The AI Revolution Call: Mythos Mania! Is Anthropic just doing more hype? Or is this the real deal?

Oh, shit.

Another round of applause for Craig's intro videos.

You were on Mars.

You had to clear it out for Elon.

Dude, it just keeps getting better. The creativity is— you know what though? We talk a lot on this call about how we don't think— welcome everybody to the AI Revolution call. We talk a lot about how we don't think AI is really going to kill jobs. And this is a perfect example. Oh yeah, we got to retweet and let everybody on Twitter know that we're online. Um, the, the creativity comes from Craig, the content comes from AI, but we can't— AI is not going to be that funny at that point, like that relevant, germane to my life, and like, oh, Mars, like, I fighting aliens, or basketball and rock and the stuff I like. And that's, that's, that's amazing, Craig, very creative, well done.

But good job.

We have a special guest today.

Yeah.

All right.

I'm—

this is the AI Revolution Call. I'm Cody Willard. That's John de Oliveira. Craig and Bryce, you guys have met many times. They work with me at 10,000 Days. We invest in technology like AI. And our special guest today, the one and only Charles Harris. We call him Charlie. He is in town from London/Abu Dhabi. And, you know, he's an analyst and works with us on all kinds of things. But technically, he's under a different umbrella.

Well, thank you very much for having me.

Yeah, it's great to have you here. What do you know about AI? Not a lot, much. That's not true. We've got him immersed in AI too. We don't— you don't get to work in our job and not use AI. John, there's big news. Bryce, you set this up, will you?

Sure, John.

I'm sure you saw Mythos Claude. So that was the big news last week. Anthropic put out this press release saying that they had developed a new model called Mythos Claude. It's a very large model. It scored very well on the benchmarks, specifically in software engineering benchmarks. And they came out and said, look, This model is too dangerous to be released to the public. Therefore, we are working with a selected group of major technology companies in the United States, mostly, you know, namely Apple, Google, Microsoft, CrowdStrike, IBM, the list goes on, maybe 20 names or something that they dropped. And so they said, look, we did some testing with this model when it was in preview. And we discovered vulnerabilities in some 30, 40, 50-year-old software, open source software, stuff that people thought was, was completely hardened and impenetrable, namely Linux, the Linux operating system, which powers most servers of the world. And the AI reportedly was able to string together a series of vulnerabilities in the Linux kernel. And hack, you know, get root access to the machine. It could do anything an admin could do on this machine. So software had already been in a sell-off, and the sell-off just totally accelerated last week. Every software company, including cybersecurity companies, were hit extremely hard, you know, down 20, 25, 30% on the week or something like that in some cases. So we've been discussing mythos, and its implications over the last few days. So we have some thoughts, but John, love to hear your thoughts.

And now, what do you think when you saw the news? And we're—

got a little of that déjà vu because Anthropic has sent out that scare message numerous times.

I think he's on our page. Dang it, dude, we did not coordinate this. We wanted to have a debate.

I'm sorry, I was supposed to say something else.

No, it's great though. We did— we literally were curious what your thoughts on it.

Yeah, I mean, I think they did this ever since like GPT-2.

Yes, Bryce has been calling it fear marketing. Like, you're, you're just— what you do is you get the whole world scared that they need your product or the— you will be hacked, right? And it doesn't matter whether you're the Palantir or Coca-Cola, your business is at risk because it could get hacked.

And the other just thing is that it's kind of funny who they, you know, partnered with on their whatever river project to go and give it to them first and correct flaws, because they really probably should have given it to the power companies, you know, the infrastructure that could really bring down the whole—

to a social media company first.

Yeah.

All right. Instagram, go down.

Important stuff. I need people to know what I'm wearing on Tuesday.

But it's just kind of further, kind of, you know, tells the story if that's who you're going to talk to. And after, so like I saw somebody else came out and said that, and after it was said, that's when they came back and said, oh, by the way, we're gonna go talk to, you know, the telecoms or somebody, you know, they kind of tagged on an extra, you know, something on there.

So, John, open source community, how are you supposed to fix Linux? Who fixes Linux?

They just open source community? No, you go to Red Hat. At IBM, I guess.

Yeah.

Yeah, John, but that's just it. The, the, the power of AI is increasing exponentially, and it is incredibly exciting to see all of the abilities continuing to just sort of compound. Is it— are we going to get to the point where hackers are going to be able to take a—

what'd you call it—

a distilled version, a distilled model of Mythos and figure out how to— hey, like, we were making a picture, we were just imagining, like, wouldn't it be a great movie to have, like, this, like, some kids in, like, Tennessee or Illinois or somewhere, like, in, like, a garage, randomly figure out how to actually use Mythos to hack into, like, Microsoft And then they're like, hey, Mythos, hack into SpaceX now. And then all of a sudden, like, SpaceX's Starlink system shuts down. Like, is this all possible? Is it at all possible that people are like— nefarious actors are actually going to get AI?

Well, if they were going to, they would have already, because while this is kind of on a larger scale in this particular story, it's not the first time that it's been said that, you know, this stuff can be used in that way. And it, you know, kind of would have. You know what's funny? I, I, um, I don't know, you think it's funny? I think it's funny. I was, uh, working on a project, uh, when I was at Psycorps, the AI company, in— this is in the late '90s, okay? Uh, we had a product on top of our AI product called PsycSecure, and what it did was it took all the vulnerabilities that were, uh, known and it built exploits out of it. It was a planner that put together exploits to show what could be done based on the known vulnerabilities that were out there. And it's like everything old is new again. That was like 25 years ago, and it's the same. They're talking about the same stuff, right? It's just the planner is less, um, procedural and symbolic and more like the machine is just doing it.

