Breakeven Brothers

[Breakeven Brothers Podcast Theme Music]

Cool. Alright, we are live. Happy Wednesday, Brad. We made it halfway through the week. We are back with episode 35 of the Breakeven Brothers podcast. What's going on, dude?

Not much, but happy new year to everybody out there. This is the first episode of 2026, and we made it. We started in what, 2024?

Oh, that's a good question. It must have been summer 2024. June 26th, 2024. That's crazy. That was a good episode, "Quitting Big Tech." Man, how we've come full circle. If you haven't listened to the first episode of our podcast, you definitely should give it a listen. We might have a few more "likes" and "ums" in there—and we still have some—but it's one for the books.

But yeah, doing well. New year, lots of changes, lots of AI advancements too, honestly. I think every week I'm sending Ben bookmarks, and it's my reminder to go back and look at them for the podcast. So I was just catching up on things, but there have been lots of changes. One thing that's been hitting the headlines recently, which Ben actually sent to me earlier today, was the Tailwind Labs drama. And so for those who might or might not be familiar, Adam Wathan, the creator of Tailwind, has changed pretty much all of web development by creating Tailwind CSS, kind of revolutionizing CSS in an easy way. There's been some fire, some scrutiny.

Essentially, someone created a PR on Tailwind CSS saying, "Can we add specific instructions for agents?" This would allow agents to essentially parse the Tailwind docs better. So they call it an `llms.txt` file. Instead of an agent going and fetching HTML from Tailwind, it would actually hit this single file which just has all the markdown of all the documentation. And Adam posted a comment yesterday rejecting the offer to add this file because the Tailwind business is struggling. In that comment, he had mentioned that 75% of their team was laid off.

Yeah, big news. Hate to see it. I for one have used Tailwind CSS for all my projects, and I even use their open-source software for their UI components called Headless UI. So their engineers created that, published it, it works great, no complaints, but yeah, a little tough to see. Adam is a smart guy, obviously, and for him to go through this business change with AI, it's the start of a shift. So I'm curious to hear your thoughts.

Yeah, I mean, it was... I come at it from a different lens because obviously, I'm not a traditional programmer by nature. But you know, being familiar with Tailwind, it always gave me fits dealing with the front end. That's not any reflection of Tailwind; it's just a reflection of me and my skills. But that was the library to use for so long, and so many front ends these days are built with it. And I think from my perspective, looking at that company and looking at Adam, who's pretty active on Twitter (on X), it was actually super surprising to see. And he was pretty honest, I think, in the GitHub comment. I think he mentions that they had to lay off 75% of the team yesterday, and this was posted yesterday, so it must have been on Monday, January 5th. And then also saying that Tailwind is growing faster than ever and it's bigger than it's ever been, yet our revenue is down close to 80%. Which is just a shocking number.

To lose 75% of your staff, to have revenue down 80% while you're technically growing—and I'm sure in large part due to all the "vibe coding" and, you know, just the mainstream appeal of building things using Tailwind. But for me, it was like, wow, someone who's as smart as Adam, who has this business that's been around for a long time that just probably printed money at certain points in the journey, I'm sure... This quickly can be turned upside down. And you know, it's interesting because I was thinking about this when you and I were talking about when we first started the podcast. And I'm sure in June 2024, things were probably not this way with Tailwind. Things were probably still pretty profitable, still doing well, I'm sure. And it's just that quickly, in that short amount of time, in 18 months or whatever it is, that we've arrived here at this point where people are vibe coding, people are doing more and more stuff with AI.

But now we're starting to see some real... and I think for you and me, it's more tangible because we don't know these people personally, but we've followed them for a long time, been connected to them for a long time. And seeing the journey of the last six, seven years, it's crazy to see it happen to people that you kind of know or follow. It feels different than, like, "Oh, General Motors lays off 10,000 people." That's horrible and you feel bad for them, but to see it happen on a smaller scale with people that you kind of know, it's like, "Wow, this is really here, and here we are in 2026."

Yeah, it's a weird divide because it powers so much of the internet. I think the team size was four people, if I was reading his comments correctly, so just to get more context, 75% is three people. Still significant for their workforce, but it is crazy to think that something that was so successful and something I recommended so much—which is their Tailwind UI kit, which I used to build Split My Expenses; the UI code there is literally built with Adam's product—becomes more or less obsolete overnight once these AI models hit a critical threshold.

And I know for a long time Adam has wanted and talked about building a SaaS, and I think he had tried a few things. Whatever, it might not have worked out the way he wanted to. So there were attempts, I think, to steer the business in a different direction, but my guess is 2026 rolled around—business reviews, doing your taxes, figuring out profit and loss statements—then it kind of hits you. Like, AI is only getting better. If this is what my revenue looks like, if this is what my costs are, it's not looking too good.

So I'm glad he course-corrected. I applaud his transparency and honesty, being in a tough position where he's rejecting what seems to be a good contribution from the open-source community due to the fact that the business is struggling. So it's kind of at that divide where what's good for everybody else is this PR gets merged. What's good for him and his business is that it doesn't get merged. And so you get this weird divide.

