Breakeven Brothers

Alrighty, we are back with episode 28, almost at 30, of the Breakeven Brothers podcast.

How's everything going, Brad?

Excellent. We need a sponsor, so for those listening out there, for episode 30, I think in the 30s we're going to get a sponsor. I'm going to be doing an ad read. I mean, we pitched Claude Code, we pitched Cursor—like, literally knocking at the door for a sponsorship. So for people out there listening, our viewer base is growing. I think we're a great candidate for a sponsor read, if I do say so myself.

I have breaking news because I'm not sure if you know, but we do have a sponsor for this podcast. It's this app called Split My Expenses, and if you haven't used it, go check it out. It's really good. If you ever need a bill-splitting app, check it out. It's the best on the market. So, thank you for sponsoring this podcast, Split My Expenses.

Yeah, I appreciate that. I heard they're coming out with the mobile app soon. So if you've been not wanting to use it because the mobile app isn't there, you have no more excuses. It's coming out, I don't know, in the next few weeks with an asterisk of, you know, we'll see how things go, but we're making good progress. I've been grinding with Claude Code and Codex, all the darn tools, so yeah, stoked about that.

Is that being built on React Native, do you think? The mobile app?

Yeah.

Okay. Okay, cool.

Yeah.

Cool, that must be a great app then. Awesome. All right, well let's just jump in. We have an action-packed agenda here. I wanted to start with a study or an article that came out related to a study from MIT. I'm not sure if you saw this.

Oh gosh, I have seen it.

Yeah, so the study was that—or the headline, I should say—was that 95% of organizations studied in this MIT study got zero return on their AI investment. So, the quick version, and then I want to get your thoughts on it, but basically, they studied 300 public AI initiatives. And among those, 95% showed zero return on investment. And this had a trickle-down effect or impact in terms of the market and investors' belief in whether we are in an AI hype bubble.

And I think that was also furthered by Sam Altman. I didn't see where he said this, but I guess he made a comment that there is a bubble, but he said specifically in privately held AI companies. So I think, like, just generally, probably AI SaaS startups, not publicly traded companies like, you know, Google and Meta and all that. It didn't seem like that's what he was referring to, but nonetheless, after this MIT study came out, I think there was a noticeable downward impact on the stock prices for some of these companies like Nvidia. But what's your take on that? Is that surprising to you, or is that in line with what you'd expect?

Well, it's actually the second study that's come out to make AI seem not as good for productivity as people would expect. So there was one two to three months ago. We probably talked about it on the podcast, but it had described 20 or so engineers using AI and them not feeling as productive as they'd expect to be using AI tooling. I think that was using old AI models, and "old" being, you know, they change very frequently.

So that study, it came out with a damning headline just like this one did. And I looked at it and people talked about it. It was kind of dismissed, given that there was a large jump from, like, the 3.5 Anthropic model to the 4.0. So you look at that study and you think, "Oh, maybe it's just use cases." I think they had sourced people from Big Tech, and the AI coding tools also aren't as good for a large code base and difficult historical problems within code.

But this study came out, and again, it has a jarring headline of, you know, "AI isn't as good as it's touted to be across all these avenues." And I haven't looked at it yet, so I need to go look at the details because I think when you look at the details, then sometimes it falls apart. But my initial reaction is my guess is that these people probably aren't, you know, using the bleeding-edge models. They probably aren't writing the best prompts. And again, I say this without having looked at it, so I'll have to look at it and report back on the next pod, but this is my assumption. Because I use these tools, and they're an insane level-up. Like, the amount of code that's been written for my mobile app with React Native using Claude Code has been astronomical, and I've only gotten better at it. And the reason I get better at it is because I'm on Twitter looking at how to use these tools from others, you know, sharing their tips and tricks. And I'm using the tools myself, so I know where this lies. So I guess, long story short, my hunch is that these people are using AI not in the most productive ways, but I need to go take a look at the study to really confirm that hypothesis.

Yeah, well let me hit you with some good points and good questions. Let me hit you with a couple of different parts from the article. This is from Axios, who published an article about the MIT study. So they specifically, and these are about AI pilots. So I think take that with, you know, it's important information because a lot of times pilots, you know, they're not guaranteed to be successful, right? It's like a trial. But still, 95% showing zero return is pretty damning.

So what they found was that the majority of these initiatives and these pilots that were being studied were in the marketing and sales departments. And what MIT found—and this is no surprise to me when I read this—is that the application of AI in back-office functions generates significantly better ROI than in marketing and sales. But I think a lot of times companies see, you know, like a customer chatbot basically, or they see like a lead-gen AI that can go out and get leads automatically. I think that's where a lot of eyeballs and time is spent. But, you know, speaking as someone in accounting, there's so much opportunity to automate the back end and other functions like accounting. So when I saw that from the study, I was like, that completely tracks, because there's tons of opportunity.

But the other part that I wanted to bring up, because you keyed in on this actually pretty well, is that the study found that the lack of return on investment was not due to model capabilities. So it said that often times the model was completely smart enough and capable enough to do the task, but the reason—and this again, this is from Axios, this isn't me—the reason is that there was a learning gap with the people within these organizations moving these products forward who did not understand how to use AI properly or design workflows. Which I think is kind of what you just said, basically, of like, there's a gap. And AI has moved so fast that there's a huge gap between what it's capable of and what decision-makers are aware of its capabilities to go ahead and put forth, you know, whether it's new products or whether it's AI agents that are helping automate certain parts of the business. So I thought that was really interesting and again, completely what you touched on already: these models are capable. It's not that they're not capable, but it's the plumbing that goes on that needs to be hooked up right.

Yeah, there's a ton of skill in actually prompting these things well. Specifically with GPT-5 that came out, there's been new knowledge that GPT-5 is really darn good, but it's good with a good prompt. And we had touched on that in the last podcast, but I think it's only become more prevalent when people are using it in Cursor, in Cursor CLI, and various forms and methods. These models are getting really good, but as we talk about "good," it gets sliced and diced into various flavors of speed, intelligence, and steerability. And I think Claude Code does a good job at understanding what you want to do, but doesn't take you too literally. Whereas I think GPT-5 takes you literally and does exactly what you want it to do. So if you're good at describing that, success. If you're not good at describing that, you won't like it.

And I think when people are using these tools, specifically outside of engineering, they're not following the latest best practices that OpenAI posts, like a cookbook or a prompt optimizer. No one's going to read that outside of engineering, and even inside engineering, not a lot of people read that. So it's very much that I think the frontier of intelligence and AI models is there. Do we know how to use it and unlock its full potential? Definitely not. And I think if I was doing research in the AI space, like the folks that published this article, I think it'd be a really fascinating way to say, "I want to research AI's impact on productivity." I think the study and the results, and probably the way they went about it—it's different when you're in the know and you're not in the know. And I hate to partition it into those two segments, but I feel like there are a lot of fast-moving pieces in AI, not only the model releases but just how to use these tools effectively and have enough time to actually use them. You have to spend a lot of effort and time, and you know, I invest time and effort and sometimes it doesn't pay out, sometimes it does. But if I was included in that study, I think I'd be a strong voice for saying, "Hey, this has increased my productivity." I'm not included in the study, so obviously I don't have a say, but I don't know. I think there should be a more targeted way to get folks who are much more in the know about the tools and know how to use them and get those people in a specific study. I think maybe if we take a look at the study in general and say, "Does the average person know how to wield AI?" I think that gives you a clear answer of "no," and I don't think I'm surprised to hear that.

