AI Agents Destroying Internet Security (Anthropic's Leak, ClawCon Tokyo, OpenAI's $852B Valuation)

Summary

More or Less – April 3, 2026

Episode: AI Agents Destroying Internet Security (Anthropic's Leak, ClawCon Tokyo, OpenAI's $852B Valuation)

Hosts: Dave Morin, Jessica Lessin, Brit Morin, Sam Lessin

---

Episode Overview

This episode dives deep into the accelerating chaos created by AI agents in cyberspace, the security risks exposed by a recent Anthropic leak, the nuances of agent-driven software development, and the evolving landscape of big tech (highlighted by OpenAI’s staggering $852B valuation). Broadcasting from Tokyo after their appearance at ClawCon, the crew debates whether the promise of AI is empowering or apocalyptic, the cracks in Internet cybersecurity models, and what the new era means for both software businesses and individual creators.

---

Key Discussion Points & Insights

1. ClawCon Tokyo: An Insider’s Look

[00:50 – 6:38]

• Vibe & Experience: ClawCon Tokyo, attended by over 1000 people (most in costume), displayed the intersection of Japanese tech culture and global AI trends.
  • “Japanese love their cosplay.” – Brit [01:58]
  • “Zen environment... No one had to be asked to take their seats. They're already in their seats 10 minutes before.” – Brit [03:38]
• Real-Time Translation Breakthrough: Japan’s largest LLM provider ran live translations, seen as a business/technical leap.
  • “A real breakthrough in technical and business communication.” – Dave [02:14]
• Unique Japanese Perspectives: More optimism about AI than in the US, shaped by demographic urgency and a pragmatic culture.
  • “AI is quite positive here versus in the US it's, like, so negative...we want to know when something is made with AI or not, but we don't, like, hate it.” – Brit [04:05]
  • “They're running out of people... They just need agents to survive.” – Sam & Brit [06:14]
• Emerging Research: The Nintendo-connected ‘Artificial Life Institute’ explores evolutionary computing: “Can you understand life well enough to imbue it into a computer?” – Dave [05:00]

2. The Anthropic “Claude Code” Leak & Security Nightmares

[07:06 – 13:59]

• Leak or PR? Discussion of recent “leaks” at Anthropic; some regarded as marketing stunts, the code leak seen as an actual blunder exposing Typescript frontend code instead of core secrets.
  • “The Claude code thing I think is like a legit fuck up.” – Sam [08:23]
  • “There is some interesting stuff in there in terms of how it works, which I thought like you generally wouldn't want other people to necessarily know.” – Sam [08:56]
• Supply Chain Weakness: Vulnerabilities stem from the deep dependency chain in software—AI agents now compound this risk.
  • “It's wild...this small piece of code some random dude on the Internet maintains is like the linchpin of the entire Internet.” – Sam [09:33]
• AI Agents Accelerate Risk:
  • Agents now autonomously seek, combine, and deploy software from the internet at velocity, without human diligence.
  • “You're moving faster, you're checking less...the cost is that there's all this implicit risk and subversion.” – Sam [10:44]
• Security vs. Speed Tradeoff:
  • “You’ll be able to charge an enormous premium for security and privacy. But the problem is...your stuff's going to suck compared to the stuff that moves fast.” – Sam [11:36]
  • “We’re kind of witnessing...the final Internet disaster of AI.” – Sam [12:17]
• Collapse of Trust:
  • “You're just going to have all these agents running around doing crazy shit...no one will trust anything and the whole thing collapses.” – Sam [12:27]

3. Real-World Example: Jessica’s Bot Security Dilemma

[13:59 – 18:32]

• Practical Security Headaches:
  • Jessica describes her increasing willingness to sacrifice security for convenience while integrating agents like Claude into her workflow (allowing permissions, skipping due diligence).
  • “Before I know it, I'm just installing dev toolkits... Did you press always allow?” – Brit [15:17]
• Bundled vs. Unbundled AI Interfaces:
  • After struggles, Jessica reconsiders if AI agents should be central interfaces, or if their features should be embedded natively in each service.
  • “After my experience...I'm not sure it needs to be all together in one place.” – Jessica [17:57]
• Sam’s Take:
  • “First of all, I sandbox this stuff aggressively...I don't run any of this stuff on my actual computer.” – Sam [21:29]
  • He demands total freedom for builders, even if it means taking on security risk personally.

4. The “Raw Metal” Future: No More Rules

[21:29 – 27:56]

• Collapse of Platform Control:
  • With models and agents capable of coding and orchestrating directly “to the metal,” usage restrictions by individual services are futile.
  • “As long as someone in the world provides some digital service, it will be found and you can instantly use it, which means that there can be no rules.” – Sam [27:44]
  • “One person sharing all the weights...away from like anyone being able to do any of this anywhere.” – Dave [27:56]
• Software Customization & One-off Tools:
  • The rise of “vibe coding” and personal apps—both Sam and Dave recount rebuilding existing services like “granola” out of frustration.
  • “You just want like customizations.” – Brit [28:41]
  • “People are going to be able to just request the software...and it will just materialize onto the Internet.” – Dave [31:04]

5. What Happens to Software Businesses?

[31:06 – 32:11]

• The End of Traditional Software Models:
  • With infinite customization and AI-generated tools, classic software business “moats” dissolve.
  • “That's why I'm so bullish on software and so negative on software businesses.” – Sam [32:04]

6. OpenAI’s $852B Valuation & Tech Funding Hype

[32:11 – 39:20]

• Retail Investor Influx:
  • New “synthetic” venture fund models allow small investors to buy into private companies like OpenAI, SpaceX (via platforms like Robinhood).
• Skepticism on Valuations:
  • Sam derides OpenAI’s valuation:
    • “This is like a ridiculous. This is just ridiculous. I mean, it's like whatever...The story is people don't know where to put any money, and money's cheap, right?” – Sam [33:31]
    • “It’s a commodity business with no defensibility...I don’t understand how any of these companies make sense.” – Sam [34:04]
• Anthropic Growth Projections:
  • Dave reviews a bullish future valuation and revenue chart for Anthropic, which Sam sharply dismisses.
  • “48 billion in EBITDA is like a very large number, that does not justify the valuation that we're talking about.” – Sam [36:27]
  • “People need to store their money somewhere and they need to store it in hope...this is the hope thing at scale...from a fundamental business perspective, none of this makes any financial sense.” – Sam [37:04]

7. Local Inference: Hype or Hope?

[37:39 – 39:20]

• Dave's Bet:
  • He backs “local inference” (AI run on edge devices, not just in the cloud) as a possible differentiator.
  • “We’re going maximum Deep in open claw...how is that.” – Dave [37:56]
• Sam’s Counter:
  • Still, hyper-flexibility and ease of switching between platforms implies no real lock-in or profit moat.