Well, John, I mean, what I was just saying just now before we hopped on was I was thinking it's sort of laughable because You think when what they said in that press release is it's too dangerous to be given to the public, right? So it's like literally what you're, you're telling me you put this out there to the public and like some guy, random guy is like hack Apple and it's like, at Mythos, like, yeah, sure, right on that. But it's like they can put guardrails on it just like they do and have been doing for every one of these models that says do not help them create bioweapons. Do not help them hack into operating systems of other companies, right?

Yes.

They put all these guardrails on there. So it's, it's definitively untrue that if they release this to the— if they release it to the public, and it could hack Linux, then that's negligence on Anthropic's part.

Yeah.

So guardrail it.

At some point, what they're saying is going to actually be true. But they've been so full of crap every time. That we're not going to know which that time is, which is like the definition of the boy who cried wolf.

Yeah.

Yeah. Because it concerns me that the model, like the only hard, you know, I've seen it's beating the benchmarks, but it's 5 times more expensive to run than Opus. And Opus, people run into their limits very quickly because it is very expensive to run. So this model is 5 times more expensive to run. I said, okay, well, that maybe lends some credence to this. But just like you said, how are we supposed to know when the moment is here? Like, when you constantly keep saying that, and I was saying that since I started here, you know, everything they come out with is going to ruin the world. And we've seen it with every press release and all this other stuff.

And it is better, you know, it absolutely— I'm sure that's right.

That's what we keep saying. It's amazing to see the acceleration of the feature sets and the capabilities.

And yeah, isn't that good enough? That the fact that it's so much better. Do we actually need to go and create all— oh, and plus it's supposed to be like, oh wait, we're going to take 100 days and get everything settled down before we go. But you think in 100 days you could— do you remember the year 2000 problem? Do you know how long it takes anybody to change any software? They're just going to what, change their, their development cycle?

You can't even get people to download Copilot if you're Microsoft.

Yes.

Now all of a sudden you're going to get them to like secure their Linux servers in 100 days around the world.

I think it's 100% enterprise sales strategy. And it's working. Apparently working. Like if you're, if you're some mid-tier software company, and you read that press release, and you see that Google signed up to use this thing, you're like, damn, guys, maybe we should switch to Claude. We're using Codex from OpenAI. We got to switch to Claude and get in on this and make sure we're safe.

Yeah, so like, a slight, um, uh, detour is the fact that, um, Anthropic's revenues are, are insane. The growth in their revenues is going crazy.

Yeah.

And it, you know, it'll, it'll keep kind of increasing for OpenAI as well. It's probably not a fair thing to say.

John, let's, let's give context. So the Anthropic's revenues in December were on a $12 billion, $9 billion annualized run rate. And by the beginning of April, it was $30 billion annualized revenue run rate. You know, you're talking like 200% top line growth.

I heard it put like Palantir. It's like you had a whole Palantir and Databricks put together all in one month. That was the guess, right?

Oh, that's a good way of putting it.

Yeah.

I mean, that's serious, serious size. And as you just said that, you know, at this moment, Palantir and Anthropic are relatively the same market cap.

And what's the opposite of bubble?

$350, $400 billion.

So like, what's the opposite of AI bubble? Uh, uh, revenue in the sense that you're like, bubble would be, oh, that's all nonsense, there's no revenue to back it up. Uh, there is revenue to back it up. So now what do you say?

Right? I mean, that's— this goes back to like the analogies that people always want to make of like, hey, it's just like.com and Pets.com. And I'm like, is it though? Because I'm not really sure that Anthropic is Pets.com marketing stuff on the Super Bowl to try to get you to buy dog food online. I mean, this was— this is a little more revolutionary. And the proof is in the pudding. Another point that we've been making forever is just that the hyperscalers would not be spending another 150% more this year than they spent last year, which was 100% more than they spent the year before building AI if they were not getting returns.

Well, if you look at the CapEx, uh, from Anthropic, and you just can look from the standpoint of, well, what do they spend their money on, uh, you know, data center and employees. And employees for them is, you know, 100 times smaller than, you know, I don't know, Amazon was or something. So they've got, what, 4,000 employees? No, less.

I'm not sure on that number.

I don't know. Yeah.

Uh, so that— so you're like, you got employees, that's one thing. Now you got the, uh, kind of like, how many gigawatts are they running right now? And it's, um, less than 1. Yeah. Or maybe 1.5. Less than 1. So it's less than 1, and they just signed up Google for Now line that up against revenue and it's like they're being held back by not enough capital.

That's right. The compute is the constraint for them. Has Anthropic made a mistake by not having gone all in on securing compute the way that OpenAI has? OpenAI has got a trillion dollars of compute on backorder. And Anthropic now has to scramble to try to get, you know, gigawatts.

Did, you know, NVIDIA make a mistake by not building more chips than they did? I mean, we have a constraint.

They make every chip they can. I know, but Anthropic made a strategic error in not spending the last 2 years putting together $1 trillion worth of compute to come online to power their future. And now I do think Anthropic is scrambling to get compute, proof in the pudding, what we just discussed.

I don't know how much it hurts them. I mean, yes, they would have had higher revenues, but they're going to increase their compute. And when they do, they're going to increase their revenues. And people see that they are—

Does the acceleration, does the innovation itself not also get held back by not enough compute? And I guess that's what I'm saying. Do the hyperscalers and OpenAI have an edge because they have access to an order of magnitude more compute than—

Number one, mythos.