Yeah, and also, I think it's how people use it and what people's expectations are of it. In the article, there was a quote, I think it's from a Wharton professor, and I've seen him—I think his name is Ethan Mollick, something like that. Apologies, I don't have it up in front of me. But the article quoted him, and I think he's one of those top voices in AI. And he said that basically the lack of adoption is because people are asking ChatGPT to do parts of their work and they're having trouble with the AI showing its work consistently or it's not being reproducible.

And those are things that—I'm not trying to toot my own horn—but those are things that we've identified on this podcast as things you shouldn't be using ChatGPT for. And I think there's a difference between AI being able to help you with admin-type work, like drafting better emails or managing your calendar—great. But then there's also a different point of automating workflows. Like if you have well-defined processes that you just need to get done, like it's a series of steps that you just need to do, it's repeatable, it's very cut and dry. That's something AI can do, but you're not going to be able to probably do it just by—at least right now, and I don't know if you ever should—just having it be prompted. I think that's better suited for a workflow. And you know, I've talked over and over again about things like LangGraph and, you know, using OpenAI's agent development kit to create actual workflows as tools. And so I feel like that's something that I don't see a lot of. I think a lot of the discourse is around prompting and, you know, just pretty much prompting. Whenever I see corporate materials around how to use AI, it's always about prompting. And I think prompting is an important piece, but that's not everything. We don't see MCP talked about a lot. We don't see, again, workflows being talked about a lot. And I think that will come, but the pace of AI has just so significantly outpaced the ability for people to keep up with it, you know what I mean? And I think the study shows that.

It's difficult for us too. I mean, I spend a lot of time on it, and there are people that spend a lot more time than me on it who are probably a lot better at it, and I try to follow those people on Twitter and get those insights. But yeah, it does paint a picture of expectations being sky-high. The pace is extremely fast and cutting-edge. And if you just take people who maybe aren't as on the bleeding edge, you probably won't get the best results. And it does matter, depending on the function and the task and the expectations.

So I think with all that said, I expect a few more of these studies to come out, and I expect it to affect stock prices and expectations. I think there's a lot of people who kind of sit in this AI antagonist camp, and I think when those articles come out, you know, that is like tangible proof for them to say, "Oh, you know, AI sucks and it's not that good." And yeah, AI's not great for everything. I've had my issues with it. I've yelled at Claude, ChatGPT, any assistant. I've had my fair share of debacles with them, but they clearly are a level-up. And if someone took that away from me today and told me I couldn't use it ever again, I don't know, that would be a huge, huge loss. And I think that alone, the premise of when you use it, you're like, "Yeah, it's pretty good." But if you take it away, that feeling of, "Oh, I'd never be able to use Claude again or ChatGPT again," that feels like a big loss. I think maybe it's not the perfect way to frame it, but for me, that value is clearly there. So do I agree with the study? I mean, maybe it's just framed in a different way. I think it'd be great to have a power-user AI study and maybe talk about how high the productivity is being increased versus if it's being increased at all. Because I think if you choose people who are closer to the tooling, have more experience with it, it's always a "yes," but how tall is that "yes"? How much weight does that "yes" carry? You know, is it 20%, 40%? That would be an interesting study that I'd like to see.

Yeah. Two other things on this, and then we can move on to our next item. But one, I think it's important to define what's at stake. And the reason I say that is because in the Axios article, it mentions that tech investment, like capital expenditures on tech investment and AI, is the highest it's been since 2000, which we all know what happened in 2000. So there's just so much expectation, you know, and so much belief that these investments will have some kind of future payoff down the road. And, you know, I think when we think about stock prices and studies like this, it's about what is that real expectation? Like, what's hype and what's grounded, actual future payoff that we're going to see.

And then the second thing that I would just want to mention too, because you said the "AI antagonist," it's worth mentioning—and Axios again does a great job of pointing this out—that the study was published by MIT and they have a certain group, it's called Nanda. It's like network agent, something something. I don't know exactly what it stands for, but basically saying that there is a vested interest. You know, MIT wants people to use this Nanda protocol. I don't know if you'd call it that. So, you know, it's just worth mentioning. It's not that the study's bad and not that the study is invalid, but it's always good to point out where different camps are in terms of what they're pushing out there and how it relates to the information that they're putting out to the public. So yeah, that was really interesting.

So yeah, I'll take a look at that. And one thing I wanted to chat about was—we had kind of interfaced on Twitter about it—but there's been a little bit of an increase in selling Cursor rules. And I feel like this ties in nicely with that paper that was released because it feels like you need the knowledge to be able to make AI effective and useful for your task. At least for engineering, there's a lot of effort to put in to understand how to wield the tools. We've talked about prompting, we've talked a bit about Cursor rules. I've seen a trend recently where both the Tailwind creator, Adam Wathan, and this guy I follow who does a bunch of Vue stuff whose name I can't remember. But essentially, both of these people are leaders in their frameworks. So they, you know, own Tailwind. This guy does a ton of stuff for Vue. So Vue.js is, you know, the counterpart to React.js.

And essentially, both of these people, again, have a ton of clout, have a ton of expertise in their frameworks, and they're packaging and selling Cursor rules, so to speak, to make your experience using these frameworks a little bit better. I think it's interesting because the price point, at least for the guy who was selling the Vue rules, was like $9. And $9 is not a lot, you know, when you think of creating a course or other kind of educational content, those are usually, you know, $100, $200, $300. Seeing the price point so low felt interesting, but also smart to me, where I was kind of chatting with Ben offline, it feels like one of those things I would buy and never use. So it's very interesting seeing these domain experts come in, try to make money off of AI, but also package something that could be very valuable because Cursor rules are somewhat valuable if they're written right. Other times they can have a lot of fluff, but these are people who are extreme experts at their craft, writing rules that are tailored for their framework. So I thought it was interesting. $9 is very cheap. I almost bought it and I didn't because I kind of got stuck in this loop of like, "Oh, it's only $9, but would I actually use it?" So I don't know. Do you think this will be a trend going forward?

You know, at first, I didn't care for it, to be honest with you. I saw it and I just felt like Cursor rules, in my experience, don't make that much of a difference to me. And I'm not a super-user in terms of vibe coding. I'm not a fan of vibe coding. I tend to try and get all the real important details out myself and then maybe just use AI to fill in some gaps. But that's just me, and I'm probably a way slower developer because of that.

