Breakeven Brothers

Cool. Alrighty, we are back with episode, gosh, 25 of the Breakeven Brothers podcast. It's a beautiful summer evening in Arizona. It's a peachy... I don't know... 104 degrees right now at 10 o'clock at night, so that's crazy. Um, Brad, how's it going?

How are you doing?

Excellent. Just came back from a weekend in Vancouver. I would highly recommend it to people who are interested in visiting Canada. I think Vancouver hits all the marks: great food, great transportation, a high variety of things to do—sightseeing, activities, outdoors—and pretty easy to just book things. There weren't long waits. Again, I think maybe I already mentioned this, the food is freaking good. I was just in Japan not that long ago. I had sushi at sunset in Vancouver on the beach, and the sushi was cheap and really good. So yeah, feeling energized, and it was just a great time. I was there for four days, from Friday to Monday. Jam-packed schedule, really long days. I feel like I did it all, in a good sense, but yeah, it was excellent. If you're thinking about traveling to Canada, Vancouver is a great spot. And I stayed in Coal Harbour. I think the hotel was in an excellent location, so I would highly, highly recommend that.

Yeah, that's cool. One thing I didn't ask you—we were talking about it a little bit before we started recording—but one thing I didn't ask you is when I was there last time, the Canadian dollar was stronger than the American dollar, or the U.S. dollar, and so things felt more expensive. I think as of right now—I haven't looked in a while, but I know for a time at least—the U.S. dollar was stronger than the Canadian dollar, and so did you feel that? Everything you said was "cheap," so that's kind of what caught my attention, but did it feel like things were cheaper? Like you felt like you were getting more bang for your buck?

Definitely. Yeah, I think right now it's, uh, one Canadian dollar is like 0.7 U.S. dollars, so you definitely get like a, you know, 30% discount. And so food was cheaper. I honestly felt it was a little bit cheaper overall, to begin with, but I think the currency does give you a little bit of a slight edge. Not as cheap as, like, Asian countries potentially, but pretty good quality for, you know, being a neighboring country. And yeah, pretty nice, very, very cost-effective. And yeah, the food—freaking good prices, pretty good. I think the one thing that I was surprised cost a bit of money was they had this suspension bridge park, which is a very manufactured-feeling tourist attraction where they have a long suspension bridge about 250 feet or, yeah, I think 250 feet, or maybe like 200 feet high. And I think the ticket was like 70 CAD, or like 50 U.S. dollars, and it's not that crazy. I mean, it's cool, don't get me wrong, but that was like the one thing that I felt... like, compared to getting sushi for two people for, I don't know, 40 CAD, getting a ticket of 70 CAD per person for three hours at a park—cool exhibit, cool park—uh, but that part was more expensive than I thought. But yeah, overall very cheap, uh, very good food, very good quality. Transit was really cheap too. Um, so yeah, I would highly recommend. Thanks for, uh, for pushing us that way. I know Ben's been talking to me about it for at least like two or three years and it had been on my list. So we finally, uh, made the decision to go.

Yeah. Well, you know, you and your wife, you guys are big travelers. So the fact that I had been somewhere that you guys had not, I was like, "Oh, cool. I can contribute to this conversation a little bit." But yeah, that's really cool. One thing I remember—and we can move on from Canada—but one thing I remember thinking when I was there was obviously things being more expensive, again, because of the dollar situation, but then also, too, the food being much fresher. And when I say that, it's because I think when I was there, I'd either just graduated or was about to graduate college. I wasn't making much money and I was kind of just, you know, doing it on a budget. And I remember we went to Chipotle and, uh, one, I remember it being more expensive. So I remember it being like two burritos, like 20 bucks. And like at that time, that was like, "Whoa, I remember this being like 15 bucks back at home." Again, this is like 10 years ago. Um, and then, two, I remember it tasting so much better. And I think, you know, my wife and I—there's no proof in this—but we just thought that it's the cheese. Like the cheese didn't have all the... you know, you hear about when people go to Europe, they lose weight eating the same foods because, you know, in Europe, they don't have all the, you know, seed oils and whatever the case may be. But like, I remember being like, "Oh, this is actually way better." Like the food is actually better too, um, compared to the same version in the U.S., you know, if you're having like a fast-food chain.

So I didn't have any cheesy dishes, but I can definitely plus-one that. I think the food quality is better, cheaper, and the service is good. I feel like they really care about people. Again, it's comparing Japan or Korea to Canada. They're not as outwardly friendly as Japanese folks, but they're up there. They care, and good service all around for all the places that we went to. We ate out quite a bit at that time. Totally would recommend if it's on your list.

Yeah, that's cool. Awesome. All right. Well, glad you're back. Glad you're recharged. One of the first things that we want to talk about in this episode is some recent news surrounding, I'd say, talent. And we touched on this a little bit last episode when we talked about the, you know, I'm going to say, quote, "poaching" of OpenAI researchers, employees, whatever you want to call them, over to Meta for like really large sums of money, really large compensation packages. So we touched on this a little bit last episode, but what's really interesting now is there's been some additional folks going over there. So I think just as of today, I think I saw an article, two folks were going from OpenAI over to Meta, to the Meta Super Intelligence team that Mark Zuckerberg's putting together with the Scale AI guy, Alexander Wang, running it. And what's really interesting to me is a couple of questions that kind of come out of this. I think one is, um, you know, when something of this magnitude, like these kinds of headlines, um, one thing that comes to mind, at least to me, is where does Meta kind of see this playing out? Because ultimately, to pay that kind of money—and we're talking rumored to be, you know, $100 million offers, like all-in total comp—you know, to be paying that kind of money, there has to be an ROI for that. So right now, if you think about where Meta is, you know, they're very much in the space where they have the Llama models, the open-source kind of local models that you can run, but they don't really—they're not really in the same league as the foundational models, like, you know, OpenAI and Gemini, you know, Claude. And so what do you make of what Meta is trying to do with that Meta Super Intelligence team and just the news that these, you know—I keep saying, I want to say "execs" because that's kind of what they're getting paid as. I don't think they're really executives, they're more like researchers and engineers.

Yeah, because they get paid more than that. Like an Apple employee, I think we talked last podcast a little bit about this, but an Apple employee got poached for like $300 million. OpenAI, $100 million, $200 million. It's in the large sums of money. But since our last podcast when we talked about it, I think there are two updates. One is it continues to happen. It's not something that existed once and is over. And then two, there has been a little bit of discussion about Meta transitioning to a closed-source model. So previously, I think Meta was just trying to get general intelligence out there and take a lot of money, funnel it in, and have something that everyone can use to kind of set the playing field. Now, I think they've accumulated, you know, research talent from all these competing AI labs that it could be time for Meta to take a step into, you know, a commercialized, closed-source model, kind of like OpenAI and Anthropic. So I'm not sure. I think, you know, last podcast, we talked about them coming up with unique social experiences. I'm not sure that's the case anymore. And with the AI hype and AI craze, things change very rapidly. This feels like an AI talent war, so to speak. Like people are getting poached left and right, either, you know, from OpenAI to Meta, from Meta to OpenAI, etc. But I think Meta definitely has the biggest paychecks right now, at least for AI talent. So lots of people are going there.

