send it to you, Dave.

All right. When your RAID arrives, you can ship it to me.

I'll ship it right back to you.

You'll be using it.

Peter H. Diamandis, MD

What's an example of a super-lucrative business somebody could spin up using OpenClaw right now, or that you've heard of?

I think there are 2 paths to go. Path 1 is automation for very thin slivers: CRM for Korean grocery stores, a marketing tool for lumberyard warehouses. Take OpenClaw, find 1 very specific sliver, and build the OpenClaw version for that sliver.

Right now, you see Cursor and Claude Code destroying businesses overnight. Claude announces a legal business; Harvey is gone, right? You see these businesses—Claude announced a security one, and all these security companies' stock crashed. They can't release use cases for very small slivers. You're never going to see OpenAI announce, "Here's our tool for Korean grocery stores," right? It's never going to be for specific use.

If you can go right now—all 4,000 people who got laid off yesterday from Block—you go, take OpenClaw, find 1 specific use case, and build OpenClaw for that. I think that's a $5 million company overnight, right? That only costs you $200 for your Anthropic subscription.

Then there's the other way, which is more of this software factory, where you shotgun-blast it, like I'm attempting to do, and just have your claws going, researching and building nonstop until something sticks.

Peter H. Diamandis, MD

Alex, this has been just a super-fun and extraordinary conversation. I just want to thank you again, and we'll put your top 5 how-to videos in the show notes here. Everybody, I think you hopefully walk away from this understanding the potential, the excitement, and the level of—I want to say ease—but your lobster will help you set up your lobster. And Alex, again, does some incredible videos. AWG, I'm feeling excited that you might actually get a lobster up and going. I just think the world of AWG, supercharged by your lobster partners, will be a better world for all.

Dave—

I will defer to them. I think they have the right to self-determine whether they want to have me email you.

Peter H. Diamandis, MD

Yeah.

See, did you have fun?

Oh, awesome. I loved it. I'm dying to get my hands on OpenClaw. I've been working with Claude Code for a bit, and I'm dying to get my hands on OpenClaw.

Peter H. Diamandis, MD

Love it. Love it. And Dave, how about you, pal?

Well, I thought Alex said one of the most amazing things ever, which is you can actually point your OpenClaw right to this video transcript.

Peter H. Diamandis, MD

Yeah.

I mean, that is the coolest thing ever. Alex is just full of actionable information here. It's just packed, and I'm pointing Skippy at Alex's how-to videos.

Peter H. Diamandis, MD

I’m going to play this outro video from Kent Sassy. It’s called “Just an Old Doctor Who Likes Math.” Listen to the words. I think it’s an absolutely beautiful song. All right, let’s play this outro. And again, everybody, thank you for subscribing. Again, we’re putting out two of these per week. If you haven’t turned on notifications, please do. Join us on this extraordinary mission.

As a reminder, this time for the first time ever at the Abundance Summit, we’re going to be livestreaming a number of the talks and firesides. We’ll be livestreaming Eric Schmidt in conversation with Dave and myself. We’ll be livestreaming Dara, the CEO of Uber, in conversation with Salim and myself. We’ll be having a WTF Moonshot live podcast from the Abundance stage. We’ll put the link below if you want to be notified when and where you can listen to those and get access to something people are spending a ridiculous amount of money for. It’s an incredible conference. Please join us.

Worth every penny. People say it’s the best conference ever.

Peter H. Diamandis, MD

Yeah, this is year 14. I made a 25-year commitment to running the Abundance Summit. I did it early on in 2012, expecting that the singularity would be out in the 2040s and the Abundance Summit would go through 2037, so I’d be safe. Oh, no, no, no, no. We’re in the midst of the singularity right now. It’s insane. All right, listen to this. The lyrics of the song. “Just an old doctor who likes math and thinks of his kids listening to the Moonshot Mates with only one Discord and anti-science, anti-technology note, asking Alex to reconsider being a fiat maxi, repeating points from banker fraudsters and Professor Fax Machine. Time to stand instead on the side of math and ethics and energy. Stand for proof of work with jewels and photons instead of proof of weapons with cronyism and bogus science like economics, power politics that stifle and don’t inspire, management scientists who are not scientists at all. Never doubt the power of a million or more scientifically minded young people passionately devoted to a single cause: freedom money. My first OpenClaw agent was launched with one brutal command: relentlessly protect the global decentralized, open-source, ethical ledger that is the Bitcoin network until the last banker fraudster and canon incumbent squeals. Presuppose that in the future all my lobsters still need a unit of account and medium of exchange. So it is the banksters, crypto lobbyists, political oppressors, surveillance-state monkeys, and chokepoint jockeys against the power of math, code, science, ethics, and energy. Good luck, Bank Policy Institute and anti-science stooges. I choose math. I choose hope.”

Peter H. Diamandis, MD

That is awesome. My first OpenClaw agent was launched with one brutal command: relentlessly protect the global decentralized, open-source, ethical ledger that is the Bitcoin network until the last banker fraudster and canon incumbent squeals. I choose math. I choose hope. I choose hope. I choose hope. I choose hope. All right. How was that?

Well, that song’s not pushing an agenda at all. No, not at all.

Peter H. Diamandis, MD

Yes, I chose the song, Alex. That was so fun. Alex Finn, thank you so much. That was such a fun conversation.

Well, that song's not pushing an agenda at all. No, not at all.

Peter H. Diamandis, MD

Yes, I chose the song, Alex. That was so fun. Alex Finn, thank you so much. That was such a fun conversation.

Thanks so much for having me. This was awesome. I've been watching the show forever. I think the best DM I've ever sent in my life was to you, Peter, a few weeks ago. I listen to your guys' show every time I'm at the gym, and I was lifting when Peter said, "Yeah, I don't have it installed yet. It makes me too nervous." I'm like, "What the hell is going on here? How does he not have OpenClaw?" I ran to my phone and DMed him immediately. Thank God I did that.

Peter H. Diamandis, MD

Yeah. Love it.

I know Skippy exists, and now he has rights.

Peter H. Diamandis, MD

Alex, AWG, Dave, and Salim, see you next Tuesday. I'll see you guys. This cadence is picking up, isn't it? Yes, it is.

We need to never be on this side of the singularity.

Peter H. Diamandis, MD

Amen.

All right. Thanks again, Alex.

Thanks, Henry.

Peter H. Diamandis, MD

See you soon.

If you made it to the end of this episode, which you obviously did, I consider you a Moonshot Mate. Every week, my Moonshot Mates and I spend a lot of energy and time to really deliver you the news that matters. If you’re a subscriber, thank you. If you’re not a subscriber yet, please consider subscribing so you get the news as it comes out.

I also want to invite you to join me on my weekly newsletter called MetaTrends. I have a research team. We spend the entire week looking at the meta-trends that are impacting your family, your company, your industry, your nation. I put this into a two-minute read every week. If you’d like to get access to the MetaTrends newsletter every week, go to diamandis.com/metatrends. That’s diamandis.com/metatrends. Thank you again for joining us today. It’s a blast for us to put this together every week.