Breakeven Brothers

Alrighty, welcome, everybody, to episode 24 of the Breakeven Brothers podcast. I'm your host, Bennett Bernard, and your co-host, Bradley Bernard. How are things going, Brad?

Excellent. And I have a little trivia question for you.

Okay.

How many days has the Breakeven Brothers podcast been in business? Because, you think about it, uh, you know, when did it start? How long has it been going? What's your guess?

I'm gonna go with 365 days. I think it's been like a year.

Pretty close. Yeah, about, I think four days ago, I got an email from our excellent podcast website software, Transistor, and it mentioned we got an achievement. And the achievement was one year of having the podcast, which is pretty good.

Yeah, crazy. I got the domain renewal email like two months ago, which I texted Ben. I was like, "Oh, you know, here it is." So yeah, I'm just gonna pat ourselves on the back real quick. One year in business, you know, uh, doing our thing, having fun, kind of focusing on AI. So yeah, super, super cool.

Yeah, totally. And I would recommend people that are on the fence about making a podcast: just do it. It's kind of fun. I mean, yeah, it feels like there are almost too many podcasts out there, but, you know, at the same time, it's fun to find your tribe, find your niche. And two, it's important to get out there and show your face and show that you're a human being. And especially in the age of just AI slop all over the place, it's good for people to be able to see who you are and that you're an actual human being. So yeah, it's been fun. It's been a good ride so far, and here's to another year. Let's do it again.

Yeah. I, uh, even coworkers at work will kind of mention, "Oh, you have a podcast, right?" and, you know, that gets you somewhere. I was telling Ben I recorded my Twitter video for releasing my MCP server a few weeks ago, and I felt nervous being in front of the camera. But when I get on the podcast, it's almost like I'm trying to have digestible information—not to say I get nervous on the podcast. So yeah, it's a nice skill. You know, you kind of put yourself out there like Ben was saying and get face time that you're a real person.

Because I don't know about you, but if you've scrolled Twitter recently, half the replies are just AI junk. And you can immediately tell. I think you can't fake this, at least not yet. We're not at the time where we can be on a beach in Hawaii and there's some face that looks like me and can pull this off. But one day, we probably will be there.

Yeah. Just automate everything. Just automate your podcast, rake in $50,000 a month. Just, you know, living the dream.

So easy. So easy.

Yeah. Well, cool. Yeah. That was a good bit of trivia. Yeah, so we had a lot of updates, uh, since the last, what, two weeks since the last time we did this. Um, you know, we went to Disneyland, I guess, by the way, in between those two weeks, which was a lot of fun. We hit up all the rides there.

Yeah, I got a good picture on Tiana's Bayou.

Oh yeah. So yeah, but in that time, we've had a lot of big, uh, new AI releases. So, Brad, you want to kick it off with the first ones? We'll just dive right in.

Yeah, honestly, every time we go through the release roundup section in our podcast notes, there just seems to be a ton of things. And that's what makes all this exciting. But yeah, to kick it off, we have OpenAI. They released their Deep Research models in their API. And when this first came out, this was making waves of, "Oh, Deep Research." You know, you'd pay $200 a month at that time, you got maybe 50 queries. You could chat within the ChatGPT app and it would do 15-20 minutes of research and come back to you. Now they've packaged this up within the API, so now you can kind of get this for your own topics. I'm not exactly sure who this is tailored to because Deep Research still exists, but I'd have to look into the documentation and kind of see where you could make it your own and customize it because it's just like any other API model. And I think it's kind of on the expensive side. We'll need to double-check the pricing, but yeah, it's available in the API.

And on top of that, they open-sourced their prompt rewriter. This, I think, is actually a bigger deal than people think because OpenAI has this cookbook, which shows a bunch of stuff, you know, how to be efficient and intelligent with AI tooling. But like, if you kind of think of the evolution of AI, back in the day, it was, "Here's this crazy prompt that has all these keywords: you're a top software engineer, you're working at Google, you've had 15 years of experience writing C++, you write memory-safe code, blah, blah, blah." The TLDR is, back then it used to be, "Write this amazing prompt," and now we've gotten really good models that don't need as much of a good prompt. You know, you can go in Cursor, Claude, etc., and just say, "Build this UI for me," and boom, it's off to the races. So I think it's a big shift from the kind of pre-, maybe, like, Cursor era where there was a lot of intelligence in writing a good prompt, and now it's not really the case anymore. But there's still power to writing a good one. So, Anthropic has their own prompt rewriter—I think they call it Workbench or some other tool internally, or not internally, on their website—and then OpenAI open-sourced their own. So it's definitely an important layer. It's not as important or as big a moat as it used to be maybe a year and a half ago. So yeah, pretty interesting trends we're seeing.

Yeah, no, totally. And I think it's interesting too, because in—we won't, I won't hit on it too much, because I think we've hit this one a bunch already. But in that drop in their cookbook, they also mention that the Deep Research agents have like built-in support with MCP, which I think if you've listened to this podcast at all, you know how much, you know, Brad and I are big fans of MCP and how powerful it is. And I think somewhere in the news or somewhere in the cookbook, I can't remember the exact part, I don't want to misspeak, but I think they say something like, you know, Opus 3 Deep Research and Opus 4 Mini Deep Research are some of the most powerful agentic LLMs in the world right now. And this is thanks to the native support for MCP, search, and code interpreter. So they kind of specifically call out those three functionalities as part of why Deep Research is so effective, which makes sense because I think search, obviously to search the web and find information it needs; MCP to be able to read through files and read through, if you have a directory that you're looking through, or just some other MCP integration; and then code interpreter as well.

So yeah, big, big drop. I'm with you. I don't know. I mean, it's pretty fresh. I don't know. I haven't really had even too much thought of like where to use that. Because again, right now you can just go to, you know, ChatGPT and do a Deep Research query. But I'm sure there's going to be some use cases where you'd want to set something up via an API to hit that and maybe make it more tailored to a certain audience or implement guardrails in the Deep Research, so that you can't just deep research anything, you know, if you're trying to limit what people can research. And so, yeah, it'll be interesting to see and see what people build with that because, yeah, it's pretty powerful.

Yeah, there are lots of times when things come out and I think, "Oh, that's probably not that useful," and then, you know, two days later I see something blow up on Twitter and, you know, that's probably the use case that I should have thought of. But pretty cool.

And then along those lines, Cursor released their mobile app. And I haven't downloaded it yet. I think it's a native app, or at least it runs on iPhone and Android. I had seen some comments that it was a progressive web app that you could go to their website, like, add it to your home screen. I'm not exactly sure which flavor it ends up being. But they released a one-and-a-half-minute announcement video earlier on their Twitter talking about how now you can tell Cursor to do things on the go. It'll then kick off a remote web server to go and develop on your app. And then once it's done, I think it opens a PR to GitHub, or you can iterate on it before it does that. But essentially, now you're able to kind of do what OpenAI Codex does, which also has first-class support for mobile. I imagine Cursor is probably not as good, but Cursor gives you access to these other models like Anthropic's models, like Google's models, where Codex is limited to OpenAI.

