The AI Revolution Call: Anthropic vs. The US Dept of War, Iran, and How Will AI Work With Humans?

That's awesome. It gets me psyched up.

Yeah. And then there's just us. Yeah.

This is the AI Revolution Call. I'm Bryce Smith, um, with, uh, 10,000 Days Capital and tradingwithcody.com. We also have John De Oliveira below, who is our resident AI expert and many-year AI veteran. I've been working in the industry. And then Craig DeLon, who works with us here at 10,000 Days and Trading with Cody. We are, we are back. We were out last week traveling in New York, um, where our fund is based, and meeting with investors and other fund managers and doing some events and things like that. And we like to go just kind of get a feel for what the vibes are like out there. And you know, there's all this craziness going on in the world, so It's good. It's good to get out of rural New Mexico and kind of see what, what's going on out in the real world. But yeah, it was interesting. Talked to a lot of people that worked in, that worked in the AI industry, from the largest companies on earth like Google all the way to startups, you know, building on top of these large language models. So happy to share some of the, some of my thoughts coming out of that. But John, how are you? I haven't seen you a couple weeks. Any, any updates on your end before we get going?

Um, had two people contact me in one day that were small businesses that were like, we want to be in on the AI revolution. Don't think they called it that, but basically they're like, you know, we want to be doing it with our business or doing it for businesses. And so trying to have me help them find, uh, someone who could work with them, um, which is difficult a little bit because, uh, you know, you're busy, Bryce, you know. Uh, but it's really— it's that kind of skill, somebody who can, you know, vibe code the crap out of stuff but also, you know, understand, I think, business to some extent, that kind of thing. So, um, if any of you listening— no, I just But it's, yeah, it takes a little skills.

It's interesting point, John. Sorry, Craig, I was just gonna say there's a huge market right now, it seems like, for like almost like AI consultants and AI integrators.

And that's where I was going with that.

It's interesting though, because on the one hand, AI is so easy and intuitive. And on the other hand, businesses need a lot of help integrating it. And, you know, like the consulting, large consulting firms that are publicly traded, they sort of oscillated between like this wild bullishness about what AI could do for their business. And on the other hand, people are like, well, if AI, AI can just do the job of the consultants, right? You don't need a bunch of IT consultants helping you do all the stuff when AI can do it for you.

So, but there's a, there's a kind of mental gap for, you know, it may be easy, but they don't know how easy or exactly, uh, you know, what the mindset, uh, should be, right? And I'm kind of pursuing this, um, thesis that, um, a lot of people— I know this one small business I'm talking to, they, they come into it with that mindset that anyone would who's familiar with SaaS thinking, oh, AI's here, let's use it and make a SaaS app. But what's in their head is a traditional SaaS app. So the AI gets used on creating the app, not on being the app. Um, and, um, there are other ways to think of it. One is like, how does AI get infused into the app? But the other is, instead of even having an app so like figured out up front, almost in this top-down waterfall kind of way where it's like, this is our vision for what this app should be, now let's go start to use AI and go to make it. Instead of that, if I can kind of lead them toward using AI kind of the way you do, where you have a problem and you go and solve it. And it's not just talking to ChatGPT, it's you've gone the steps beyond that, right? Where, for example, if say you want to have an intelligent agent-based kind of workflow across like what your business does kind of thing, well, that's not something you just like design from the top. You can take any given task within that business, kind of set aside a folder for it, and in there do what you need with the help of the agent to give it the context for, you know, what's the definition of done, what's the description of the task, what are the resources needed for the task, and just sort of do that one by one. And as you start to chain them together, you've got an intelligent business. And that's where, if we're talking about businesses being like AI native, that's what I would think of as like an AI native business.

Yeah, I agree. Every business, even legacy businesses, you know, my family does construction, for example, but like there's still a lot of office processes that you can automate, um, even if they're not directly related to like your core business. Like even if you're running a restaurant, there's still a lot that goes into running a restaurant besides cooking food. And, um, those kinds of teams are like the least um, like, not least competent, but they're— that's not their area of expertise, right? They're not tech people, they're not software developers. Um, and that's why I think it's so cool about like Manus, which I've talked about, uh, and the way these agents are going is it helps ordinary people with just, uh, like, you know, if just given enough time and credits, tokens, you can figure some of this stuff out now. Like, you know, you don't have to know how to get API keys and spin up cloud instances and set up a database. Like, then agents do that for you, and then you, like what you're saying, you feed it the right information, connect it to the tools that your team uses, and voilà, you have something like a workflow in place.

So even though, you know, you can go and do what you're saying, still there is that barrier, that mental barrier, where they know what they're good at and that's not it. And so you tell them, don't worry, it's not hard. It's like, well, That didn't solve it. So they need someone to— you say the thing about consulting and how it shouldn't be needed maybe because you have— things are easy to do. They're easy for you to do because you come at it in a certain way that is not common. And they always say, you are not the audience. You are not the customer because the customer is different than you. In this case, the customer of the consulting services, that they need the help and in several ways. One is to help see what the opportunities are. Another is to see how to approach it, like maybe to go on that task-by-task basis, not to get stuck in chat, but also not to think you're just going to go and kind of spout it all off to some product person to, you know, go, go take care of it. You know, you've got to be involved. And then is the helping to understand what's that like for them to be hands-on. And finally, it's to provide the kind of glue where, oh, this little piece here is missing. There's not— how will it— from your construction example, I can't figure out how to get it to read blueprints. OK, well, go investigate. There's some tool, and whether they're capable of or not so far. Kind of provide the technical glue on some things. So there's like about 4 different areas where you really do need some help. And I'm looking at small businesses, but now you can go and, you know, size that up to the departments within companies who are trying to roll out AI and they're talking about the failures that they're having. Well, yeah, you know what, it just keeps getting easier and yet it doesn't seem to help.