But at first, I didn't like it. But the way that you described it actually made more sense to me, in the sense that you're paying for really specialized knowledge. And it's not just generic rules in terms of Cursor rules. It's very tailored to, you know, in this case, Vue. I think his name is Eduardo. I just looked at the tweet. And to me, the only people that are going to be buying that are people that are really, you know, dead set on building something with Vue. And so if it makes their job easier, makes their life easier, then $9 is actually probably a pretty good buy. And you know, hopefully it's a good product and the rules are good in the sense that it does what it's promised to do. And I can see that gaining traction in other areas too. And it's just that specialized touch. And for $9, maybe that's worth it. I think it's actually pretty interesting. It reminds me of Udemy courses because I've seen one that's on sale for like 10 bucks. I'm like, "Hey, I'll do that," you know? And even if I watch like...

Do you actually watch it?

You know, I got one on FastAPI recently, actually, because there are some things that I don't really understand about FastAPI. I've never used it as a production-type application. And so I wanted to do that, but you know, I bought the course and there's stuff that's like basic Python. I don't need to watch that. There's stuff about setting up a virtual environment. I don't need to watch that. So, you know, you might only get half of the value or half of the course might be relevant to you, but for 10 bucks, why not? And I feel that's how I feel about this.

The price point is extremely low and attractive. One thing that had recently come out is Laravel released Laravel Boost, which is kind of an opposite approach. It's packaging a free, open-source rules system. So I guess how to explain it is, in Laravel, you can use different packages, like Composer packages, and Laravel evolves over time. So there's a Laravel 10, 11, and 12, and across these different versions, the framework acts differently. So the Laravel team developed a package called Boost, which essentially looks at your project and sees the dependencies you have, and then will pull in appropriate Cursor rules or Claude Code rules for the specific packages you're using. Because AI always struggles with SDK upgrades and version upgrades. They made it easier to say, "Hey, if you're using version three, I'm going to have these rules. If you're using version four, I'll have these other rules." And so they give you an easy platform to say, "I'm going to install this Boost package. It's going to go look at my Laravel project, see the dependencies, and then auto-detect the versions." You know, it could be old, could be new, and then have a better footing where AI can understand your project a little bit better, because these things do change over time.

So we're talking about Tailwind and, you know, Eduardo—or I guess Adam and Eduardo, both in Tailwind and Vue—using their expertise to craft something. And then we're talking about Laravel, which is kind of the opposite approach of crowd-sourcing these rules and taking a look at your project. And I think that approach feels right, but it's not really the craft and the expertise of one person behind it where you know they have the chops because they built the framework. Like you couldn't argue that Adam doesn't have a great understanding of Tailwind because he built it, you know? And same with Eduardo, he built half the Vue stuff. So it's really interesting to see this dynamic of people trying to monetize in the AI space. And, you know, no one's going to try to compete with Claude Code or try to make a better model than Anthropic's model. So where do you make money as a creator or maybe a lead engineer on a project where people want to get some knowledge out of you? And how do you provide it? Maybe that ends up being Cursor rules, because I could see more framework leads coming out with, "Hey, these are the rules I would write because I know the framework in and out," which, you know, could just be extra money.

Yeah, and I see that they have affiliate links, so we'll definitely think it's a great product and you should go buy it. No, I'm just kidding. Um, that's cool though. Yeah, you know, I think it's something that, in this, you know, I don't want to say AI era—I think that's too dramatic—but in this time where there's just so much AI and it's making jobs change pretty drastically in certain ways, you know, for someone that has specialized knowledge in an area to have... You know, I'm sure he has a job, I'm sure he has a primary income source, but to have another source of income... You know, he just—I don't know how much time is involved in managing Cursor rules. But, you know, I'm sure the idea is that he publishes the Cursor rules as a product, people buy it, and he can be one-to-many in terms of customer touch and all that kind of stuff, right? And so I think it's smart. I think it's smart on his part. And I think again, the people that he's probably targeting are not just your average front-end developer. It's probably people that are really like diehards in making things in Vue. And again, nine bucks, why not? Why not take a shot at it, you know what I mean?

So yeah, I would love to see the stats on how many people bought it. And even on the Tailwind side, I think they've also been looking to make more money in a recurring way, but this is definitely not recurring.

I think Tailwind is a one-time purchase. I know Eduardo's rules are a one-time thing, but it's interesting seeing people try to fit monetization in with AI because, I don't know, I feel like I would buy Laravel rules from Taylor Otwell if he put them out. Maybe, you know, that's just putting it out there. If you're listening to the pod, you know, you got some Laravel Cursor rules, $10, I would buy it.

Yeah. Yeah, it's interesting. We will see.

Yeah.

Cool. So we talked a little bit about the MIT study. And in that study, I mentioned that the stakes are really high because the tech capital expenditures have been the highest they've been since 2000. Well, kind of in related news, we talked last month, I think, about Meta's billion-dollar offers and how they're poaching top talent with these crazy lucrative deals. Well, apparently, and reportedly, that is being put on ice by Meta, and there is a hiring freeze specifically in the Superintelligence AI division. Um, and so it's interesting because it seems like they opened that door, maybe got the talent that they wanted, and then slammed it shut. So Brad, I'm sorry to say that there might not be any billion-dollar offers headed your way from Meta.

I was just going to say. But one note I have to say: "put on ice." Man, you should be a news reporter. That was clean.

I am, dude. I am. Here we are, right? I'm reporting the news. So yeah, so basically it's, again, being reported that any exceptions to the rule will have to be specifically approved by Alexander Wang, who we've talked about, who was part of that acqui-hire. So it's interesting to see. Again, I think the timing of it's very interesting because, you know, we had the MIT study, we're seeing this hiring freeze go on with Meta. Are people starting to tighten the belt a little bit and start to be maybe more grounded in terms of what they're going to, again, expect out of AI and how much money it's going to cost and how much money it's worth putting into it, you know?

Yeah, I mean, I'd be surprised. They hired so many people. If they were able to pull off taking this large chunk from, you know, OpenAI, Anthropic, even DeepMind, and bringing all those people to Meta, having them work well together... Everyone has their own chunk of the pie. Sounds pretty tough to do that. And, you know, I've worked at Meta. There are a lot of people at Meta. It was a very performance-driven culture. People did projects for performance, and it was very much a topic of conversation throughout the year, and there's a lot of pressure and a lot of scrutiny on why you are doing things.

And I imagine in this age where people are getting paid hundreds of millions to be, you know, acquired as an asset, that performance culture must really add up here. And I think the people who probably joined will be under an insane amount of pressure, but they're getting paid buckets. And I think on top of that, I've also seen tweets, I think within the past two days, that some people who have joined Meta are actually leaving. I don't know if you've seen too many of those, but I think I saw at least two that were like, "You know, I joined Meta for the mission of the superintelligence lab, and they're going to do good stuff, but I'm going to go do my own thing." I don't think they really disclosed why. My guess is they're probably going back to OpenAI or wherever they may have come from or starting their own company.

