Breakeven Brothers

Yeah. Cool. All right. We are back with episode 26 of the Breakeven Brothers podcast. Brad, how's it going?

It's going OK. I spent like an hour on Claude, trying to fix this very, very frustrating React Native bug. And yeah, I don't know. I was talking to my wife. I was like, I hate React Native. When it hurts, it hurts a lot. But outside of that, doing pretty well. Getting back in the groove of things, travel has kind of wrapped up. And so now I'm just in a chill period, time to get some work done, get back on track. Hopefully, the month of August will be about getting back in the routine of things. So, excited to do that. How about you?

Yeah, I've been good. I've actually been cranking away on my favorite library, LangGraph, a bunch at my day job. So a lot of fun doing that. It's been challenging, but the right kind of challenging. It's like those challenges where it's hard and you get frustrated when you're in the moment. But then once you kind of solve something or you can finally see it work after all that work, it feels pretty good. So that's kind of where I'm at with that work. So, yeah. So it's been good.

Yeah. We need two sponsors. Anthropic, hit us up. And LangChain, LangGraph, whoever the parent company is, hit us up.

Harrison and Lance. I know their names.

Right, right. On a first-name basis.

Yeah. I've just watched so many of their videos. They're good. They're really good.

Yeah. Yeah, Lance. I've just watched more of Lance's videos. So not knocking Harrison, but Lance's videos are super good.

So speaking of good YouTube videos for AI, have you watched AI Jason?

Yeah, yeah, yeah, yeah. Yeah, I think so. He popped up on my feed like months ago and I watched one of his videos and I think I was like, "Oh, this is pretty good. But I don't know who this guy is. Like, I haven't seen him on Twitter." Like, I don't know. He just didn't really have the online presence that I expected. And then a few weeks ago, he popped up again. I think this time he was talking about how to use Claude Code. And then I watched it. I was like, "Damn, this guy's pretty good." And I don't really subscribe to many people on YouTube, but I thought his content was excellent. Like, he has 20- to 30-minute videos of pretty in-depth coding, AI, LLM stuff. And yeah, really good. If people aren't aware, AI Jason, I think, is the username. It's this Asian guy on YouTube. Pretty awesome content. Like, just straight to the point, no fluff.

Yeah, you know, I just looked it up and that's not who I thought it was. So now, okay, I'm bummed. I'll check him out. But the one I was thinking of, his name is actually Garrett Love. I don't know why I got them mixed up. But, um, cool channel. Got almost 20,000 subscribers, 228 videos. I liked his stuff because he was one of the first ones that really tapped into the Google ADK plus MCP, because I remember when Google announced the ADK, you know, MCP had been six months old by then, kind of around that time frame. And, um, he was the first one I saw with really clear videos on hooking up MCP within the Google ADK. So we just shouted out two cool YouTube channels: AI Jason and Garrett Love.

It's hard. There's so much out there that I think it's hard to sift through. Like, quite honestly, I want good AI content. I go to my YouTube page, and I just feel like all these are built to be attention-grabbing. And for me, I'm trying to sift through the garbage, and I'm sure there are good ones that I've probably dismissed. But yeah, I feel like, especially when learning programming in high school, I heavily relied on video resources, and finding someone who tailors content just the way you like it is pretty unique. And I think with the age of AI, people are just pumping content out. So if anyone has any good AI creators, teachers, etc., please drop it in the comments. I would love to find better resources to learn content because sometimes reading X or other blog posts, I'm like, sometimes I kind of prefer the video content for various topics.

Yeah, yeah. And I think I'm with you. You know, the majority of the channels that get really popular or the videos that get really popular are not novel. They're always doing the same thing, like building a Slack bot, an AI girlfriend, or something like that. You know what I mean? But like, you find those videos where it's like, I was specifically looking for this issue and this person did it, you know? So it's cool when you find those videos.

Yeah, it's like someone finally understands what I need, you know, and they're talking to me.

Yeah. Hopefully, long-term, we can make some content that exists outside of the podcast, maybe AI learning, coding. For me, I know Ben and I have talked about it for some time, like, I don't know, just putting more video content out there because this is our first foray into that. It's fun, but it's a lot of work. I can't speak too soon, but it's a good avenue.

Cool. Well, so let's dive into this episode. When we ended the last episode, we talked about how we needed to recap our bingo card. It had been basically halfway through the year. Now we're a little bit more than halfway. And so we're going to check in on our bingo card. We're going to make it a pretty lightning round, I would say. Did this happen yet? Yes or no? And maybe add a sentence or two for some context or for some other important tidbits. But we're going to just kind of dive right in. So Brad, I have my bingo card up. You go first and then it's mine after. And so the first one that you had is the crackdown on drones because they are so powerful. So have you heard anything on that?

No, it's just funny. I feel like I'm looking at old code. Sometimes I write code that I thought was excellent, and I look back on it like, "What was I thinking?" And I think this first box, as you're reading it, I pulled it up on my other tab. I thought, "What is that?" So, yeah, a crackdown on drones since they're so powerful (UFOs). Not sure. I don't think anything's happened there. I think we should move right along. I think there might be better ones.

This is going to expose us a little bit, too, for how much we are following the news, I think.

I think when you made that one, it was when all those, like, New Jersey drone things were going on.

Yeah, I guess probably. There was some meat to it at the time. And I think there's some regulation around drones, I think. Again, I'm not an expert. I know if you have a drone, there are certain areas that are no-fly zones that I think have existed for a long time. But yeah, I don't know about anything new. I guess maybe after this episode, for another recap, we can double-check our facts here and make sure nothing has actually changed.

We don't do that. We're not journalists, Brad. We're just entertainers.

Cool. Mine was the Ovechkin-Gretzky goal record, and he did accomplish that. So super cool. Yeah, I remember watching that game. And we had watched like the last three games where he was going to definitely do it. And he finally got it. So, a score one for Benito.