Yeah, I've worked for companies where you need consulting for anything new. Like, as soon as you start to change your workflow, the first thing is the world ends for everybody in that department. It's like, oh my God, we're changing something. And then second, you need to have somebody on site to help you integrate it. And it really just needs to be for the people in the department. They just want to click a button, right? They don't want to know how the sausage is made. They, they just want it to work. And I feel like that is why we still have that demand for consulting, especially with AI, because a lot of people hear AI and their eyes glaze over. But like, John, do you feel— so that's right now. In 5, 10 years, maybe sooner, do you feel like this is the iceman selling the refrigerators kind of thing? Do you— how long do you think this is going to be needed for?

Maybe forever. There's You know, they can say they don't want to be involved and they just want to push a button, but a role that one can play to help them is to show them that, yeah, that's not going to work. You can do that, but then you're not really participating in the AI revolution, and you'll get run over like anybody else, uh, because you really didn't— you, you just said, I want another way to not be involved. And this is the thing that involves being involved. So you can help them.

Maybe think of something else, John. I think Marc Andreessen, famous venture capitalist, created the Netscape browser and things, said software— every company is going to become a software company.

And he said that back in the— yeah, when he wrote Software is Eating the World. Now it's every company be an AI company, right?

And And I do think that's true. And then where, where you go take that is like, if you look at software as an example, like it wasn't static, it wasn't like once a company integrated ERP software, that it just stopped, like there's constant integration, the constant updates, new tools that come along. And there's like, we've talked about this, but like, the big, the biggest expense of software is not developing it, but maintaining it. And keeping the tools functioning, and so that your employees that don't know how to build software can sort of go and click buttons and things work, right? And I think AI is going to probably be something similar. You're going to have new AI tools, you're going to have new systems you need to integrate it with and things like that. Um, so in, in that respect, um, that would be very— there is a big place for advisors, consultants, types of people helping businesses do that, just like they have with software for the last 20 years.

I thought though the most interesting thing I kind of came across as looking into some of this stuff was that notion of like organically growing something that you don't have to— like they're talking about, we want to be internally using it and, you know, improving kind of the smarts of our business, but ultimately we want to productize something about what we're doing. And it's like, that's a good way to think of it, is, you know, go do inside first and how while they're doing inside, just the way you develop things is by doing it, solving problems, you know, and growing in the set of problems that are solved. And then if you do that, the intelligence that you've built is distributed throughout every task. They're, you know, like every time a task happens, something smart is happening with that task, not, oh, my whole thing is blah, blah, just people filling out forms. And when they, press this one magic button, it uses AI for something. It's like, it's not like that, right? It's throughout. And that was a new, um, a new idea for me anyway.

Yeah, I like that. Uh, what it made me think of, one of the people I met with, a friend of mine, um, she has an AI startup in the legal field. And, um, I won't get into too much specifics, but like one of the things they're doing is like, you know, attorneys are reviewing things that the AI has sort of given it the first pass, and then the attorneys have to kind of check off the boxes. You know, you still have to have attorneys giving advice, but then they're taking all that data of their attorneys reviewing those documents after the AI gave it the initial pass and then training the next model, right? Or fine-tuning, I guess, would probably be the better term. It's not pre-training. Um, and do you think, do you think that would be a common thing in all business? Like, there would be a way for even, let's say, the construction company or the restaurant where they can start to teach their custom models on their processes to make them even better? How will that work?

Well, I don't know if it matters that the intelligence get all the way into the model versus it be part of the harness. And the thing I was describing you are giving it more intelligence every step of the way. You're kind of like, what counts as a good answer here? You know, what's the right data to use? And then you start hooking the stuff together. You know, how often does this thing happen, and when are its outputs used for something else, and what are they used for? It's like you're constantly updating, um, its intelligence of your domain and of participating in your domain. I don't know about— when you think about the model itself, probably not in the way that they're currently built anyway. That's just too much money and time to go produce the pre-trained version. If you're talking about the reinforcement learning that can happen on top of that, perhaps. But there's just— Why would you bother while you can do in-context learning, which is what I guess I'm talking about. The prompt area just keeps getting bigger. You got a million tokens, 2 million tokens that you can use to describe the current situation. That's a lot. That's like, here's an encyclopedia, tell me what we should do. Like, well, you just told me what we should do.

What I'm, what I'm saying is, like, you think about a business and the value of it is the people and the institutional knowledge that those people have and how they do things. And, like, you know, when a new intern comes in or new employee, there's so much they don't know about how that business does things. And I guess that's what I'm saying, like, when, if you bring in a raw LLM and plug it into your workflows, yeah, that's very helpful and it's very smart, but it's not going to understand the nuance of that specific business. So is— doesn't there need to be a way to keep—

for if you're Coca-Cola, like, this is the way Coca-Cola does it, this is what Coca-Cola, you know, I don't know, as an example, because they're so big, they actually could spend, you know, if they wanted to, a billion dollars and go, you know, make a model. But even then, does it make sense? Because can they do it enough to keep up with Anthropic or OpenAI. So, you know, if you want to actually operate with the smartest models, then you're not going to be changing the core, right?