But I think in general, they're spending so much on data centers; the Nvidia GPU spend is off the charts. The acquisition cost for these, you know, quote-unquote "assets of intelligence"—of these people training the models, post-training, etc., all the deep research on creating a good AI model—all that's adding up to a lot of money, and you can only do so much with that. Like, you can't just keep spending and spending. And Meta had done that for Reality Labs for many, many years. But at this point, the Meta stock is in a really good spot, you know, 700-plus. Zuck has this new kind of energy and charisma and is really putting his eye on AI. But how far will that go? When will it fall apart? You know, we've seen so much news, but how is it going to tee up?

And I think now we see the first glimpse of, "Hey, it's not perfect." They've got to figure things out. You can't just hire all these geniuses and put them in a room and have greatness happen. I'm sure we'll see great things, but it's not as simple as that. You need to have, you know, structure, order, a roadmap—all these things that are critical to delivering success with lots of people. They're going to lose people, they're going to have a hiring freeze, there's going to be chaos. I expect good things to come out of it over time, but I'm not super surprised to see that they want to halt things, get things under control, realign themselves, and probably move forward. I expect them to keep hiring, maybe in three to six months, but for now, yeah, probably just a short pause is my guess.

Yeah, and it's interesting because, you know, I think a couple... one of these questions is going to be really open-ended, and I'm just curious what you think about it. But I did want to point out that just because you throw a boatload of money at something doesn't mean it's going to work out. And I think—and this isn't to be critical of Meta, it could happen to any company—but, you know, there are a couple of more notable recent, maybe... I don't want to say "fails," that's probably too harsh of a word, but tries that didn't work out. So I think the Metaverse just didn't hit the ground like maybe they originally thought it would. And then the other one that comes to mind immediately is—I'm not sure if you even remember this, I guess you were probably there when this was happening—but Libra, the stablecoin project. So those are two, and I'm sure they didn't put as much money into Libra as they did into the Metaverse, and as they are with AI and Superintelligence. But, you know, it just goes to show that you can throw a ton of money at something and it doesn't mean it's going to get traction and be successful.

But the other point that is interesting is, you know, those offers are so, again, they're just so distant to my world. And frankly, I don't think I'd ever want that kind of offer and the pressure that comes with it. But at a certain number, does it not matter anymore? And therefore it just becomes like, you know, if you're just doing something for the next dollar, it kind of goes back to that mission and purpose, right? Like, do you see yourself with Meta's mission, or do you feel like you're a missionary like OpenAI claims to be? And so I wonder if people are finding that... you know, if the people that are leaving Meta reportedly are finding that, "Hey, this hundred-million-dollar increase isn't worth it. I don't feel that tangible increase in my worth." And, you know, maybe that's why people are leaving. What do you make of that? Like at some point, you know, you've heard the phrase "more money, more problems." Do those big offers and the expectations that come with it—and maybe you don't really see yourself super-grounded in Meta's vision—does that start to eat at you? Even though you're getting paid an amazing amount, does that become not worth it in the long run?

I do think Silicon Valley has a different perspective on pay, and there is a larger focus and sense of worth based on take-home pay. And so I imagine, you know, not to paint these people a certain way, but it's just the culture of being in Silicon Valley that has more emphasis on that area. I think people derive a lot of, again, self-worth from how much they get paid. And once you get to a certain number, yeah, it's not really meaningful. Like at a certain amount, you know, if these people are getting paid 10 million at OpenAI and they got paid 100 million at Meta, I would argue that that's not meaningful to their overall life quality. Lots of goals in life, for the most part, are very achievable at $10 million a year.

So yeah, I think to answer your question, it's definitely different in Silicon Valley. However, I think they would probably agree that it's not meaningful. You know, if you 10x your pay from 10 million to 100 million, you know, it's a pat on the back and you're in the hype cycle of the Meta superintelligence team, but I think the pay on paper is enticing, but it's not the end-all, be-all. My guess is it was more the vision and seeing people next to you go to Meta and thinking, "Oh damn, I'm missing the boat. They're going to go create something great." And then they come along for the ride. And again, if it's a hundred-million-dollar offer, it's probably 25 million a year, which is also insanely high, but they're not getting that money immediately. I'm sure there's a sign-on bonus that gets clawed back if they leave.

So in the end, I think for people who have joined Meta for the mission and potentially for the pay, if they see things are awry and just chaos, I imagine it's very easy to go back home. You know, OpenAI wants them. It's very easy to say, "Hey, you know, I apologize for the chaos. I'm back here with your mission," and you get a glimpse of what Meta is like. Not to say it's a bad place—I enjoyed my time there—but I could imagine with the intense focus on AI, it can probably feel kind of crazy. I can only imagine with that much money and that many important people who all think they're important—who they probably are—but it's just probably the talent density and the intense focus from the company that creates an environment that might be chaotic and not the most fun place to be. And I would expect at OpenAI it's probably also chaotic and crazy, but they're not flinging around news articles about them paying people 100 million and 200 million. Like, you probably have OpenAI stock that's worth a lot, but it's a different beast.

And so I think the grand scheme of things is, I'm glad the people who have joined Meta have the courage to leave because, you know, you could feel like you're stepping away from something big. But if they came from somewhere where they were already making big things, like OpenAI or Anthropic, it probably feels like they're coming right back home. And we had talked about some of the Anthropic folks going to Cursor, and then they left like two weeks later. So if there's anything I've learned about the AI space, it's that these companies and these employees are very transferable. Like, the people who will hire them—it's an extreme skill, an extreme asset for these people. And they can go wherever they please. And so the door is always open on each side, it sounds like.

Yeah, and it's hyper-concentrated in Silicon Valley. You know, it's been Silicon Valley for the last, you know, 20-plus years now, but specifically within AI, I can't think of a time where I've seen such a concentration, like within five zip codes, of this is where everything's happening. And, you know, you're on-site or hybrid; it's not a bunch of remote work typically at these places because you're probably expected to be there and be clocking the hours. But yeah, it's definitely a fascinating talent space in that hyper-niche area around AI development.

So yeah, it'd be sad if the age of the Meta offers is over. I was hoping to snag one of those, maybe for 200 mil, but maybe next time when they start hiring, this podcast will push us up there and we'll be a dual package, two people for 100 mil. That'll be okay with me.

Yeah, you'll look back on it and say that was a steal, honestly, seriously.