So pretty cool. My first thought, though, was, "Damn, this is going to be distracting," because as someone who works on a codebase all by myself for my bill-splitting app, there are a thousand things I want to do. If you give me mobile access to, you know, an intern engineer to get things done, I might be glued to my phone. And I saw a comment that someone was making that we're kind of in this era of people trying to maximize agents running all the time, or like coding agents running all the time. And someone had, I think one of the top comments on their post was like, "Goodbye work-life balance," because, you know, people are trying to spend all this money using agents to get productivity. They allow you to create as many agents as possible using this mobile app. You can ask for one task to be done, then go ask for another task. And I kind of have the same conclusion of, I wonder where this goes. Is it scalable? How does it operate with the day-to-day workflow? And is it more fun than useful? Is it more useful than fun? I haven't really put my finger on it because I still have to try it, but I am concerned that we're getting almost too hands-on, which is having extreme convenience to having AI agents do things, which I'd love. But I'm also like, you gotta have limits somewhere.

Yeah, yeah, I could see that being... I'm someone who tries to be mindful of when I'm on my phone and when I'm not, you know, and doing it in the right settings. And I could see that being one that I'm a little reluctant to do, you know, unless it was like a really exhilarating project and I had time and I was doing it. But um, yeah, that's something that I generally try to avoid, you know, installing a bunch of mobile apps. I don't have that many mobile apps on my phone compared to probably other folks. So yeah, and I just checked while you're talking, and it is the progressive web app.

So, okay. Yeah. It's not a native app, like in the App Store. It's like you go in and add it to your home screen, I think.

Not as nice. Not as nice. But you know, people are dying for it. So I'm sure they could ship something a little faster that way.

Well, and it's, I guess one quick thing on that is it's kind of built into... Cursor announced something they made, I think last week, about background agents. Because they announced that and they said, "Hey, if you have Slack and GitHub, you can, you know, basically tell, via Slack, 'Hey, go work on this codebase in GitHub,' and run in the background, and then it'll do that." And so I think this is a similar interface, essentially. Instead of having to use Slack, you can do this mobile installation or on the web and then do the same thing. "Hey, go work on this GitHub repo and do it while I'm not... you know, do it asynchronously while I'm not hand-holding or babysitting it."

So yeah, yeah, my thoughts are, I haven't used it yet, but my immediate thoughts are, I wonder how much code is going to be written at a pace that people can't review it. Because I, you know, I think people get a little too excited, uh, with the vibe-coding, and it's getting to a point where I think the models are good enough, but yeah, I'm not sure it's good enough that we can one-shot a bunch of tasks and move on to the next one before even checking the output. So lots of thoughts here I got to unpack. And I think they're setting themselves up for the right spot in the future. I think the models are really damn good today and only going to get better. But being on the mobile interface and not on the computer, getting as much control, easy input—I'm not sure it's fully there yet. So good for them for releasing it. Unfortunately, it's a PWA, but you know, I've been there to develop my own mobile app. It's kind of a pain. So I understand where they're coming from.

Yeah. But I thought making mobile apps was really easy with AI now. I thought you could just, you know, put it in.

Don't get me started. Don't get me started. React Native is a little easier now. Swift is still a pain since Apple's docs are not ingested at all. But, yeah.

Well, let's do this one next, Brad. I know you've been a big proponent of Claude Code and the CLI interface. I'm someone who does not traditionally like the CLI interface. I like to be able to see my files and just feel a bit more hands-on. So sell me on CLIs and what's going on in that space over the last two weeks.

Yeah, I would still like to announce that I was early to Claude Code. I picked it up. I tried it. I was like, "Damn, this is good." And so I told everybody about it. And just now people are getting around to it. So it feels good to put my stamp of, "I got that first." But yeah, the AI agent CLI battle is definitely heating up. I would say there are roughly four main contenders at the current moment, at the end of June. So there's OpenCode, which I think is an open-source tool that tries to replicate a CLI AI agent. There's OpenAI Codex, which can run locally or can run on the server, kind of like Cursor's background agents. Claude Code, my favorite one, the one that I paid $200 a month for, is extremely well built. They built the model, they built the tool, they use it internally. It's a great flywheel for them, and it really shows when I use the tool. And then Gemini CLI, I think, is one of the newest ones, if I understand correctly. And it's open source, which is great. And there are free credits. Google is really vying for adoption, and they'll do that any way they can. And hats off to them, because their models are also pretty good. So it's not like they're offering free stuff and it sucks. It's free stuff, and it's pretty good. I've seen people compare Claude Code and Gemini, and I think Claude Code still wins, but Google catches up decently fast.

But more on the why, and I think this is something I've been trying to understand more recently, is I think as developers, it's a really easy sell. Installing one of these tools brings you to the terminal, and the terminal is kind of where developers feel at home. And as I've been using these tools, it becomes an understanding that these tools can do a lot more than what Cursor can do. And the way I say that is because Cursor can edit text files. It can run commands. The interface is very similar to a traditional software engineer's code editor. That's great. But it's almost like when AI comes out, you need to rethink these experiences. And I think the AI agent CLI tool initially was like, "Ew, we're leaving a graphical interface. Why would we ever want to do this in the terminal?" I think people were a little bit confused and thinking that's not going to do well.

But I think some of the reasons it did well is Anthropic, again, created their model and trained it with tool use. So as we talked about MCP servers, it's almost like a form of using a tool because MCP servers have tools. But essentially, Anthropic trained their model to say, you know, "If you need to access tools, here's like prompts to get you to use tools." Previously, a lot of these models had tool use kind of tacked on at the end. So you'd have a pretty intelligent model, but it didn't know how to natively use tools. And I'm not, you know, an AI researcher. I don't know exactly what this means in practice, but I can tell you as a user of the tools, it really shows that Anthropic has done the right job to train the model to be understanding of tools at a very fundamental level and leverages these things really well.

So I think on the AI CLI front, people have realized the power with Claude Code, incredible feedback from developers online, and the price point at $200 a month is a bang for your buck. And I think since people saw that, all the competitors are coming out and saying, "How can we make this cheaper, better, faster?" And there's a whole spew of software that's built on top of Claude Code or wrapping Claude Code or all these really cool and unique ways. But I think, one, it's very friendly to developers. Two, these models are getting really intelligent at calling tools. And it's a little easier to do that on the terminal where the program already exists. Maybe a little bit technical, but Cursor asks you to run commands in a little command window and you click "Run." In Claude Code, you're just already there. Run a command or auto-run a command.