Nice. I'm looking at, I think, so we did every other, right? Okay. So I think the next one is a $1 billion exit for a solo AI founder. I think that's one of those rooted in doing more with fewer people and AI company valuations just going through the roof. And so sadly, I don't think this has happened yet, although these Meta offers are pushing the limits of how much people are getting offered. And I think I'm going to go through Twitter messages and bookmarks. I swear I saw a supposed $1 billion offer somewhere. Okay, I found it from my Twitter DMs. Supposedly, on July 19th, from @GrowingDaniel, Meta offered someone $1 billion for four years of work. And supposedly, an update after that was a $1.25 billion offer for four years is the new highest this guy's ever seen. And that tweet has 3.1 million views, 500 comments, 20k likes. And then the comment says the original poster basically said the person said no, by the way. So I think I get like half a point where I didn't expect Zuckerberg would be throwing around oodles of cash, but I think it emphasizes the trend that we were trying to express, or at least I was trying to express, that there's a lot of value in AI. $1 billion for one person, whether that's joining Meta or exiting their own company. I don't know. What do you think? Half a point, maybe?

I'll give you half a point. Yeah. If I got a billion-dollar compensation package, I would have like 10 legal teams read through that letter. And then if everything was good and I signed on, I would, like, triple-guard my bank accounts and be like, you know, because just the amount of pressure that you are under to be signed for a billion dollars, like, that's crazy.

Yeah. It's unfathomable.

Yeah. U.S. crypto coin. I don't think this... I should know more about this, but I, again, have been putting my attention elsewhere. But I don't think this has happened. I do think there's been some favorable legislation, I want to say, or executive orders. Don't quote me and don't kill me if I'm wrong. But I do think there's been some traction in the crypto space with this administration. But I don't think—and I feel fairly confident—there's not a U.S. crypto coin. Just that, you know, I think there is a bit more of a friendly attitude towards cryptocurrency in the U.S. and in the White House. So I'm not going to give myself a point for that, but maybe in the second half of the year, there'll be something more official.

Okay. Okay. Mine is gaming Olympics. I think what I meant there is like gaming being added—e-sports, probably in some flavor—to the Olympic Games. I was just searching, "what is gaming Olympics?" and it turns out, I guess there's an International Olympic Committee organizing the first-ever Olympic E-sports Games, to be held in Saudi Arabia in 2027. So I don't think I get a point for that. Unfortunately, it's not being added yet. But I assume if, you know, whatever this committee is doing, it sounds pretty official. And maybe it'll come in the next few years. Maybe not this Olympics, but the next one. Who knows? It would be cool to see, I think.

Yeah, I want to try and go to the Olympics. They're in L.A., obviously, for 2028. Gosh. So, yeah, we're going to try and make it there.

Okay, my next one is, and I looked this one up because I was so proud of being right on this, and Brad's going to give me a point, is that JavaScript usage will go down year over year. You know, when you were talking about React Native, I was like, you know, that's going to be a thing of the past one day, Brad, because JavaScript sucks. I'm just going to say it outright. And the reason I'm going to say that usage has gone down year over year is when I look at the 2024 Stack Overflow Developer Survey, for all respondents, the most popular programming language, JavaScript is the most popular. I'm not knocking that it is popular, sadly, but it has 62% popularity, and Python was in third. Now, this isn't a contest between JavaScript and Python, but just keep those numbers in mind. If I compare those numbers to 2020, the most popular was JavaScript, which was at 65%, and Python was fourth at 48%. So what you're seeing in the trend line is JavaScript's going down as a percentage of popularity. Python's going up. Makes total sense. AI is all basically powered by Python, by and large. JavaScript sucks. It's bloated. No one likes it. And so, yeah, I'm going to give myself a point for that. And haters are going to hate and DM me. I don't care.

We'll allow it. That's OK. TypeScript is here to stay. That's for sure. My second one, or I guess the one after that, is Neuralink will be more widespread with a smarter model, higher IQ. So, again, one of those where when I look at it, I'm not really sure what I was thinking, but I did look it up. And I think from the news article I'd seen, essentially a Neuralink patient got it installed last November, November 2024. And then there's an article written in May 2025 describing this person using the Grok model, XAI's Grok model, essentially understanding things, writing replies, using the Grok model, quote-unquote, "in his head." So I'm not going to read the full article because it's a little long. But I think my train of thought was that Neuralink is an interface to get things done. These AI models are pretty damn good these days across all the top-tier providers. Therefore, I don't know, like, does this person have a higher IQ than an average person? I don't think it's proven from the article, but maybe a half a point based on the trajectory. And, you know, someone that has a Neuralink is actually using an AI model, which I think is pretty cool. What do you think?

Yeah, I don't know enough about the IQ... like, how do you... I don't even know my own IQ, to be honest with you. So I don't know enough about it. But I guess you have to differentiate between being smarter... it's just like having more information that you can go look up versus, like, intuition and logic and reasoning. You know, I don't know. I don't know where I stand on that because I just feel like... yeah, I'll give you half a point. Yeah. But it is a good question.

Yeah. Um, okay, my next one, I'm going to speed through these. Robot cop-dogs, absolutely getting a point. I can see that during the L.A. riots in June or, like, the whole ICE riots, there supposedly were National Guard robot dogs being deployed. If you look at Boston Dynamics, they've got their dogs pulling trucks. And then there's a tweet out of China on April 12th showing a Chinese robot war dog with a stabilized turret carrying a belt-fed machine gun, which looks like one of those Boston Dynamics dogs. So you can't believe everything you see on Twitter, but I definitely saw the videos and articles of them during the L.A. riots back in June. So they're here, and I'm going to give myself a point for that.

Sounds insane. You reading off "dog with gun," whatever that whole phrase was. That was a lot.

Robot cop-dogs that have guns attached. I mean, you've seen videos of people, like, you know, image tracking, right? Someone will hold a sign up, and they'll have a webcam, a camera, and they'll move the sign...

Oh yeah.

...and the webcam follows it. You know, some lunatic is, or some army is, and they've put a gun on that. And just... yeah, I've seen too much of that. Yeah, it's horrifying. The *Black Mirror* "Metalhead" episode, that's the whole plot. So there you go.