The other kind of big AI news is reports that Apple is considering using Google Gemini for the underpinning of Siri, the next iteration of Siri. And we've talked before about Apple and the lack of AI, and it seems like it's been a miss. And so, a couple details on that. This is being reported by, I think, Mac Rumors, Tom's Guide, 9to5Mac, and TechRadar. So it's not 100% confirmed, but these sources do cite that basically Apple is considering—it's doing basically an A/B test of a proprietary Apple AI model versus Gemini. And they're calling it Lynwood, which is Apple's model, and then Glenwood, which is Gemini. But still, they're finding that Glenwood, the Gemini model, is outperforming Apple's model. And so it's something that they are reportedly thinking about using for their own product. Which to me is just pretty wild.

I think, said another way, for so long, probably the last 20 years, in the hardware space, it was Apple iPhones versus Android, you know, Google's Android. And there was, in my opinion—I'm a Mac person—like Apple had really good domination on that front. But you can see that they've just missed the boat a little bit on AI. And I think in one of the articles, I'll try and find it, they said—and I think they found a direct quote, or they got a quote from someone that worked at Apple—that supposedly a couple years ago, Apple didn't think that generative AI would really be a big thing. And if that is true, that kind of goes into the history books as maybe one of the poorer takes of the last decade, and maybe that explains why they're so much on the back foot. But what do you make of that? To me, it's fascinating because it's almost... I don't want to say "humble pie" because I'm not saying that Apple lacks humility, but they've just had such a dominant position for so long that having to go to Google and use their AI model... it's got to hurt a bit. It's got to be a bit of a sour taste.

Yeah, I mean, Google puts out good AI models. We've been singing their praises for Gemini, you know, 2.0, 2.5, 2.5 Pro, Flash, Lite, whatever. They're pretty darn good. They're very cost-effective. And they have an extremely good free tier, which Apple would obviously blow through. You know, they released that paper what, like two or three months ago, talking about how hallucinations happen and how they don't believe that LLMs are the key to unlocking greater intelligence. I think they were saying it needs to come in some different flavor. So maybe they're kind of not investing in AI and wanting to use Google out of the box for Siri, just as a short-term measure. But maybe they're going to spend more money investing in some sort of new, non-transformer AI model, which again, there were AI models before LLMs came out, but LLMs unlocked a whole bunch of use cases with text.

So maybe Apple is spending their chips on something more next-level. And Apple, you know, as they work on these devices like the Vision Pro or the Apple Watch, I think they're created like five to 10 years in the making. Just a ton of hardware engineering, a ton of resources thrown at it in secrecy. Who knows? Maybe the AI division of Apple is doing something in secret that is really impressive, hopefully. I have my doubts, but it could be the case that they are working on the next level and realize that they're late to the game on the LLM side and are just looking to have Google fill that gap in the meantime because, yeah, again, the Google models are great and very well-priced, I guess.

Well, yeah, and that actually reminds me. It's one thing I forgot to bring up on the MIT study, not to come back to that a third time, but there was one important point that was addressed. And it's that companies that bought AI tools were far more successful than those that developed their own proprietary model. Now, that was again for the 300 different pilots, but that makes sense in the idea that these, let's say four—you know, Google, Anthropic, OpenAI and, you know, maybe just three.

Deepseek.

Yeah, you know, so maybe just three, but they have such a lead over everybody else. Now within those three, it's very competitive, but they have such a lead, and so trying to develop to catch up to them or build your own version of it just seems like a fool's errand. And that you should just be using these models at this point until, you know, maybe in the long run you can eventually build something that's more niche for your needs, but again, they're so good right now.

And it reminds me of this Apple situation where, you know, Siri is, in my opinion, completely useless. To be honest with you. I don't use Siri. I'm not sure if you do.

No.

Um, and Apple Intelligence is almost like an oxymoron because I've never seen it be useful in any kind of intelligent ask. So, you know, they almost have to do that. But, um, yeah, so they're making a decision several weeks from now. But yeah, it's something that supposedly—and I'll try and find the article that talked about this—there is an internal battle within, I think, the Siri department between two leaders. And one's a bit more, you know, what's that phrase, "move fast and break things" or whatever. And the other one's much more methodical and, you know, maybe slow to the punch. And that power struggle, I guess, has led it to be, um, kind of aimless in the Siri department. And that was a quote from the article, not something I made up.

So yeah, that's it. Those are tough words right there. Wow.

Yeah, yeah.

Apple's been under a lot of heat recently.

Yeah, because it was... I'll find the article. I want to make sure we link it because it said that the AI team was nicknamed internally as "aimless" because they just didn't have any, yeah. So things are not going well there. But I am confident that once they get it right, they will, you know, because all these different models and all these different things we're talking about are not really hardware-specific. Maybe you could say Google has a great one-up on everybody else. You know, they have a one-up on OpenAI, they have a one-up on Anthropic because those two don't have hardware, you know, they're just strictly models, whereas Google has the phones, and Gemini is in—if you have an Android, you have Gemini basically built in. And so, I do think once Apple gets it right, it's going to be sweet, you know, it's going to be... yeah, it's going to connect to your watch, connect to Vision Pro. Like, it's going to be good. They just need to get there, and I think they will.

That's what everyone thought for Apple Intelligence, though, to be fair. I mean, they've sold that thing hard for the last... I think it was the last iPhone launch, maybe it was two iPhone launches ago. Probably like two years ago. It was like, "If you get iPhone 14 or something, this has Apple Intelligence." And that was the selling pitch and the marketing pitch, then Apple Intelligence fell so flat. Everyone has dismissed it. So at this point, the expectations for anything Apple AI are rock bottom. If they would integrate Gemini, that would level them up immediately. I'm just slightly concerned that, you know, Apple would want to actually go that route and sign a contract with Google to commit to spending, you know, who knows how much money on that, billions of dollars a year.

Yeah, one important detail, then we can move on. In the article, it's being reported that it's Gemini, but a custom version of Gemini that will run on Apple's servers, on Apple's infrastructure. So they would still have the commitment to privacy and on-device processing.

That makes sense.

Yeah, so it's not... yeah. So one important detail we should mention.

Apple Intelligence, they did create a pretty complex system for privacy where—I don't really know how to describe it succinctly—but they essentially have some various cryptography security mechanisms in place so that when you ask Apple Intelligence to look at your health data or other things, the more complicated tasks that process data will be offloaded to servers, but these servers are cryptographically secured so that your data won't be leaked and it knows it's talking to this server because it's sending that personal information from your device to the cloud. And it sounds like they're keeping that structure and they're making Gemini play fair and safe within that architecture they developed. So that's cool. We'll have to see where that ends up. I'm very curious to see if they sign a big fat check and then people love Apple's AI, or if it ends up coming back to bite them. We'll have to see.

Yeah, we'll have to see.

Cool. Well, one thing I saw very recently is there's a new terminal in town called Ghostty. And Ghostty has made waves on Twitter, being this extremely well-put-together terminal. So if you use the macOS terminal that comes out of the box, it's not that great, very clunky. There are plenty of terminals out there—Zed, Warp, you name it. So Ghostty is one that's being created as an alternative, and their creator, who is very, very famous, Mitchell Hashimoto—got to make sure I got the name right—he created a ton of open-source software like, I think it was Vagrant, he created HashiCorp, Terraform, all these kinds of very fundamental software pieces that are used in the modern day and age. This guy created and pioneered and pushed them forward. So it only makes sense that he would have created a very impressive terminal.