---

Quotes & Memorable Moments

• On Network Security Collapse:

  • “It's always been wild...this small piece of code some random dude on the Internet maintains is like the linchpin of the entire Internet.” – Sam [09:33]
  • “You're just going to have all these agents running around doing crazy shit and including things. They have no idea what they're doing and...no one will trust anything and the whole thing collapses.” – Sam [12:27]

• On Builder Power:

  • “I want the rawest tools available and then take responsibility for sandboxing them myself.” – Sam [22:16]

• On the Software Business Model:

  • “That's why I'm so bullish on software and so negative on software businesses. Because I'm just like...These are just tokens.” – Sam [32:04]

• On Individual Empowerment:

  • “The amount people are using software is not going down...people are going to continue to want software like they want music. And the fact that more people can make this than ever before I think is a really good thing.” – Dave [44:02]

---

Actionable Takeaways & Closing Thoughts

[39:56 – 46:34]

• Sam’s Call to Action:

  • “The number one thing you must be doing is aggressively playing with everything...you should be playing with everything and having your own opinions.” – Sam [40:11]

• Dave’s Perspective:

  • “This stuff is just incredibly empowering...watching people who are normies able to create their own software.” – Dave [42:31]
  • “Does that mean there’s going to still be as many trillion dollar companies? Maybe not. But...you can build deca-million, hundred million, billion dollar companies—and a lot more of them.” – Dave [44:02]

• Semantic Battleground: Who is a “Software Developer”?

  • “The number of software developers goes down. Just the number of people who...build software the way I write essays goes up.” – Sam [44:59]
  • “Do you still feel...that the number of software developers is going to go up or just the number of people who...build the way they write an essay?” – Sam [44:59]
  • “Whatever we call them, we call them. But I think it's...very great.” – Dave [46:27]

---

Important Timestamps for Segments

• [00:50] – ClawCon Tokyo, Japanese AI take
• [07:06] – Anthropic leak, marketing vs. real security breaches
• [10:44] – Security model breakdown with AI agents
• [13:59] – Jessica’s “Bot Corner” and permission fatigue
• [21:29] – Developer’s view: sandboxing agents and “raw metal” computing
• [27:44] – Collapse of service “rules,” the futility of platform limits
• [31:04] – The future of software: personal applications, “vibe coding”
• [32:11] – OpenAI’s $852B valuation and the logic (or lack thereof) behind it
• [37:39] – Local inference as competitive edge
• [39:56] – Advice for listeners; empowerment, chaos, and creativity

---

Summary for Skimmers

• The internet's security paradigm is imploding as AI agents autonomously assemble software from vast, unvetted libraries—creating both wild new capabilities and exposure to invisible threats.
• Japan’s positive, pragmatic view on AI contrasts with western anxiety—partly because of demographic pressures and a culture of practical adaptation.
• “Vibe coding” and personal software will explode, driven by agents and accessible AI—obliterating traditional software business moats, but opening creative agency to all.
• Valuations and capital flows into AI companies are “narrative-driven” and not grounded in business reality; the abundance mindset makes hope (and hype) into an asset class.
• Main takeaways: Play with the tools now, experiment, and recognize you don’t need to be technical to build software in this new era; the future may not belong to software companies, but to empowered creators.

---

End of Summary.

Transcript

[A]

Basically, as long as someone in the world provides some digital service, it will be found and you can instantly use it. Which means that there can be no rules. You're just going to like have all these agents running around doing crazy shit and including things. They have no idea what they're doing. And they'll check the APIs, be like, yeah, that thing works and it sort of will, but there's going to be like a, like something hidden in it and like everyone's going to do it and no one will trust anything. And the whole thing collapses. I need something that does this, I need something that does that. And just pulling shit out of GitHub off the Internet, more or less.

[B]

Is it no or is it yes?

[C]

We'll debate the text. That's best when we get More Or

[B]

Less Dave and Grip plus Ham and

[D]

Jess put it all right to the test.

[A]

More or less.

[C]

Why? Hello, friends. Welcome to More or Less Morins. Where are you? It looks dark and stormy.

[B]

It's because it's morning time on the other side of the world. We are in Japan. So for an annual trip we do with some of our offline founders, but also because we got to put together Claw Con Tokyo, the first ever open claw event in Asia, by the way,

[D]

with our dear friend Michael Galpert.

[A]

Galpert's there. I love Galpert.

[B]

Yeah, Galpert.

[D]

Aaron Satig's over here.

[C]

Wait, Aaron sitting there? You got Aaron Sitter got a plane. That's like the headline news.

[B]

He loves Japan. It's like the design mecca. It's easy to get designers to come to Japan.

[A]

It's easier to get Aaron to meet you somewhere outside the US than inside the US I've actually seen him far more outside the United States than inside. Hard to get him to the peninsula easily to get him on a 12 hour flight.

[C]

I want the three takeaways more. And what is the vibe in Kyoto? Give us the. The chi.

[D]

So we had clock on Tokyo, I guess it was the night, night before last and we had over a thousand people show up in, in Shibuya, mostly

[B]

in costume, mostly in full lobster. Japanese love their cosplay.

[D]

Yeah, we had great costumes, but also a Tokyo fashion and you know, they put together a really beautiful event. They had, you know, a lineup of fantastic local speakers. Everything from, you know, your large language models. The biggest large language model provider in Japan actually provided a really amazing real time translation feature. So we had a, you know, massive screen that showed the content that people were talking about. But then along both sides of it you had real time Japanese and English translation happening with the best large language model from Japan, which was actually a really cool experience. Like, I've never seen that happen before in real time at the speed at which it was happening was. It was a real, in my view, like a breakthrough in technical and business communication. It was really, really neat. So that was cool. We did a Fireside chat. One of the top podcasters in Japan talked to both Peter and I on stage about what's going on.

[C]

What was the toughest question you got? Dave?

[D]

What was the. I don't know, Brit. What was the toughest question?

[B]

They're pretty straightforward question.