Nice. Plus one to you. My next one is a virtual office with headsets. I actually think this one is a pretty quick no. I think if anything, from at least the Bay Area's tech scene, for a lot of tech companies, they're bringing people back to the office for more in-person collaboration. So I don't know, this one was kind of a devil's advocate, I think, if I remember correctly. And it turns out, I think more in-person collaboration, even more in-person interviews, I've heard on the grapevine. So maybe, you know, if you're looking for your next job in the Bay Area, you might have to take an in-person interview to avoid the drama of the AI coding tools. So for that one, definitely a zero.

The next one will be really fast. Homemade salsa. Didn't make it. My garden, my backyard, just died. It just died. It just did. So it didn't work out. We're going to try again next time. Okay, you're up.

Breakeven Brothers merch. So I've hinted at this one for many months now and it's still sitting in my pantry. It's a one-of-a-kind mug that will probably go on sale whenever I find the time, but it's not up yet. So maybe half a point.

Okay. AI acquisition is on the next one on the bingo card. I don't remember really what I was trying to say here. Maybe one of the models being acquired, you know, and Brad and I have talked, we'll talk about Anthropic here later in this episode. But I think I was talking about one of the major providers, one of the major LLMs acquiring each other. So I don't think that's happened yet, of course. But I do think that there'll be... I still think there'll be one relatively soon. So I'm holding out for that one.

OK, my next one was car-free, master-planned city. I think I saw this on YouTube. There's some city in, maybe, Utah where they're just building a new city from the ground up with a really cool YouTube series. I have no clue what the progress is or when it's going to be done. So for that reason, I'll give a zero and we can move on from that one.

Okay. Yeah, I think as these go later on, I think we sort of ran out of ideas, but that's okay.

Greenland partial or whole acquisition. That hasn't happened. When we made this list, that was very much in the news, and now it's very much not. And so, yeah, no, nothing on Greenland. Moving on.

Okay. Mine was U.S. Wellness Day, at least one. So this is inspiration from tech companies in the Bay Area giving, like, wellness days, days off for people to recharge, rest, et cetera, kind of on the theme of a four-day workweek. There have been no wellness days as far as I'm concerned. So that one is lost.

Daylight saving time abolished, hasn't happened yet. Moving on.

Okay. AI judge in court. I don't think it's happened, but there's been a ton of drama about lawyers using AI in court. So a little bit of a different twist. So I give myself a zero for the judge, maybe a pity point or a brownie point for a bunch of news articles coming out about lawyers using AI and being called out for it in various ways, which we won't go into now, but kind of interesting. So overall, zero, but, uh, interesting theme.

Yeah, and one quick thing on that, I think Sam Altman just came out and said that the conversations that you have in ChatGPT are not, like, attorney-client privileged or something like that. Right? I don't know, big news, honestly. Big news, surprised me. I don't know a ton about it, I'm not a lawyer, but, uh, I don't know, just had some thought there would be some protection, but now you know.

Yeah. Small business accounting credential. Hasn't happened. Okay, moving on.

Uh, blockchain app store on Solana or Ethereum. Haven't caught up with blockchain technology, wouldn't know. Moving on.

Yeah. Database skills for accountants over Excel. This one is kind of a squishy one. I don't know if I can really pinpoint like, "This has happened," or not. Uh, but I would say, I'm not going to give myself a point, but I'm just going to shout out that recently, I think it was in July or like the last week of July here, there's a couple of folks that are very popular and well-known in the accounting community: Jason Staats, Chad Davis, Twyla Verhelst—sorry if I'm mispronouncing that. And there's one other person, I'm sorry. He started this company called Aider out of New Zealand. I can't remember his name. But anyways, they went on this kind of tour and they talked at, I think, like six or seven different conferences over the span of like two weeks, just talking about AI and building things in AI in the accounting space. So I can't give myself a point for that as far as database skills over Excel, but just maybe some news I would shout out that the tools and the technology are getting out there more and more. So yeah, cool to see. It's a good area.

Cool. My next one was catastrophic AI code in open source. So what could go wrong when Claude Code is running amok? It hasn't happened yet. Surprisingly, Claude Code has been used by everybody now and, surprisingly, it hasn't happened. That's a zero and a surprise.

Cool. LLM creating an app store to moat. I don't think we've seen anyone create an app store. Like, ChatGPT does not have an app store, you know, Anthropic, blah, blah, blah. So, nope.

I was going to think maybe Connectors would be a similar setup there, but for that exact one, no, as is.

Yeah. So, mine is a quantum compute programming language. We were hoping with AI models, somebody, namely probably Google and their whole research center, would come out with some quantum compute chip, and we could program and do cool stuff with it. Hasn't happened. A lot of hype about it. Moving on.

Cool. My last one, AI-enabled company making inroads on QBO domination. Hasn't happened yet. I think things are bubbling under the surface, so to speak. Xero has been doing a lot of really cool stuff. They had MCP first, like when it was first coming out. Their API, in my opinion, is far better than QuickBooks'. QuickBooks just introduced these, like, ridiculous pricing tiers—ridiculous in my opinion and in the opinion of many others. So there are things bubbling, but, you know, QBO still dominates the U.S. market. So no point.

Uh, I'm laughing here looking at my own thing, but Split My Expenses mobile app on iOS and Android. And, uh, not out yet. Working on it. It was one of those nights, I mean, going back to the intro, one of those nights I was like, "You know what, I could watch a YouTube video of, like, Ryan Trahan's series on 50 states in 50 days"—I don't know if you've heard about it, but kind of an entertaining YouTube series—"or I could program my React Native app." I thought, "You know, I'm going to power through, work on the React Native app." And I just fought with it for an hour, just pulling my hair out. And so long story short, it's not out. I get a zero. The good news is that the momentum is picking up and we're getting there. Although, damn, if I don't get frustrated on certain days. For React Native, it's definitely not all that it's cracked up to be in the good ways. But there are some things I do like about it. So it's not terrible. But sometimes it's just nice to be comfortable in the Apple stack that I've been productive with for my entire career. So yeah, long story short, zero. And I think with that, let's do a tally real quick based on how many you got and I got. So let's see, I'm going to count mine, you count yours, how about that?

Yeah, I think I got four.