But long story short, Mitchell posted recently on the Ghostty tool, the CLI tool, saying, "AI tooling must be disclosed for contributions." And his riff on it is that he loves AI. Like, this is from his own words: "You know, I'm a fan of AI tooling. I use it myself all the time." But he's getting an increase in open-source contributions where people are taking the Ghostty CLI and just creating junk code. And we have the term for that, you know, "AI slop." He used the term, saying, "You know, I'm getting a lot of AI slop, and for me to review code that AI has written, it puts a lot of burden on me. Because of that, I want people to disclose that they've used AI for a PR." One of his key comments is, "You know, I like helping people get their PR, like their change, merged into my repository. You know, it's cool to get outside help, but I don't want to do that if I know it's just an AI on the other side. You know, it's not a real human trying to get this to work. This is Claude Code, you know, Codex, Cursor making a change, and he's not going to want to ask for changes, then have someone go and copy his message, take it to AI, and iterate on it that way." Like, he wants to actually help people.

So it hit the internet in waves. He posted this about last week, and his comment has 500 upvotes and only three downvotes, and then 11 laughs, eight tada celebrations, and 100 hearts. So I think when it landed on Twitter, it was like, "Oh my gosh, this is a big change." The reaction to this GitHub pull request that he just opened—I think he modified the readme. Oh, he modified the contributing guidelines. So he put text saying if you use Claude Code, you have to say in a disclosure, "This PR was written primarily by Claude Code," or, "I consulted ChatGPT to understand the codebase, but the solution was fully authored by myself." So, like, two different variants of that. But it's very interesting because I don't manage open-source software, but I use Claude Code five days out of the week, and it writes a ton of code. Some of that code sucks, some of it's good. And I have the knowledge to know when that is good and when it is not good. And I can clearly tell that he is frustrated with people coming to his repository creating changes that are not good, and they don't have the skill to know it's not good. So he says, "Hey, I'm done with this. I'm going to put the line in the sand and say, if you're going to contribute, please disclose this." And I think it's a change for the better. It's very interesting.

Yeah, I agree. And AI slop is a problem. I think there's going to be a tremendous opportunity for engineers to come in and just fix all the horrible, vibe-coded crap that goes on. So there's plenty of work for engineers in the future years. But my favorite thing about that announcement and the tweet was just the tweet replies were incredible. Like, just the amount of snark, maybe is the right word, or just backlash is so insane because, you know, one of the examples was people were saying, "Oh, you just don't want people to submit with AI code." And he's literally like, "I didn't say that and that's the wrong takeaway." I think is what he said. "I never said that and that's the wrong takeaway." And there are so many other ones that he didn't even respond to where it's just basically people being like, "Oh, why is this even being discussed? Are we going to have people disclose their IDEs too? Like, where's the line?" And it's just like, just read the PR, read what he's saying. It's not what he's saying. So whenever I want to get discouraged about humanity, I just go to Twitter and look at replies to wholehearted, completely innocent, honestly logical posts like this and just see people's responses and go, "Oh yeah, there's still lots of brain rot out there," I guess, for lack of a better word, because yeah, it's crazy.

I think people really latched on to that headline, you know, "AI tooling must be disclosed for contributions." But if you actually read his communication, his thought process about it, and how he describes it in his actual changes, it's very straightforward. I think if you read that as a headline, your first reaction might be like, "Oh, that's odd," or "I don't agree with that." But then when you read it, it totally makes sense. So it goes to show that lots of people do not read things and will only take a headline for a headline. I'm curious if Mitchell's repository looks a little bit better now or if people are like—you know, because we saw all those poor responses or just odd responses—people are attacking him in any way to just continue down that path of being like, "I wrote this myself," and it's clear AI slop.

So it's very, very interesting. I expect it to follow. Like honestly, I wonder if Laravel, for example, a very popular open-source framework for PHP, if they have that same problem where people are just coming in, writing junk code. My guess is probably not. I think—I'm going to take a look right now—but I think Ghostty is written in Zig. Yeah, so it's written in Zig. I don't know if you've heard of the language Zig, but it's kind of a nuanced language, not nuanced in the fact that it's hard, but just not a ton of people know about it, and AI is really good at understanding any language. So I think he probably has a more complicated problem than others writing this open-source software in a language that not a lot of people understand, so they lean on AI. And writing a terminal—I guess his official description of Ghostty is a "cross-platform terminal emulator." So, you know, a handful of words there, but essentially a really complicated part of the computing stack with a complicated language that not a lot of people know. You know, I imagine it's AI slop all around for any contributions. So he probably had to put his foot down, rightfully so, due to the facts of what he's working with.

Yeah, just to call out, I think he's responded to these things fantastically. And I love reading his responses on Twitter. Again, they're just hilarious to me. Someone said "why," and then he said, "Sorry, this only applies to people who know how to read. As a first step, you have to learn how to read though. The PR contents explain it all. Reading is cool." And he replies, and he replies immediately after that. He says, "Oh, shit, you can't read this message." So, fantastic response by Mitchell and way to go. We're cheering you on.

That's pretty good. I didn't see that, but I just saw it now.

Yeah, fantastic. That's honestly better than the first message. So yeah, way to go. Cool. I had one last tidbit I wanted to chat about. And that is actually from my old stomping ground, PwC. So PwC, for those that maybe are on the engineering path and aren't familiar with them, they are a Big Four accounting firm, and they do audit and tax, but they also do advisory services as well. And they're a huge global company. And they made headlines, I guess, or published an article, I think this was actually a month or two ago. I'm not sure exactly, but it's made its rounds to me in the Twitter and LinkedIn algorithms recently. So I thought it was interesting because the headline is, "In three years, junior accountants will be doing the job of managers."

And so it's like, "Okay, all right, well let's see what's involved there." And basically, it's from the AI lead at PwC, and she's asserting that AI will be doing all the routine, mundane data entry tasks and that the junior accountants will be responsible for just reviewing that work and critically thinking about the outputs that they get. And what the article also stated was that because of this, they're going to have to retrain how they train, or reformat how they train their new hires. And again, speaking as a former new hire back in 2014, of course, there was no AI. And so it was much different. And so it'd be interesting to see, because they said they're going to emphasize things like deeper critical thinking and professional skepticism over technical accounting skills, because I think the presumption is that you should, one, already know that, but then two, that the AI is going to be doing that work for you.

Um, so I thought that was interesting because, you know, we've talked about "manager mode" before on the podcast, about how if you are really using AI to do meaningful parts of your work, you have to review it. Like you can't, speaking of Ghostty and that announcement, you can't just have Claude Code run and then just go and ship to production. It's not there yet, at least.