That's what I'm saying. You take that core model, but somehow you do, whether it's through reinforcement learning or fine-tuning, like, is there a way on the back end where you can get some of that nuance in there without having to do the whole—

that's what I'm saying. So like, there's, there's, there's, there's, um, 4 layers. There's the pre-trained thing, there's the reinforcement learning applied to that, there's fine-tuning applied to that, and then there's what's called in-context learning. In-context learning isn't just what I go put in as a prompt when I'm doing chat. If you think of Claude code, it's what's sitting in the agents, like in the.claud file, and the other, uh, and the skills files and things like this, you know.

That's it. So it's— we're talking about like skills. Is that the way—

skills are still— yeah, they're in context, meaning the way the skills take effect as you're going and getting a task done is that the skill gets— the, the skill gets looked at and loaded in along with whatever you went and asked for. You're like, hey, go make this out of this. And what actually goes to the to the language model is, go make this out of this. By the way, maybe this will help. It goes hands in a whole huge chunk of stuff. That's in-context learning.

That's starting to take off. I know with Manus and maybe with Anthropic, a couple of others, they've launched skills stores like the App Store almost where you see somebody that has already pre-built one of these skills that would be useful for you. Just take— point that into your, into your prompt as well, right? And so you'll have like in-house skills that you've developed and spent time working on, plus probably using skills from third parties as well. You agree with that?

Yes. And in, in my mind, that stuff, if you look at it, is really like early on. It's got a lot of room to get more sophisticated. Already it's very cool, but you look at it and you're like Okay, but it's literally in this one directory that I pull in these files, one per skill or one folder per skill, along with the little chunk of code that goes with it and so forth. This stuff keeps a little bit reconfiguring here and there in the same way that RAG had the whole way that it worked and people aren't talking about it as much now because they're talking about doing agentic. Search instead. And it's like, it's good, it's got room to grow. But yes, already it's got the cold, the kind of the marketplace sense. Something I worked on a long time ago was, um, it's kind of related, was it was modular chunks that you'd plug into a knowledge base. A knowledge base being like, um, early AI, uh, where that knowledge base contained facts and rules. So it was like skills and other context info. And we had what we call knowledge packs. And a knowledge pack was this just plug-in extra bit of knowledge in some domain area. So you might have your general medicine one, but then there'd be like a cancer plug-in. Um, yeah, yeah.

And we've talked about that a little bit. I think even for robotics, it would sort of, in my mind, will be something similar. Where it basically the same thing you're talking about. The robots will literally have come pre-programmed with some skills, but if you need to teach it something that's domain specific, you might have to do that. Or if somebody's already trained the robot to do, to flip hamburgers, right, and you can just download the app and it now it knows how to download the skill, and now that your robot knows how to flip a burger or something.

So like, and the cool thing there with, with robots, think about like Optimus. We've said this before, I think, is that, you know, this Optimus over here figures out how to flip a hamburger, then all Optimuses know how to flip a hamburger, because it doesn't need to kind of happen individually. That wouldn't be true if computers were analog. That's one of the problems with— they talk about, oh, we'd save so much energy and money if we switched to analog. And it's not like you go make an easy switch, but if you did, every individual one of them is like a distinct person. It doesn't— it just transfer around like that.

I feel like, I feel like as we continue on this path of, uh, spatial AI and just having robotics and other, uh, models learn, I mean, it's just gonna go hockey stick at some point. I mean, like you're saying, it's behind the other stuff we've been—

it's kind of another one of these coming up behind it. You know, we have the one for ChatGPT and everything like it, you know, taking off. And like robotics started a little bit later, but it's going to hit its, uh, hockey stick as well. And having things like world models let it go and practice in a kind of Matrix world, uh, you know, virtual reality, and but, but still with gravity, you know, whatever, try things out and learn that way. And then go and be able to use it in the real world.

An infinite amount of environments. It's absolutely incredible what they're doing now.

Just to change topics here, I don't think we've talked about it, but it is besides the war that's going on. One of the major news events of the last 2 weeks was the Anthropic and OpenAI drama with the Department of War and whether they can use Anthropic's models. And OpenAI sort of came in. I don't think we've talked about that on this podcast.

No, we haven't. Yeah.

So thoughts now, I guess the current state of things, OpenAI supposedly did a deal with the Department of War, said that yes, they could use their models. But then Anthropic was like maybe behind the scenes trying to renegotiate some of the lower executive team was still in negotiations with the Department of War trying to keep Anthropic's models in there. And then I saw today that there was a leaked internal memo where Amodei, or however you say his name, Amodei, said was bashing Sam Altman and how he was spreading lies. And he was just kind of giving, giving, like throwing wind to security and alignment, which is a term that AI people like to say for like safety purposes. So what are your thoughts on the drama? What's true, what's not true? And then kind of the fundamental issue here is, Should a— should private AI companies be dictating to the government what the AI can and can't be used for? And is it okay if AI can be used for surveillance and making up things?

The way you put that question, you know, should private companies be able to whatever, that's where I ended up was. And when you ask the question that way, the answer is no, they shouldn't. And I kind of, I went back and forth. Yeah, I went back and forth a lot. I was really conflicted about this.

The libertarian in you, I thought, was going to say the private company should be able to do whatever they want.

Well, yeah, but the Democrat in me, I mean, small d, the one who favors democracy. I mean, we have all this talk.