Oh, and the third thing that I just remembered is that Zuckerberg, I think, went on a news publication and mentioned that Meta is building this massive AI factory of just having tons of AI chips and GPUs. I think that part kind of plays into the mission of, you know, you imagine you're talking to a top researcher from OpenAI. It's, "Oh, we'll pay you great money. We'll give you like unlimited resources for you to tinker on things," because as a researcher, engineer, you know, experimenter, don't you want that? And then three, it's probably like, you're going to be with other smart people. OpenAI definitely has a pretty impressive roster that they have today, but pulling, you know, top leaders from Google, OpenAI, Anthropic—it's just been this talent mishmash of just getting people to join these different companies. I think Meta has a pretty strategic and compelling offer, all things considered, with high cash, unlimited resources, and people who are just as smart as you, hand-plucked from other companies. So I don't know what they'll do with that. Hopefully, they make a big impact. It'll probably be either social—adding it to Facebook, Instagram, the family of apps over there—or two, going closed-source and trying to compete with the Anthropic and OpenAIs of the world. I'm not sure where they'll end up, but those are the two real strategies for them right now.

Yeah. And what you're referring to, I think they call it an AI supercluster. I think they've named it Prometheus and Hyperion. Um, yeah, and I'm not—I think they said that they're using it to train like their own, um, large language models in the future, and I think it's expected to go live in 2026, the first one. So, um, because I think we touched on that a little bit last episode—I think it was last episode—about like the energy requirements needed to fuel those kinds of things, which is kind of coming into play. So one thing I found interesting, Brad, I guess on the topic of like talent, is this: when you read any TechCrunch or Wired article, they're using the term "poaching" a lot. And we've used that term too. And it's really interesting because I think "poaching"—and maybe this is just me, I want to kind of get your thoughts on it—is kind of like a negative term. Like, it's illegal to poach, you know, like animals and stuff like that, right? Um, why can't they say "recruited"? And some articles do, but I think there's been a lot of emphasis on the "poaching" and almost like an acquisition-type language. And to kind of reinforce that, you know, I'll say negative connotation with these moves, the OpenAI team—I think the individual's name is Mark Chen. I think he's like the chief research officer. Don't quote me on that, but his name is Mark. And he had like a memo sent out to employees after one of the rounds of people going over to Meta, where he said—and I think the article quoted that—"it feels like someone broke into our house and stole something." Now, obviously, I understand that it's a big loss for an org to lose smart people. That's always the case. But in this scenario, it seemed like, you know, these are people that have their own free will to go work somewhere else. They get paid good money. From what I understand from the articles I've read, OpenAI countered these offers, at least to some extent. But what do you think about OpenAI kind of casting this as like Meta's doing something unethical or "distasteful," I think is one of the words that they've used? And Sam Altman, the CEO of OpenAI, said that "missionaries"—what did he say? He said, "missionaries prevail over mercenaries," something like that, and that OpenAI is on a mission and Meta's doing something else. What do you think about that kind of painting of the sides that OpenAI is trying to do with what Meta is doing?

I think Meta comes out swinging, and I think that's the "poaching"—these massive offers that are hard to decline. And two, the timeline in which these individuals are given to make a choice. I think some of the reporting came out recently that they're given literally hours. I've been in job negotiations; I've never had an hours timeline, but that sounds really stressful. And I'm also not getting paid $100 million. So you could put those together: $100 million in four hours and make a decision. That feels like a ton of pressure from your head honcho at OpenAI, or one of the head honchos. That, I think, is why it's characterized as poaching. Because: important person, strict deadline, and high pay, it puts you in a spot where it's hard to make a rational decision. You kind of just have to go with what you think. And when you see people around you taking this package, you think, "Oh, maybe I'm missing out. Maybe I could get paid a lot and enjoy my time."

But I do like how you bring up Sam's quote because I think right now, OpenAI can't really compete with these offers because they're not public. They are being evaluated at a valuation that is probably disclosed every so often. I'm not a startup guy, but I think Sam had pitched it as, "Yeah, it's the mission. It's the people who stay here." And two, "I think our company will be worth a lot." I think he had mentioned the value in terms of dollars but wasn't able to provide a concrete, "Yeah, we're paying you $100 million," because valuations change and competition comes and goes, as we've seen in the AI space. But overall, I think it's the pressure, the pay, and just seeing people around you kind of not crack, but make that decision to go. And that makes you kind of feel like you're on a sinking ship. So I wonder if these people have any regrets? I mean, probably not since they get paid a lot, but you never know until you get there. Maybe Meta turns out to be a crapshoot and all these smart people don't have the authority, or maybe there are too many smart people in the room to make an impact. Hard to say, but I think from the reporting, it was very clear that Meta put them under tight control and a tight timeline, probably to force a decision to get them to join. That's very strategic recruiting, to get them to be like, "Hey, you have no time and here's your juicy offer, take it or leave it." And you know, clearly, it made an impact.

Yeah, yeah. And I saw a couple of other things from OpenAI and Sam that were interesting. I think they, you know, said that they're kind of more mission-based and, you know, they're here, um, kind of building something different than Meta is, kind of what they've tried to paint. And I did see the quote where he talked about how, I think he said, there's more upside in OpenAI stock than Meta. That was obviously his opinion. But, you know, of course, he probably said that to the people that were thinking about leaving or had offers in hand and was trying to kind of leverage, you know, "OpenAI's going to be worth more down the road." But I do find the OpenAI framing of this to be a little interesting in terms of like "unethical" and, again, "distasteful" and "poaching" and "stealing," right? Because, you know, and maybe you find this ironic or not, I certainly do, because they have a kind of a unique and some would say borderline unethical, like, entity structure where they're like a nonprofit and they're private. And, you know, no one's got their hands clean of everything. But I did find that interesting to kind of say that when they are not paying any—from what I understand, again, private company, but from what I understand, they have a nonprofit status that lets them not pay taxes. And surely they're raking in, you know, billions of dollars in revenue, um, you figure. So I found that to be a little rich, but I don't know what your thoughts are on that.

There's been a lot of talk about OpenAI's structure and how it's like non-profit going for-profit, being sued by Elon for this whole setup of how they've structured the company. I don't even know where they're at today. Clearly, they're making tons of money. They're still private as far as I'm concerned, but I don't know about the non-profit arm versus the for-profit company structure. I think it has a lot of drama when it comes up, but there's so much other drama that that is kind of a backseat in terms of the AI headline news for the week. So yeah, a little interesting. I think OpenAI definitely wants to cast shame on the Meta recruiting approaches because they just probably can't compete, at least with real cash money, compared to a private valuation. But yeah, definitely a little bit sensational in terms of headlines of "coming into our house and stealing our people." I don't know. Good to get some headlines. I think that's what they want, to draw attention to make Meta look worse and therefore keep more of their talent. It's definitely right now a time if you're an AI employer, probably to check in with your employees, make sure they're valued, make sure they're happy. If not, they'll probably end up at Meta in the next three or four weeks.

Yeah. Hey, Meta, you know my number. I'm just kidding. But yeah, it's interesting. It's funny that I think Sam is the one that broke or that specifically mentioned the dollar amount of these offers, like the $100 million. I'm pretty sure that came from Sam on a podcast he was doing.