Yeah, it's not there yet.

And so there's going to be a need to have people review. And so I think that's smart that they're calling it out. But this is the first time I've seen such a public statement from the Big Four accounting firms about how it's going to change entry-level roles at PwC. What I found to be interesting though was that it said—and you know, I've heard this phrase a lot—that it's not going to be fewer jobs, it's just that people are going to be able to be faster and more efficient at their jobs. And to me, it's almost like that can't be true to a degree, because if you can be faster and more efficient, then you need fewer people to do the same task, right? Um, but, you know, they maintain that it's still going to be the same, but you're just going to be more of a manager when you come out of school and that you're going to be focused on reviewing the outputs of AI and stuff like that.

To me, it's interesting. It just seems like—and I largely agree with the statement—I think, again, people need to be more trained on review now, and the actual act of doing will be less important. But I just don't buy that at a large scale, you're going to hire new grads out of college and just have them review AI output. Like, to me, something doesn't add up with that. And I guess it stems from my maybe lack of a clear path of where that kind of accounting path goes in the long run with AI. You know, what I do believe is that people that can build, or at least partially build, AI systems stand a much better, excuse me, like a much better foothold in the job market than people that just consume AI. And I think I've said that before, where if you're just consuming AI, that can be anybody, right? If you're just taking the output of an AI, why would they need you? And I think that's where, sure, you might be a CPA, you might be a trained accountant, but there are a lot of those out there. And we might need fewer of them because of the automations and AI that are going on.

So to me, I'm not sure what to make of all that yet, but I do think it makes sense that they're going that track. But what do you see in terms of engineering? Of course, coding has been one of the areas that's just taken off. Like, you can generate so much code so much faster, and it's such good quality. You know, do you think that changes how engineers will be coming out of school and what the job market, what skills will be sought after for engineers in the job market as new hires?

Yeah, I mean, first off, it sounds extremely interesting because I've heard of this phenomenon where younger people using ChatGPT will use it for things you wouldn't expect, where it'd be like, "Oh, I go to Burger King and I ask ChatGPT to choose the food for me." Maybe a weird example, but it feels like they're delegating a lot more to ChatGPT than one would think. You know, I use it for house tasks or researching coding frameworks, other various things that I feel like I don't know about, versus like, "Hey, should I do something that seems trivial, maybe to you or me?" I think to them, they kind of offload to ChatGPT because they think, "Oh, you know, this is a super-intelligent being and it's clearly much smarter than I am. So why would I make the decision?"

So it's interesting that it's positioned that way, as a manager, because I think truthfully, the folks that are using AI in this capacity that I described will probably be much more in tune with how to use it. And again, consuming AI versus writing a prompt and building workflows is vastly different, but I think people will be a little bit confused initially. Like, will they be ready for it? I think maybe they will be because they use it so much already. At least, you know, for the few folks who are very ChatGPT-focused.

I think for coding, I don't know, I would hate to say a new grad for coding could just use Cursor, because I couldn't tell you how many times I have asked it to do something, saw the AI slop, and said, "None of that, you know, that's not what we want." So to me, that's impossible. Again, I think the models will get better, but as we get this divide between, you know, Opus, the grand thinking model of Anthropic that does really well, or Sonnet that does really well, and then we have GPT-5, which again, writes really good code but you need to know how to prompt it. It feels like Claude's version of the model is more beginner-friendly. It feels like GPT-5, where it knows exactly what it needs to do if you can describe it, is very advanced-user-friendly. You would never give a new grad GPT-5 and let it run rampant on your code base because they won't know when it's going wrong. They won't be able to spot that. There are fewer guardrails in place, whereas I think Claude models are a bit more understanding of what you actually want and will do those changes, where I think GPT-5 is a bit more like a very precision knife.

And I think when I think about new grads being prepared, I feel like it does shift to reviewing things. Like, you're reviewing a lot more code because AI is writing it, but since you're reviewing it, you need to actually understand it. And if you're not coming in with that baseline understanding, you can't review it. And you get to that point where you're faster, but you need to know the skills, so you can't really get away with taking a shortcut. So I wonder in accounting if you could do that potentially, but to me it feels like to review the work, you must know the work. And at that point, you are becoming a manager of some sort, but you still need the fundamental skills there.

Yeah, to me, I think it needs to change—and I would imagine this for engineering too, but it'll be curious—the change needs to start with education more. Like to me, even when I came out of school with a bachelor's in accounting and I was on the Big Four track, you get on the job and you don't know anything, and you kind of have to learn on the job. And that's kind of expected, I think, or at least that was the expectation. Like, "Hey, you just need to have the foundation, but then you're going to learn on the job how to actually do things." And there's a big gap between the practical knowledge of what your employer needs you to do and the theoretical knowledge that might underpin what you're doing. But theory is a lot different than practice, often.

And, you know, I think in hindsight, if I were to go back and make the accounting curriculum today, I think you would need to put so much more emphasis on systems and data and learning how to set these things up, like how a general ledger gets stored. Like I had one class on accounting information systems, and it was a great teacher, nothing wrong with the class structure, but it was focused on Microsoft Access, which I never saw after that class. And there wasn't really any SQL. And, you know, so there was just a big opportunity back then. I'm sure it might be better now, I'm sure it is, but it just seems to me like the education needs to be so much more grounded in practical skills for you to come out of school and be ready to go, you know. Otherwise, you would just want to have a more seasoned person in those roles. That's the part where I'm kind of, as I sit and think about it, I'm like, "I don't know what that will look like in five years. Like, what will the education look like? And then how does that change things for people that are just graduating to get those roles?"

Yeah, I think it shows, at least in the computer science side of software engineering jobs, there's not really a market for new grads, and it's because AI has gotten to the point where it can write code at the new-grad level. I mean, it can definitely write far past that, but on average, I would pin it at that, maybe at the minimum, so to speak. But they don't have the skill, like you mentioned, of on-the-job learning and understanding of when things go wrong and how they go wrong. And because they don't have that, they're not as in demand. Not all companies are hiring new grads anymore, which is a stark contrast from when I graduated college, when it was very much like, "We need you, we need everybody, we need everybody now."

And as we're in the current age of AI, that is not the case. It's a complete 180, and it's kind of sad to say because as I've helped new grads prepare for interviews and such, I very much pushed them towards using AI because that is the future. You need to be positioned for the future, and that is what the landscape is for software engineering. So I imagine the roles being different in the future and how that will affect new grads versus more senior accountants and engineers. It's going to follow a similar path in terms of, you need to know this stuff and you need to know how to use AI. How you package that in the education system so that you can get hired in the appropriate context, I'm not sure, because it kind of comes to that dilemma of, "Oh, you need five years of experience for a new-grad role." You know, like a stupid job requirement description like that where it's like, that's impossible. You know, how do you do that? How do you pull that off? So I think we'll end up in a spot where there'll have to be changes, and it could be that you work with AI and that becomes a proving ground to say, "You know, I'm not only a new grad, but I'm a well-versed AI new grad," and that becomes something where they're like, "Oh, we want to hire you," not the traditional new grad who has the degree but no AI experience.