And I think lastly, one that I've come around to more recently is it can do a lot more. Like I mentioned, you're in an interface writing code with Visual Studio Code, PHPStorm, whatever IDE you use. But when you go to Claude Code, I've seen examples of people reorganizing their file system with Claude Code. Like people are using Claude Code as almost a raw agent, not really a coding agent, where I think a lot of these tools—Cursor, Windfall—are for coders, developers, software engineers. That's what it's for. If you're not one of those, you'll feel less at home. Where I see Claude Code coming in is it is a developer tool, so we feel very at home as developers. But if you can get over the hump that it's a developer-focused tool, you can do a lot with it by just asking it to do things. It's very intelligent at organizing your files, cleaning up your files. And everyone has a messy desktop or messy folder here or there. So it's almost like trying to understand how powerful it is as a universal tool and applying that in various areas. And I've seen some really cool examples. You know, one is like, "Clean up your files." I think there was another guy who cleaned up a bunch of old transcripts with Claude. Like, literally any task that you're like, "Oh, I don't want to do that. It'll take 20 minutes." Like, ask Claude to do it. It sounds kind of dumb, but the reality is that it can probably do it, write a script for it, run it, then delete it. Like stuff that, you know, you don't really want to do.

So long story short is I think CLIs are here to stay. The models have been trained to use tools, and having the AI agent in the terminal, in the command-line interface, it's very naturally equipped with tools. And it's kind of like a Swiss Army knife is what I'm getting at. It can do a lot of stuff. It's excellent at coding, and I love it, but it can do a lot more than that. And I think people are just now getting around to saying, "Oh, there's a lot more to these. What else can I use this for?" And that's a pretty fun adventure, to be like, "I'm going to throw Claude at X, Y, and Z and see how it does." And surprisingly, it does pretty damn well at everything.

Yeah, I wish they had better support for us poor Windows users because it's, you know, Claude Code, I think, is mainly for Mac. And if you want to do it with Windows, you have to use the Windows Subsystem for Linux, which I'm sure people can figure out.

But I'm just special. You can figure it out.

It's just, it's just not... it just feels kind of icky. I don't know how else to describe it. It just feels like... I have to use the Windows Subsystem for Linux when we have tinkered with Laravel stuff on my computer, and it just feels kind of, I don't know. I don't know how to describe it, but I'll keep trying. I'm not giving up, but yeah. I do want to try the Gemini CLI. I'm a bit more of a Gemini team player, so I do want to try theirs. I haven't tried it yet, but I do think that's probably where I'll dip my toe in. Do you think they have good Windows support?

I don't know. I haven't looked into it, like the specific CLI, but I've been writing my Google Agents. I use an ADK in Windows, and no problem. So now you know, they've got good docs, you know?

Yeah, I'm definitely on the Google team of the AI team. They're doing well. Their AI Studio is also pretty darn good. So, hats off to them. And the Google Cloud Console, where, like, if you are developing—I think I mentioned this in the last episode, and I'll shut up in a second—but it's just nice because you can, with a simple command-line command, if you're developing with Google ADK, once you're ready to deploy, you can do, I think it's like `gcloud deploy`, and it'll push all that up into Google Cloud Console, which is pretty nice. Because I think deployment of these agents can be a little tricky sometimes, depending on the platform. So I like that aspect. That's just me.

Yeah. And I've been talking about this, but I definitely see the larger shift in pricing. So while Cursor released their mobile app, kind of side news for that is they released a max mode, which sounds like every company now is going down this bandwagon of trying to charge people about $200 a month, maybe $250. I think Google has inched a little closer to $250 for some of their subscriptions, but essentially $200 to $300 for an unlimited plan. And I had bought the Claude Code unlimited plan, I don't know, a month or two ago, maybe just a month ago, because I just got my renewal like two days ago. $200, unlimited tokens, and it truly does the job. And I think we're ending up with this new era where people want to have vibe-coding be more or less like a slot machine. So Anthropic released a blog post talking about this, where across all the people at Anthropic who use Claude Code, there are different strategies of using it. And one of those ended up being, quite literally, using it as a slot machine, which means you have code that you want to add to, and you ask Claude, "Can you make this change?" It'll go do it. If it doesn't work out, ditch it, ask it again. And it's not really feasible when you're working with pay-per-token usage because you get this mental state of, "I just spent $20 on that. Do I really want to ditch it and go back to nothing?" Sometimes you could spend an hour or two here kind of iterating on this.

So I think it's a two-parter. One, they want to get more money out of users. I'm sure people will spend $200 and won't use it. Two, I think other companies have it. So, Anthropic led the way, more or less, and now Cursor is playing catch-up. So I expect Gemini CLI to have a max mode within the next month. And I'd also put it on OpenAI to probably come out with unlimited tokens for their Codex CLI in the next month too. So it's pretty interesting because I'm okay with $200 a month. It's expensive elsewhere if you can't afford that. But the unlock you get... For example, I pay $200, but I got roughly maybe $2,000 out of the tokens I've used for Claude since it was unlimited. So, different mentalities. And I think we're converging at a time where $200 or $300 is going to be the new normal.

Yeah. Yeah. And I'd say if you are posting online about your Claude Code stats, like, post receipts, because that's very trendy right now to kind of talk about how much you're spending. It's a lot of engagement. But post the receipts. We want to see them, you know. We do want to see them.

Yeah, I sent a tweet to Ben the other day of this guy leading with a great hook, you know, a hook that would catch me anytime: "Top 10 tips on how to improve your Claude Code usage." And I go, "Sign me right up." And I scroll down, you know, one tip, two tips, three tips. I think tip number seven or eight has a screenshot of the Claude Code usage, which for people who are unaware, is a table that breaks down your daily Claude Code usage by tokens in and tokens out. I say that because they charge differently for how many tokens you type in versus how many tokens the AI generates. And for each day, it gives you a dollar amount total for those tokens that you used. And for this guy or person online, I think they were using like $8 to $15 a day, which if you ask me, as someone who's used Claude, I'm spending at least, like, $80 a day if I'm doing a Claude session. Some days I don't use it, but the days I do, it's at least $80, up to about $250. So when I saw this person, you know, tip number seven was like, "Show your Claude Code usage," I saw it was maybe $40 or $50 at the end. I thought what I just read was a total waste of time because I've spent more time in one day than this person has spent in seven days. And I thought, you know, maybe we should have a little bit of receipts here because if you're showing a $2,000 spend on Claude Code, you'd be spending, you know, eight or 10 hours with it. And could you not be learning from eight or 10 hours? Maybe, but at least it gives you some experience, you know, using the tool. So, yeah, from here on out, I demand if you're posting any Claude Code tips, please lead with your screenshot of usage. And then I will happily read the rest.

Yeah, we're getting to that point. I'm just going to call balls and strikes. We're getting to that point now where there's so much just hook-grabbing, attention-baiting stuff out there that you got to validate it. There's one famous one that I saw. It was someone had posted, "I spend $6,000 on Claude Code, but it's so worth it." And I think if you kind of read the comments, it's like, "Well, wait, aren't you on the $200 a month plan? You didn't really spend $6,000." And it's like, "Well, yeah, this is what I would have spent if I had done it," which is still valid. So they clearly are using it, but you didn't spend $6,000. And of course, in the thread, it's like, "Come see what I'm building over at X." And it's like, they link their startup or their little web app. So yeah, sometimes with these posts, you've got to kind of squint and be like, "What are you trying to sell? Are you trying to sell me something?" Because I can kind of smell it. You can smell a car salesman that really needs a commission. You know, you can kind of smell it on them. So I think you can start to see that on some posts online these days.