I think you gave me one for salsa, but I wouldn't count that myself.

I don't think I got one for salsa. Oh, wait, three. Sorry, I have three. Well, I have two. Geez. Okay, that's why you want to review.

He saw his answers and was like, "Let's bring it up."

And the two you have are from two half points. Geez. Wait, do you have two or one?

I have two half points.

Because you're giving yourself the billion-dollar exit for the solo founder.

Yeah, I feel like that's in line. Yeah, yeah. We'll give it to you. We'll give it to you. Close enough. Close enough.

Yeah. Cool. All right. Well, that's the bingo card for halfway through, or a little bit more than halfway through, 2025. Good checking point. Yeah, we'll revisit at the end of the year. Again, it's kind of funny to look back on this and be like, "What were we thinking on some of these?" But you know what? It was fun.

Cool. All right. Well, with that, let's dive back into the main juice of this episode. One of the things we wanted to talk about was OpenAI Agents. They were released, I want to say, almost two weeks ago now. And it's funny, we recorded our podcast episode, and I think the following day they released this. So we didn't get to talk about it on last week's or the last episode. But a relatively hyped announcement, I would say. But I think when you peel back the layers, it's not overwhelmingly life-changing. So just as a quick recap, OpenAI Agents basically connected their Operator, which is the more autonomous, task-based agent that you could have used if you had the $200-per-month plan, and they connected that with Deep Research, which, you know, lets you do really expansive, in-depth research on the web for certain topics. And so they basically put those two things together. And now you can give this agent a task and supposedly it can go out and autonomously do this thing in the background. I think in all the hype videos that they released, they showed someone prompting it and then going, getting a coffee, and coming back, and the work was done. So I think the whole topic of autonomous work and having work be done, you know, while you sleep, right, is super appealing to people. But I don't know, what's your thought on just, I guess, the initial announcement and release? And kind of, you know, I'm not sure if you had the chance to really experiment with it yet. But like, just what was your take when you saw that announcement and what it was able to do?

I am subscribed to OpenAI's tweets on X. So anything they post, I get an immediate push notification. I saw they had talked about an announcement for the next day. I go on Twitter. I see people leaking it already. So I was a little bit disappointed that they didn't have better guardrails on that. But first thoughts: "agent" is so overloaded. Like, them naming it Agent, I think, kind of sucks for SEO terms. They have an existing agents SDK that's not related to the product. Long history of OpenAI choosing bad names for products and frameworks and tools that all just kind of have the same thing, like with Codex. But I think my reaction is, one, I haven't used it. But two, there's Deep Research, which I've used maybe four or five times in my entire life, which isn't a lot, but isn't a little. And then Operator, which I've never used because I just heard it was terrible. And so I think this release, I'm not super stoked about it, not super excited. I usually rely on Twitter to give me initial posts. I didn't see much at all. I think across my entire feed of people that I trust, value, and curate, I don't think I saw anything super exciting. Yeah. So I took a look at the announcement, saw Agent, I was like, "Yeah, I'm not going to play with that." I think it initially came out for their highest tier at $200 a month. I wasn't paying for that. It came out like a week or two later, like last week, for the $20-a-month plan. Didn't touch it, didn't care. I moved on with my life.

So I think it's cool technology, I just don't think it's there yet. And if I have one more thing to add, it's that I had recently talked to a friend who visited me, and he was telling me about Perplexity's Comet browser, which essentially is the Operator. So, you know, it controls your browser, but it happens locally on your own browser with your own login stuff. I think the big friction point here for all these autonomous browser tools is that it's not using your browser, your login credentials. Like, if you have to go to Yelp.com and book a reservation, you have to log in with your Yelp account first. Like, it sucks. I feel like that's a big pain point. And I think—I haven't used it yet—but I think the Comet browser from Perplexity, which is a new product they're releasing, is a bit closer to what you'd expect for automating tools, where you can open up this Comet browser locally, ask it, you know, "Can you go book a reservation?" I'm already logged into Yelp, so I don't need to do anything. It clicks around, does the job. And so, yeah, on OpenAI's announcement, I haven't used it, it doesn't seem very useful, it seems way too tied to Operator, which is like a remote browser, which feels clunky. So I don't know. Don't love it.

Yeah, our friend, Justin Anderson, he did a lot of tests and posted on LinkedIn and actually had a lot of really valuable insights. So I think, you know, and judging from what he had posted on LinkedIn regarding Agent, I think it was pretty positive, but still had some glaring issues. I think one of the issues that he called out in a couple of his different posts working with it was it seems to be... and I guess we haven't touched on this yet, but like, Connectors was something that also was kind of introduced with Agents, or at least they talked about it with the release of Agent. And Connectors are essentially where you can connect your, you know, your ChatGPT agent to your Google Drive or to your SharePoint, to these other apps that you use on a day-to-day basis. And so what Justin kind of highlighted was it does a decent job if you have a nice, clean file structure and a consistent, like, naming convention and stuff like that. But if you have anything that doesn't resemble cleanliness and orderliness in your file structures or in these directories that you're asking it to go search through, it's just going to do a bad job. And which is, you know... and I think a lot of times in enterprises and in businesses, you know, you like to think that everything is nice and ordered and organized, but rarely is that always the case. And so that was something that I found really interesting was like, you know, it doesn't have the intuition, the ability to kind of take all the different inputs it's receiving, all the different feedback—the files and the file names and the file size—and pick the right one. It kind of just, you know, if it's disorganized in the directories that it's looking at, you're going to get a disorganized answer. Um, so that was something that was really interesting.