Yeah. And I think it's not enough, to me, to just try and get better at prompting, because that's what I see whenever I see AI training, it's always just basically prompt training. And I think it's important, but I think there's so much more to AI and being able to use AI than just prompting. I know, don't get me wrong, it's important. Like in the article with PwC, they also talked about KPMG, which is another one of the Big Four, that they have a retreat for their new grads for two days. They go to their training ground in Florida and they learn about AI. And when I kind of went and read through the article, it was pretty much just prompting, you know, and how to prompt it differently. And the use cases that they cited were drafting emails and doing industry research, which, okay, that's cool, but there's nothing novel about that. And, you know, it's a good starting point, but there's so much more to it. And I think once you stop at prompting, that's a mistake. There's still a lot more to learn.

And I actually had some people ask me, "What should I learn about AI? Should I do a course on AI and all that kind of stuff?" And I told them, I was like, "I think honestly your time is better spent learning SQL, like learning how data is stored and how data is managed." That's just my own personal opinion, not speaking for other accountants, but you know, because to me, learning about AI... I think it's more important to understand the systems that are in place. Like, how is the data stored? How can you query it? Of course, you can have AI help write your SQL queries, but just knowing the foundation of that, to me, is more important than knowing how to massage ChatGPT to give you an answer in a certain way. That was kind of my podium or my soapbox, if you will. So, you know, it's going to be an interesting time in education for sure, and for new hires. So we will see. The future is fast approaching us. It's already here.

Yeah. Cool. Well, that was a lot. I know we talked about a whole lot there, but I'm ready to dive into bookmarks. And the one that I have is from Peter Steinberger. I think I've brought him up before, but he is, kind of in my head, the lead on AI tooling and AI workflows for engineers. So he has spent oodles of time just tinkering with every known AI tool, AI IDE out there. And he kind of builds in public or talks about this stuff in public, saying what works and what doesn't. And it's great to follow his tweets. I have a lot of intention in following his updates, but sometimes it's overwhelming and he talks about a lot of things. So what he does is he writes blog posts maybe one or two a week. So a decently high frequency. One he put out actually today was his current AI dev workflow, and he uses Ghostty, Claude Code, and minimal tooling, which is what he calls maximum productivity.

And he underlines "less is more." And I love that because I think as you go through the AI tool rabbit hole, you say, "How can I write this crazy prompt? How can I use sub-agents? How can I use blah, blah, blah, blah, blah?" I think you get to that point where it's hard to measure your productivity gain using all these little tips and tricks and little things. It feels overwhelming and not very productive. Like, you spend more time setting up the tools and trying to squeak out 1% here versus just getting things done. And yeah, he essentially boiled it down to just use Claude Code, use a good terminal, and review the code. Other than that, you don't need any complex setup or tools or layers and layers of stuff. No worktrees, no crazy things like that. Like literally just use the tools, you know, yap to your computer using AI voice, and you're pretty much good to go.

I like that because I truly trust his opinion. He's built a bunch of stuff. He's sold companies. He's very much an engineer at heart, and he's spent a ton of time with the tools. So whatever he says is not really the gospel, but I trust it in the fact that he's spent the effort and the time. And as we've talked this whole podcast, it's about having to spend the time and effort to learn these things. And as you use them, you learn more about them. I think that's where I lean on his expertise: he's spent more time than most doing this. And due to that, he knows what works and what doesn't. And he has that engineering background. So all things considered, give it a read. Basically, he distilled that simple is better, and here's what he uses. Um, I think it's a fantastic resource for people to just follow and get up to speed quickly.

Yeah, yeah. I saw he had a nice post about the MCP servers and how much token space they take up.

Yeah, a ton.

Because MCP, I love MCP. Like, I think it's awesome, but it's good to know, especially as agents come out more and they start to use MCP and you have limited context windows. Then yeah, that was good stuff.

So cool. I have a late audible. I actually had a different bookmark, but I swapped it out for this one because I just saw it today, and I think it's really interesting. It is a tweet from someone named Marc Colbrugu. I don't know how to say his last name.

He's very much in the Levels.fyi category. I've seen this guy a ton.

Yeah, must have some kind of Nordic name with the little "o" with the two dots above it. So apologies, I butchered your name. But um, Mark, we'll just call him Mark. So he put a screenshot of Cloudflare that shows there's an option now to block AI bots by default. And so I thought it was interesting because, of course, the whole reason we are here, I guess, is because for years, these models web-scraped all the content on the web and built up—that's the foundation of the AI models. I'm generalizing a bunch, but that's essentially how it happened. And, you know, we talked about how the models are trained really well on PHP because they were scraped on those kinds of sources and stuff like that.

So it's interesting to see the walling-off that might start to occur. And basically Cloudflare, which is pretty widely used—like, it underpins a lot of sites, right? So we'll see how much more that gets turned on, this "block AI scrapers," and see how that impacts maybe businesses or just engineers that are reliant on AI scrapers. But more holistically, I do—and we've talked about this before—like at some point, I can see these different models starting to limit access to certain other areas. So for example, you can connect your Google Drive with ChatGPT, but at some point, will Google say, "Uh-uh, we're not letting that connection exist anymore because we need you to use Gemini"? And that, to me, I could see it going there. I don't know if it will. It reminds me of the Apple App Store gamesmanship, maybe for lack of a better word, where you have to play by their rules to be on the App Store. I wonder at some point... it seems like AI is very open and you're able to use different models and connect to different things, but will that start to close off? And to me, this Cloudflare one is interesting because it's the first time I've seen it really block AI scrapers, you know, and yeah, it's interesting.

I think there's been discussion recently about Cloudflare blocking Perplexity, which we've talked about. It's kind of a weird state where all these AI chat apps like ChatGPT or Claude or Perplexity are all scraping the crap out of the internet right now using various methods. And Cloudflare underpins, you know, at least half the internet, is my guess. And they are kind of sitting in the way of, do we allow this or do we not? And they are the bandwidth people. So as people request data from your site, it goes through Cloudflare. They pay money for that through their servers and infrastructure. So they kind of have an incentive to reduce AI bots to save themselves money, but it hurts the experience. And yeah, it's kind of a weird contention right now where they own the internet, they have the control to almost turn this off. But should they? Why would they want to do that? Do they want to just save money? Is there something else in play here? Like a vested interest. It's hard to say. It's kind of odd right now.

Yeah. Yeah. It'll be interesting to see as it gets more and more competitive and, you know, they start doing some gamesmanship between the different providers and the competitors in this arena. So yeah, it'll be cool to see.

Cool.

All right. Well, let's wrap that up. Good stuff, Brad. And, uh, we'll do it all again next time.

Awesome. Sounds good. See ya.

See ya.