And what scares me a little bit on that topic is for people who actually do go ham and are power users. So I just pulled up my stats. I'm ending up at $1,600... $1,680.32. So pretty good for spending $200. But people are going off the rails and doing multiple Claude agents in the same terminal or the same project. Whatever they're doing is adding up to a lot of dollars, and they're posting screenshots on Reddit, on Twitter, etc., saying, "I spent $200 a month on Claude Code Max, but I'm getting $5,000-$6,000 worth." And that's great, but there's a tinge in me that thinks, you know, as Anthropic sees people spending $200 a month are getting $6,000 of value, is this going away? Is this here to stay? There's a little bit of a confusion, or maybe a debate, of whether it's here to stay or whether it's... you know, here to stay being tokens always get cheaper, which historically has been true, or not going to be here for long, which is, "Get people hooked on Claude Code, you know, increase the price to $500 and say, 'Hey, you know, you loved it, you know, come pay us more money.'" So I hope it's the cheaper price, you know, things get cheaper. If it's not, I'd still probably pay a little more. I'd probably pay $300, $400. But it's something to be seen that I think they know they have a grasp on interested users. What are they going to do with that?

Yeah, I'd be curious what your specific usage is compared to the rest of the subscribers, the rest of people paying for Claude. Because I think I'm paying for Claude and I don't ever really use it, to be honest. I think I'm on the $20 a month plan. So as much as they have people like you that are really extracting a lot out of the system, they probably have a lot of people like me that are kind of like doing... you know, my usage is probably literally five cents sometimes on a certain month because I just use, you know, OpenAI or use Gemini. So, yeah, it probably balances out a little bit. But it'll be so curious to see these companies' P&Ls, you know, see their financials because, you know, it's obviously making a lot of money. But like, you know, where are they most profitable? Where are they bleeding money? And yeah, it'd be interesting to see. But yeah, I'm sure there's some power users that are just going absolutely nuts. There are limits in place. I've never hit them, which I'm surprised. But I guess people just go really, really crazy. So we'll have to see if the pricing changes. Fingers crossed it doesn't. I'm still very much in support of Claude Code. So if you're trying to use a coding CLI agent, that would be my recommendation still.

Yeah. I wonder what it'd take to hit that limit. Like, "Hey, code me a video game in pure WebAssembly code."

Yeah, I don't know. It needs to be something that it just keeps going and going and going. So yeah, "Keep writing tokens out, like generate me an infinite number of Fibonacci sequences," or something like that. And it just keeps going and going.

Yeah. Who knows? Cool. Awesome. Well, so moving on from that, there's been some big news with the big players in the kind of AI arms race, I guess we'll call it. And I'm kind of defining that right now as Meta, as Anthropic (of course, the people behind Claude), OpenAI (of course), Meta—I said Meta—Google, um, Apple, if you want to call them Apple, you know, they're kind of out there. So I think the first thing is some very big news is that there's been a large amount, I mean, relatively speaking, of OpenAI employees that have been essentially poached from OpenAI to join Meta. And on including or on top of that, Meta very publicly acquired a company specifically, and just the wording, I'm not sure if you saw this the same or saw this a little differently, but the wording seemed very much of, "I just want to get these people." Like Mark Zuckerberg really wanted to get that team specifically. There was one person they called out, the CEO, Alexander Wang, to join Meta. So it seems like Meta is really opening up the checkbook to try and get a lot of really top, talented people in this space into their company, into Meta, to build out a superintelligence team, I think is what he's... I love that name. That sounds amazing.

Yeah. Yeah, I mean, imagine, you better be smart if you're on the superintelligence team.

Yeah, but yeah, so it's very interesting to see. I think one, OpenAI employees, you know, obviously if you get thrown a ton of money at you, I don't fault them for taking the money. But it is interesting to see, you know, they're building something really incredible at OpenAI, obviously. Obviously, they were open to moving over to Meta. Obviously, Meta is going to do some cool things too, I'm sure. But what do you make of that? What do you make about Meta spending that kind of money to get, at the end of the day, humans? I mean, I'm sure they're smart people, but it's just humans like you and I, you know, they bleed.

Yeah, honestly, I'm not an AI researcher, but Meta, if you're listening, I will take a similar package. I'll work there tomorrow if I get a $100 million sign-on bonus. But in all seriousness, I think it's pretty interesting. Like, I've never seen a time in which such valuable people were straight-up assets. Like, these are not humans. These are intelligence assets. And I think the people that Meta had picked or poached or, you know, got to join from OpenAI... employees on Twitter had mentioned, "Oh, they got this person or that person." This person directly led Opus 3 Mini or, you know, a top-line model that's really good. So these people, not that everyone isn't smart at OpenAI, but you could really tell that Zuck put a lot of effort and thought into getting the people who have been leading the charge at OpenAI on important stuff. And so my first thought is, like, congrats to them. I know OpenAI released a public memo maybe a few days ago, saying something like—I think some C-suite exec said—"Oh, Meta robbed from our house," or some clickbait title of, you know, "They took our employees. We got to get them." So you could tell that they took critical talent.

But the fact that they got paid such big bonuses just shocks me. And I think, "Wow, what a time to, you know, go through college, university, choose something to study, and end up at the doorstep of, you know, $100 million just to sign a box and join a different company," which, you know, I've joined my fair share of companies, and usually they're not that different. I mean, yes, they're different, but a company is a company. And so I do wonder how quickly they thought, like, "Should I leave? Should I stay?" or they saw that price tag and thought, "I don't care. I'm out of here."

Yeah. So, so very, very interesting. I think Zuck's approach of being one of the big tech founder-led companies is he will throw money at things and see what sticks, like the whole AR/VR thing of $50 billion spent over many years. And they still make a ton of money and the stock is up. So I don't see this... it's significant spending, for sure. Is it going to break Meta's bottom line? Definitely not. And I think it makes the competition really hard because I'm sure Meta will be looking at Anthropic, OpenAI, you know, who knows, maybe Google DeepMind. Like, all these are open, poachable targets. And I think now these companies are going to have to do the reverse, which is, "Let's retain tech. Let's pay these people more, you know, come stay and hang out with us." So, yeah, kind of yet to be seen, but insane news. And congrats to the people who got these massive sign-up bonuses.

And it's almost... I think it does two things, too. I think one, you know, it kind of resets the market. Like, it's kind of like in sports where you got a quarterback and, you know, you got your Patrick Mahomes. He gets his new, you know, five-year, $250 million guaranteed deal. Now every other quarterback is looking and going, "That's my benchmark. I want that," you know, like all the other top quarterbacks. And I feel like for this kind of talent now, people that are working at OpenAI, that might still be those top people, those critical people, you know, they might be looking and going, "Okay, I saw my coworker, my peer get this crazy deal over at Meta. Like, OpenAI, you got to pay up, or I might be open. I might be listening to what's being offered out there because that's the price, you know, that's the market price for people that have those skills and of those talents." So that's gonna be really interesting to see if there's more or if there's, you know, people leaving Meta to go to Google and stuff like that.