Yeah, that was like two months or three months before it all happened, and people said that was complete BS, and then it ended up being true.

And Meta kind of denied it. They said—I think there was a Q&A they did internally that Wired was reporting—that, you know, they said it's not like $100 million cash. They do, it's more complicated than that. So they're not outright denying it, but they're kind of, you know, they're not being as firm as what Sam had said. And it's interesting because a lot of what OpenAI has done, they do a lot of things publicly, I think strategically, and they say things publicly strategically. And it's interesting to think why—and I don't know, I don't really have any speculation—but why Sam would just come outright and say that, that they're losing employees to Meta for these fantastic offers. If you're someone getting that offer, you're probably pretty excited. It's interesting for him to disclose that. And I don't know if they're kind of playing the victim card or trying to get the sympathy vote. I don't know. But I'm sure there's a reason behind it. It's not just haphazard. So I'd be curious to watch.

It made headlines back then. However, people quickly dismissed it, saying that's ridiculous. And now $100 million is like the lower end. We've seen $300 million reportedly for an Apple lead researcher. So, yeah, it's crazy. It's definitely a pay I've never expected from these folks. I know this is a specialized skill set, but, man, the competition is heating up. The pay is heating up. This is only the start. I don't know where this is going to go. If AI proves to be the powerful, you know, tooling, coding agent that we're expecting it to be, I'm hoping Meta can make a large impact given their new kind of roster of super intelligence. But we'll have to wait and see. Really, you know, it takes time to build these cool things. It takes time to train. It takes time to even get the hardware to train these models, which sounds like they're building. So all that combined, I imagine 2026 would be a big year for them. And I've checked their stock, still at the $700 mark. Great for them. After all this news, I'd expect it to pop off a little bit. It sounds like the market reaction is, "Let's wait and see," or maybe this was already priced in a few days before. I don't know, but it wasn't as big of a reaction as I expected, given we just collected intelligent folks to run the super intelligence team.

Yeah, well, and they—I mean, it might be the recognition that they had to shell out a lot of money to get them, you know, and that will show up somewhere on the P&L. But then also, too, I think there's a recognition that Meta's a little bit behind. They're behind Google.

Yeah, definitely.

And so you can see this as a brilliant strategic move, and they're getting the top talent. But you can also see this—and again, OpenAI has said this—as kind of like a desperate move. And so you can kind of pick. You're probably in one of those two camps on how you think about it. So, yeah, we'll see if it all pays out for them, or plays out, however you want to say it.

We will see. I'm excited. Hopefully, they make cool stuff that we use.

Yeah. So we can jump into releases. It's been a crazy two weeks, things being released left and right. One of the biggest releases, I would say, in the past few days has been Kimi's K2 model. And so this is a big one for a whole bunch of reasons. One, it's open-source and open-weights. So this means that you could technically run it on your own computer. However, it's 1 trillion parameters, which is a large model. I think if you download it from Hugging Face, it's roughly 900 gigabytes. So making sure you can download and run that on your local Mac—no way, no chance. But the biggest thing is that it scores really high on benchmarks. So we've had Anthropic release their awesome, like, Claude 4 models, which are Sonnet and Opus, which do really, really well and are powering the craze of Claude Code. The big deal about this new model is that it also is really good at tool calling. So when we look at models today, there's, you know, Gemini 1.5 Flash, 1.5 Pro, there's OpenAI's GPT-4o, et cetera. And then there's Claude. And outside of that, there's not too much going on, but even within those three, I would say Gemini is not that good at tool calling. So Claude has built a really good model for tool calling, basically training the model to do that from the ground up. And I think this is the first model that has replicated the success of tool calling. So we're basically calling this new Kimi K2 model very "agentic." Like there's almost like a flavor profile for these new models. One is high intelligence and the other is agentic. And Claude Code has really shown that if you can make an agentic model, it can do incredible things. You can almost take it as, in the next evolution for OpenAI, Google, and Anthropic, would I rather have a highly successful agentic AI that maybe is not as smart, or a new AI model that comes out that's smarter but less successful with tool calls? And to me, it seems like we're getting to the point where that tool calling is the bread and butter. And if you're not scoring high in benchmarks that measure that, your model might not get adopted because all these models are getting plugged into Cursor, other AI IDEs. And the main factor for getting these experiences to work and work well is the reliability of tool calling. So this is a huge release—open-source, one trillion parameters, very large. It's a non-thinking model as well. So it doesn't do the whole thinking prologue that kind of started with DeepSeek, where you ask it a question, you see this internal monologue. And it still scores really, really high. So very, very exciting. I think the model itself probably won't be run that much since it's so large. But what the research shows, how they trained the model, all that will be used to kind of create the next generation of models after it. It's very, very similar to DeepSeek, where no one's really using DeepSeek today, but the fact that it was released open-source, open-weights, had kind of a revolution in scores where it kind of met the OpenAI family at the time, like in December. This is that same large moment of open-source, high intelligence with tool calling. So really, really exciting.

Well, and it's interesting. I think one thing that just immediately sprung to mind was I remember when DeepSeek came out, it was, "Oh, this is the OpenAI killer, the ChatGPT killer," because it's, you know, free to use and it's just as good, if not better, blah, blah, blah. But I agree that I haven't really heard much, um, you know, much usage out of DeepSeek, um, since that announcement, really. I'm sure people are using it here and there, but like not to the same level commercially that, um, the other ones are being used. But I totally echo, I think the tool calling thing is, you know, especially in—and I'll just say in my line of work, in accounting—because a lot of times we have to do something a certain way. We don't want that to be abstract. We've talked a lot about this in other podcast episodes where like, I want the tool, the AI, I want it to do the specific thing. I don't want you to interpret my instructions and do it a different way. And so having tool calling and having that be precise and having the tools be well-written and defined—that's where I think, again, for accounting, super useful. And I'd rather have that versus some brilliant PhD-level model that I don't need that level of intelligence. I need you to understand my query and then go figure out the right place to do that.

It's interesting. I was tinkering around with LangGraph, and, um, I had written a simple, you know, agent, we'll call it, that has two tools. I had one that was—it would sort, um, the letters or the characters in a word alphabetically. So if you took "zebra," it should do A, B, E, R, Z, you know what I mean? And so it would do that. And then the other tool call was like, "check if it's a palindrome." And so if you give it "racecar," it will say, "Yep, this is a palindrome." If you give it something else, it'll say, "No, this is not a palindrome." And what was interesting—and I was just tinkering with it, I didn't spend a ton of time with it—but it was interesting because depending on the model I was using, I think it just defaulted to 4o Mini, um, it would sometimes call the tool, and I could see in LangGraph, um, you know, the tool call and the response. But then I think I had switched it over to Gemini at some point, um, and it was doing those checks, or it was doing those things without calling the tool. It was just checking the palindrome. It wasn't calling the tool, um, you know, palindrome checker or, you know, alphabetical sorter. And, um, it was interesting 'cause I didn't really change like the system prompt at all. So to your point about like, you know, different models kind of interpreting tools differently and how they use tools, um, it's kind of, you got to find what works for you in your own situation, you know? And, um, again, having like the trace to kind of see that LangGraph, the agent was calling those tools versus not, was helpful because if I wasn't really looking at that, I wouldn't know, like, necessarily how the agent got to the answer it gave me. Um, which I think is important again, as people start building out more and more custom, you know, agents with AI, it's like, make sure you understand how the AI is getting that answer, not just what the answer is and taking that as gospel, you know, because, you know, we've talked about hallucinations a lot, but sometimes they'll give you the right answer, but how it got there, you know, you probably want to understand how it got there to some degree.