[C]

I mean, it was probably fluff.

[D]

Well, it's early here, Jess.

[B]

Yeah, it's interesting being here. It's like a month or two behind, you know, like, obviously super hyped about openclaw and all.

[A]

And respectful. They're just respectful.

[B]

So respectful. Oh, my God, Sam.

[D]

That was actually the funniest thing for all of the Americans were like. Like, Michael Galpert was like, ready to go in there to, like, energize the crowd, like a rally. And everybody's so deferential and chill. And it was like a very Zen environment.

[B]

No one had to be asked to take their seats because they're already in their seats. They're in their seats 10 minutes before it started. No one's talking. Everyone's being patient and kind. It ended exactly on time.

[A]

Dave, you're looking a little Asian. Did you get a haircut?

[D]

I'm just not wearing a hat, Sam.

[C]

That's what Dave's hair looks like. Okay, so you guys are out there. Any other, like, what's. What's, like the ground of Japanese tech right now.

[B]

Like, what.

[C]

What's everyone talking about over there?

[B]

We actually talked to the lead. One of the lead actors is from Shogun and also some of the Nintendo family we got to meet with yesterday, which was very cool. And, like, everyone's perspective on AI is quite positive here versus in the US it's, like, so negative. You know, AI is less popular than ICE in the US and like, all kinds of terrible things. And in Japan, they're just like, oh, yeah, like, it seems it's fine. Like, obviously we want to know when something is made with AI or not, but we don't, like, hate it. We think it's just part of how we are evolving with technology.

[D]

Yeah. One of the cool use cases that we heard when we were talking with the actor from Shogun, he was the guy that played the sort of Judas character. In Shogun, the guy who kind of betrays his master and ultimately in the end ends up dying. So he was kind of the lovable villain. But they were talking about how one of the use cases that people are using things for is when you're developing a new script, being able to embody the character into a contextual window in the language model so that you can then explore what the character might do in any given situation to make your acting better. Which I thought was a really cool use case. That was neat. And then on more the Frontier, Jess, we have very close relationship with the Nintendo family. They're one of our partners at Offline, and they have a. A project here in Kyoto called the Artificial Life Institute. And so they're focused on evolutionary algorithms and how to embody life into computing, which is kind of a question beyond language models is, you know, can you, can you understand life well enough to imbue it into a computer? And so they have a bunch of very interesting research going on around that and some great partners at the University of Tokyo and, you know, various tech companies in Japan. And so that, that's actually a very interesting area of, I think, unique research that Kyoto is very interested in right now.

[C]

Fascinating.

[A]

Well, to be clear, the Japanese are also much more okay with this because they're running out of people.

[B]

They just need agents to survive. They're like, we must evolve into agents so that we survive.

[A]

They're like, america doesn't have a population problem yet, but Japan totally does. Like, yeah, we need robots. We've been in robots. And we want AI immediately because we're running out of humans. Right. So, like, I think that's part of why the vibe is so different there on this stuff.

[D]

I don't know, I think that it's just, I mean, yeah, I guess that's the funny take on it, but I think they're just more utilitarian about how they think about technology that are just like, oh, yeah, this is a good new tool. We should integrate it and we should do great research. And they're much more pragmatic and matter of fact about it and that it's just like this thing that's coming and of course we should use it and of course we should figure it out and like, let's be really great at it. Right? Like, that's kind of more the vibe here.

[C]

Got it. Okay, well, I'm gonna break precedent and let you guys pick the first news that we jump into. So consider this a Russian roulette of headlines of all things AI. Okay? We have a Claude code leak. OpenAI closing. What is it? $122 billion round at some 800 and something.

[B]

852 billion now.

[A]

$852 billion. Is that pre or post?

[C]

We've got Iran threatening US data centers.

[D]

In the US or, or globally.

[C]

In the Middle East. In the Middle East. So let's start there. We've got the fundamental massive, massive, massive fundraise cloud code. Second leak from Anthropic in like a week. The first was the details of Mythos, their new model and then this Iran story.

[D]

Sam, this sure looks like a marketing strategy to me. Not a, not a leak strategy.

[C]

You think it's a marketing strategy?

[D]

Yes.

[C]

Okay, say more, Dave.

[A]

Well, I agree with Dave. The first one's probably marketing. So powerful he. Oh no, yeah, don't. Yeah, like that's all. Any. Anytime the really, really powerful thing is leaked. It's not leaked.

[D]

It's.

[A]

That's, that's fine.

[D]

But I mean, that's kind of the same thing with the Claude code thing.

[A]

Well, no, the Claude code thing I think is like a legit fuck up. Right? Like you don't really. Your source code.

[D]

Is it the source code? I thought it was just a bunch of typescript related to the front end or something.

[C]

Yeah, it's not like core like model weights and all that kind of stuff.

[B]

Wait, can. Yeah, can someone explain this? Because I've been in Japan. I don't know what happened.

[A]

Isn't it the app? I mean, I didn't look at this in detail obviously, but like I thought they. What leaked was like the source code for the, like the code endpoint. It's not obviously the core engines code. It's like the desktop installable leaked.

[D]

Yeah, it's just all the typescript. Right.

[A]

But it's still. There is some interesting stuff in there in terms of how it works, which I thought like you generally wouldn't want other people to necessarily know. Like I saw some this morning. I woke up to like every single Twitter post with some hot take on what you learned by reading the core code, which is all probably generated by Claude eating itself. But it seemed like mildly interesting. I don't think you leaked that on purpose. I mean my, my big take on the leaks is simple, which is like there is a bunch of exploits going on where people are like, effectively there was this like. And there's another NPM upstream package thing where someone got some keys and injected some shit.

[D]

Yeah, that was actually the bigger. That was the bigger thing on the Internet yesterday, which was Like a serious security leak.

[A]

It's always been wild to me that like, this is the thing is like, it's always been wild to me that you have all these maintain containers of these, like, projects that get included and included and included, you know, into these projects that end up this small piece of code that some random dude on the Internet maintains is like the linchpin of the entire Internet, right? And then like, something goes wrong with that, right? Or someone gets some key and all of a sudden that's compromised and everything is compromised. And like, this has happened forever. I do think that what's happening with AI this is just going to get so much worse because all of a sudden everyone dangerously skips the permissions because why? And like, you're not even evaluating.

[C]

This is my experience. I was going to get into Jessica's bot corner today, but you're jumping right to it.