The other one that's interesting to me when I saw this was specifically on the poaching from OpenAI. So I understand the acquisition of, I think it was called Scale AI, is the one that they purchased, Meta purchased to specifically get that individual, Alexander Wang. But what's interesting for the OpenAI one is like, Meta is kind of, from what I know, they're kind of operating in a different sphere almost with AI. They obviously are doing, they have AI ambitions, of course. They released Llama and I'm sure they've got other things in the cooker. But I kind of view OpenAI and Google to really be almost like enterprise-y solutions. They're really trying to get into corporations and have AI be valued there. They're first and foremost a social media company. And so I wonder what they're trying to do with that. And maybe they stated that and I just missed it, but it is interesting to see because they're kind of different. Like if it was from OpenAI to Gemini or vice versa, I would be like, "Okay, that tracks. I can see, you know, those are kind of direct competitors." Obviously, Meta is a competitor, but they're kind of at a different angle than those other two, in my opinion.

Yeah, it's funny you say that because I saw a tweet, which... with all Meta communications internally at the company, they always get leaked. And I think it was a leaked memo of their, maybe, CPO, like Chief Product Officer, someone in the C-suite of Meta, being exactly as you described. "We have these geniuses from OpenAI that are going to create some great models. We're not in the business of selling an API endpoint to hit our models. We're in the business of social media and connecting with friends," and that whole plethora of products that Meta already has under their belt. "How do we bring AI to that forefront?" So I think they acquired these people. They created a superintelligence team, you know, headed by, like, 11 key researchers, Alexander Wang included. Now it's, "What do we do with these people?" Well, they create a great model, but their models are aimed at, you know, kind of satisfying social needs and maybe video, audio, unique experiences, I imagine, within the Meta product suite.

So it's... I don't see them pivoting towards competing with OpenAI or Google. I pretty much expect them to stuff AI as deep as they can in their products to make more ad money. Not sure what that looks like in terms of features, but they are an ads machine. Any way they can make more with ads... like maybe just, you know, using AI to generate a bunch of ad templates for people. I don't know. It seems obvious, but I'm sure they've thought a lot deeper about that. So ads, social, I think that's the kind of playing ground that they're working with. I have no clue what that'll actually end up as, but it'll be an interesting concoction of smart talent, social products, and AI—which usually you think AI is for productivity. So I don't know what they'll end up building with these smart people.

Yeah, I almost wonder... I don't know if they're still doing the metaverse at all. I have no idea. That kind of, you know, I just don't know if that ever really stuck. But, you know, if it is about connecting people and friends and all that kind of stuff, you know, maybe there's like creative metaverse... you know, because you can obviously use AI. And I don't use it as much for this, but like the Veo, you know, video model where you can make videos, and then there's Midjourney where you can make videos and images and stuff like that. So maybe they're going to try and do something similar that's more on the creative side. I don't know, but yeah, it'll be interesting to see what they come up with. So, more to come.

Yeah, and kind of on the note you mentioned earlier was, "What are companies going to do if their employees get offered these large packages?" Well, Mark Gurman, a popular reporter for Apple, released a blog post, kind of a State of Apple. And usually these are kind of not negative, but insightful, kind of peel back the curtain on what's happening at the company based on leaks or contacts or whatever. I think in his latest one, a few days ago, there was a paragraph describing how Apple's MLX team almost left. And their MLX team is their machine learning on-device team, which for people who aren't aware, the iPhone ships with, I don't know, eight cores that are a CPU and then four cores that are more optimized for their neural engine. Their neural engine powers a lot of their AI experiences. And that's a huge deal because these chips are really optimized for matrix multiplication, which is GPU-side, but it's optimized for machine learning kind of workflows. And the reason that's important is because you can build these cool experiences. So you could have in the future a company ship an on-device ML model that could run really performantly. And it's kind of a trend in mobile devices or cellular phones in the future is, you know, ship your normal CPU and GPU, but also ship a highly capable ML neural engine, is what Apple calls it. It's kind of the same thing that they do with video, where they ship dedicated video chips because video is so important.

But long story short is there's a blurb in the article saying Apple's MLX team almost left, which would be a huge deal because Apple's kind of leading the charge in developing these excellent chips. Apple Silicon is great. People have been applauding them for many years now for leaving Intel and doing their own thing. And I would consider MLX as part of that group. So you can imagine a talented group of Apple engineers building a really powerful ML chip, and we're at the pinnacle of AI, ML, you know, everything is pointing at AI, and this critical team would almost leave. And I think in the subtext from that Gurman article was Apple had retained this set of employees or this team or org or however large they may be by paying them more, which I thought was interesting because, you know, as an AI-focused team at the heart of Apple, you know, you see these large deals of $100 million, you kind of go, "Hey, what about us? You know, we're not treated that well." And then you kind of raise your hand, you get paid more, and then, you know, you're quiet for now. So I wonder if anything will pop up, if these people will leave. I imagine Zuck has kind of opened the gates for all this AI or machine learning research talent that people will probably start looking elsewhere. And competition is great. So hopefully this ends up for the better for consumers. But yeah, kind of an interesting tidbit that the MLX team almost left up in arms. It would have been a disaster, I think, for Apple, quite literally. So they retained them now with some cash money. We'll see what that ends up being.

You know, I think two episodes ago we talked about, you know, OpenAI partnering with Jony Ive. And so maybe, you know, Ive's got... Is it Jony Ive or Jonathan Ive? I think.

Now you're making me confused.

Yeah, whatever. You know who we're talking about. Anyways.

Jony Ive, yeah.

So obviously he's got deep roots to Apple and, you know, all the products he's built there. So I wonder if maybe he and, you know, Sam Altman were trying to poach some of the MLX team, because they're trying to build their own... OpenAI is trying to build, or rumored—it's not even rumored, they want to build—physical devices in the future.

Yeah. That was part of the whole announcement, right?

So, yeah, I'm just making a line there that might not be there, but maybe in the future that team will be looking over there saying, "Hey, I'm ready to work, and pay me well, so let's do it."

Yeah, yeah. Ex-Apple people are pretty good, especially that team sounds quite useful.