Yeah, and Gemini 1.5 Pro was the leading model for a short stint of time, kind of before these new agentic models came out with tool calling. So they have their time and place, and I think if you're not calling tools, you're okay with coming out with a better model. But I think right now, my guess is a lot of these leading AI companies are coming out with, you know, hitting a baseline of 98-99% tool calling accuracy first and then trying to increase that intelligence. And I have no background in making an advanced AI model, so that's probably a gross simplification. But again, I think if OpenAI came out with a 15% smarter model, but 5% worse tool calling, I probably would not pick it up. And I think that's kind of what we looked at. We can talk about Grok a little bit, but Grok-4 came out, hit the top of the leaderboard on a few intelligence evaluations, but they're just not that good at tool calling, so it doesn't pull me away from Claude Code and Anthropic models.

But we can just wrap on Kimi. A few other interesting details is that they have a modified MIT license. In that license is essentially MIT at the top, and at the bottom, they have a disclaimer saying if you use our AI model in commercial products with more than 100 million monthly active users or $20 million in monthly recurring revenue, then you have to say that you use our model. So this kind of exists from the DeepSeek land where DeepSeek was open-source, open-weights. However, their license made it explicit that I think if you had over 600 million monthly active users or X amount of dollars, you had to disclose it, or I think maybe you couldn't even use it. I can't remember the exact stipulations, but there's a clear divide in that if you're not a big corporation with lots of users or making lots of money, it's completely free and open to use. But once you cross that threshold where you use it on a large scale, then you either have to attribute it, which would drive a lot more recognition, or you can't use it at all, I think, in DeepSeek's case. So a very interesting license, and it clearly has a separation, a very clear-cut line of "can you use it or not, and these are the restrictions behind it."

Yeah, that's interesting. I don't know of anything—I'm speaking out of some ignorance here—but traditionally, when I think of open source, I think of like, you know, truly open source, like in terms of like, I don't know, FastAPI, Pandas, you know, like Laravel, right? I think it's open source. There's no limit on how much you use. It's not like if your Laravel project does $200 million a month, you got to start paying, you know, the Laravel foundation or whatever, you know. So that's interesting. It's kind of like a free... Is it enforceable? We don't really know. I think that was some of the discussion. Does it exist there? Who's going to enforce it? Who's going to self-report? Not really sure. But on top of that, it's a relatively cheap inference cost. So I think today the rough standard is like Anthropic is serving 1 million input tokens at, I think, $1. And then output for their same amount is like $15. Almost like a 1-to-15 ratio. And I think their pricing was like 20 cents per million on input, I think like $1.50 for output. So roughly like 10x cheaper on both sides. And it gets very close to Anthropic's performance. So really exciting—high intelligence, tool calling, open-source with a caveat, and hosted and cheap. It is a China model, so you have to be careful of what you send to it. But at the current status, pretty exciting. I think it'll be used to up-level the next models that come after it. Again, it doesn't have a reasoning mode. It doesn't do, I think, images, but it's kind of the groundwork for a next generation, next evolution of models. So you can try it out, I think, on kimi.chat. Let me check real quick. No, it's not that. I'll have to find it in the show notes. Or maybe it's kimi.com. Yeah, okay, kimi.com. If you go to that, you're presented with a chat window. Unfortunately, it's not as good as Claude and ChatGPT because those are very refined and polished chat interfaces. But if you want to give it a run, go to kimi.com, type in a query, and check it out.

Nice. Always something new.

Cool. Yeah. And then on top of that, we talked about Grok-4 a little bit. I won't go too deep into it, but it was released by X's AI team, called xAI. Released about a week ago. They had a live stream talking about it at night, and it hit top of the leaderboards on intelligence, which was a first for the xAI team. So hats off to them. But again, as we've talked about just minutes ago, not necessarily the number one, end-all, be-all of me adopting a model. If I can't use tool calling effectively, it kind of falls apart. So impressive results nonetheless, but I think I'll probably pass on that model. And their pricing, I think, is the same as Grok-3. So I think it's in the ballpark of like Anthropic's Sonnet and Claude models. Not as cheap as Kimi. The one callout I think from Grok-4 is they didn't do a lot of safety training. So when the model first came out a few days ago, there were a bunch of reports of it being able to instruct on how to make a nuke, how to make world-ending drugs, how to make a plague. Very scary things that I'm even scared to type in as like a quote-unquote safety test or a red team. I don't know what data goes on, but these people clearly have tried it and checked it, and they've posted screenshots that are redacted, essentially Grok-4 searching the internet, finding Wikipedia pages about chemicals, warfare, et cetera, and composing a pretty detailed answer of how you can do some nefarious things. So since then, they've come out and provided some safety restrictions, basically trying to identify if it's like self-harm, warfare, et cetera. And they've started blocking those, but it's kind of after the fact. Things that may have already been talked about have already been talked about. So a little bit of a miss there on their end. I think they don't care too much about safety, where other AI labs care a lot more about it.

Yeah, yeah, everyone's got their own little flavor of what they prioritize and build for versus, you know, others. I just had a—we're calling an audible here on the podcast real fast—I just had a... this is probably just like a, I don't know, an obvious idea, but kind of also like just sci-fi and and and dumb even. But you know, Neuralink, I think is what it's called, the little brain chip that goes inside people's brains that's being experimented with. You know, we've heard stories, and I think that's very experimental. Obviously, they're—I think they only have like five or six people that have gotten the chip, um, and they're really good at playing video games, you know, where they can like headshot people in a video game like pros because they have this chip in their brain. But like, it'd be really fascinating if they could like just upload, you know, one of these models to one of those chips and all of a sudden this person knows everything.

I don't know if that's the plan. I think that is the plan eventually, but they need to connect the brain-to-human interface a little bit better. I think you can use your brain to move things or move/control objects, but I don't know if you can talk to it, get an AI response, or how that all works. But getting a chip in there is the first step. Now unlocking it to do other things, hopefully they've thought of that—and I'm sure they have—uh, would be pretty cool.

Which like, it's obviously not that—it's impressive and crazy, you know, I shouldn't say it's not that cool—but like, you know, right now you can take out your phone and look up anything in the world and get an answer. And is that really that different than thinking it and then having the answer in your head? Um, but yeah, it just seems like those two, you know, timelines converge nicely of like AI and Neuralink or whatever. And now it's kind of like we're at this point where that chip can be plugged into the internet, so to speak, and dropped one of these models. Then that's pretty powerful.

If you think about it, they're both Elon companies. Grok-3, I think, was integrated within the Tesla vehicle. So Grok-4 is smarter, more capable. I wouldn't be surprised if Grok-4 makes it in the vehicle. But he has an impressive army of companies, all doing different things, all slightly related or could benefit off the other companies. So yeah, hopefully they have a nice synergy together and Neuralink can access some of these things. But it would be crazy to walk around with someone who got Grok-4 in their head, who got like PhD-level research across, you know, thousands of world-level topics that you could just walk by something and say, "How does that work?" And like, boom, immediate answer in your head, running off a chip or something, or a chip connected to your phone. I don't know. Pretty cool, though.