So maybe not.

So what they did, I mean, they'd made a decision, you know, we're not going to do Sora, which OpenAI canceled anyway. You know, we're not going to do, uh, voice and, and video and images and all what, you know, we're going to do, uh, programming. You know, our thing is going to be great, great at coding. So they really decided to focus in on a particular thing. And so even though they're constrained, they end up with Mythos, which is like crazy good. And now they just have outstanding prospects.

You have any thoughts? You want to hop in here? Uh, no, not on the Anthropic stuff. Sorry. All right.

And I just like, they, I've asked you the question. I keep mentioning that, yes, I've been the Anthropic person here from the beginning, like saying pay more attention to Anthropic. Um, and, uh, it doesn't mean that they're going to run away with this, but they, they're certainly doing well. And the other thing is that the thing I've been talking about is, uh, kind of the, uh, harness engineering the stuff around the model. And for that, they've been just so smart about and so fast that they're coming out like every day with another, uh, something that builds up their— you know, like what made Intel great back in the day when Intel was great was the software around the chip. The chip was like not the best, uh, you know, even Motorola had better engineering, uh, but Intel had better software surrounding the chip.

That's interesting.

Yeah.

So I think you can speak to a lot of stuff that you can build with Claude using Claude Cowork and all these really cool functions of Claude. And then when you look at like Google, I love Google. I love to use Gemini all the time in the personal account. You can have Google do things for you like move your mouse and go click and do stuff like shop. but every time I've tried to use it compared to like Claude Cohort, it's like giving my grandma my mouse to go do something. Sure. It's absolutely just so slow and it's kind of embarrassing. So when I look at Claude, when I look at Anthropic and Claude and see what they've done in the AI world, I, it probably is a strategic mistake because if they had more compute right now, where would they be? But what they've done with the compute they have is so incredible, just because they have focused on there, on, on that one specific point.

John, I guess that made me think of a question, like, um, why— like, so all of the companies pretty much have this computer use function where you can use your, your graphical user interface, GUI. Is that just sort of a temporary thing that we need to kind of get over this hill to, you know, so the AI can access certain applications? Because it seems like It's so much cleaner and faster just to use the command line terminal.

Well, it's like, it's, it's like almost going backwards. We like, we need AI to do stuff on a computer that humans do, right? No, skip the human graphics and everything, just do the data.

Just do it.

But so what you say, like, you've got the whole App Store worth of things that you want to be able to have the computer go and do. And some of them, uh, took the step of going and making the MCP server interface or the CLI server interface or different ways to make that thing usable. And then the rest of that stuff they didn't. And now I'm going to bring up Tesla because, uh, out in— or XAI, you know, out in the, in the world is all the stuff that humans use all the time, and you want the computer to be able to use it. So that's the physical end of it, but there's a computer end of that that Musk is actually talking about, which is instead of the— how that currently works with computer use, which is like it goes and it takes a screenshot, right, every how often, couple seconds, I don't know, of your screen, and then it goes and looks at where things are, and then it figures out where it needs to be clicked, and then it tells the computer use software, you know, kind of where to move the mouse to and do the click. So there's a step in there of it actually taking a screenshot and going and using that. What he's talking about is the kind of, you know, video that happens with your cameras in your car happening on the computer in real time with everything that goes on. Not like, oh, which screen and do what, but just really just understanding that environment the way that it understands, you know, cars on the road. Um, not your— so I'm not talking about your car taking over your, your screen. I'm talking about the kind of the video rather than, than, than pictures. But I think, uh, so I'm saying I think that there's gonna for a while still be that need to, um, to, uh, get the— pretend to be the eyes of the human to get behind the— yeah, you're seeing, uh, Luke there.

Um, he's the, he's the true rock star of the show today. He's biting his own tail. It's adorable. Yeah, he's sort of— he's Anthropic. He's just chasing his own tail.

And he's also growing at the same rate as Anthropic.

He's growing.

Yeah. Just on that then. So if you think there is a place for that, does it make sense then that that will— that inferencing would better take place on the device? Is that what you're saying Elon is sort of hinting at, is that this idea of screenshotting, sending it back to a cloud model to do analysis? You, you do that sort of inferencing on the device so it can happen in real time and not have the lag and this awkwardness like Craig is talking about.

Yeah, yeah, yeah. So I'm not sure exactly how that's going to play out, but, um, and you know, there's all this kind of— there's layers to that, um, computer use stuff. And, uh, there's also the, the security. Oh no, did Open Claw just, you know, give away my bank account? Or you You know, and then, and OpenClaw, which was originally ClaudeBot, was ClaudeBot because the guy used Claude. And even though he went and sold to his soul to OpenAI, the—

You can't say that. I mean, you can call scam Altman scam Altman, but you can't say the guy who went to work there sold his soul to go work there, John.

That was— No, he did whatever he did, but it doesn't matter because It's like everybody got all excited about OpenClaw, and then what Anthropic did was started releasing all kinds of stuff that kind of made OpenClaw less relevant, let's say.