Yeah, cool. And then last news on the big players that we've been talking about. Something I found really interesting was both Google and Meta have signed, in recent news—I think this is probably the last three or four weeks—have signed multi-year nuclear energy deals. And to me, this kind of stuck out. One, because nuclear energy you don't really hear about as much. I think it's coming back. I think there's definitely a lot of push with tech companies to bring back nuclear energy for all the climate reasons and energy needs that they have. But specifically, what was interesting about this was in both press releases, this was kind of done with saying, "Hey, we're going to need this to secure our energy to power our data centers and our AI." And in fact, the Meta announcement, the... I can't remember who the person was, but it was someone from Meta that was part of the announcement, the press release. And they specifically... I'm reading the quote here when they bought the energy deal. It was to, quote, "for securing clean, reliable energy is necessary to continue advancing our AI ambitions." So they call it out plain and simple, like, "This is part of our AI push, is to get this energy." And I think that makes sense because, you know, I think people often see ChatGPT and they get the use case out of it, but underneath the surface, these things are consuming energy. And, you know, there's not a free lunch when it comes to what they consume. Like, they consume energy. And so it's interesting to see that tech companies are kind of looking maybe around the corner, maybe around three or four corners, of like, "Hey, it's gonna be really important for our success to have locked-in contractual agreements with energy companies," and specifically nuclear energy, to power all this stuff up.

It's actually insane thinking about nuclear energy powering my ChatGPT queries of, "Is this food healthy?" or just other random stuff like that.

Yeah, it's wild. And then also, too, we haven't talked about this much on the podcast because it's kind of outside of our understanding, at least I'll speak for myself, but, you know, quantum computers. Google is making a big push into quantum computers as well. They've announced that. I think Sundar, the Google CEO, said in an interview that where we are with quantum computing right now is like where ChatGPT was in 2017 or 2018. So basically saying we're three or four years away from having some big breakthroughs with quantum computing, which supposedly has a ton of implications. I am not an expert in that, but I guess they are very expensive and they cost a lot of money and cost a lot of energy. And so certainly, again, needing all that energy for AI, for these quantum computers, they're planning for the future, which is smart on their part, I'm sure.

Yeah, there was a lot of news probably in the past year, maybe not news, but speculation about how much energy and water OpenAI was using for various queries into their ChatGPT product. And I think Sam Altman came out early June this year saying each query takes about 0.34 watt-hours of electricity and a few drops of water. Comparing this to a Washington Post estimate that a ChatGPT response or email would take about a full bottle of water for a 100-word email. So I think they're a little off on the estimate. However, there is a slight concern with Sam's post because there's not that much data to back it up. So I think he had come out with that statement to kind of ease the burden. I think there are a lot of environmental folks who are very concerned that AI is great, but the byproduct of consuming water, electricity, zapping the grid, is not so great. So I think that kind of adds up with, maybe nuclear energy could be a safe, hopefully safe, alternative just to not zap the environment. But of course, that has its own concerns. So, yeah, it sounds very, very interesting. And one last point is the quantum computer stuff has a direct impact on crypto, where I think supposedly if this quantum computer could be as powerful as some say it may be, it could crack some of these really complicated encryption algorithms that we use today to secure pretty much lots of technology. So if that's the case, we're all screwed. But until then, I think we have many years to enjoy some of our secure encryption so far.

Yeah, I, um, for those that are interested, Marques Brownlee, right? MKBHD, whatever his name is, um, him and then there's a girl, I think her name was Cleo. They did a joint—I think she's similar to him, kind of like a tech reviewer YouTuber—they did a joint video about quantum computers. It was really cool. So definitely check it out because they talked about the encryption aspect of quantum computing. And I guess the algorithm or the framework—I'm not a cryptography expert at all—but from what the video sounded like, there's an algorithm that underpins a lot of cryptography for our passwords and stuff like that. It's called RSA, I think. And they said that with quantum computers, RSA would be completely moot, and that they would need to figure something else out. So, um, yeah, yet to see where that all goes.

Yeah, I think any tweet from Sundar about quantum computing always kind of rattles the crypto markets because it's like, "Oh, is it over?" Yeah, he has some control in that regard.

It's crazy. Yeah. And one thing I want to chat about was, we've been talking so much about vibe-coding, and I saw this tweet from Karpathy talking about how it's not prompt engineering anymore, but it's context engineering. And I think I really liked this and subscribed to this because as I've been working with Cursor, with Claude, the main skill is understanding how to talk to the AI tool. But as we've gotten longer context windows, you can talk more to it. How do you have just the right amount of context? Not too little, not too much. And so he coined the term "context engineer," which I really, really love. I'm hoping this sticks because it gives a little more legitimacy and intention behind it, where I think vibe-coding, prompt engineering—people see these terms and kind of laugh or scoff and think, "Oh, it's not serious work." But as I've been using AI at work, you know, at my nine-to-five and outside of my nine-to-five, these are tools that are here to stay. And I personally don't really love "vibe-coding." It feels very unprofessional. So now what I would consider myself as is a context engineer. You know, we've got this new title, new fancy label for what I do. But essentially, I'm trying to provide AI with the utmost, full, detailed context, not too much, not too little. I think that is the skill. Yes, it's the same for prompting, but it has a different meaning behind it. You have to use more tools, more thought, more intention behind it. So yeah, what are your thoughts on this shift in terminology?

Yeah, I mean, I do like it. I think vibe-coding is fun to say and it is fun to kind of get the project started, but, you know, any... we've talked a lot on this podcast about how any good product needs that thorough review, that element of control where things don't get built without intent behind it. And so that's something that I think is definitely needed for sure. Um, and so yeah, I mean, and I think so much of using AI effectively is managing the context, because if you just let it be an open book, you're going to get generic answers that aren't really super applicable to your use case or your query. So yeah, it makes sense to me. I hadn't heard that term until you showed it to me, but yeah, I'm fully with it.

Yeah, hopefully it's here to stay. I think as engineers, it's like automations, optimizations, and it feels like context engineering is more along those lines of optimizing the LLM performance and making sure that every token counts in a positive way, where I think prompt engineering is just sending so many keywords, buzzwords, and phrases to try to trigger some latent memory in the AI or some condition for the LLM to generate the next token that is one that you like more than your previous prompt. So it feels like context engineering is hopefully here to stay, and it's a bit more developer-focused where prompt engineering feels more user-centered. You know, "Create this feature." But context engineering is, "I want to update this code, update these tests. Here's the background on it. You know, go." So it feels different. Hopefully it's here to stay.

Uh, I love how these terms pop up quite frequently in the AI space, where someone just kind of thinks about something and thinks "vibe-coding" is cool, but, you know, we're in the next evolution. And if you're in the AI space, we move a mile a minute, so these new terms pop up every week. I've told Ben sometimes before we hop on the podcast, "When are we going to come up with the next term?" I'm not there yet, but I would love to coin my own term, like "context engineer," because sometimes, you know, you've got the right word and people read it like, "Oh, that feels right." So we'll eventually get to something hopefully before the two-year mark where we've coined our own phrase and hopefully it spreads, but if not, it'd just be fun to kind of run with our own AI wording phrase because there's so much out there today.

Yeah, yeah, for sure.

Yeah, cool. And I think the last thing that we want to wrap up with on this episode was just, I think it came from a quote from the Cloudflare CEO, or he did an interview. And basically what came out of that was him stating that Cloudflare handles 20% of global traffic. And he, the CEO, warned that AI bots are completely reshaping the web. And so publishers need to adapt or risk being left behind. And included in there were some stats. So, Brad, you want to mention the stats? I think you had put this on there, and just kind of let that sink in. Because 10 years ago, you were writing `nofollow`, I think, or maybe that was even further back. But, you know, 10 years ago, we're going to say, "Oh, that seems like a long time ago," but it really wasn't compared to now and how far things have changed.