Here's the plan. And I'm going to say this and we can pin this for three to four years when it's come true. The strategy that Elon's going for is he's got his robotics company. I think it's called Optimus.

That's something like that. I didn't know about that one, but I'm not surprised.

Yeah, they got little humanoid robots, right? And so it's like iRobot. Very unimpressive right now. They did a demo and it's whatever. But that is something that he's working on. He's got, he's got the, um, SpaceX, right? So here's the strategy. Everyone's gonna want a Neuralink in their brain at some point because that's just what people do. They want to have access to all those things. They want to talk telekinetically with their buddies over the internet in their brain. I don't know. But they're all gonna get that chip and it's gonna be running Grok because that's his model, his baby. And then what's gonna happen is the humanoids, the robots—he's enough—not he, but like someone's going to flip a switch and all that brain that is, or the chip that's controlling our brain and Neuralink, is going to now basically force us to do work. Like technology. Like we're the Tesla fleet. We don't have free will.

Yeah, yeah. We don't have free will anymore. And we work for the humanoids and for like the Grok ecosystem to build out these rocket ships for SpaceX to go to Mars. And then we're going to go as a unified hive mind of half-human, half-robots going to Mars. I'm telling you right now. That's happening.

I love that idea. It sounds very, very scary. It would be very hard to convince billions of people to get a Neuralink. But with his track record, who knows what's possible? Say, "Hey, it's free. You can come in here. You'd be so much smarter. Promotions."

Yeah. Yeah. "Don't you want to make more money?" "Yeah, absolutely." And then they got you and they flip the switch and you can't take it out yourself. You don't know how to do that. You don't know how to do open brain surgery. So yeah, it's like a tattoo.

We should reach out to Black Mirror. I think it'd be great to have a fake Elon Musk with all these companies, and people really get it. They're like, "Wow, you know, Ben's vision is now on Netflix," and uh, it hits pretty hard. So yeah.

You heard it here first. Yeah, yeah, exactly. Exactly.

Cool. Well, audible success. I love your idea. Moving on to Claude Code. So this, basically two days ago, I think Claude came out, or Anthropic came out, with like Windows access. And I was in the airport yesterday flying home, and I was in an airport lounge. I was throwing Claude at everything because if you know me, I spent a few days away from the computer in downtime. I'll think, "Oh, you know, how will that work? How would that work?" So I got back to my computer finally, went to the lounge, opened it up and started, you know, Claude-coding like I normally do. And surprisingly enough, I had the banner that pops up when you use Claude Code saying you're approaching your limit. And I saw that sooner than normal. And I thought, "That's a little odd." You know, as someone who's used it since the beginning of June, which again, I'll pat myself on the back, I was early. And I was early not only to using it but knowing that it was going to be big. So great job to me. But more realistically, I knew that was early. And so I just pushed along. I think people have kind of assimilated it to a low-battery screenshot on an iPhone, where you kind of see this important piece of display and you kind of think, "Oh, that's interesting." So I sent a picture to Ben saying, "Oh, wow, like my limit feels pretty low today. I haven't done that much work." And then about an hour later, I get a full block from Claude. So I ask it to do something, and it says, "You've hit the limit." And I've never had that happen before in a whole month and a half of using Claude Code. And I pay for their top plan. I've never had it fully block me. So I didn't really get it. I tried switching models. So by default, I use Opus, which is their slower but more intelligent model. I went inside Claude Code and used their /model command to change it to Sonnet. Then I asked the AI agent, I said, "Continue exactly where you left off." And it said, "No, you've hit the limit." And I go, "Damn, I thought I paid for the unlimited plan," which is quoted at 20x the normal usage. And I had never even gotten that warning very often, maybe like twice or three times, and definitely within the past like two and a half weeks. So I texted Ben. I said, "Wow, I'm like completely locked out. I've hit my limits. I have a reset timer of, I think, at 10 p.m., it's unlocked." So I was blown away because we've talked about on the podcast, Claude Code is seriously the best coding AI agent and best AI agent in general because of the tool use and because Anthropic builds the tool and the model. They have a great feedback loop there. But it gets to that point where we've talked about people posting these screenshots, boasting about their usage. Everyone and their mother are signing up. Now we're at a point where Claude Code, you know, hopefully it doesn't get saturated. It's feeling saturated. It sucks, you know. As someone who was early to the game, uh, I'm now, you know, competing with all these others. I don't know what to think of it. It's not, not great.

All good things come to an end, Bradley. You know, we talked about on the podcast before that like the more that you promote how much you're just like juicing the model and juicing your usage, and you're not alone, you know. People have really kind of, you know, we talked about how, you know, "post your stats or don't post them." Yeah, I mean, um, but yeah, they must have been—I mean, yeah, it's interesting because you had the max model, which, or max plan, which means unlimited. "Unlimited" means unlimited, last I checked. So like, I don't know how that's even allowed, to be honest with you.

Yeah, there's like a star when you sign up, saying I think there's like some usage limits, which they do outline, but they have clearly shrunk that window without any communication. And Cursor had a giant pricing shift kerfuffle that kind of ruined people's trust. Anthropic with Claude Code, they're not at that level yet where I'm not like, "I paid for this, I'm not getting it." I felt like a tinge of that, where before I was free-roaming. I could do whatever I wanted, it felt like. I think what people have been doing now is there's been a layer of software on top of Claude Code to make it easier to manage multiple Claude Code instances. Basically, UIs that will take one single prompt and kind of like span it across three or four Claude Code sessions because there's some randomness when it generates the code with an LLM. So now I think people have really optimized multi-Claude Code usage and are just draining the pool of GPUs from others. And due to that, I think they're shortening the limits on Claude Max plans, even the top $200 plan, really to have more capacity for others. Because if I boot up Claude Code and I can't use it at all, that sucks. If I can have an hour or two of productive coding, that's great. What I'd really like is four or five hours, but I think I'm getting limited in that in-between area because other people are using it. So I kind of blame myself a little bit, maybe others who have built tools to use Claude Code more, which is a good thing, but there are only so many GPUs. It's not an infinite resource. Hardware is powering all this. So it's kind of in a weird spot. I love it. I don't like the recent changes. I don't know where they're going to go from here, given they just released Windows and that might spawn a whole new community of developers to pick it up and use it. So if Anthropic is listening, maybe make the Claude Max 20x plan that's $200 a little bit more unlimited. I don't really know where to put it. I mean, I think we're hoping we get better models and more effective tooling, but if I get locked out after what felt like a really productive session, right at the peak of me making changes, it kind of sucked. Honestly, it kind of sucked. And one other thing I saw recently was I think Claude Code signups were disabled. So you can clearly tell something was in the water. Usage is high, people love it, like not only me but lots of people on Twitter. What are they going to do? And Anthropic came out to confirm that they don't downgrade the model intelligence when it's under peak load. They say they don't alter anything. There have been rumors going around that like Claude is better on the weekdays or during the workday or on the weekends, whatever. Timing aside, I think people are really concerned that OpenAI and other model providers are secretly, behind the scenes, downgrading models just to save costs. I think they came out to confirm that they don't do that, which is great. But they still have a problem on their hands. I'm very curious how much money these power users like myself, or maybe even others who use it a lot more, how much money they're losing from that crew and how much money they're making off the crew that buys it for $200 and doesn't use it at all. Given the recent scaling back of limits, you'd be hard-pressed to think that these power users are really killing the budget, and it might be time for changes. Knock on wood, it doesn't affect me, but maybe the glory period of Claude Code may be coming to an end. I hate to say it.