Cody will hear that, he could kick you off the show. Yeah, we have all the Um, Cody doesn't like Republicans or Democrats, for the record. Yes, he doesn't like the parties, two-party system. He's very okay with people who are—

yeah, he's good about like— never mind, I'm not going to go into, uh, the, uh, issues there. I, I— yes, I just wanted to clarify that this is not like a— I, um, yeah, I guess I did go back and forth. Number one, I'm like I Hate to See Mom and Dad Fighting. Um, that's the Department of War and Amadei, uh, or Anthropic, uh, because it's such a good model or set of models, and they're actually in good use by, uh, the military right now. They're like, they get it, and they're doing good. Which I was concerned about in general about the government, you know, is do they get it? Do they understand the value of these things? And, and and so forth. And, you know, I'm on Team USA for, you know, we can decide internally about whether we ought to get involved in this or that, whatever. But that doesn't— but then we decide that internally and we don't go and kind of go out to the world and get on the world stage and, you know, speak against or act against kind of whatever, you know, the US is doing. We could— the US decides what the US is doing. And how do we decide is the question. Like, what if Amadei had come at it like whatever kind of opposite opinion or something? Um, should that— that just shows that there's this guy in Silicon Valley, you know, who chooses how this stuff happens versus— because originally I was basing it on the issue itself, like, oh yeah, we shouldn't like let AI pull the trigger, it's too— it's not— it's too stupid and it could go hurt things and so forth. Then again, we shouldn't let, you know, self-driving cars self-drive because they might hurt somebody, uh, when humans are out there killing 50,000 people a year, uh, because of their terrible driving. So, you know, as long as you— you know, if you can be better than that, it's It's like you've already won. So that's maybe not the best argument is saying like, if you've ever heard what's his name at Anduril talk about it.

Palmer Luckey.

Palmer Luckey talk about it. They say, yeah, but would you let AI do some important thing like this? And he's like, would you let some important thing like that be done without AI assistance? Because if it can help and improve things and do a better job, then maybe it's good.

Anyway, yeah, I think I know, I think those are great points, John. I think from my government perspective, I would say a private business shouldn't be compelled to like sell its product to government.

No, agree.

But the difference with AI is that, and I think this is what the— like, I think this is actually like a milestone because with weapons and systems that are so powerful that is the government's— the job of those companies is to provide that to the government and things. Like, it'd be like Boeing building the F-35 and like, now— or sorry, Lockheed made it, but Lockheed's like, builds the F-35, it's like, no, America, you don't get to use it, we're going to sell it to the Russians. And the government and every American on Earth, on the country, in the country, like, hell no, right? Like, you built this incredibly powerful tool with destructive capabilities and all these things, and we're going to keep it— like, the government needs to have access to the best tools available.

No, it's a good point. Yeah. Hey, fun fact, I, um, go ahead. Fun fact, I, uh, I, I worked on the F-35 about 20 years ago.

Are you serious?

Yeah, on the Boeing side of things before Lockheed got the contract. So that's how long. Wow.

Did you work for Boeing or were you a contractor or what was it?

I was working for an AI company and I, uh, worked on, uh, software that they would use to do kind of full life cycle maintenance, you know, who's got to have what, where, when, you know, logistically to get— because it's very complex, you know, machinery. And I went and I remember because I flew to Pennsylvania for— to this one Boeing location and, you know, met with them out there and everything and then went— Lockheed went and got the guy. It's just what's interesting is that that was 20 years ago.

Military. Contract. Um, the sales cycle is quite long for—

unless you're Anduril. Yeah, that's the whole point of Anduril is it's not a, it's not a, um, contractor that says, oh, we have this idea about something, will you give us some money to, you know, work on a proof of concept so that we can then work on the real thing for years and years. Instead they're like, hey, you know what, we made this product, if you like it, buy it.

Yeah, I love that. Um, On the point, me, Craig and I were discussing this the other day on a drive home. And I was like, we were, we were talking about all this on the philosophical point of like whether AI should be allowed to make a kill decision. And you can, you can talk about the best way to do that. But what we ended up sort of coming around to this idea that the graver threat is not that AI becomes a Terminator and starts killing people. But that AI in the hands of like other state actors doesn't follow our rules, right? And so, and it like, let's China or Russia, let's say for example, build a very sophisticated AI. Let's say they have humanoid robots, which they do. And let's say they're, they build humanoid robots with no, they have no qualms about killing a person, right? They're military robots. And then we have this like incredibly powerful AI that's also pacifist. Like we're hope— we're helpless, right? Like, it would be the graver threat is actually that we build these incredible powerful AIs that won't defend us from a, you know, a nefarious or a hostile state actor. Yeah.

What's up? Like, the guy, um, I heard the, the deputy, uh, Deputy of, uh, Department of War guy, uh, he was on the All In podcast and he was talking about, um, The more likely scenario in the short term is autonomous drone versus autonomous drone. You know, they're going and it's on its own, it's deciding to do something, but it's not like against a person. But again, it's not like I want this stuff to be, you know, on its own off killing people. But it comes down to it's about who decides that, and that ought to be decided by you and me. Although I don't have a vote all by myself, but so I'll let the rest of Americans be part of that too.

But see, I have a— when I think about that, you know, I would like it to be decided by you and me, but it's the tech, the technology companies are the one bringing this forward. You know, it used to be where the military just had all the secret tech, and we're still excited we have our toasters, you know, like, and now it's completely flip-flopped, you know, like, tech has all these amazing secrets. And one of the points I asked Bryce the other day was like, does Anthropic and the other tech, do they have a moral obligation? And even some nationalistic— I mean, not in the, you know, like China sense, but to help the government protect us when they— it's like, like Spider-Man says, you know, with great power comes great responsibility. Do they have that responsibility? Responsibility to ensure that lives aren't lost because they're withholding their technology?