[A]

Well, just like this is like, this is the way that all ends.

[D]

I think this is a super real thing. You know, we've got, obviously Peter Steinberger is over here with us, we've got some of the top maintainers from Open Claw, but we also Guillermo from Vercel's over here. And you know, this is like the premier question in security of the Internet right now is that.

[A]

But this is not the thing about it is not a new question. It's just the speed.

[C]

Can we define the question for those of us who don't know what these letters mean? Let's start there. What's the issue?

[A]

The issue is, is that when you're using a piece of software, it almost certainly references inside it shitloads of other software, tons of people developed on the Internet. And when you start using your thing, you're implicitly including all of the other software written by all these people. You don't know who they are. They're just like random keys on the Internet that end up. And someone trusted it and that person trusted someone, and that person trusted someone, and that person trusted six other people. So if you exploit any of those, you start including stuff in your software that you have no freaking idea. It's like six levels of trust removed that if you change two lines, you can steal everything, right? And like, this is how it's always worked and it's always been intellectually. Like, huh, that's kind of scary. But it's kind of implicitly worked. And like, the reality is, is the incentives were kind of fine. And by the way, it did break a bunch, but it was like kind of okay, the velocity was low enough and if you were like, including a new package, there was like a process around you, like, do I trust this? Do I not trust this? Like, what are the implications? Someone was on the hook. Now you have AI agents running around looking for capabilities. I need something that does this, I need something that does that. And just pulling shit out of GitHub off the Internet, right? And using it. And because you're going so much faster, it's like just driving a car faster, you're moving faster, you're checking less, you're able to do way more. But the cost is that there's all this implicit risk and, like, subversion that's going to, like, happen as a result, right? It's kind of like going around and just grabbing random pieces of writing off the Internet and praying. And I think the reality is, is, like, people, this is not a new problem. Everyone understands this at some level. And, like, what you're actually seeing it play out. And what it comes down to is everyone's going to say, no, no, we need, like, more secure xyz. But then other people and more people are like, I don't care. I just want to go as fast as possible, right? And, like, I think the as fast as possible crowd will generally win because they'll be able to get more cool stuff done faster. But it will create a huge security. It's a huge problem. The. The flip side is there will be like, the secure core where, like, someone really checked all the packages, right? And like, someone knows what's in them. And like, you actually the problem. And, like, and you can say, well, I will help us check more. And that's. That's sort of true, but it's so much more expensive and slower that there's going to be this tension. And I agree with Dave. I think all the money is not going to be made on who can run the most tokens fastest or build the coolest shit, because everyone will go infinitely fast on that. You'll be able to charge an enormous premium for security and privacy. But the problem is the cost is going to be. Your stuff's going to suck compared to the stuff that moves fast. It's going to be slower, it's going to be 10 times more expensive, and most people won't care, right? Per usual. And so I just think that's what's going to happen. And like, it's, it's. I mean, it's. I mean, I do think that we're kind of witnessing, like, the final Internet disaster of AI. The first Internet disaster of AI is like, Bye bye web. And like, bye bye ad models and like, bye bye the entire business model for the whole thing. The second cataclysm of AI is going to be you're just going to like, have all these agents running around doing crazy shit and including things. They have no idea what they're doing and they'll check the APIs, like, yeah, that thing works and it sort of will, but there's going to be like a, like something hidden in it it and like everyone's going to do it and no one will trust anything and the whole thing collapses.

[C]

So I want to talk about Dave. I just want to share personal experience for a sec. And then you, as someone who's probably studying this deeply with open AI, with openclaw, would love to know.

[B]

Right?

[C]

So Jessica's journey through Claude code yesterday is really a case in point of how these vulnerabilities can easily creep in. Sam set me up with a new version of Claude code, working really well, has access to a lot of things. I decide I'm going to be a ninja over here and connect all my Google services via an MCP server. Great. Claude's going to walk me through every step of this. I am largely following Claude's instructions, but every step or so, there's a new thing that I have to accept or skip. And in the beginning I'm like, I'm going to be diligent about this and call the wonderful head of engineering at the information and get permission for him to accept such steps, because that's what a reasonable person does. But then I'm like, that's taken too long and there's really way too many steps and I just want to get my email set up. Then I'm like, before I know it, I'm just installing dev toolkits. Like, I'm a whatever and I feel like a rock star over there.

[B]

Did you press always allow?

[D]

Oh, no, I recognize.

[C]

No. So first it was talking to a trained, experienced professional who runs an engineering team. And I'm like, you know, Mike's busy. He has better things to do to help me. So then I'm like, you know who could answer this for me? Claude. So I boot up Claude in my cowork and I just send each message to Claude.

[D]

I went, claude, famous last words.

[C]

They're honestly logged in in different accounts. Not that it matters, right? And Claude is like. At first, Claude's like, absolutely. Don't do that. Because that's what Claude does. Unless this is a server you set up. I'm like, yeah, yeah, No, I set it up and they're like, great, go. So basically I feel so proud because now Claude Code can tell me when my next meeting is, when I could really look at my calendar. So that's where we are in the cycle. But I had the same reaction, like in January when I tried Cowork and I wrote an article about this, so there's a record. I was like, oh, I'm not comfortable with giving all these permissions. And I instantly installed it, you know, March 31st, I'm just pressing one, you know, approve, approve until the cows come home. And what. And this is really, I'm curious what you think about this. Smarter people in this. I've just decided I can't have anything important in any of these services because,

[A]

well, that's why you're, that's why you have a completely separate instance set up in the cloud running this stuff and you're not running it on your computer. Right. Like we've already sandboxed it out. But you can't put all the stuff that's on your computer into the sandbox. Or then there's just the sand you pissed in the sandbox.

[B]

But then what use is it? She's just limiting her use in that case. Right.

[C]

So here's where I really ended up. I went and I was like, it's going to be a security nightmare. People like me are going to wreak havoc on all these things because they're lazy and, you know, impatient. Then I was like, well, I'm also not going to put anything in email, I'm not going to put anything like I'm going to assume that things could be leaked and then more relevant to like these broader debates happening now. I was like, I'm not yet sold on the interface being everything aggregated under these new AI chatbots versus just building these agent experiences in the services I already use. Because where I ended up after all of this is like all this drama. And I'm literally asking Claude Code now I know I'm not being very advanced, but I'm asking Claude Code to do some analysis on my calendar, some analysis on my email when I really wanted is Google to do that analysis for me because it would have saved me like 19 steps and concerns about what permissions I'm giving. And so I, I know we've been talking about this on the pod and I apologize for the length of this, but it was really been eye opening, the bundling, unbundling. Right. Is the primary interface going to be one of these new chatbots or Are we going to see this, like, agent functionality rolled out across all of these services? And after my experience yesterday, I'm like, you know, if I could just be in all these other environments and get this level of analysis, I'm not sure it needs to be like, all together in one place. I've done. Dave, tell me why I'm wrong.