I'm not sure. I have two thoughts on that. I think the first one is, if they came to you and said, "Hey, it's unlimited, unlimited, like really unlimited is $500 a month," would you still buy it?

Probably. That'd probably be my ceiling, maybe like $400, $500. I think it would actually devote me to using it more where $200, at least in my current setup, I'm like, I definitely get a lot of usage out of it. But $500 feels like a steep 2x pricing. I'd probably pay for it though if it was truly unlimited. Anything past that, I feel like you have to be a real business. Right now, I use it on the side. My side business isn't raking in the cash. So that's kind of my limit, I would say.

Yeah, yeah. And I think the other point that's interesting is, you know, Claude Code, and I would just say Claude more generally, is definitely less popular than—well, ChatGPT is the most popular in OpenAI, of course. But I don't know where Gemini's popularity stands. Obviously, they are making up ground, and they've made up a ton of ground. But in terms of just overall usage, Claude and ChatGPT, I feel like were the kind of the first two on the scene that really made a dent. But I guess where I'm going with this is I think Claude is probably mostly used by like engineers, whereas like ChatGPT and Gemini probably have a more broad use case. So I would wager that—and I'm just purely speculating, I don't have any details—but I would wager that the Claude users are less than ChatGPT, and I would even probably say Gemini. But the usage is more because engineers like yourself and others are just slamming that thing because it is the best model, and everyone agrees that that is the best model for coding. It's not even really contested. And so you probably have a scenario where you have fewer users than the other names maybe, but you have users that really use it because obviously you're writing codebases and the pricing is based on token output. And so if you're generating whole files and website architecture, then that's obviously a huge amount of tokens that it's generating and therefore expensive to run. So it's interesting to kind of compare it to those other more broad-based models that have a more general appeal, I guess.

Yeah, like I use ChatGPT for various questions and will use it maybe once a day at maximum. But Claude Code, when I get the steam and momentum and like focus time to do two to three hours, I can start a task on my mobile app, I can start a task on my web app. I can ask it to plan, which it goes, reads a bunch of code, and creates a plan. I've recently switched to this "dangerously skip permissions" flag which doesn't ask me if it should do anything, it just does it. So it's fully autonomous, thanks a lot, and it probably costs them a decent amount of money. And I think all this Twitter hype of Claude Code, which I've contributed slightly but I don't have a ton of followers—so if you're listening, go follow me on X—but there's been so much, I sell it but others have sold it way, way more. And I think developers who love something really put their mouth behind it. They'll say, "Hey, out of five tools," and in the AI race, all these tools are kind of interchangeable. So when Claude Code is here to stay, at least for the past two months, everyone's talking about it. Like you'll see it on your feed if you use it or not. But it's clear that when developers like something, they'll talk about it. And Anthropic has won clearly on the front of the agentic CLI, which we kind of talked about last podcast. But yeah, I'm a bit concerned to see what the future is. It's not a complete red flag; it's an orange flag in my book where we'll have to see if it regresses, where they buy more GPUs and I can use it more, or it gets worse. I have a shorter and shorter session time, less token output. So we'll have to see. I'll report back in the next pod how things look, so I'll definitely be using it. But just a little bit of a warning for those who have been listening to our podcast episodes. I'm still a supporter, officially. But things are changing, and I think this kind of exists for the ecosystem at large. Like we saw Cursor had a pricing issue, which they kind of incorrectly priced things. People got really confused. Part of that is I think these companies are trying to make more money, turn a profit, kind of finalize on pricing because it used to just be $20 or $200. I think we're getting to that age where you asked me $500. I can really see that becoming a reality, given all these things considered, where I'm getting a lot of value for $200. If they scale it up, now that I've seen the power, I would do it. If they told me $500 to start with, I'd say, "Oh, I don't know." So I think either they've been planning it or they're running into an issue, but changes are coming is my prediction in the next two to three months. Hopefully, those changes are for the better.

Yeah. And I'd be curious, I guess this is maybe a call to action for our listeners. What price would you—and I guess this is specific to Claude Code users—but for those that are using Claude Code, what price would you walk away from? Would you say it's too much? So Brad already said he's locked up for $500. But some people are using it and they're getting $5,000 worth of usage when they look at their stats. Would you pay that much? Or would you say, "No, I'm going to walk away and try Gemini CLI once it hits $800"? So again, listeners, let us know in the comments or shoot Brad or I a Twitter message because yeah, I'd be curious what people's kind of, um, ceiling is on that price, just given how much you get out of it right now. It's one of those things where it's like there's a limit to how much I'm willing to pay for this, but maybe I can do a little bit more if that means I really get a true unlimited. Um, again, I don't know how their "unlimited" is not unlimited, but you know, whatever, beside the point.

Yeah, and then, uh, we'll quickly call out that Amazon released Q, their kind of competitor to Cursor. One of their unique features from their product announcement was kind of spec-driven development. So as we've seen AI tools evolve over kind of the past six to twelve months, a lot of the, you know, prompt engineering has fallen away and now we're getting to this spec-driven development where we have this giant kind of PDF or product spec outline to say what the feature is, what the requirements are—essentially the job of a product manager to define a new feature at a software company. So I think their AI IDE has a built-in workflow for spec-driven development, as they've coined the phrase. I haven't used it. I think Amazon has a few tools and offerings in the AI IDE or agent space. So this is just another competitor in the mix. Yeah, we'll have to report back when we use it.

But to jump into our next one, it ties directly into Windsor. So we kind of had the Windsor-Cursor side-by-side battle. You know, these two companies in AI IDEs were literally head-to-head, uh, and what we had saw from Twitter was basically, you know, month for month it was, "I use Cursor, I would never touch Windsor." "Oh my god, I use Cursor and picked up Windsor, I would never go back to Cursor." Uh, so all these things considered, there's a huge kind of news, uh, the past week that the Windsor CEO and a few of their executives are being kind of acqui-hired or maybe just poached, as we've talked about, and joining Google, which is a big deal. And in this kind of midst of announcements, initially, they had kind of left their employees behind, everyone that was outside of this Google deal. And over the past 24 hours after that happened, a bunch of bad press came out saying their CEO did a bad job, their leadership should negotiate a better package for those left behind. And, you know, partly on Google where they're taking the Meta approach, which is, you know, "I'm not going to acquire the company, but I'm going to acquire the key stakeholders." And so this is a new trend in AI, you know, acquisitions. But thankfully, after that, in a span of news is that the rest of the Windsor company, for those who did not join Google or get hand-selected, are actually being acquired by Cognition Labs, which is the creator of Devon. So the kind of initial autonomous coding agent we've talked about, Devon, before. But yeah, a bit of crazy news is one is Windsor is kind of departing. Their key leadership folks are departing, which is a big change. Who knows how much they got paid from Google? Two is, wow, they left the rest of the crew behind, a huge mistake. And then I think three, they made up for that, which is great, but probably after a lot of backlash and communication. So a whole whirlwind there, and we're clearly getting into an interesting age where companies at big tech are clearly looking to poach people. And I think with all of these changes, I'm not sure Windsor is really here to stay. So maybe it's now just going to be Cursor and Q. I don't think it's a good look for them.