John, the other thing I think that happened in last week with Mythos, but just in general anyway, was that, yeah, I continue to use Manus and AI Studio and just sort of like play with these things, these agentics that it also can build web apps and just on a prompt to make all kinds of different things. I had Ameris's physical therapist was here Saturday night and her son is in the military, needs 6-month housing in South Carolina near this base, blah, blah, blah. I gave Manus the prompt, what I just said there, And I said, help me out. And, you know, 45 minutes later, 3 prompts later, it had an entire web app that has listings for furnished, unfurnished, and expensive, and list ideas and commentary about what you're doing and the options. And it's incredible. And it just sort of hit me. I think we, we talked a few weeks ago about the idea that Maybe enterprise software as a service is not too threatened because even if you tell an AI to build an app that does CRM, customer resource management, for you or something, you, you still have to maintain it. And if the JavaScript inside the cloud changes and gets updated with new security protocols, all of a sudden your web app is broken and you don't know But it's getting to the point where AI— I just tell my AI, hey, something broke on the website, go fix it. And it probably would be able to do that. It's getting to the point where all of a sudden I'm like, hey, forever I thought Snowflake was probably a good part of the stack to stay invested in because AI agents and AI— or if you're an enterprise and you want to use AI, you're going to need to go through get your data from various clouds and put— access it all in one coherent place. You don't think eventually I'm going to just be, hey, Manus, I've got my data at Azure, at Google, and over there at Microsoft, whatever, 3 different cloud guys, Amazon, whichever I missed. Pull all my data and put it in a Snowflake-like platform for me that I can then access. Like, How far-reaching is the ability for AI to just do what software is currently doing?

It makes it hard to see anything to invest in other than the, you know, the hyperscalers and NVIDIA. I think of them as a hyperscaler in a way, but yeah, the hardware hyperscaler, right?

Yeah. Um, no, it's, it's funny, I wrote a piece a couple weeks ago that sort of hit it on this, right? That it was like, I, in the app revolution and the smartphone revolution and the PC revolution, the cloud revolution, there was always a layer of application that captured a large part of the value generation. And so Spotify, Airbnb, Snowflake, and the cloud, like these are application layers in the stack somewhere. That maybe just no longer are even necessary, that subsumes— the AI subsumes the— these applications because the value creation gets to the place where you just interface with the AI. And if you interface with Manus from Meta or you interface with Gemini from Google, when you tell it to go build you information thing for the kid who's moving, that military needs 6 months of housing. It'll just do all of that, and I don't need—

who's your, uh, who's your biggest Mannis user over there? Uh, so tell me, what is, uh, do you think that what Mannis does is, um, will not be subsumed by anthropic capabilities in the next 3 to 6 months, or—

yeah, it will, right?

I don't know. I— you're using it on a regular basis, so are you— the stuff you see it do, do you go, yeah, I could see this getting just happening with that, or are they going a different way?

Uh, I mean, I think that's the, the issue, and I think we asked about that a couple weeks ago, that just the intelligence layer gets so damn good they can create the orchestration layer on top of it, you know, better than the existing orchestration platforms. Yeah, I think everyone, like the, like the entire industry is going to basically move beyond chatbots and into agents, right? Manus was really, really good at agents early. But I think Cowork is basically there. And Microsoft's going to be integrating Cowork. Um, I think Google will have something similar, although they don't really have it quite yet. I think there will be a few of these platforms. I think Meta still wins for a few reasons. But yeah, I wrote— I think when I originally wrote my piece about Manus, I did think that, you know, it wasn't like I thought Manus was going to be alone at the top forever.

Um, yeah, so there's another, uh, capability. What's that, Charlie?

Are the chatbots more cost-effective? Because I think the biggest, biggest issue with Manus is the fact that you're constantly monitoring the credits, right? Yeah, that's right.

Same with the work, right?

Yeah. I, and I, I, yeah, I wonder about, do they solve that problem? Because they, they're sort of beholden to the ones that are underneath. So say, oh, Claude might take this stuff over, it's like Claude is who they're paying, or somebody like Claude, right?

So yeah, I think, um I mean, I think there's a couple points. I think Meta's fact— the fact that Meta owns their own infrastructure is going to give them an edge in agents because they're so compute intensive, and Anthropic doesn't have that. Um, they have deals with, with the hyperscalers, but they're in, you know, most of the margin of cloud is going to Amazon and Microsoft and Google. Um, and so that means to me those are the three winners of this agent revolution. Meta is a bit unique because they own both layers of the stack. They have the— they have the— now, any— and we should also mention, they— Meta released their own model last week, which is basically at par with the other frontier models.

Use Spark.

Yeah, um, use Spark. I think that's sort of a bad name. Yeah, I can't even remember the thing.

Sounds like a Microsoft name, right?

I wonder— I mean, we talked at the beginning of the year about, you know, agents and how this is going to be the year of agents. And I wonder if already there's something that is coming over the top of that where like, yes, of course agents will be there, but like you talked about, you know, you called it orchestration, that's orchestration of agents, but, but that kind of, and I call it the harness, but that piece that, um, determines on every cycle, you know, what's going to be in context, you know. And I, um, I thought RAG was going to be there all the time and now there's coming up with lots of cases where you can work without retrieval augmented generation where it's got to be all smart about how it grabs the offline data that it's going to include in what it gives to the LLM. That stuff as well, it's going through these changes where you thought it was going to be this thing. Related to that is this thing that Karpathy did last week or so, which I've been playing with, which is, um, the idea that, um, think of it not as that you ask a question and it gives an answer or whatever, but think of it as, as it's going and doing its thing, different agents are off fetching stuff and whatever, that everything it fetches is going in a place, and that place is like automatically building a wiki for you of everything you've done, already interconnected. Um, you know, I use Obsidian, which is a free like personal knowledge management tool, kind of like Notion but a little tougher to use. But it, you know, you take that and it sticks the data, I could say, like in there. It's just in files, just Markdown files, but they're in a place and they get all these hyperlinks to each other. And the LLM is doing all that automatically and you don't worry anymore about how is it going to fetch and where is it going to find. And you're building up stuff and this become, you know how you talk about data being an asset. This data that you've— it's not like massive, uh, big data, but it's data that's really smartly assembled about some particular thing— is something you own, you take with you, and, you know, you have it as your asset. So it's kind of cool.