[D]

I mean, I think it can be both and right. Like, if we're in the era of, of agents and it's sort of like when the web took off, everybody was like, oh, I can build a website, I can build a web server, I can have a database connected to my web server and I can build an application. And that application can be a CRM or a CMS for running a news website. Right. And people started to figure out, like, I need to have these different architectures for security around. You know, all of my customers can access my web server through my website and they can read my news articles, and in order to read some of them, you have to pay for it. So I need to put up a gate for that. And we went through this whole build out of how web servers need to be architected and clouds need to be architected in order to have, like, safe, secure experiences for users and the enterprises that generate them. Right. Like, I think we're in the same era with agents right now, which is that people are experimenting with agents. You know, like, if you remember in the early days of the web, people were putting web servers on the computer underneath their desk and you would like, come to their website and that website was being run off the computer sitting under their desk.

[A]

I mean, some of us still do that.

[D]

Yeah. I mean, some of us do, but we're in that era with agents where people are putting them on their desk, they're putting them in the cloud. There's all these questions about what's the best architecture and then what's the best interface to interact with them. So, you know, Jess, I have, you know, I have. I do not allow my email, for example, to be read by my core agent. Like, my core agent is super locked down, you know, has really tight security around it. But then I have other agents that do specific things that I want to get done. I've actually found that the GOG GOG Google Workspace plugin that Pete built for OpenClaw is extremely powerful. It's way more powerful, it's way faster than the official Google one. And so I have an agent entirely set up just for my email. And that agent reads my email project, manages it like A great project manager or chief of staff would. And it actually talks to my main agent using English in a group chat about what I should know. You know, that's going on in my email. And my assistant can actually also talk to that, my email agent and get all of the use cases she needs to get done. Right. And so that architecture is like a security architecture that enables me to feel confident that I've got things sandboxed in the places that I want, and I can still interact through my main agent for all of the interface that I want. And I've got the sort of the workflows and the things that I want going on all set up nicely in my core agent. And then I know that it's secure over here. Right. Whether or not that architecture works for you, like, doesn't matter.

[A]

So I don't know. I have a different take, which is like, first of all, I sandbox this stuff aggressively. So, like, I actually don't run any of this stuff on my actual computer. Right. Like, I have, like, instances running in the cloud, and, like, I'm very clear about what I do and don't give it. But here's the interesting thing. I actually got super mad at Claude this last week because I was doing something and it was getting super mad at me because I'm just, like, pasting RAW API keys into it. And for the first time, I was like, don't do that. I was like, I don't care. Just take the keys and use them. And I'm like, that's not. Ooh. I don't want this locked down. Like, I know why you want to lock this down. You're not wrong. But I'm intentionally using stuff that I don't mind. I'm like, I'm smart enough to be like, it's fine in this case, right? And so it's just this interesting thing where there's this race condition. I want the rawest tools available and then take responsibility for sandboxing them myself and, like, knowing what I'm doing.

[B]

But you're, like, not a normal person.

[A]

No, no. But I think this is interesting to architecting these things. I think there's a one version of the world. It'll be interesting how it plays out. And I personally want the most completely unrestricted. Don't you dare try to lock me down. Version of the world that I will then completely sandbox and take responsibility for separate from my actual core identity. Now, on the flip side, Dave, to your point, I'll tell you, my thing is I would never Give it access to my email, but I do need it to be able to draft and send emails for me. So I just have a totally parallel universe set up where it can draft and send things that are from me and then, like, CC my actual email. But, like, I treat it as like a separate universe. I'll tell you a funny thing, though, it actually screwed up. Yesterday I asked by CRM, I was like, hey, we're doing this creator bot school. Go find the 20 to 30 most important people that creators would want to meet with in New York. I'm sorry, in San Francisco, and invite them if they want to stop by. And like, you know, it pulls up our friend Adam A. And whatever, and it's like. And I'm like, great, just send them each a draft. Like, create a draft to go out from. Not even my real email, from my secondary email that you're allowed to touch, effectively inviting them to this thing and CC my real email. Right? And hilariously, it screwed up and instead of sending drafts, just sent them all. And I was like, ooh. I replied to each of them. So, like you said, like, probably 30, 40 of the most important people in Silicon Valley notes inviting this. And they weren't bad, they weren't terrible.

[D]

It probably just was like, what would Sam do? Sam would fucking send it.

[A]

Just send it. So then hilariously, I then, I then replied because I was cc'd on them. I replied to all them, like, I'm really sorry my bot just fired this prematurely, but it is a legit invite.

[D]

So, like, look, I think one thing I want to double click on real quick is that I think Sam's right in that this kind of goes back to what started this whole conversation, which is, you know, what's happening here from a security perspective, that's profound. Like, Jess, you asked this question, like, what is the question? We're now at a place where the models can just speak any language in the entire software engineering stack. And they can move up and down the stack. Models can speak assembly, they can speak typescript, they can speak HTML, they can speak everything in between. And so Sam's kind of right here, which is that the raw compute, the raw power of a single computer box is now the thing that you want. You want it to just talk to the metal directly and do exactly what you want it to do in a highly sandboxed environment. And then you want to, like, chain those together in ways that are interesting. And so that's, I think the thing that's happening right now is that a lot of the way that we built software which was like, we're going to build all these little shims in between and there's going to be all this open source software and I'm going to use yours and you're going to use mine, and it's going to make my life easier. Like, a lot of that stuff goes away when these models can just speak all the languages and you can build anything against the compute. And I think that is breaking the entire cybersecurity model of not just the Internet, but like the operating system, the whole thing. And so we're like, that's the reckoning moment.

[A]

So here's the other interesting thing about running the raw metal and like the moment we're at with everything can speak to everything so easily is there are no more limits you can set on anything. So I'll give you a really an example. I'll be a little risque for the sake of being risque. You know, there's been all of these, all. I'm getting eye roll already. Jess. We gotta, we gotta, we gotta get some clicks on the podcast. So here's the thing.