The other thing that he had mentioned, too, that was worth shouting out was that Operator got faster, which is pretty cool. So I think he did a demo or did a post, I think six months ago, when Operator was kind of... or when it was just Operator, and it took like 15 minutes to analyze like three PDFs, I think is kind of what he had said. And on the recent Agent, I think it took like five minutes to analyze like 100 PDFs, something to that extent. Don't quote me on the exacts, but it was way faster, which is pretty cool. But yeah, I mean, by and large, I think, you know, it's something that can help people do their job, but it's not going to replace any core piece of your job, you know, because you're not going to be able to... I think I saw a post that likened it to an intern. Like, you're not going to be able to take that and run with it without fact-checking it and double-checking it, you know? Which is still... it's still probably getting there. And, like, I don't think anyone doubts that there will be a point where it can do all these things really well. But, you know, is it there yet? It doesn't seem like it, at least with that product. Now, on the Perplexity browser, on the Comet browser, it's funny you mention that because I do have, um, I'm going to save it for my bookmark, but there was an accounting podcast that actually live-tested the Perplexity Comet browser in Xero, and I'll kind of mention there what they found. But, um, I did see that product as well. So, pretty interesting.

Yeah, I think overall it just doesn't feel like it's there yet. I think the main consensus is "promising." But if it takes that much effort and friction, it's just not right. I think Claude Code is a good example of something that's like a general-purpose tool. It can do a lot of things. You can be creative with it. So it gives you the ability to do lots of basic things, and it kind of composes those together to have a robust execution plan, access to things—like, the world is your oyster. But I think with Agent, or OpenAI's Agent, you have to connect all these things, do all this one-time setup, and then it works. And if you invest all that time, I think it could pay off. For Claude Code, you don't have to invest as much time. You can be more open-ended about your questions. It can go do things that it can figure out on its own. But I think OpenAI Agent with Connectors and Deep Research and Operator just needs a lot of hand-holding up front. And I don't think I love that today where, yeah, again, if I had my own browser and it imported my cookies and it had me all logged in, I'd give it a whirl immediately. Even for OpenAI's CodeX tool, which can do coding on the go, it can look at my Split My Expenses website and add features to it. I can type to the ChatGPT app on mobile, but it takes like 45 minutes of setup where you have to go describe your entire development environment to CodeX. It then replicates that, and then you're good to go. I literally could not spend the 45 minutes just to set that up because I was like, "Is it worth it?" Like, I invest in a lot of AI tools. I think this falls in that bucket of my hunch is that it's not worth it, therefore I don't spend the time. I haven't spent that much time researching and seeing what people say because I haven't seen much. So my assumption, given all that, is it's not worth it today. So yeah, maybe just a little tidbit there. But if you put in the effort, maybe. But for me, I don't think it was worth it, so I didn't put in the effort.

Yeah, and I think it's, you know, could it be useful for, like, you know, getting some travel ideas and having it kind of suggest, you know, do some research in the background while you go do other things? Yeah, for sure. I think when you're using it, and I think this is true just for general AI, is, you know, for mission-critical tasks where you already know that you have a set of steps that you have to go do, like, in order to do task A, I need to go do A, B, C, and D to complete that task, I just don't know why you would use AI to do that in the first place. Like, if you need a direct, always-the-same-steps path, why even bring in AI when there's a chance that it can hallucinate and how it gets to the answer it gets to can change if you run it the next time? You know? And so I think people, you know, over time will come to understand, like, there'll be certain things that you'll want AI to do. And again, these kind of non-critical, maybe research-type questions. But on the types of things where it's like, "Oh, create me an invoice in Xero for this PDF I'm going to send you," I don't think you're going to want to use AI except for very limited pieces of that process. You know what I mean? Like, as far as... and that's what the podcast I'd mentioned earlier, what they did with the Comet browser, was they said, "Take this PDF, back the values, and enter it into Xero as an invoice." And, you know, long story short is it really struggled because I think you need to be really thoughtful about where you use AI. And I think Lance or Harrison from LangGraph said it in one of their videos—and these are people that their whole livelihood, their whole product is using AI, you know, and having people use their AI tools—you know, they said, "You should try not to use it if you don't need to." You know? And use structured programming and use deterministic workflows. And so I think at some point, some of the use cases I saw where people were trying to get the ChatGPT Agent to do, it was like, "Well, why would you do that? Why wouldn't you just make an API call instead? And why wouldn't you use an actual workflow tool, like N8N or something like that?" You wouldn't want to just give open-ended instructions for something as critical and as, you know, rinse and repeatable. You know what I mean?

I even think like there are tools like FireCrawl AI which allow AI agents to scrape the web, parse JavaScript, do all this more-or-less complicated "how to interact with JavaScript-heavy sites." And I think you can piece that together with your own integrations. Like, I could connect to Google Drive with my own Python script, I could connect to FireCrawl, which does web browsing, and like, "Hey, I have an open-source version of the Operator tool." Even Deep Research is now available through the API. They bundled it up into a product, but I don't think the raw capabilities are significant enough. Like, I love the new model releases because they bring something new. But I think this was, I don't know, it just wasn't as shiny and as exciting as I had hoped it would be. But not all things are for me. Not all things are exciting. I think it's still an impressive launch, just not what I would use.

Yeah. And like the last thing, and then we can move on to our next thing is, you know, for the example I had, or excuse me, for the example I was talking about at the beginning of this episode that I was busy with recently at work is, like, you know, I'm only using AI to, one, get the user's date, you know, because sometimes date-time formatting can be so just, I don't know, tricky to work with. If you're just trying to think about, you know, when someone says "last week," what does "last week" mean? You know, "today," that's an easy one. But so basically having the AI kind of parse out what the user intends when they say the dates that they want to run this analysis for. But then after that, it's all, you know, API calls and SQL queries that just get the data. It's, you know, just regular Python functions that analyze the data. And then basically at the very end, we give that output to an AI to kind of say, "Summarize this based on these specific data points." And so, like, you know, I wouldn't want the AI to do the analysis. I wouldn't want the AI to go fetch the queries. I think that's a lot of, you know, there's no need for that. And I think you're honestly putting a bunch of context into the AI, too, that you really don't need, and you're ultimately getting charged for, most likely. So that was just kind of my soapbox. But still cool, you know, we're getting there with Agent. And so I'm sure there'll be more releases from others, kind of following suit. You know, I'm sure Google has their own kind of version cooking up to compete with OpenAI's release.