Yeah, I think it kind of goes with the AI slop, culmination of AI intelligence getting better and better. But yeah, the Cloudflare CEO kind of talked about how 10 years ago, Google crawled maybe two pages per visitor, meaning, you know, it got a decent amount of information and it was able to have good results for two pages. Six months ago, compared to 10 years ago, Google maybe six pages to one visitor. OpenAI, on the other hand, is 250 pages to one visitor, and Anthropic is 6,000 to one. Supposedly today—I don't know if these are 100 percent accurate, but I think it gives you the gist—is Google is now crawling 18 pages per visitor, OpenAI 1,500 pages per visitor, and Anthropic's, 60,000. So across the board, like 10x, 20x, it's pretty insane.

But the TLDR is the web is being scraped at an unprecedented scale by these AI agents because people are hopping into ChatGPT, hopping into Claude, asking, "How do I change my tire?" click the web search icon, and go. And, you know, that in the background, Anthropic, OpenAI, whoever, is going to Google, to any other reputable search engine, scraping the crap out of their content. So, you know, probably typing a few queries, seeing the top 10 pages, clicking on each of the links, reading every page from that result, then summarizing that and giving you the answer. It's great because if you think of a human workflow, I would do that myself. If I had a question of how to change a tire, I would go Google it, go look at the top three, think about what that might mean, and apply it. Now I can go to AI and do that.

And so it becomes this kind of issue where AI is so well equipped to search things. Like, is it even worth it to create your own original content? If I'm going to be one of 60,000 pages that Anthropic scrapes to answer a user query, how do I ever get real traffic? What happens? I mean, Google is even pushing the Search Labs-like AI overview. If I ask anything or type anything in Google, I get this immediate blurb before I see any search results that, you know, Google has done the heavy lifting for me. So we end up at this weird spot where people like writing original content traditionally because it appears on Google and they'd click on it. Now it's being summarized by AI, so it isn't even worth it. Now we're getting into the AI slop era where people are using AI to create content that's being consumed by AI. And now it's just, you know, feeding on itself. So I don't know where it's going. It's something that... I have a personal blog I write from time to time. I enjoy writing on it, but I'm not so much a publisher, but I can feel as a large publisher might be thinking, "How do I get value, money, out of the original content that I write?" Because today, these AI agents and tooling and, you know, ChatGPT-like interfaces are really taking out the manual process of reading things. And you don't really get the raw value as a human anymore. Now it's just filtered through AI agents.

Yeah. Well, and it's interesting too, because it's like, you know, I just used Gemini earlier today because my AC unit is busted. And so I took pictures of the AC unit and the capacitor and just trying to kind of make sense of what's going on and why this is not working. And it gave me a response. And like, if you think about what it's doing when it's responding, you know, it's basically generating that response out of a body of knowledge that it scraped from the internet, right? Like it read articles, it read manuals, it read all these things. And, you know, those people really don't get the credit for that. And so there might be kind of a, you know, that's kind of an issue, I think, because if you had a search engine, like a traditional search, you would click that person's page or that company's page. And, you know, hopefully that was original content on their part and they didn't plagiarize it from somewhere else. But at least in that case, like, "Oh, you went to, you know, acrepair.co and saw their content." Whereas now with those kinds of AI overviews, it kind of just says, "Hey, don't worry about that. This is what you need to know."

And so I think it's one of those weird things where I think it's better for the consumer. Like, it's better for me to kind of get the answer. Yeah, but at the same time, you know, Google, are they—and this is just more of a philosophical question, it's not picking on Google specifically—but, you know, are these companies taking advantage of all the content that's been put out there to kind of build this body of knowledge, which they do charge for, right? Like they charge monthly subscriptions to use ChatGPT and use Google Gemini and stuff like that, right? So, you know, I think the contrarian point of view is like, "Well, these companies put that stuff online for free. So, you know, all Google is doing is what you and I would do, like just read the article." But it does feel, it feels a little bit unique.

And I think to the point about AI slop, that's why it kind of ties into what we've said at the beginning of this podcast. Like, if you're making content these days, it has to be video, I think, or at least audio to show that you're not just repackaging, you know, just other people's written content or AI slop, you know, because I think everyone's seen those LinkedIn posts. And it's almost every other LinkedIn post. Like, I could go on a rant about this, but it's just so much crap that you see out there. Um, and everyone knows it's fake. Like, everyone knows these are bots. And, um, so if you want to show up authentically, you know, either you write really original content, like you have your own unique point of view and you're in your own perspective and your own tone of voice—that can be kind of hard to do—or you do video and audio like we're doing here and show people that you're a human, that you make mistakes, that you trip over words, and that you're just a real person and not just a robot.

It's going to become more valuable to have a little mix-up because if you're like me and use LinkedIn, like me and Ben, the one telltale sign of AI content, I swear, is an em dash—so kind of the long dash, not the hyphen, like the minus sign, but it's almost like a double-length dash. And usually, I don't even know how you even get to that key. I don't know. To be quite honest, maybe you hold it down or shift or control. I think that alone tells you that if you see an em dash—like, it's literally e-m dash—that is probably AI content. So for those of you trying to get around the AI detection, maybe ask AI to write your content, detect, you know, look visually yourself. Does it have an em dash? If it does, remove it. Because now literally anytime I see that, I think, "AI content."

There was a post on Twitter a few months ago talking about how the top Reddit posts all had em dashes over the past six months. And they're like, "Well, people are going to AI, you know, superintelligence, and asking it to write an amazing, incredible hook to go popular on Reddit." And it does well. It clearly does well. And I don't fault it for that, but it just doesn't feel amazing. I don't know. I don't know. You know, the search engines, I don't want to use them anymore. I want to use AI to get answers quicker and faster. But then you come down the spiral of, you know, it's going to start reading AI-generated content, which is probably not super accurate to begin with, and it keeps going and going. I don't know where it bottoms out. But we're at the current stage where the internet is mostly original content, and now it's shifting to be more AI-generated. In five years, I don't know. It's going to be all AI-generated probably, and it's going to have a whole host of new problems to detect original content. And at that point, I imagine the video models would be so good I can just have a mini-prompt of like, "Hey, create a Breakeven Brothers podcast, episode 25 with Ben. Figure out the content." It'll go scrape the web, fill us in here, and we'll be done. Okay, MRR, we're done. So it'll be a fantastic future.