[B]

We never think Sam would say something risque.

[A]

Here's the risque thing. There's been this hilarious game which we all know about, like OpenAI and Claude and Grok and like what they do allow and they don't allow and all this type of stuff with the image generation and you know, just again, in Sam experimental mode, you know, I was making images with like one of their APIs and it was like, content moderated. And I was like, well, that's stupid. So then I was like, hey, Claude, what's the best way to like have a completely unrestricted image generation pipeline? And it was like, here just like boot a GPU in this cloud, an H100, you pay 2 bucks an hour and like download the software and like run it and you can just have completely unrestricted access to image generation. I was like, normally that's really annoying and hard to set up, right? Even six months ago, like, I don't want to rent a raw GPU and figure out the six packages to install to like do this just because, like, I'm annoyed that I got content limited on some stupid API. Now I'm like, just set it all up for me, right? And it works, right? And so like this idea that like any one service can lock down, like what it will and won't do is ridiculous because as long as anything else on the Internet will do it, like, it'll just happen. You just ask the Thing to like, you do a much more complicated setup which costs you nothing now and is not like a time stock to do whatever you want. Right. And so I think about that's true with email sending, that's true with anything where like any one of these agents tries to be like, well, if you're using Grok, you can't do blank. You're like, yeah, I can, I can do whatever I want because I have like access to the raw compute in a machine that will spend hundreds of hours setting up anything down to the metal I want. It's an important paradigm shift because it's like before, like the idea that a service can be like historically be like, well, the thing you're using is blank and technically there's a way to do something else. It's really hard and technical to figure it out.

[D]

Yeah, none of this matters anymore.

[A]

Now it doesn't matter. And so basically, as long as someone in the world provides some digital service, it will be found and you can instantly use it, which means that there can be no rules.

[D]

Yeah. And by the way, this also invalidates the entire marketing model of Claude and OpenAI to some extent has been this. Everyone should fear these models. And like, you know, at the end of the day, like we're like what, one person sharing all the weights of a model on the Internet away from like anyone being able to do any of this anywhere. Which is just like, that is what's happening. Right. And so we have to have a different kind of conversation about what's going on right now.

[A]

I, I even Dave, here's another example. Practically as a builder, like I basically just re implemented my own granola because I got annoyed with granola.

[D]

That's funny. I was doing the same thing.

[C]

Guys, what's wrong with all these services? Like just use the services.

[B]

You just want like customizations.

[D]

It sucks.

[A]

I have a great one. I'll release it. It's called record or request clip. It's up. It's sick. It's actually running right now and doing the entire multi party translation, like transcription, summarization. It's actually quite good. But here's the funny thing about this. I was like, okay, I, I need to like stop using granola. Like this is like annoying me with its features. So I like build my own version. So start building it. And then I'm like, cool, like I need the transcription and summarization and like use it. It's like, great, we'll use ChatGPT's whisper whatever to do it. I don't even think about it, it's done. But then it starts limiting me and I'm like, that's not right. So I'm like, fine, like, what's a better service to use, right? That's going to actually do this properly, right? And it hilariously, it pointed me at assembly AI, which, by the way, we received investors and I put 100k into their first round, but they're crushing it and it's perfect. And so I'm like, great, this is cheaper. This does multiparty, full transcription of everything. It's like, way better. So again, this goes back to, like, there's no way for anyone to make real money on any of these things. It's like, if I don't like one service or it cuts me off, it's like, there's no implementation cost. I'm just like, do the better one.

[B]

But, like, done. But Sam, just to push you on that, like, again, I think the theme in this episode about you and I love all this experimentation is like, the theme is chaos.

[C]

I'd like to bring us back to,

[B]

you know, I think it's like, Sam goes unhinged and, like, which is fine for Sam because he knows technology and how to manage and how to build new things, but, like, for the 98% of people who are just going to stick with granola, because even though they can, like, maybe vibe code a thing, they just don't fucking care. It's good enough. They've got a good brand. They trust it. You know, Like, I actually disagree with you because I think that there are going to be a lot of services and pieces of software that maintain their stickiness because normies just want what they want and they don't want to change and they just, like, want it to be good enough.

[D]

I don't know. The through line is like, I was walking through Kyoto last night with Guillermo and we were talking about the. We were here together, like, year ago and a year ago, like, he gave this example. You go back in his Twitter and look at it like, hey, guys, I vibe coded this. You know, there's this annoying thing that I try to do, which is like, change JSON into, like, convert it into a different format. And all of the tools on the Internet are really difficult for this and they're filled with ads, and so I just made my own. And he got super negative feedback on the Internet. He's like, this is the era of personal software. Everyone's going to make their own tools. And a year ago, people were like, that's not what's happening. That's like not what's going to happen. It was like this super negative takes from everybody on the Internet. Where are we today? Like, the use case Sam just said is like pretty insane. Like he wanted his own version of granola. He just built it. It took him a day. And Sam's a super high agency builder. Right. The through line on this is that people are going to be able to just request the software that they want in like one to two years from now and it will just materialize onto the Internet. And I do think that's like actually an extremely important trend that is it's now just happening. And I think that's like a big, big change in the mental model of where we're going because, like, it's going to happen. And I think normies are actually going to request things.

[A]

That's why I'm so bullish on software and so negative on software businesses. Because I'm just like, I don't like, just. These are just tokens.

[C]

Okay, friends, you can keep on this topic, but I am going to bounce and I suggest you go to OpenAI's Munster fundraising complete with $3 billion in raises from retail investors. What does that say about the prospects for an IPO?

[B]

3 billion? Out of like how much? How many billions? How many billions went in?

[C]

Still $3 billion, but out of like a hundred.

[A]

That's right, because 90, the 1% controls 99% of the wealth. Right. So they put in 99% of the dollars and retail puts in 20 bucks.

[C]

When is the last time retail put money? I think you're missing the forest for the trees. When is the last time retail put money into a private technology round?

[B]

How did the retailers put. How did they put money in?

[D]

Well, now they are through these new synthetic instruments. Right. Like, isn't that one of the big. One of the big stories of the last two weeks is you've got these like crazy new synthetic venture funds that are being traded on Robinhood and I can't remember the name of the other one that contains the stock.

[B]

Right. And they're investing in like all the top 10 tech companies. Yeah, SpaceX and OpenAI and just buying Allocation. Right.