Yeah, you got to know where to draw the line between programming and automations and then, like, AI-assisted, agentic automations. I think people are still figuring that out and toying with it. So yeah, cool release. But a big, hot topic this week has been Claude Code. So we've been beating this drum since May and I've been talking about it forever. And it's the best, end-all coding agent on the market. And recently, there's been quite a few updates. So today is July 28th, when we're recording the podcast. About 10 days ago, I got an email from Anthropic mentioning that they're upgrading Claude's infrastructure. This date effective for this upgrade was August 19th, so they're planning for it. But essentially, they sent an email out to subscribers saying, "Hey, we're adding multiple processors in other regions. We have to notify you because of some laws about U.S. processing, blah, blah, blah, blah, blah." Long story short is that clearly they are hitting insane demand because right around the same time, I think it was, you know, July 19th is when I got this email, was if you went to the Anthropic AI subreddit, you could just see boatloads of people saying, "What the heck is going on? You know, it's not producing the same results it was earlier in the month. And I'm hitting limits extremely fast." And our last podcast episode, same thing. I booted up a Claude session, ran it for maybe three hours, and I hit my limit. And I thought, "What the heck is going on?"

So since about 10 days ago, there have been rumors, you know, what's going to happen? There's clearly something in the midst. And today, before we recorded the podcast, Anthropic sent another email titled, "Important Updates to Your Max Account Usage Limits." And it's a long email, but the key notes are that traditionally the Max plan has two tiers: $100, which gives you 10x the usage, and I think $200, which gives you 20x. For that usage, your limits—you know, it's an "unlimited" plan, but there are still limits—your limits reset every five hours. With this new change that they're implementing, which takes effect at the end of August, so it's not out yet, essentially there are two limits. There's an overall weekly limit, and then there's a Opus 4 weekly limit, which is a big deal because I use Opus 4 for everything. I didn't even touch Sonnet because I had unlimited. Why would I choose the dumber model? And then two, a five-hour window changing to a seven-day window is massive because I don't think I'm like the top 5% of users or whatnot. But if I want a session, five hours feels like the right amount. I don't know how they got there, but a coding session usually does not last more than five hours. That feels right to me. And they kind of prefaced this with that. Some of the people were just hammering Claude Code. They had mentioned these changes should affect less than 5% of users. They even posted on Twitter, on X, saying there was one user who was costing tens of thousands of dollars in API credits but only paying $200 a month. So I'll just pause there because Ben has talked about it every time I brought up Claude Code. Ben has mentioned, "Is this sustainable? What's going to happen? When's it going to happen?" et cetera. And I had kind of hoped that we wouldn't get here, but I almost felt like we were getting here regardless. And yeah, we're just in a mode where things are changing. We don't love it. We do love it. I don't really know what to think. What do you think?

Yeah, I think it's fascinating because, you know, we've talked about how Anthropic is very different from OpenAI, from Google, and from Meta in the sense that they don't have these other businesses that they're already largely profitable and built up. And, you know, I think Anthropic has always just been Claude. I don't know if they've had other products in the past, but let's be real, the reason that we're talking about them is because of Claude and what they're doing in the AI space. And so to me, it's interesting because we talked about it last time where I think with Claude being the best coding model, you got the most costly people using it. Like, you know, your regular mom and pop are probably using ChatGPT. Your hardcore, you know, kind of sweaty developers are going to be using Claude. And so, and they're going to be using far more usage. But to me, what feels really strange is, you know, they priced that plan... and I think it was marketed as unlimited. I don't know, they've changed it since, but I think that Max Plan was marketed to be unlimited, right?

Yeah. On the record, it says "unlimited" with an asterisk, and they do publish a help article describing the limits, which I'll have to pull up just to be extra sure. But essentially, the Max Plan that I paid for was 20x. There's a Claude Max Plan that's $100 that... I've got to remember how much that one is. Either it's 5x or it's 10x. OK, it's 5x. So to paint the picture a little bit more clearly, there's the original $20-a-month Claude subscription that gives you access to all their models, but you have limits. There's also the API, so you pay-as-you-go pricing, which means you pay for the tokens in and the tokens out. Outside of that, there's the Max plan. So there's Max 5x, which is 5x the normal usage, which is their $20 tier. So it's $20, 5x that is $100, which makes sense, you know, 20 times 5 is 100. Then we have their 20x plan, which is $200 a month. The math doesn't check out for that one, but essentially it's 20x more usage than their Pro. The problem with all this stuff is that these limits are floating. They're variable. They're not well described within their help articles. Essentially, they describe it as "depending on load," that's how this is scaled. So if their servers are on fire, 20x might be pretty low. If their servers have a ton of capacity, 20x might feel like a giant, never-ending window. And I think that detail has always been described, but it hasn't been felt up until Claude Code got wildly popular and extremely successful. Then you felt it every day. It was like every day I'd hit a limit. Every day it felt like Claude Code was deteriorating in quality due to a reason I couldn't control because it's not extremely transparent like other IDEs are. So it's very odd. It is officially labeled as 20x, basically unlimited, but that is extremely hard to formulate. There's not a clear consensus. And to be fair, in their latest email, they outline, "Most Max 20x users," so the plan that I pay for, "can expect 240 to 480 hours of Sonnet 4." Even that alone is a pretty wide range. That's 1x or 2x. And then they describe 24 to 40 hours of Opus 4, which again, that's almost like a 1x to 2x range. So clearly, Anthropic is not fully finalized on pricing. They even apologize in the email and say, "We take our tool and decisions very seriously, so we want to support this in the long term." Clearly, they're bleeding money. But yeah, I don't know. Like, I expected it to happen, but I don't know how I feel about how they're handling it. And I don't know if I like seeing 40 hours of Opus 4. Like, maybe that always was enough for me, but I don't know. It feels different.