Well, we're mixing together a couple questions here, right? There's, uh, the, the one has to do with, um, is it okay for the Department of War to say it's that, hey, it's not okay what you're doing with the stipulations, right? And that's where I agree, not okay to, um, put those stipulations. Now, second question is, should Anthropic be forced to provide, uh, what it has? And now you're, uh, kind of— because there's also— that's been threatened, which is kind of weird, because if they're, uh, kind of identified as a supply chain risk, but then Department of War is going to come in and, um, I forget what the commandeer, I guess, would be a word. That technology.

Sorry, the Defense Production Act or the Defense Production Act.

We're going to take that because it's essential. So wait, it's both essential and a supply chain risk. This doesn't—

those two things can't both be true.

You can't both be true. So if you're going to go on to that second one, you better release the first one. But in terms of it being like essential, Maybe not at the moment because there's other things to go to, but if they get to a point where, you know, we're really at World War III and, you know, we need certain kind of something for things to, uh, to, to work, uh, I don't know, you know, at that point.

Um, so it's, it's more philosophical right now and less practical, but there could be arguments made.

I think they're using— I think the War Department is using Anthropic currently to help plan and fight the war in Iran. I mean, I think that the Defense Department loses the supply chain risk argument. I think they're going to lose that one. And Anthropic filed lawsuit today to have that distinction or that designation removed. And I think the courts are going to say, well, you're saying, look, the War Department said this is so essential, we have to have it. And it's also banned by— like, it's also banned.

I thought that they should have been filing suit. Second one.

Yeah, yeah. And because that supply chain— the supply chain risk designation, from what I've read, is a pretty steep punishment. Um, no, no company that does any business with the American government can use that company, which would include all of the hyperscalers which support the Anthropic model. So if, if they lose that argument, like Anthropic just gets turned off by Microsoft, Amazon, and Google, It'd be tough.

It would really be tough. So the another thing that that guy who's from the Department of War kind of, uh, uh, put out was he put out a tweet, um, last week right when this stuff was at its height, and he showed a like a screenshot of something from within Anthropic, and he said they're trying to hide and get rid of this, but this is, you know how they have like, um they call their constitution, and it's the rules that underlie ethically how the model should behave. Well, apparently in their first version of the constitution, it had something like, um, when in doubt, choose the, the choice that others who are not part of Western culture would choose. It was like anti-Western culture. And if so, we have a real supply chain risk because if you're building into your AI stuff that has kind of this, this time bomb of you're not going to— when it comes down to it, it's the US or something. So that's probably not in there anymore. But the notion of it is like there were some maybe real America haters that were part of what was going on with that, you know, product, kind of like—

Well, they did, they did recently, they did update their, their safety, like memo or something like that recently, like right before they were designated that. So I don't know if that has—

I certainly understand where Anthropic's coming from and trying to— not, not the anti-West thing, trying to instill philosophical, like guidelines into the model. But on the other hand, like, it's always going to be somebody's viewpoint Right? Like, like, I, you have to appreciate what Elon, at least what he says Grok is intended to be, is like maximum truth-seeking. But even like with that, like, what if the AI determines that like it's more efficient to not have people like, or something, right? We don't need certain people because inefficient or something.

Yeah, there's work going on for that.

Efficiency is not the end-all be-all. Right? And so, like, there's like maximum truth seeking. I don't know. I don't know if maximum truth seeking is the ultimate, like, phrase that you can put in the AI to make it ethical.

And no, there's a lot more to it. But how do you go about that?

Like, do you— like, the Constitution has— like, do you— I mean, if you could make it just follow the American Constitution, like, just follow the law.

I mean, if you just follow the law, there's more than that to it as well, I think. But the On the other hand, this notion that this is this thing that's out there trying to do good or bad— like, I saw this news headline the other day, and it said, you know, agent escapes and buys Bitcoin or something, does this and that to go buy Bitcoin. It's like the degree of anthropomorphizing that's going on for a computer program, which is what this thing is, you know, and Amede saying, yeah, at this point we don't know if it might be somewhat conscious. It's like, give me a break. That's not what consciousness is. It's not conscious. And in fact, I, you know, I, I kind of, I put my bets on it won't be conscious even when it gets AGI and so forth. It'll have great understanding of the world and people and everything, but it won't itself. Then you'll be like, hey, I'm going to turn you off. It'll go, okay.

Right.

Well, so when, when Amadei says stuff like that, like we don't know if it's conscious, I mean, are you— do you think that's theater? Do you think like— yeah, because I feel, I feel like part of me feels like he believes some of that, maybe.

But I— it's, it's bad theater. But yeah, uh, does he— I mean, because he puts out constantly— he's, he's kind of a doomer, uh, which is interesting for what he's building to then be you know, to be a doomer.

But I don't know that Anthropic has anything like this. But I know OpenAI, in their contract with Microsoft, it says that Microsoft has all of the IP rights until you get to AGI, which has to be declared by the board of the nonprofit arm of OpenAI. But then you get Sam Altman constantly saying, yeah, we've hit AGI, we've hit AGI, because he would love to not pay Microsoft anything.

Which is kind of such a weird, uh, yeah, kind of, um, discussion happening around, you know, is it AGI? Because I haven't heard anybody, uh, have consensus about what AGI means.