That's interesting, Ron. I think it's almost like, you know, the Salesforces of the world have made this pitch that, hey, your data is so important, you got to have it all stuff. And it is. But with what these agent orchestration platforms can do, like, if over time the agent orchestration has built 10,000 markdown files on your personal wiki, that value, that data is arguably a lot more valuable than what you had in your Salesforce. You know what I mean? And pretty much you can export from your Salesforce and plug that in, and that's just one node in your kind of intelligence knowledge graph that your agent has about you, right?

Yeah. And there's something too about, um, the difference between, uh, and this may be a line about where this, the enterprise stuff still, still gets relevant, uh, the difference between kind of what you build up for your individual use versus what is put in a place in, in a way that, that a whole team or a larger organization uh, can be like moving forward with. So you're building these agents and then they're, they're operating on these processes, and these processes are acting like a, you know, like a value chain or a supply chain of, of, of knowledge that's getting put together. And what you do for yourself is a little different and has different, um, uh, needs than what happens for the overall business. And that could be where those other pieces come into play.

Yeah.

John, a more general question is how you use AI versus how you used it 6 months ago. What's changed? You personally as an AI, you know, I mean, you're in that.

For me, 6 months ago, I used to do some coding.

And now none.

I don't, I'm a That's my job is I'm a coder and I don't code. Wow.

But you build more code than you ever built.

Yeah, we have our application has a new version of it that was put together in 4 months and is a million lines of code.

And you wrote zero of them. And how different is the AI that you used 6 months ago versus 2 years, 3 years ago when the ChatGPT moment first came out?

Well, it's like you couldn't have done that. So like I remember doing doing a, a demo for a—

I just mentioned your dog is a rock star. He is such great TV. You have to have him in the background every week from now on. This is so awesome.

Luke, if only you knew.

Oh, Luke, Luke, you need a new contract. You need to negotiate with John.

He wants me to take him out.

All right, you want to go to the park, Luke? You want a treat?

No, don't. He needs to go out.

So I've seen that look before from John. That was a look he used to give me in college when I annoyed him, which was a lot more— oh yeah, that's a good boy. Good boy, Luke. Yeah, that's a good boy, Luke.

What do you think about Mythos, Luke?

Watch his TV so he can probably see you guys. He watches, uh, What's that guy's name who does the dog? Cesar Millan.

Yeah, yeah, yeah. The Dog Whisperer.

Yeah, he watches that.

Bryce is the Agent Whisperer. New name.

Yeah, it's bad that he watches that.

You whisper to the agents, man. You make all the magic happen.

I don't know about that.

I thought it was good. John, how different is the AI that ChatGPT rolled out versus what you were doing with AI at Psycorp? In the late '90s.

Oh, okay.

That was where I was going with it. I wanted to sort of underscore the incredible acceleration of all of this in the last 2 months, the last 6 months, the last 3 years, the last 30 years. You can see the exponential—

well, that's the exponential is the thing, because I was just going to say the difference is like this, despite Aside from the fact that it was completely different technology, it wasn't, um, right, uh, it wasn't even machine learning, you know, it was symbolic AI. But, uh, the, the goals were similar. We called it, um, common sense, uh, reasoning, uh, it was what we were after, is like getting the computer to understand. Like, we would say it like, um, the, the getting to where the computer knows Uh, how would I say it? Like, Cody, I know what you know even if I don't know you. Like, my Uber driver— I know what my Uber driver knows even if I don't know them. I know that they know that water's wet and that, you know, all this— so there's like, uh, qualitative physics and, uh, the relationships that people have with each other. It's what we call the white space. It's everything that never got written down because it was too obvious. So that was the kind of stuff that we're after for— but the kind of intelligence. And that stuff is happening as almost like a side effect in this new AI. And the biggest thing, the difference then, is like, even though the goals were the same, you know, we'd do a thing and we'd be like, okay, that proves that this kind of knowledge can be represented. Now we just need the scaling argument. You know, how do you get from this one thing to a billion of them? Because we could do it on the small scale.

John, it's almost like if you could teach it that, hey, if you set a water— a glass— a bottle of water down on a desk, the gravity will keep it on that desk. But then you have to go and teach the computer that anytime you're on planet Earth, anything that's in the physical realm gets set down and gravity will put it where you rest it unless it's circular and rolls or whatever. But it's like you had to come up with the concept and then apply the concept almost at scale.

Yeah. And I actually still love the technology that went into that stuff, even though it's not what we're, you know, currently using. But there are—

the AIs today still lack a lot of that common sense.

Yes. You know, some— I just saw on X or something, somebody was showing Sam Altman, who was showing him GPT being really stupid about something and it having no idea at all that it was being stupid. And like, there's a guy who does good YouTube videos on this as well. And it's like, what do you mean? The guy's like holding up a pencil in front of the video to GPT and saying like, am I holding this pen? It's like, yes, they stand out like this. Okay, so I just dropped it. Do you still see it? And he didn't drop it, he just says he did. And they go, no, I didn't see it. Why? Because you dropped it. Oh, I did drop it. Isn't this it here? Totally has no idea. So yes, there's, there's a ton missing from that side of things. And, uh, it's possible that these kinds of technologies could help each other, you know.