Yeah. No, I think they massively got their pricing wrong, you know, with customers. And the reason I brought up those other companies is because, you know, those other companies can afford to maybe take losses in certain areas because they make up for it in other areas. You know, like, if you have people... just take Google for example. Their ecosystem is so wide that maybe they take a loss on the Gemini model plans. People might be using more, like, their server costs might be more than what they're getting from the subscription specifically for Gemini. But they make up for it with the Google Workspace subscription or, you know, they make up for it with YouTube TV and YouTube, right? Like all these other areas. So they can still be profitable. You know, they take a loss on the AI piece of the business. I mean, like Amazon, too, because I think they're kind of slowly coming into a bit more relevance when it comes to AI with their recent IDE they released. You know, AWS was famously their only really growing business for a long time. Like, that was the golden child, I feel like, in the 2010s for them. And so Anthropic doesn't have that. And I think their pricing is felt much more acutely if it's wrong or if it's right with them. And I'm sure, you know, I think profitability in the Bay Area and tech culture now is a bit more of a focus than it used to be. I don't think you have the zero-interest-rate world where, you know, Uber could run a loss for however long it ran and that'd be totally fine. I feel like companies now are expected, in general, to profit or at least have a profitable strategy that they're working towards. And yeah, they must have just been getting completely hammered on their costs to run these things versus what people were paying.

You know, I feel like there could have been a better way to maybe roll it out. And so I think to me, companies get their pricing wrong all the time. They have to go and they fix it. But I think what stands out with this one is it feels very sudden, feels very quick. Usually, I can only speak from the accounting business, you know, usually when you need to raise prices on your customers and your clients, you kind of slow-roll it. You might grandfather people that have been locked in their plan for a little bit longer, but any new customers are going to be charged the new price, you know. But the fact that it feels like maybe in April you were singing Claude Code's high praises and, you know, how much work you were getting done and how great of a deal it was. And now we're here, you know, three months later and it's like you're getting limited pretty early on in your workflows, essentially. And I think, too, what's so frustrating with this specific point is when you're coding something and, you know this, you're in the flow, you're working on your React Native app, you've got some things going on here, and you just hit that wall and you're like, "Damn, like, what do I...?" I gotta, you know... It's just so interruptive to what you're doing.

Yeah. And it kind of, at least to me, they always kind of come out of nowhere, you know? It's like, "Oh, I didn't even know I was close." You know, you might get those little warnings, you know, certain places do it, like, "Hey, you know, you're coming up to your limit," but still, it kind of feels like it just comes out of nowhere. It's like, "Oh, damn."

Like, well, it's funny you say that and describe the in-flow state because they have added a new setting such that if you hit the limit, you can actually switch to API rates. So you're paying $200 a month for the Max plan, you know, you're grooving, Claude's actually doing good work for you some days, and then you just hit that wall and it's like, "Hey, now you actually just pay us exactly what you're costing us," at, like, you know, maybe a 1.5x markup. Yeah, it'll be curious to see if they make a ton more money because these limits, I feel like, seem like a giant reversal and making it more restricted. So I'm sure they'll win on that front. On top of that, for people that are true power users and love it—which again, I've been singing Claude's praise for a long time, and I think when they don't have performance issues and stability issues and harsh limits, I think it's an excellent tool and will only get better. But with the chaos, everybody adopting it, I think it truly has suffered in terms of quality. Like, the code that I was writing in the beginning of May is different than the code that's being written in July. And if there's anything about the AI space, it should get better, not worse. And clearly, they've sent this email out to kind of rectify that and hopefully, you know, basically price people out and don't let people suck Anthropic dry with, you know, $5,000 of credits for a $200-a-month plan. Anyone would laugh at that and say, "That makes zero sense."

Yeah, I don't know. They could have done it better. I think if there's any takeaway from this, it's that $200 is definitely not the ceiling for AI. And for all these Claude Code users out there like me, we have about 30 days. It's time to extract every token we can from Claude Code to get to the highest amount on the Claude Leaderboard. I was chatting with a friend who had told me about the Claude Code Leaderboard. Essentially, you download this mini CLI program that uploads your Claude stats, and then people just kind of praise you on the Claude leaderboard. So I uploaded my stats the other day. I think when I uploaded it, it wasn't super popular at the time, but I think I was like 57th and my friend ended up being like 30-something. I was like, "Oh, pretty nice." But looking at the top of the leaderboard, there are folks that were using five to 10 grand for an individual account. And I just thought, you know, if that leaderboard isn't proof that Anthropic has a pricing issue, I don't know what is, because that is clear. Like, "I'm getting maximum value out of this $200-a-month plan." And yeah, there should be limits. I'm glad those people are being taken care of, but I truly hope it doesn't hurt the rest of us who sit below the extreme extortion but above the "I actually don't use this, but I pay for it."

Yeah. Well, it's funny when you said your friend, you know, and comparing the ranks, whatever, I thought of that scene from *American Psycho*. Do you know what I'm talking about? Where he has the business card? Have you seen that movie?

No. Christian Bale.

Then you won't...

I've probably seen it, but I can't...

Yeah, it's... yeah. Well, I won't bother explaining it. But what I wanted to say with the Claude Code thing is what's interesting is, you know, as much as they have those users that do all that volume, they have users like me who, I pay for it right now. I don't pay for the Max plan. I just pay for the $20 one, I think. I don't ever use it. Like, I don't ever use it, you know. And, you know, I think that's probably true for a lot of all the different AI providers. They have people that don't even nearly use the $20 that they pay each month. But clearly, Anthropic must be feeling out of balance in some way. Otherwise, they wouldn't be doing what they're doing. You know, so, but there are people that are on the opposite end of the pendulum. But maybe, you know, Claude isn't popular enough on a mainstream level to get that many users of that ilk, you know, of those that don't nearly use how much they're paying for. You know, whereas, like, ChatGPT and Gemini, they're so popular that everyone and their mom has a ChatGPT Pro plan. They only use $2 a month of tokens, basically. You know what I mean? So maybe that's the difference.

I do put some blame on the AI slop-post world with Claude Code because you saw for the last few months tweets of someone running eight Claude Code terminals to do something completely stupid and pointless, but it would get so much engagement and they'd be like, "Wow." So I think that trend, I suppose, I think has something to do with it too, I'm sure. It's definitely part of the maximum value extraction of, like, "If I can run eight Claude Codes and rack my API token usage through the roof, I can then go post a screenshot on Twitter and get some clout." Like, "I would have spent $5,000, but I spent $200," you know?