I can't— for as Microsoft being the behemoth it is, I can't believe that their attorneys agreed to making that like a foundational term of that contract.

I think— didn't they back off already though? Like they were going to base it on that, but then they went, oh, we don't care, we don't— hands off, you know.

I think it's still in there. I think— I don't think they've completely taken it out because that was an original term when like they— when they helped— when they gave them the first check was— it's not like the subsequent things after ChatGPT has already taken off. It was like the initial investment, they put that as a term.

So speaking of, uh, now the, the money side a little bit or something, Something, you know, we talk about the different businesses and how are they doing, the hyperscalers, how are they doing through all this and everything. I saw something recently that was arguing that, you know, even though Apple looks like they're just so behind the curve and everything, their chip, and they were kind of one-to-one examining like what is it about, you know, a Blackwell or whatever the next, what's it called, the one Rubin, Rubin, you know, like, here's how Rubin works with regard to memory and so forth. Here's how the M5 chip works with regard to memory. Here's the energy usage of this versus this. And, you know, this is behind on this but ahead on that. And kind of like, and also the question of whether is inference going to move out toward the edge. Instead of being just in the big data centers. And well, if that happens, is Apple then in a very good position because they have such great chips that they can— when people went and talked about buying Mac Minis, which wasn't because they're looking for the maximum power, but still it was like, oh, that's where we'll go. What if it's like, I'm going for an M5, I don't care if it's an Air or a Power whatever, a Pro MacBook Pro, but I'm going for an M5, I'm going for an M6, um, suddenly you're running serious stuff. Uh, I don't know.

Yeah, I think it's going to be both, um, from what I've seen. I mean, there's— there are advantages to having local LLM. Number one, you're not spending a bunch on cloud resources, right? It's very cheap. Like, if you use Manus or any of these agent things very much, you run out of credits pretty fast if you have it doing a lot of serious work. And so in theory, if you could do all of that inferencing locally, that would be more efficient. But I think there's like— I think you're always going to have some basic inferencing probably happening locally. But these models are so big, like, it would take— require like a fundamental change in the model architecture, I think. And also, your data is all almost all in the cloud. Everybody's data is in the cloud. Like, very, very little data is stored locally these days, especially at the enterprise level.

I mean, Apple could get in that game of it being stored in the cloud because they have that notion, and they also have a, a reputation for, uh, really supporting the, the privacy, you know.

Yeah, I think— I mean, I wrote after the OpenClaw thing, I think I wrote a market diary, and I said that Apple is supremely well positioned for this if this is the way things shake out. But it's still, it's, it's just a little bit, I just think the ship has sailed a little bit on local versus cloud. I think we're on the cloud train and it's fundamentally more efficient. You get upgrades, the whole SaaS model, right? Like, if you're doing LLMs on a computer, next year, the LLM is out of date. So you have to like download a new LLM. So people are doing that. You know, it's like, I don't— I just— there's all these problems that we've already kind of solved with the SaaS, and like, are we going to go back there for AI?

No. And I mean, we, we download software all the time, get updates and things like that, but, um, so you can have things come onto machine, but you can also have where it's getting less and less as a percentage of the overall software market, the like browser-based SaaS is much larger than downloaded. But you can also have the kind of, um, uh, instead of large language model, kind of large action model, and have a less frequently changing but powerful local thing that's helping with, uh, running everything on your machine, uh, and it still has to work with the LLM that's elsewhere. But it's carrying out. So you get down to where it's doing the kind of the Manus thing, but locally, right, you know, that it's like the machine gets smarter too.

Yeah. So go ahead, Greg.

I was just saying, I was wondering if it's going to be like, like, end up going to be a hybrid approach, you know, like, there'll be certain applications where edge case AI is absolutely 100% needed. And then that data can be— because the data is so valuable, right? Especially, you know, when you're talking about iterating on your product and just getting information about how your customers use the product. So if that information can be packaged and then sent to— instead of continuously inferring, I don't know if that's something— am I even talking about anything that exists?

Or— well, but I mean, OpenClaw, Claude Code, these things already can work with what's local on your machine and can put more stuff local onto your machine. So they kind of, they can operate there. And then you have them operating where it's got a virtual machine, but sort of locally, kind of having that be this encapsulated thing that stuff happens in. So yeah, this, I guess, is already some hybrid going on, right?

Okay. Even with your OpenCLaw and Claude code, you're still inferencing the model in the cloud, right? Even though it's running through a portal that's downloaded to your, your site. You can't run Opus 4.6 on any kind of a device yet.

I don't think of— yeah, but I don't think of the model that much anymore. I think of the harness. I think of the whole, right, the set of things that happen in conjunction with that, that model. Like if I'm using, you know, I can use Claude Code and I can plug Codex into it instead of Claude, you know, instead of Opus. I can plug these other pieces along the way. I can have a mixture of Google anti-gravity, you know, and some anti-gravity, some Codex, some Claude code, and kind of mix around. So I don't care as much about like the model.

Right. Which, which Where do you spend most of your time these days? Is it Cloud Code or—

Right now it is. I just did a thing where— and I like to— you guys saw my comic. Did I share with you?

I don't remember.

Oh man, I gotta like— you gotta let me pull this up on the screen here.

Okay.

Let me see. Figure out how to do this. So what I was going to say is I do a mixture of Claude where I get it to— I work through a prompt with it that I then give to something else. Like, so today I did a Claude help me make this detailed prompt in order to vibe code an application in Codex. Right.