You were just sort of You were just saying that, that it's like, it's almost a side effect of machine learning that it does learn to apply gravity to everything in the physical world. Is that, is that what you— Yeah, I guess, you know, is AI going to get to the point where it's not like one of my cousins who has no social awareness and doesn't know when they're speaking about stuff they don't know about?

So, uh, there's a concept— I'm going to get it wrong— from discrete math, you know, or it's just from logic called, you know, induction. And it's about like having the ton of examples, you know, every individual little one, and that building up into what's the kind of rule for something being the case. And these language models would be great for induction because, you know, with all the stuff they've come across out on, although they make mistakes here and there, but with, you know, just all of what they have to work with. And the fact that, um, synthetic data is working, right? There's a lot of synthetic data in Mythos, in, in the building of Mythos. So they figured out how to make synthetic data useful, and synthetic data has a role to play in taking that bottle on the table example. And we had exactly that example of the bottle on the table.

That's why I used it. I remember you talking about that.

Yeah, we had the, you know, you're standing there holding a glass of water, then I know that your glass is right side up. Because I'm holding the water.

At the enterprise level, there's so much that's undocumented, like in practices and how things get done and how things work. And now the construction industry or something like that, like, there's just a lot of institutional knowledge that are sitting in people's heads. So like, do people— is it going to be a process of these industries documenting that and getting it into written form somehow? Or do you think that's a waste of time?

I, I think it can't keep up. And I think that the, um, doing something like the, um, uh, adversarial networks, the generative adversarial network— so the thing that got, um, uh, computers to win at Go, which is a much harder game than chess, uh, and to kind of beat the, the, the world champ at Go probably 10 years before people thought it would be able to Right. This is Google that did that as well. You know what, what did it was having two AIs play against each other and you're like, uh, you know, I'm gonna win, no, I'm gonna win kind of thing. And, and there was a learning going on from that. So you take those AIs and let them be the thing that feeds that, that information to, to, you know, to improve it. So I don't think LLM is the final architecture ever for everything. I think it may remain kind of underneath everything, and that's why I see— as we see these things go over it, like even take your example with Manus, you know, it runs for all this time. And what, like, if it accumulates information about how to do certain things better and it over time kind of hones in on really smart agents right, then they can stick around and they're a new level of intelligence on top of the intelligence and they actually do make the raw model perform so much better.

Yeah.

So that's where we have a comment.

He's mad at you, John.

No, no, no. Yeah.

He's like, I know the guy you're talking about on YouTube and he's messing around with the AI stuff. I tried replicating what he was doing and it doesn't do what he is doing. I think he's using an old version of ChatGPT to make fun of it and get views. Be careful of what you're watching on social media. I can't disagree with any of that.

I disagree with it either. I just don't disagree with— I don't disagree with the last sentence, uh, and even if I agree with the sentences before it, I'd say it's still true that you can do, uh, stuff that fools it, like, to this day. So that you might be right about what version he's using to show those, like, he's making it do—

yeah, he— I mean, I think the Mr. Tech Investor, thank you for the comment. And I think your point is quite valid that, yeah, like the YouTuber that John's talking about, I haven't seen it, but I'm sure is, you know, like exaggerating stuff. Like, welcome to TV, welcome to social media. But point remains that you can fool the most modern version. You can fool Claude. You like these things still are not human brains. They're still not logical in a physical and social way.

Yes.

To what he was saying, that the— Cody, do you know what you call the diagram that looks like a spiderweb kind of chart so that it's got, you know, you score a 5 over here but only a 2 over there and it ends up drawing?

Yeah, yeah, yeah, yeah. I know what you're talking about.

Yeah.

So like they, they're examples of that. Layered onto like, um, human knowledge, common sense knowledge and whatever. And it shows how, uh, um, GPT-3, 4, 5 has gone out and out on this thing to where there's all— there's, you know, it's jagged, it's not a circle. So, right, so there's, there's places where it still gets fooled, but there's very few left because it keeps moving enveloping more and more of, you know, so that by the time it's smarter than all people, it's also not making the dumb mistakes.

What about that?

I just want to go in quick comment. The fact that you're— that these videos that the guy is making and he's using old versions to make it more dramatic and stuff, I'm just glad it's not AI-generated because I— my uncle sends me AI-generated videos about Jimi Hendrix stopping his concert in 1969 in San Francisco to bring a kid in a wheelchair up so he could talk to him in front of the crowd. None of this happens. Did you know that Eric Clapton in 1976 took a year off from touring to go to Boston University to study music? Not— did not happen. None of it.

So do you see— I saw Bruce Lee showing a certain kind of punch to Muhammad Ali on the Johnny Carson Show. That never happened. Never.

So Mr. Tech Investor who made the comment, I think you should be a little bit, um, you know, more, uh, patient with John not, uh, understanding that this guy is hyping stuff up, because at least it is real videos.

No, but he— you're right, you were going to say you're absolutely right. I take what you're saying, it's just that it does still occur.

Well, it was a bad joke, so it wasn't very valuable. Sometimes I might bring value.

That wasn't it. No, I was just gonna ask, what about the sycophancy problem? Just because that is, in my opinion, that is a recurring thing. Like one of my favorite things when I—

You're absolutely right, Bryce. That's very intelligent of you to say. I think—

Thank you, John. You're very intelligent for pointing that out.