[D]

I don't know. Sam, do you know what these things are called?

[A]

I don't really care.

[B]

Do you care about the valuation of OpenAI being 850 billion now?

[A]

I just don't. I mean, like, they're just. I just. This is like a ridiculous. This is just ridiculous. I mean, it's like whatever. It doesn't. Like, if you want to buy 20 cents of infrastructure per dollar in a thing with, like, increasing competition, where it turns out the story of whoever gets the AGI first doesn't win. God bless. I mean, I think the story is people don't know where to put any money, and money's cheap, right? But I don't know, like, it's fine. I have no opinion. It's like, I don't understand. I mean, the only opinion I would have is, like, it's so clear to me.

[D]

This is the. All right, everybody's annoyed with the topics we're talking about episode, so let's move on.

[A]

Well, just like, I just. It's like, so awes obvious that the, like, basically, hilariously, matrix multiplication is matrix multiplication. And Dave, we've talked about this. It's like whoever has cheap energy eventually wins. Is a commodity business with no defensibility. It's not even. Not locally regulated. It won't be until someone makes some security argument. And there's no. There's no capacitor issue. Like, it's just infinitely available globally, and you can vibe code whatever the hell you want. And so God bless. I mean, I don't understand how any of these companies make sense.

[D]

Did you see that CO2 chart, which was like their underwriting of Anthropic?

[A]

Yeah, I think I did. I like their charts. That's a Karen. I think that's a Karen Maroney thing, which is awesome. The C backstop.

[D]

Oh, interesting.

[A]

Very windy.

[D]

Good old Karen.

[B]

What did the charts say, Dave?

[A]

That, like, once you have agentic, the basically number go up once the agent spends its own money. I mean, basically. Here's the funny thing. It's like the big bottleneck on AI spend and AI stuff in general was that you had to have human input. And they're like the interfaces. And so what Claude code does and what OpenClaw does and what open route, like, all these things are doing, and now Codex is just like, hey, let's get rid of the bottleneck, which is the human. So we just can spin circles as much as we want. Then Sam can all of a sudden not hit accept, accept, accept, or deal with whatever limitations exist in chat. ChatGPT's interface. I can just be like, build me granola and change the limits and make it more shareable. And it does it, and it's great and it works really well, which means there's no.

[B]

Did you actually prompt it to build a replica of granola?

[A]

No, I had a slightly different take on what I wanted, which is actually totally sweet. You guys are going to love it.

[D]

Here's the chart.

[A]

Why are you on Kucoin? We're talking about crypto.

[D]

I just was looking for this anthropic COTU chart.

[A]

And you found it on a crypto website.

[B]

Explain this chart. And especially for people that can't see it visually right now.

[A]

Translate it. You can.

[B]

You could request clip. Can you get on that?

[A]

Probably actually will translate it. Yeah.

[D]

Today, the valuation, or at the point they were doing this investment, you know, anthropic's valuation was 360 billion and that they're on track to a $2.4 billion value or $2.4 trillion valuation by 2031, and that their EBITDA will go from negative 14 billion today per year to 48 billion per year in 2030 and 78 billion per year in 2031. Like. Sam, do you. Do you buy this?

[A]

No. And even if I did buy this, the funny part about that is, while 48 billion in EBITDA is like a very large number, that does not justify the valuation that we're talking about.

[B]

Well, 18 billion. What did they say in that chart? It was 18 billion in 2026 revenue with 14 billion loss. And their valuation's at 360 billion right now of 18 billion.

[A]

It's like Facebook does 40 billion a quarter in EBITDA. Right. And now.

[D]

Right.

[A]

Like, so I'm just, like, I just. It's. This is all just, like, narrative driven. And look, I get it. People need to store their money somewhere and they need to store it in hope.

[D]

Right?

[A]

Is the upshot. Especially when they don't know what else to put it in. Right. You're going to either put it in gold or hope. And, like, this is the hope thing at scale. And people need things for hope at scale. So, like, that's fine, but it's like, from a fundamental business perspective, none of this makes any financial sense. And it's not going to. Right. It's just a question of, like, how that plays out. It could be that because people need to put money in hope, it doesn't actually matter. Like, hope go up and everyone does fine. Right. But it's a hell of a narrative to believe. It is quite religious.

[B]

Dave, do you have a counter to that or do you. Are you more aligned with Sam these days?

[D]

I don't. I think the wild card is local inference. Like, I, you know, my. My bet's been made here. Like, I've taken an extremely long bet on local inference by, you know, we're going. We're going maximum Deep in the. In open claw and into, you know, how is that.

[A]

But just to push you. Is that local? It's. That's like just inference in the inference anywhere. It doesn't matter.

[D]

I know, but I think it's both like we're going maximum deep and taking a maximum long position in local inference and openclaw because we just believe.

[A]

Are you buying Apple stock then? They're the local inference company.

[D]

I already own a lot of Apple stock, you know, like, so that's your bet.

[A]

I mean, you know.

[D]

Well, I think that's our bet. But we're also, Sam, we're going to seed, we're going to seed a ton of companies in this direction. And like we think that this is like, I don't think it's a question of or like a lot of people frame this as it's hyperscalers and cloud AI or local inference. I think it's. And I think it's going to be both.

[A]

I just think it's whatever fit. Yeah, definitely. I mean like I. This is again, I'll go back to my image generation blockage. I have like an image generation issue in a cloud service. So I'm like, all right, just like put it locally on my Mac mini with 24 gigs of RAM and it's fine. It's actually better. And I'm like, yeah, I want even better. So like I'm go back to an H100, right? And like the reality is the fact that I can just sit there and in three commands while I'm doing other stuff do that is mind blowing. But it means there's no lock in on anything. I'm just going to sit there and price and quality optimize, right? Like there's zero. I can switch local. I can go, I can go cloud, I can go to China, I can go to whatever the hell I want, right? And like, again, it's a super powerful thing as a builder. I just don't know how anyone makes any money on it.

[B]

We're going to bring this to a close soon, but I'm curious with where things are at right now today. And Sam, you're obviously breaking a lot of the mold here. We all are in different ways. But what would you advise our audience is the key takeaway of all of this that we've been discussing? Today's episode, we've gone from security to sort of like cloud local, like future of all of this. Everything goes to zero. You know, Sam's everything goes to zero line. Like, if, if our listeners are leaving this episode. What is it that they should do this week or think about more?