You know, it's interesting because I think probably in April or May, we had talked about on the podcast because it was news, and there's a TechCrunch article talking about how OpenAI was the one that was going to be buying Windsor. And that was pretty much a done deal. And I think the number that we had heard was like $3 billion was the offer that OpenAI was making to Windsor. Again, all reported, nothing confirmed. And then it kind of died. And I thought it was basically official. And I remember recording the podcast with you and you being like, "I don't think it's official." I'm like, "Well, this TechCrunch says it's done." So I don't know. And so we never really heard anything from that point on. And then, yeah, come to find out, of course, Google. And what Google did was they, like you said, acqui-hired. And they bought some licensing and some IP. And they brought over some key people. So the way that it made it sound—and it's a little bit complicated, and I don't profess to know all the details and ins and outs—but the way it made it sound was that basically Windsor, what was left of Windsor, was like the employees and like some limited set or limited features of Windsor. Because like, again, Google acquired not just the people, but they acquired some other kind of IP from the acquisition or acqui-hire. And what's interesting is, um, more in the articles that have kind of come out since, is that I guess OpenAI wasn't able to, you know, follow through with that acquisition, or it didn't pan out, I guess, because of Microsoft's involvement in OpenAI. You know, just for refreshers, Microsoft and OpenAI have this kind of weird, I think they call it a partnership. Um, and basically the reporting is, I think it's from Wired or TechCrunch. I'll link it. I'll find it and link it. It's from InfoWorld.com. But that Microsoft, when OpenAI was in plans to buy Windsor, Microsoft was apparently like, "Okay, great, we'll have that too." Because any technology or IP that you guys develop or acquire, that's kind of part of our partnership. And I guess OpenAI was like, "No, hold the phone. That's not how this works. This is ours if we acquire it." Because of course, Microsoft has its own IDE. They've, of course, always had, gosh, Visual Studio Code and Visual Studio. And they're a competitor in that space. And they're a direct competitor to Cursor, to Windsor. And so I guess OpenAI was like, "No, this is ours." And Microsoft's like, "Well, no, this is going to fold nicely into our ecosystem with, you know, Copilot and all that." So it's interesting to kind of hear that be the kind of stopping point with OpenAI. And more to come probably on that relationship because, you know, Microsoft had a pretty heavy hand in getting Sam Altman back into OpenAI when he was fired the first time, I think, last year. So Microsoft's kind of, you know, they're there, they're part of this journey, but they're kind of sometimes in the back burner or kind of behind the scenes until a big moment happens, like Sam getting fired or them looking to acquire this technology in Windsor and causing a little bit of a hurdle. So interesting. I'm sure there'll be more events with Microsoft and OpenAI in the future.

I think there are more details to come out about the deals, specifically about the folks joining Google and about Cognition acquiring Windsor. I think there's good data right now or good reporting, but I think as these things play out, we'll get a lot more insights. But I think one last note is Cognition kind of highlighted that Windsor had, I think, 350 business customers on Windsor. And I think a disclosed amount of money they were making either annually or monthly. So they had kind of acquired a growing business in their mind. So pretty cool. I mean, it kind of shows you that these companies are very, very important. But when these key leaders are selected, not everyone can make the cut. And I think you've got to be careful about the AI company that you join because this could happen to you. And it seems like it sucked. I think the employees were pissed off, and hopefully things are slightly better. But I'm sure just that feeling of being left behind, especially when in the AI age, people are working pretty hard to kind of win and win big, is kind of a bummer to see. But hopefully things are resolved and moving forward, less of this happens. But I think the key theme here is that if you're working at an AI lab, you know, and you're important, whether that's C-suite, researcher, et cetera, big packages are there for you at any of these big tech companies.

Yeah. Just real quick, because you mentioned it. TechCrunch back in April, they reported, according to a source, of course—Windsor was a private company—that its ARR, so annual recurring revenue, was $100 million, up from $40 million in February. So from February 2024 to April 2024, when this article was written, it went from $40 million in ARR to $100 million in ARR. That's what was reported. So kind of crazy.

Solid. Yeah. Yeah. Imagine being there during that time. Pretty crazy.

You know, one thing before we kind of move on and close out with our bookmarks, you know, we talked about Q, the Amazon, you know, IDE. You know, we talked about Claude. And we talked about like Gemini CLI and Codex. We talked about these CLIs and, you know, these companies, especially these foundational models like OpenAI and Gemini making Codex, making, you know, I think Google has its own. I think it's called Project IDX. We haven't really talked about IDX, but I think IDX is like an IDE that's kind of Google's version. So I think it's going to be really interesting, especially when Amazon came into that space and I saw that announcement, where these IDEs like Cursor, like Windsor, like Devin, even Claude Code—and I know I'm going to like boil your blood here in a second, but like—they don't have the infrastructure like Google and Amazon do, if we kind of just focus on them for a second, right? Because in the past when you would write software, and let's just talk about web apps, I guess, when you'd make a web app, you would write all your code, you'd have your codebase, and then you'd go deploy that on like a service, whether it was, you know, DigitalOcean or, you know, you certainly know this a thousand times better than I do. But, like, you know, you go deploy it somewhere, right? It seems like now with AI agents, you know, you can—and I'm a Google fanboy, I've made that very clear on this podcast—you know, you can develop your AI agent in Project IDX, you know, using Google's ADK, right, their Agent Development Kit, and then you can host it in Google Cloud Console all within simple CLI commands. And it just seems a bit more like a better experience, like a better packaged experience versus "I wrote my codebase in Cursor or PyCharm or whatever that one's called, right? PyCharm." I've written all this in this specific IDE. Then I got to go bring it over here to deploy it and do all this extra stuff. It just seems like, you know, in the case for Amazon, I can write these apps in Q, and then, of course, a lot of people use AWS for deploying their websites, right? And so, like, having that connection, that more integrated end-to-end, you know, I write the code in this IDE that Amazon's a part of, then I deploy it in their system that they own, it just seems like that's going to be the way to go. And I feel like Cursor and Claude—you know, Anthropic doesn't have that, so they're kind of in the bucket of Cursor, Windsor, in my opinion. They don't have that. And so I don't know. I'm curious what you think about that. Like, does that make a difference to you? Or do you think that's kind of irrelevant to the overall process of building something?