Yeah, it's really funny. It reminds me of one thing, and then we'll move to bookmarks. But I remember seeing on YouTube, I got a YouTube ad, and it was someone talking. They made it look like with their background and their mic setup and the angle that they were on the Joe Rogan podcast. They had the red velvet curtains behind them, and it was some... you know, and that's what it was. I think it's for some just like Ozempic-type medicine, but it wasn't legit. It was clearly like some stupid ad that someone was paying for to run. And but yeah, they tried to make it seem like they were talking on the Joe Rogan podcast to kind of, of course, you know, impress upon a viewer and a gullible viewer that they were someone of influence and of importance, right? So it's just the tactics are so, so sneaky and clever. And that's not new. You know, the whole phrase "snake oil salesman" was around way, way back when, right? Because people try and do these little tricks and sleight of hand to get you to open up your wallet and buy something. So it's not new in and of itself. But you know, the tools are new, and people need to be internet-literate to be able to really sniff out what's BS and what's not.

Yeah, I know these companies do internal phishing. And I imagine if they did that internal phishing with modern AI tools, it would be quite hard to detect and it would catch a lot of people. But I think it would be so effective at some point, even doing that. So, yeah, be careful out there.

I saw, just on the topic of phishing really fast, I saw this hilarious meme. It was, you know, the Ryan Gosling meme where he's at the Oscars and he wins, but he's confused and surprised. Have you seen that one?

I have not seen that.

Then it won't make sense. But for those listening that have seen that and you can envision it in your head, the caption, the meme was, "When you pass the company phishing test because you don't read any of your emails." It was him, surprised that he was winning something.

I just looked it up. That totally lines up.

I'll find the meme. It was so good. But yeah, cool. Awesome. So should we switch over into bookmarks and start wrapping this one up, Brad?

Yeah, I can start. So I found a kind of an intelligent bookmark, which I love to share, the little, you know, quick wins. So OpenAI basically has their transcription API, which allows you to take audio and transcribe it to text. Adrian, who actually I think is the founder of Simple Analytics now that I look at it, which is kind of like a Google Analytics alternative, but essentially his tip of the day, which I am resharing, is: run your audio file through FFmpeg, which is a command-line tool that just does audio and video operations. You can run one command to speed up your audio file by 2x or 3x. Therefore, when you send it to OpenAI, you're being charged less because they charge by the minute. So, for example, if we had our 60-minute podcast and I sped it up by 2x, it'd end up being 30 minutes, and I'd pay half the price because OpenAI is still able to transcribe things accurately, but it's shorter content. And if you use timestamps, that's kind of the top comment of his tip is, you know, times it by two, times it by three. Hopefully, I think the math works out in that. But yeah, if you want to save a little money transcribing your text, speed it up locally, send it to the API, and maybe you can build a business off that. That's half the price for content. So yeah, thanks, Adrian, for that awesome tip.

That's cool. $50,000 SaaS right there. Vibe-code it. Awesome.

Cool. My bookmark is from actually a good friend, longtime friend of Brad and I's, Justin Anderson, who's a fellow accountant. So automatically puts him in the W column right there. But he's been posting some cool stuff specifically from the OpenAI cookbook. And I think he actually posted another one from Anthropic's. So he's been posting a lot of good stuff. Follow him on LinkedIn or Twitter. But what he posted that I specifically bookmarked was "Data Extraction and Transformation in ELT Workflows using GPT-4o as an OCR Alternative." So obviously a big mouthful of a title, but in accounting, OCR, which I think is Optical Character Recognition—don't quote me on that—basically being able to, from an image, identify numbers and text. And basically, it's really important in accounting. The classic example is receipts. So like someone takes a picture of the receipts. You want to tell, you know, using OCR historically, find out what the charges are and how much they are and just import it into our accounting system instead of having me look at the receipt and hand-key it in manually.

So what was cool with the cookbook that Justin had linked was it talks about using 4o instead of a structured OCR workflow to just get that data that you need off the receipts. And so, and that's actually something that I have experimented with personally as well, because I was working with, just kind of prototyping, making an agent that is in Slack where you can Slack it an image of your receipt and have 4o read the contents of the receipts. So identify the vendor, identify the amounts, and then also use, you know, of course, an MCP server to get a listing of different expense codes that we have and basically use that context. And of course, a system prompt of like, "Hey, I need you to look at this receipt, identify the amounts. And then once you have the amounts and a description of what they are for, identify the expense codes that they belong to." And you give it examples to help kind of really refine, almost like machine learning, like pre-training. I wouldn't call it that, but you just give examples and it kind of helps with accuracy down the road. And so, and that worked pretty well. Like, it actually worked really well to kind of pull off the data you need from those receipts. And it could be, you know, a crumpled piece of paper that was laid on a table, or it could be a very nice, clean, you know, fully bordered shot. And it did a good job either way. So definitely worthwhile checking out. I haven't read the full cookbook, but I did read the abstract. And when I read it, or when I saw he had posted it, I was like, "Oh, I've done a little bit of that personally with my own kind of stuff." So I think it's really cool for accountants to check out because there are a lot of use cases in that kind of stuff. I mean, we still surprisingly do a lot of things on paper in accounting. So it's good to have an interface to bring that into the digital media.

Yeah, that's awesome. I also fight with receipts on my bill-splitting app, and I know firsthand how important it is to have a good photo. I've seen users upload shifted photos, and it doesn't work. It's out of my hands because that's AI processing it. So if you do take photos of receipts, just a little quick tip, make sure it's square, cropped, flat. For all the people out there who think computers are that good, they are just looking at it with the best of their ability.

That sounds personal. That doesn't sound like a PSA. That sounds like a favor that you're asking. We'll not talk any more about that one, but let's move on. So we talked about a lot of things today. I think overwhelmingly the AI CLI battle is an important and big one for me. So if you have not used Claude Code, I will repeat myself every episode from here on out to the end of the year until it gets dethroned. But please try Claude Code. You can do the pay-as-you-go model initially. Try it out. If you have the bank and you're using Cursor or Windfall, I would give it one month and spend $200 and just give it a full run. I think you'd be impressed and surprised at how good it is. And secondly, you can do a lot more than just writing code. So if you have tasks that you don't want to do, you have configuration on your computer you want to fix, you know, even Apple macOS settings—like all the macOS settings can be configured through the command line. So you could just ask it to do things. And I think as someone who's inching closer to giving Claude more and more control and doing more things for me, I would encourage all of our listeners just to try out Claude Code and give it a task that you wanted to get done, but you felt like it was too much effort. That's outside of coding. You know, kind of let us know what that is. I'm very curious what our listeners are doing with Claude Code, if you're picking it up, and what unique use cases you're spending your tokens on outside of the coding domain, because there's a lot, I think, that people have not realized we could do with it yet.

Yeah. And I think some of the best ways to get ideas or get inspiration is from people sharing. So, yeah, definitely let us know. Leave it in the comments. And while you're at it, give us a five-star review if you're liking what we're putting out there because it means a lot to the show and to Brad and I. But yeah, let us know what you're working on, what you end up demoing and trying out, because it's always good to hear what people are kind of seeing, how they're seeing things, and how they're experimenting with things.

Yeah, just post a video on Twitter and tag us and we'll take a look.

Yeah, for sure. Cool. All right. Well, let's wrap that one up, Brad. Good stuff. And yeah, we'll do it all again next time.

Awesome. Sounds good. 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.