[A]

So what they should do this week other than if they haven't. If they've been invited, come to our creator Bot Summit, which is gonna be so fun.

[D]

I'm bummed to miss it, man.

[A]

It's gonna be so good. I'm like so excited for it this week. The number one thing you must be doing is aggressively playing with everything.

[D]

Yeah, could not agree with this more.

[A]

This is a moment where, like, you should not be reading, listening to podcasts.

[B]

Stop listening to this podcast.

[A]

I'm only half listening to this podcast. I'm.

[D]

Yeah, listen to the podcast and learn and then go play.

[A]

You, you should be playing with everything and having your own opinions and. Because the reality is, I mean, this was my best experience. The last week was that we were in Alaska. David left. So it was me and a bunch of very non technical, wonderful people and excellent skiers. And I back to back sat down with these great people and said, actually, in this case, I did use granola. I said, let's talk. They had some amb idea. So great, let's talk about it for 3:30 minutes. We sat down, we just discussed it for 30 minutes. I literally all I did was take the granola summary and the granola transcript and I stuck it in my Claude instance and I said, make this. And it produced, you cannot believe how good stuff. Like, I'm talking about fully deployed, amazing things, one of which is actually almost certainly become like, actually a very real thing that an amazing person wanted to build and just like, didn't have the technical language to do it.

[D]

But like, I guess I was there for the first part of this. Like, this was a really cool story.

[A]

Well, the second one got even better, but it booted the whole thing and it works and it's actually leveraged for them. And like, that is wild, right? Like, this is the time. It's funny. I actually think that Silicon Valley and at least, you know, let's call it your average PM/engineer in Silicon Valley, hilariously did just completely disrupt themselves. You no longer have a stranglehold on anything in building, and the building is going to go back to people who actually know what they're talking about and have community and have insight into the specific markets. I think this is like the best time to be a generalist. Buy Claude Max at least for a month. It's 200 bucks, it's totally worth it. And just go to town figuring out what you can do and if you can come to a bot school come to someone who can teach you how to do like the little shim that's missing. Because you do need a database, you do need to be able to spin up app servers. There's like a little bit missing, but it's like Dave knows this, I know this. The pieces that are missing are achingly small and simple at this point. And the second your bot can send email, send sms, remember things in a database, use an asset store and have an app server, you can do unbelievable stuff, right? And that is the moment.

[B]

Dave, what about you? What's your take for any of the listeners?

[D]

I think that this stuff is just incredibly empowering and that the use case that Sam was talking about is a dear friend of ours. And the web application that was built was something where this person was trying to get budget for this, build this thing for a couple of years. And the case that Sam explained became extremely empowering. And then once this person understood how to use these tools themselves. Sam and I got a text message yesterday from the person and they said, I just added my own feature to this and I feel so powerful, right? And like that's such an amazing thing. And like to me, like people were asking me on stage in Tokyo, like, why are you, you know, what, what, what's this, what's most inspiring to you about all of this? And that's like what I keep coming back to is the individual empowerment and the, the power that people feel to create their own. Whether it's a website or a little app for their computer or you know, even just seeing our 11 year old create his own games and put them onto his phone and put them onto his friend's phone. Like this is just fun and awesome and it's not as hard as people think for a variety of different things. And I think that's like maybe why we went so long on this topic on this POD is it's just very inspiring to watch people who are normies being able to create their own software. And I've said it many, many times on the the pod, but there's only like 10 million software developers or so that make the software for over, you know, 3 to 4 billion people that use mobile phones and software in the world. And the amount people are using software is not going down right? Like in fact the consumption of software is going up faster than it ever has in history. And so I think people are going to continue to want software like they want music. And the fact that more people can make this than ever before I think is a really good thing. Does that mean that there's going to still be as many trillion dollar companies? Maybe not. But does that mean that you can build deca million 100 million billion dollar companies and a lot more of them in the software world? Probably. And so I just think it's very, very empowering and inspiring to watch people do this.

[B]

I think that's where Sam and Dave might do disagree.

[C]

Yeah.

[A]

This is my closing question, which I think you're going to know exactly what I'm going to ask, right? Which is Dave, our big. We agree on so much. The thing we disagree on is I don't think this is like software development anymore. This is like. And to me like the answer is the number of software develop software developers goes down. Just the number of people who like, build software the way I write essays goes up. Right. Do you still feel we'll timestamp it on March 31, that the number of software developers is going to go up or just the number of people who like, build the way there's software the way they write an essay?

[D]

I just think it's a, it's like a semantic difference, right? Like, Sam, you have a fund for creators, right? Like, is a creator somebody who posts to their story on Instagram or is a creator somebody who transitions from posting videos and telling stories on their personal Instagram to becoming much more professional about it and telling story at much larger scale? And like, I don't know what you call someone who begins now to speak the language of computers and build their own software to, you know, give it to people that are like more than just their friends and start maybe perhaps building a business on it, like, that's like a semantic, a semantic battle, I guess that we can have. But to me, like, the idea that more than 10 people, 10 million people now can like make software for the, the billions of people that use it every day, like, that's a really positive thing. And whatever we call them, we call them. But like, I think it's fair enough. Very, very great.

[B]

Yeah, I think, I think we're gonna, I think we got somewhere today, gentlemen, sort of by the end of this.

[A]

Enjoy Japan.

[B]

We are gonna enjoy more of Japan. We've got some interesting stuff going on here in the next couple days. We report back further next week on this podcast. And you know, Jess, we missed, we missed the rest of Jess, but I'm sure she's going to have some big updates for us soon too. So. Thanks to everyone who listened in this episode. Thanks to everyone for liking subscribing, commenting. I had some shout outs, by the way, at Tokyo clock on for our pod from the Japanese, which was amazing. So we're global, you guys. We're going somewhere in the world.

[A]

Wild. They must love me in Japan.

[C]

Who came up?

[B]

We'll get Sam here one day.

[A]

Have fun, guys. Talk to you soon.

[B]

Thanks, everyone. We'll see you back next week for another episode of More or Less. Bye Bye. If you enjoyed this show, please leave us a virtual high five by rating it and reviewing it on Apple Podcast, Spotify, YouTube, or wherever you get your podcast. Find more information about each episode in the show notes and follow us on social media by searching for oreorless Avemorin at Lesson at J Lesson. And as for me, I'm at Brit. See you guys next time.