I think it depends on kind of the customer profile. So for me, I want high-intelligence coding agents. And Claude Code knocks that out of the park. I think those integrated solutions that you can code and deploy, that target audience is probably kind of like the low-code, no-code crew. I think that will gain a lot of money. That's a whole section of companies that have these employees and individuals who are trying to just build stuff. However, I think there's the other bucket where serious software shops that just really need complicated, intelligent reasoning models that can reason over a legacy, large codebase. I think those are a different category, and I care less that the code changes I make will make it easy to deploy. For a lot of these companies, that whole deployment is a solved problem, so we don't have to worry about that as much. So I think it's two separate categories. I think the excitement of making the barrier to deployment lower for these integrated tools, excellent. But for me personally, again, that's more or less a solved problem, not too complicated, although it takes time to understand. I think I would love to use whatever tool is the smartest, and that currently is kind of the Claude Code bucket. So I think they exist differently. I think with these easy-to-deploy solutions, they spend less time on making the most intelligent model, but more time on the end-to-end experience. But Lovable.dev, or like Lovable, the AI IDE, they kind of have that solution where you can build out like a React front end using like Next.js, using kind of key infrastructure pieces that they've hand-selected and trained their model on such that you kind of build out this templated website. The good news is that once you're done with that, it's easy to deploy. So it looks good, easy to deploy, et cetera. For most people, it makes a lot of sense. But for others that want more control, want to choose their hosting provider, all these things that are important to some, not important to others, that's where I see that divide. And I think the customer base is vastly different between those two groups.

Yeah, that's a fair point. I think Claude will get, or Anthropic will be acquired at some point. That's my hot take. By who? I'm going to say in the next 12 months.

I don't name names, Brad, but I don't know.

Put that on the bingo card.

I know. We've got to revisit our bingo card. Just for our audience, we haven't forgotten about our bingo card. Brad and I, we were going to talk about that actually this episode, but with all the announcements and talent stuff we talked about and releases, we just figured we didn't have time. So we're going to make that a priority next episode to revisit our bingo card and check in because I already looked at it a little bit as we were preparing for this episode and I was like, man, the cycle of news changes so fast that like, you know, I'll just sneak peek it. The Greenland acquisition was one of my bingo card items. It was obviously kind of a silly one, you know, but like the U.S. would take control indirectly or directly of Greenland. Kind of a fun one. But, you know, that has completely, completely moved away from the news. Even at the time we recorded that, that was like all the news. The VP and president were in Greenland at some point. And so it's funny looking back on that bingo card as if it was like years ago and it was really only like, you know, five months ago. But, you know, such is the way that news works these days. So anyways, we'll do that the next episode.

Next episode. That'd be great. Yeah.

Cool. All right. Let's dive into our bookmarks and wrap this up, Brad.

Yeah, I'll go. I'll go first. Mine's a little bit quicker because it's a bookmark, which means I haven't really gotten fully invested in it yet, but it's on my list to do. That's the main reason Twitter has the feature. Just in case anyone was confused.

It's, "I'll look at this later, but I'll never look at it until I'm on the pod."

Exactly. Um, and now I can't find it. Okay, yes, it is. So it's the news that Google—you know, I'm a fanboy—they released an MCP Toolbox. And I'm also an MCP fanboy, so if you haven't picked that up by now through episode 25, then you haven't been listening because I talk about MCP, I talk about Google quite a bit. Um, but they basically—the tweet is, um, and I'll link it, I don't know who this person is, I think it's just like one of those AI aggregator accounts, but that's what made it into my feed. So Google just released MCP Toolbox for databases to build AI agents with database access. So the line is, "Integrate your AI agent in less than 10 lines of Python code for built-in connection, pooling, and authentication with databases." So I'll just say this immediately struck a chord with me because in my line of work—and I'm sure this is true for lots of other, you know, um, I'm sure you would resonate with this to some degree—but like getting information from a database is fundamental. Like, it's like writing an email. Like, you have to do that to do something useful typically. And in accounting, you know, being able to connect an AI with like, uh, you know, our financials via a database, um, is very nice. And that's something that I think has been a hard challenge, especially for folks, um, traditional accounting folks that are not in a technical background, right? They don't work with Postgres. They don't work with MySQL. And so it's kind of a foreign language. But having an AI agent be able to connect to that database and help kind of have the user read off databases in more of a natural language than no strict SQL syntax, I think it's something that's going to be really potentially promising for accounting. And then also authentication. I've lamented how horrible—and this is probably a personal problem—I am at authentication in general, like OAuth. I've pounded my desk a few times trying to work with OAuth before. And so the fact that this is part of this package—again, I haven't looked at the package details itself, but they mentioned authentication in this. So I think it's promising if it can kind of solve those two things of getting an MCP tool to access databases and handle all the authentication that comes with that to make sure it's secure. Super exciting. So looking forward to getting in there.

Sounds cool. Definitely not short, but sounds cool. My bookmark, I promise it will be short. So Claude Code is excellent.

Nice. Hey, cheers. Well done. You got me. Bring me in, officer. Oh, man. All right, what's your bookmark? What do you got for me? I'm going to put a shot clock on you.

So Claude Code is excellent, and Kimi K2 just came out. We're seeing cheap prices and pretty good intelligence. So what this guy created is a "Claude-Code-Kimi-Grok," and to break that down is Groq, G-R-O-Q, is serving Kimi K2 for inference. So essentially what you can do is point Claude Code not at Anthropic's servers but at your own local server, which will then forward the request to Groq. So end-all be-all is that you're getting the Claude Code CLI and interface, but you're not paying the Claude Max $200 subscription. And hopefully, hopefully, you're getting decent performance with agentic coding via Claude Code. So yeah, check it out. He has a repo that you can clone. I think it's a simple Python script that boots up a server, and then you change a few environment variables and then boot up Claude, and then now you're basically using Kimi instead of the Anthropic models, hopefully for a cheaper price. So yeah, pretty cool.

Yeah, that is cool. And I'm just going to complain that there are two Groks in the AI world, one with a Q and one with a K. It's confusing. It's really annoying, and I had to look it up because I'm like, "Wait, but I've seen it..." That's not the first time I've looked that up, sad to admit. But yeah, that's cool. So Grok with a Q, not Grok with a K.

Yes, yes. Cool, alrighty. Well, that was a jam-packed episode, like always. And yeah, let us know what you guys think. Leave us a review, give us a comment. Again, takeaway: I'm really curious to see what price people would be willing to pay or what's the ceiling on what they'd be willing to pay for something like Claude Code to get the true max unlimited experience. The Rolls-Royce of token plans, if you were, or pricing plans.

Yeah. And additionally, check out kimi.com. So during the podcast, I asked it, "Tell me about the Breakeven Brothers podcast." Quite literally that query. And yeah, it found some great information, which is pretty cool. And it's all for free. You don't have to sign up. So check out the latest and greatest kind of open-source, open-weight model on kimi.com. And then give us a comment, you know, how much you would pay for Claude Code. I think mine would maybe be $500. Who knows? Hopefully, there are no price changes, but if you had to, what would that end up being? Assuming you use it and you get value out of it.

Yeah. Yeah. Pretty cool. All right. We'll wrap that up then.

Cool. Awesome, Brad. Until next time.

Until next time. See ya.

See ya.

Thank you for listening to the Breakeven Brothers podcast. If you enjoyed the episode, please leave us a five-star review on Spotify, Apple Podcasts, or wherever else you may be listening from. Also, be sure to subscribe to our show and YouTube channel so you never miss an episode. Thanks and take care.

All views and opinions by Bradley and Bennett are solely their own and unaffiliated with any external parties.