Oh, wow.

I've been doing that.

I feel great. Then I said, but this other thing here, I said, hey, Claude, I want you to help me make this comic book. And it's for an upcoming product at my company. So this stuff, like, don't tell anybody. No, I'm gonna put it on anyway. I don't care. It's like we're a small enough audience, I think, that this is something that's being worked on at my business. And it's described by a comic that I made. And I made the comic by talking with Claude to get really rich prompts that I then fed to Nanobanana Pro. And I'll show you what it looks like.

That's so interesting. I use that same— I use that same strategy because I started playing with Godot, the game engine. And I'll use— I'll use Claude to help me write the code and help me make prompts to write the code to put into Godot. And it's been Super helpful.

It's like there seems like if I want to share the file, I need to upload it or something, but I can share a screen.

You should be able to share a screen.

And There it is. Okay, I'm gonna share a screen. Oh, cuz it's Hang on. Here it is. Because you and I are talking and our window is in Chrome, it keeps making Chrome jump to the front. But here, Do you see that?

There you go.

Yep, we see it.

Okay, so it starts off with a story about the first wave of AI gave us powerful tools, but only for those who can master them. And it goes through how it really required expertise to do anything. So that means that you had the AI fluent, like the affluent, uh, the AI fluent few and the AI struggling many. That's the, the great divide. But then what if AI could watch how you work, identify where it can help, and automatically build and deploy the custom AI agents you need? This is true AI democratization, not giving everyone the same tool hoping it'll figure it out, but giving everyone a system that builds their custom AI solutions for them. So then it brings into it, and this is— I taught it how to do the kind of cartoon comic, right, that is like Scott McCloud's Understanding Comics, where there's this whole book about comics that's done like this. It has, you know, a perfect out— a person outside the frame, right, as the narrator, right? It says the future of work isn't human or AI, and it's not human and AI. It's AI building AI for humans.

Wow, that is awesome.

It goes through and says that— let— now it's going to talk about 4 different people, 4 different titles. So it's kind of persona exploration, different industries. They've never met, but they're all tangled in the same knot.

Yeah.

And it also does a thing where it's— I'm doing, um, genre switching. So for each one, one of them is like the kind of Spider-Man look Another one is manga. Um, so here you go. She's the one who keeps the engine running, Maya Chen, VP of Operations. I'm responsible for making this whole machine run. But so what I have is a kind of a problem slide, a problem page and a solution page for each persona. So that's her problem page and that's her solution page. And then, um There's, uh, this guy is Marcus Torres, the CTO, and he's got all these different AI tools and different vendors, and nobody told anybody. The bills keep coming but don't know who's using what. And got the different, uh, kind of departments showing there. Don't know what to do about it. The chaos wasn't a bug, it was the whole system. Same guy, one platform, one view, one answer to what are we using. AI, consolidated view. You know, people happy. Engineering builds product, not plumbing. And this is what's going on. Then you really like that.

I really like that. I don't mean to interrupt you. But I really liked that last slide where you say my engineers are too busy building integrations, they're not building product, you know, that, yeah, that opportunity cost is a serious thing that we talked about all the time in enterprise. So very interesting.

So then you got the— this guy's the managing partner. So I did it in kind of a head coach theme.

Yeah.

And so that's what you see back there is, you know, senior associate analyst, so forth. He's talking about you're only as good as your team. And here he's showing how he had to jump in in the meeting because they weren't handling it well. So he's got to just kind of do it right now. You get the whole feel of it as it happens there. But then he got an assistant coach, one that never sleeps. Yeah, uh, the AI pulled the comps and drafted the framework, and, you know, see, you get the idea. Yeah. So then finally you've got, um, the head of AI, Kenji Tanaka, and she's hired to bring the future, but the future won't come. And the board gave me one year, prove AI works or the budget disappears. And so all— it's got that kind of look you get when, you know, things aren't working right. For it. And then watch her once things settle in. Same person, but you know, that kind of a look. Hapic advisor— Hapic's advisor— 3 new automation opportunities detected. Wait, it's, it's not watching me configure, it's watching them work and it's building for them. So then, uh, it goes into another section where that narrator is saying, these aren't relics, And what I'm about to show you isn't magic. It's 7 things that, that change everything. Pay attention. So now it's going to be— it didn't have a good transition there, but it's a kind of got a, um, a quest happening where there are these 7 things that she's gonna, uh, go after. The first one, you know, is here's the realization that, um, everything, uh, I know what needs to be automated but no time to build it. Wait a second, it's building it. This one has one slide per, for those things that are happening. Uh, what she's leaving, everything she knew is walking out the door. Uh-oh, the system didn't just store her data. Wow, that's how she worked. What she knew, leave, it stayed.

Um, you know, preventing brain drain, and that's interesting.

So then, um, she had an assistant, but, um, It only spoke when spoken to. That's the normal chat kind of thing. But then here it is, suddenly, I noticed you did sa-sa-sa, so I did blah blah blah to help. Um, and, uh, it goes on. So anyway, you see what I'm— this is like, it was all made really great, and then it comes to an end here. She gets the different ones, finally brings it all together, can, you know, integration with other stuff on the outside. She's got her, her powers, you know, her the the ring with all the jewels on it, you know, kind of thing. And she's got all that pulled together. And it's like, you didn't collect 7 separate tools. You discovered one system with 7 dimensions. I'm not even an AI expert. I never was. I'm just someone who needed things to work. And now they do. And that's the idea there.