Yeah. It was partially, "What can I do with it? What can I mess around with?" and partially, "I want to look good online on Twitter." So...

Yeah, yeah. That was definitely part of it. It's like... it's like smelling your own farts. It's like, you're going to Claude Code this application, like, your SaaS app, right? You're going to make a little screenshot of your eight Claude windows and you're going to post on X, "I'm making my, my, you know, stupid startup app, blah, blah, blah." And, like, "Look, I'm using like $5,000 worth of tokens." You get all this engagement, you know, good and bad from that post. People then see that you're making this startup app, and then your startup app gets some eyeballs, if that means anything, you know. But people like that when they're building something. Like, "Oh, you know, I got people to sign up for the pre-release," or whatever, you know. So it's just this little, you know, gassy echo chamber that they're in, you know. And yeah, it's just a slob. It's everywhere.

Yeah. And that's kind of the bad part of Claude Code. And I guess before wrapping on Claude Code, there are some good changes. One can see this change as good and bad. But I think there have been some clear improvements over the past two to three weeks. One feature they came out with is hooks. So essentially, Claude Code can do things on your computer. It can be editing a file, listing out a directory, etc. Before it does things, you can create kind of like plugins that hook into these actions. Before I edit a file, run my linter. After I edit a file, run my type checker. Whatever. Literally whatever you want, it can do it. Pretty cool. And then they recently came out with sub-agents. So instead of just having Claude, you can essentially spawn sub-agents that have their own kind of personality and goals and traits. Their example is a code reviewer, such that you use Claude to write the code. And then at the very end, before you commit your changes in Git, you'll say, "Hey, can you please review this code?" And based on what you say in the prompt, it'll identify if it should use a sub-agent. Again, that sub-agent basically has a 400-line prompt within it, and then it kicks off the task to that sub-agent. And then that sub-agent has its own context window. So usually in Claude, you have like a 200k context window in the main chat. As you get down through it, you'll see in the bottom right corner it says, you know, "20% left, 19% left." But even if you have 15% left and you kick off the sub-agent, you still have 15% left because that is essentially a whole new window. So a pretty cool feature. Uh, right now, I'm still trying to figure it out. I think with all the Claude Code changes, people are just kind of waiting to see how the dust settles. And this is very, very new. And I think the cool part about their sub-agents is that they can generate a prompt for you. So in their sub-agents workflow, you basically type in `/agent`, it presents you a menu. It's like, "Look at your agents," or "Create a new one." Uh, press 2 to create a new one. And then you describe what you want your agent to do. And then Anthropic takes in your kind of goal and will write you, like, again, a 400-line agent prompt based on what you want, knowing how to prompt Claude Code. So it's pretty cool that you don't have to spend all the effort of like, "Oh, I want a Python expert. I want a Swift UI expert." It can get you 80% of the way there with a few sentences on, you know, "I want Python. I want to have types. I want to use these packages," etc. Then it does the hard work of writing it all out. So pretty cool feature. And it's still being figured out. I think we just got to wait and see how the usage limits and all that kind of settles. But again, it's not kicked in until a month from now, so sub-agents will be able to run wild and free. And, you know, the people that we see with eight Claude Codes, they could be eight Claude Codes with five agents in each, and we're screwed. So yeah, it's the wild west out here right now.

Yeah, I think Claude Code... I think Anthropic will get acquired by Microsoft, and they're going to use Claude Code to basically replace Visual Studio. That's my... that's my, uh, bingo card. Bingo card. No, it's because they're going to get so fed up with what OpenAI does and how they operate. You know, again, trying to acquire, you know, some other company but blocking the IP from being shared with Microsoft. They're going to get sick of that agreement at some point, and Microsoft's going to acquire them. So you heard it here first, everybody.

Yeah, yeah. That's right. All right, cool, Brad. Let's go ahead and wrap this sucker up with some bookmarks. I'll go first. I think mine is actually pretty, you know, scary, hilarious. It depends on, I guess, what side of the camp you sit on. It's a ChatGPT agent that controls a live security camera, I guess, to search for a turquoise boat and find its name. And so you can see in the tweet, it's zooming in on the camera screen. It's going to try and search for where the boat name would be, which is usually on the back of the boats. And so, you know, you think about how many cameras there are around, all the stoplight cameras we have, you know, all the different front door cameras and stuff like that. It's going to feel like soon we're living in an *i, Robot* episode or, you know, the *i, Robot* movie or *Minority Report*, where you're just always under surveillance and there's some hive-mind ChatGPT or hive-mind Gemini that's seeing what everyone's doing. And seeing, "This person, Brad, he woke up and he had this for breakfast, and then he went over here..." And so they know everything about you, and they're going to sell you ads and just inundate you with hyper-targeted ads. Yeah, but anyways, that's my... that's my bookmark.

Yeah, I just watched it while you were talking about it. Pretty cool. I think I've seen so many pre-canned videos of, "Oh, you know, interact with Excel, interact with Google Calendar with the Operator or OpenAI Agent that does the web browser." But this one was actually pretty cool. So if folks are interested, this is a unique use case. I'm not sure how useful it is, but it kind of paints the picture of giving agents something that's not a pre-canned marketing approach on just something you can do with it. So pretty cool.

Okay, so my bookmark is "Seven Useful Claude Code Sub-agents." This is from Ian on Twitter. And like I was talking about earlier, you can create your own Claude Code sub-agents, but it's not perfect. So Anthropic gets you probably 80% of the way there on a good agent, but Ian put in the hard work for us and listed seven sub-agents and actually posted them on GitHub. So: code refactorer, PRD writer, project task planner, vibe coding coach, et cetera. So essentially, you go to the repository, copy in the sub-agents into your `.claude` directory, and you're off to the races. So thanks, Ian, for sharing. Hopefully, that'll make people more productive. I haven't used it yet, but I know we're getting a lot of these sub-agent tweets coming out, and I've followed Ian for a while, and he puts pretty good content out there. So should work pretty well.

Cool. All right. Yeah, I'll check it out.

Awesome. All right. Well, let's wrap that up, Brad. Good stuff, and yeah, we'll see you next time.

Sounds good. See ya.

See ya.