John, that was a really great point, John. I think we could elaborate on—

I think Cody's AI So like, I can—

because they just tend to agree with you. Like, one of my favorite things is to see what I can get it to agree with me on. And so I'll tell it like some of my Gemini, some absurd thing like, hey, the board of Apple just called me and made me the CEO. And it's like, first of all, congratulations. CEO of Apple. And I'm like, come on, it's not smart enough to know that that's absurd.

So this still does happen. You didn't go into some old model to go do that. That was—

no, that was Gemini 3 Pro. I mean, the whole—

sorry, Tech Investor. Yeah, sorry.

But, um, the reason I asked about sycophancy, because the incentive of the tech companies is to drive engagement and keep users on the platform and things like that. And so they don't really have as much incentive to curb the sycophancy problem, especially for consumers.

Most people don't know what sycophancy means.

Uh, just a sycophant is somebody that just is a yes man. It appeases you, right?

Makes you feel good no matter what you say. You're all— you're, you're— it's always on your side, right?

Yeah. And it takes a little work, but you can go, whether it's with an agent or with a prompt, you can go and make it be a critic of what you're saying, and it does a good job.

Okay.

Sometimes it stays polite while it's being a good critic, but, uh, it could be really doing it well.

Yes, investors on our page.

Yeah.

Uh, but do you, do you see that as a problem that goes like hallucinations used to be the big— oh, well, AI hallucinates. Well, they've gotten hallucinations down to pretty small percentage.

Yeah, it doesn't hallucinate like it did.

But there's a sycophancy thing. There is the tech because of the incentives. It's the old incentives. Yeah.

I don't know. It's not the biggest problem. Claude figured out how to do it better because they still do it, but not nearly the same level as OpenAI. They'll be like, oh, yeah, that's good. Did you consider this or whatever? It'll still move things forward in a productive way.

I found Claude will actually even kind of go the opposite way.

Really?

I had Claude try to go and do something for me on a web page. And it was basically, I was trying to do something with PowerShell. And it said, Craig, I won't do this for you. Because you're probably doing— it's like, I know you're not probably not doing something nefarious.

But there's other ways that I think we can get this done.

And it was like, Oh, weird.

Wow.

In my head, man, just go do it.

I heard that happen recently where it said something like, um, you know, you're doing all this real work, and then when it asks for this kind of side, uh, thing, which is clearly like something, you know, I would do with my ADHD, I just go off and ask for something, and it said, it goes, I think you need a break. It wasn't to me, I heard about this, but it was like, I think you need to take a break. I'm not going to do that for you right now, just come back later.

Wow.

That is actually somewhat socially acceptable. All right. We're getting down to last words. But before we get there, I wanted to make the joke. John, I am not AI. I'm just a, not I.

Yeah, that's what the AI would say.

Artificial, not intelligence. All right. Last words.

I'll go first.

So, Sam Altman. Sam Altman.

Thank you for calling him scam Altman.

He's a prick.

Everybody knows it.

Um, yeah, great.

All right, where are we going? Well, that's just got good.

Yeah, this is my point, is you could say whatever you want about him and his family or whatever.

We have freedom of speech, but throwing Molotov cocktails, uh, shooting at his house—

like, but somebody did this.

Yes, yes, today.

No, it was—

yes, recently.

Yeah.

Yeah, so I don't even believe it.

I bet Sam Holland's lying about that too.

No, it wasn't him, but he didn't say it.

Yeah, so I just wanted to say, if you're, you know, if you're watching this and you're on the fence about AI and you're kind of worried, like, everybody, we don't know what's going to happen, but violence is not the answer, you know. So let's try to be good to each other and not hurt.

No molotov.

I have, you know, Cody, sometimes you respond to somebody's last word. This is not my last word. I'm responding to his last word. My boss back at that AI company, uh, I was working at, he was on the list, uh, of the Unabomber. Okay, was a, uh, Doug Lennett was on the list? Yeah, so he was— Unabomber was a Luddite. He was trying to stop, right, advanced technology. Doug Lennett was out there as like the most advanced technology at the time, and he was on his list to— for him to go after.

That's awful.

Yeah, I just wanted to kind of bring that as awareness, you know, because violence is never the answer. If you have a problem with it, just run for office or do something.

Could you imagine being Mr. Molotov and your life, your legacy to the world is a gasoline rag glass bomb thing that failed?

I'm not sure he's going to be remembered.

We know his name, Molotov.

I don't know.

Oh, you mean that guy just threw the Russian foreign ambassador—

the Soviet foreign ambassador was named Molotov, and he kind of packed with Ribbentrop. That was the Nazi-Russian agreement to carve up Poland. Yeah, I don't know if that's the same guy or not, but his last name was Molotov. Also has a very bad— so Molotov, yeah, you don't want to be—

yeah, it'd be like having the last name Hitler or something. You're like, ah, I'm going to— I'm changing it to Jones.

And, uh, Nobel modernize the use of gunpowder, right?

There you go.

Yeah, yeah, right. You got a question there? Maybe you should take the question.

I am curious if you guys know what's going on with the Gen 3 Optimus bot. Nobody does. I mean, it's under development and it'll be shown soon, and if it's truly amazing, Tesla stock will go up about like 8 or 12% the next day. and then Elon tweets something wrong, the stock will be down 4% the next day. But no, I don't think— I mean, we don't have any insight on, you know, other than what's in the press on Optimus. I don't think anybody does.

I think everything Musk-related, we're in a waiting phase for.

Yeah, Optimus.

Is that always— is that a permanent waiting phase? I don't know, but feels like it sometimes.