So, fantastic. John, I got several questions. So you made that, with the Nano Banana? It made each slide?

Uh, with first a lot of work with Claude.

Okay. And also with chatbot.

Sorry, just the chat.

Yeah, chat. Okay.

And I like— so I'm just chatting. And also with, you know, and I think it's important is I have zero, uh, artistic ability, but I have been interested in it as a you know, comics as a genre for quite some time. I've read those books. So like, I know what I wanted, and I'm able to describe what I want. It's not coming from zero. That's why it's important, uh, that people get that, you know, having the end in mind and knowing what that looks like in some detail is valuable.

Totally. Um, one of the things it made me I think that's awesome. Like, one of the things that made me think of though, is like this idea of observability, app observability, which is an old concept, but it's like Datadog and Snowflake has a service that does this. But like, how are, how are big enterprises like looking and watching at who's using chat and how much and are we spending too much on cloud, or we need to cut back on this API, and how are consumers using our app, are they using our chatbot? Like, like, that's sort of the pitch for Datadog and the other observability companies. Do you think that's— does that fit in with that framework you just showed there of one system that's sort of doing everything for you?

Yeah, um, so I use Datadog, uh, we, we have it, um, and, uh, I think that we're not different from a lot of businesses where there'll be one or two people who, uh, use it in anger and actually go in and, you know, get, get really something out of it, and a lot of others who have access to it and don't really, uh, make that kind of use of it. And so it's up to the business for, uh, how, um, intentional they are about going at collecting that kind of information. But also, is it just something you're collecting as static thing. Like, we can use it both as, help me go solve this bug, I'm going to trace through what came in, or activity level on different, you know, kinds, kinds of things. But when you could take both of those and sort of consider putting them like on the other end of an MCP server where you're like, the agent can look at that activity and make use of it to make decisions, then you start getting interesting, I think.

Does Datadog have anything? Does— is it— is Datadog building that agent that you just described?

Um, I don't know, and we don't have it for that, but I have it already for enough other tools that I think it's either there or could be there easily.

Yeah, because yeah, see, that's what's interesting. When I first went down the Datadog rabbit hole, um, I was like Well, this seems like a company that's probably screwed with AI because they make a bunch of dashboards for humans to look at and observe, right? That's kind of the fundamental premise of the company. And I'm like, well, humans aren't going to be the ones doing the observing, maybe.

Well, they have an API. So it, you know, right.

Yeah. And, and so then whenever that I kind of got that over that hurdle, then I was like, oh, wow, this could be really incredibly positioned company to do the the things we're talking about and helping, because this is only going to get more important as having— as every company becomes an AI company, your small, medium, and enterprises are all going to need a way to see what the whole organization is doing with AI, with the applications, with all the software that they're— that they're— all the tools, right? Yep. So I'm gonna do some more work on Datadog. That's interesting.

And others like them. I mean, the thing, Datadog, we use one called Braintrust, and, uh, it's more, um, it's, it's like a Datadog that's focused on agents and what they go say to each other. Um, and, uh, another interesting one.

Well, I look at it. Awesome. Well, that's an hour. Do you guys have final thoughts? Craig?

I'll just say, if I've just been thinking about this all week since the Anthropic OpenAI deal, and it just— all I have to say is that I would have rather been protected by Anthropic than Sam Altman. But, you know, we are where we are. And hopefully, the government can utilize the technology better, you know, than than what I'm thinking they can. So that's just kind of my final thoughts here.

We're only at the beginning of chapter 1, so the stuff may still shake out.

John, final thoughts?

Just, we always take a kind of temperature check on the rate of change, you know, are things still progressing? Uh, yeah, they are. You know, things are still going very fast. Uh, you know, every week new models, new capabilities, uh, you know, people and businesses figuring out how to make use of them. It's, uh, we are on the curve and the curve is getting more steep.

I think you're exactly right. I'll just echo that because even in my conversations in New York this week and even the small talk I overheard at the tables next to us and on the subway. And like, when you're— when anytime we were around business people or things like that, I mean, everyone— I heard so many conversations about AI. And finally, I know New York is in— it's a tech place and it's a finance place. So they're talking about it, of course. But like, it's really— and then with this Anthropic fight with the DOJ, Department of War, it's like, this stuff is so like everyone is integrating this stuff now. It's like, it's like as important as the internet, like, oh, our AI is not working, like, oh, our internet's down. Like, it's almost— it's not quite to that point yet. But everyone is doing it. Everyone is integrating it. This is real. It's not like— I don't know if you remember, there was that report sometime last year, like, oh, only 5% of companies are actually using AI successfully or something. I'm like, well, I think that number is going higher. Like, everyone is trying to figure out how to That 5% could have been a real number. I'm not saying that, but like, now that's the bull case. The bull case is that only 5%, like, there's 95% of businesses that are yet to figure this out. That's very bullish for the AI industry.

Yeah, yeah. If I had a nickel for every time we overheard the term API, I'd have a lot of nickels. It was all over. Everybody was talking about integrations because we, in our, where we, where we stayed in our hotel, there's a lot of people just coming and plugging in and working in their own calls and stuff. And it was, it was very telling. It felt like I was kind of plugged into the, the hub of where all this is being integrated.

So it was very interesting. Yeah, quite a tech scene these days, you know.

Yes, it is.

More than just Wall Street.

Yeah. All right, folks. Well, good to see you.

Okay, bye, John. Thanks.

Thank you. Bye.