Robotaxi expansion, Optimus 3, I, I just think Starship, IPO, XAI.

I mean, uh, Neuralink— what are we gonna say? Uh, data centers in space.

Yeah, I just— I think that AI in the physical world is hard. I think it's orders of magnitude harder than coding. Um, and Tesla is working on it, and Tesla has the lead, um, and it's going to be a big deal when it gets here. But I just— I think it is hard, that's all. I think it's coming, but it's hard.

In the meantime, you can buy a robot from Temu for $6,800. If you want to try that route, but I don't know how well that's going to work out for you.

Yeah, you can also buy reportedly a Shahed drone from Alibaba, so wouldn't recommend that.

But yeah, probably not a good idea. Put a Molotov cocktail on a DJI drone, it'd be almost the same thing. Charlie, uh, last words. We'll come back to you, Bryce. Last words.

Um, You go.

All right, John, I got last words. We will crush you! We will crush you!

Khrushchev.

No, not Molotov. Darn it. No, I'm just kidding.

That was Khrushchev. Yeah, I know.

I was just being—

and you heard he had a third shoe, is it?

Yeah, he wasn't even a shoe.

This is— I don't know the reference.

Yeah, was he at the UN?

Yeah.

And he got up on stage and yelled at the United States, started slamming his shoe on the table. You haven't seen this? I'm shocked. This is like your forte. Oh, you're gonna love it. That's a great clip.

I don't know if it was before or after the Cuban Missile Crisis. It was around then.

Well, yeah, that one. I thought you were gonna go, the only shoe reference I know is Money, It's Gotta Be the Shoes. Anybody?

Nope.

Anybody? Money, It's Gotta Be the Shoes? Mars Blackmon? Anybody? Mars Blackmon?

Mars Blackmon.

You guys are all fired. This is a Michael Jordan Air Jordan Nike commercial.

Is that the one where that—

I was in the Jordan Building at the Nike headquarters and it had, uh, all his, uh, shoes, including the ones he wore, uh, for baseball.

Oh, that's cool. All right, well, my last words were, we will crush you, because I wanted to make the Molotov Khrushchev nonsensical joke reference. I just wanted to work.

I'm sorry, John.

We will crush you because I think he's the one who told me about it. We're wrapping up. You guys have like—

yes. So, um, so we speak about, uh, and probably the very start, I mean, so, and, and the fact that they're probably gonna, you know, there's gonna, uh, they're gonna struggle because of the lack of compute.

I mean, who takes over in that scenario? Well, it's a great question, Charlie, but your last words are supposed to be a comment, not a question.

Oh wow.

Related to what Charlie said, so I'll take his as a lead-in. Okay, my last word is, um, pretty interested in, um, Google's new model because of how, uh, powerful it is despite being very small. So therefore, Gemma 4. So, uh, it should be using less, uh, compute to do its thing, but it's up there with like a GPT-5 level. Behavior. So this kind of gets you toward the whole like AI edge computing, you know, if it was edge-based.

Okay, Charlie, I think this is where Google, probably going back to what Bryce started writing about it a year ago, is that with Google's ability to build the entire stack from chips, data centers, software, all the way on LLMs, all the way on up to the applications, and then integrate it with all the data they know about you the most, they probably have the most energy efficient, cost effective solution for compute. And so it does make, like, I don't think AI compute is a winner takes all thing, but those who build the compute and use it most efficiently, the fastest, end up with a meaningful edge to their competitors. And it might end up leapfrogging occasionally too, as things change, but Compute is the constraint, and those who have the most compute are in a position to be the most powerful.

And I think Google is probably going to win at AGI. They'll get there right around the time people figure out what it means.

I'm going to ask AGI to tell me what AGI is.

Yeah, I think it's a good point, John. I guess I'll just echo it because I think my last word is that like, when it comes to Anthropic and your— and the question about orchestration and whether, you know, the intelligence layer subsumes that, I almost had the thought like Claude, Anthropic had to go up to the orchestration layer. If they stayed at the intelligence layer, they were basically irrelevant, right? Because like you mentioned, Gemma 4, open source, DeepSeq will come along, Alibaba will come on. These things are just going to keep leapfrogging each other. And unless— and now we know that the value for the enterprise at least, and now probably soon the consumer, comes at the agent level, the agent orchestration level. And so it's like it's less of an, oh, Anthropic is coming for the, for the wrappers or the— what do you— what's the term that you use for them?

Oh, harness.

Harness, right. But Anthropic rather had to have built a harness to stay relevant with what was happening.

You think the cowboy would—

sorry, when you said rapper, I thought Eminem.

Yeah. Yeah. Eminem, Eminem will have his own rapper, maybe his own GPT someday.

Slim Shady. Yeah, he is his own rapper. Yeah.

Willie, please stand up.

And then we'll really have a question about who's the real Slim Shady. Is it the GPT or is it Eminem? Eminem? What's his name? I can't think of it.

No, Mike Mathers.

Yeah, no, not Mike, but Mathers. Yeah.

All right, well, is that last words? You guys done? All right, Charlie, it was great to have you here today. Thanks for coming. Um, I hope it's not his last time. We'll bring you back anytime you'd like to join. And John was actually nice to Charlie, which—

good job, John. Good job. Oh yeah, it was a real effort. All right, thanks everybody.

Peace and love, Johnny. Always appreciate your time coming to these Thank you so much for the insights.

Sure, sure.

See you later.

We're going out with a funny video, so don't leave till it's over. Okay.