I think what it boils down to is if you're in a business where text output—specifically coding text output—is the core business model. Like what I paid for Tailwind many, many years ago, I was paying for pre-built UI components. Something that looked pretty that I could go copy and paste and put in my web app. And it worked wonderfully back then because we didn't have the Opus of today. But if that's your entire business model, it's a little tough. I mean, coding agents have come so far in six months, so far in two years, that you really need to have more diversified income streams at this point. I'm sure he's obviously realized that by now and way before this, but at some point, it's hard to come up with a new business out of thin air when things were working so well. It's like a complete 180, I'm sure, over the past two years. So definitely a tough time for him.

And on top of that, there was additional news in the same realm that Jeffrey Way, the creator of Laracasts, which we all know and love in the Laravel community, also laid off, I think, one or more of his talented instructors. Originally, Laracasts was kind of the "Netflix of Laravel content," so educational content. Over the years, he had hired really talented Laravel developers to come in and expand the platform. And he had been tweeting actually in the past few weeks about how AI is taking over and how he needs to pivot the business of educational content to include AI. And then just recently, maybe a day or two ago, like pretty much the same day as Adam's situation, one of their top educators on the platform, who is really well known in the Laravel community, was laid off. And he had described it as "abrupt," where I think he was traveling on a plane, and when he got off, he was informed he was laid off, and then he got on his next plane for a layover.

So it definitely sounded abrupt, but if you kind of look at Jeffrey's tweets over the past two weeks, my read on it is, "Wow, things are really changing," in the same bucket of what I imagine Tailwind Labs is dealing with, where if I need education content, I'm talking to AI. I'm not going to Laracasts. I'm sure there's good content there, but it pulls me out of the flow, it does different things. It's just a different era. And I think Jeffrey has not published the official numbers of decreased revenue and costs, but my guess is that revenue dip is pretty significant. So significant that he had to adjust the team size and business size to make what's best for the business happen. So unfortunate news to have talented folks from Tailwind get laid off and folks from Laracasts get laid off when both Jeffrey and Adam are extremely intelligent people. It just feels like these businesses are in a tough spot given the age of AI.

Yeah. And, you know, I was thinking about the Laracasts layoff because, you know, I'm completely speculating here. I don't have any inside information, obviously. But for it to be abrupt, I'm wondering if Jeffrey, the owner of Laracasts or the founder, knew he had to cut some people to pay the bills and stuff like that, but didn't want to do it before the holidays. And so as soon as the holiday break was over... could be. You know, like maybe that was something. Or if it's not that, if it was like, "Oh crap, I don't have enough money in the business, enough money to run payroll for all these folks for the next month, I need to cut bait now." Like, I don't know what the situation was, but for it to be abrupt, it's interesting wording because it's like, you know, was he waiting until after the holidays or was it like, "Oh my gosh, we are really bleeding cash and I can't bleed anymore. This has got to go."

Because yeah, I mean, there's no good way to do that. You know, I'm sure he feels crummy about it. It's probably just one of the cons of being a business owner and having employees, so I'm sure it sucks for him. But yeah, it is interesting, the "abrupt" comment, but I suppose no matter who gets that message, it's probably always abrupt anyway, you know? You're just taken out of... you're clearly caught by surprise. Like, "Oh, wow, okay. Yeah, this is a lot to take in all of a sudden," you know, and in his case, on a vacation, right? So, yeah, sucks. Hopefully he gets back on his feet and finds something else to do. Sounds like he was a pretty talented instructor. And yeah, that's just the way it is right now, you know? And we're just in the first week or second week of 2026, the first full week of 2026. It's supposed to be the year of agents, right? Like, this is supposed to be the year. And you know, maybe we're already starting off with a bang on that front.

Yeah, it's interesting. I think the last thing in that pillar is Stack Overflow, which has been a repeated showcase of how AI is shifting what used to be the most popular developer platform. To get information, it's completely bottomed out. I think the number of searches they've shown over the past five years went up during COVID because that's when a lot of coding was happening. And then the second AI models took market share, the graph went completely straight downwards.

So, yeah, tough time. I think there's lots of reflection in 2026. I almost think that... I was doing a bit of spending reflection myself, you know, understanding what AI tools I pay for, because there's been so much momentum in that space that I felt like I was paying for probably one too many things. And I thought, you know, going into 2026, what do I actually use? Like, I'm a big proponent of Claude Code. I've started using Codex more, which I'll talk about later, but at some point, there's like this AI honeymoon, or "AI-moon" to call it, where you're like, "Oh, you know, I'll try that, I'll try that, I'll try that." And at a certain point, some of these costs have to go, especially when you look at a business where you're trying to spend on software that makes you more productive. Well, maybe I'll try a few things. Not everything sticks.

And so I wonder if 2026 feels like the first full-focus year of adopting AI across all sectors, where it's not just engineering, but we're talking, you know, legal, accounting, HR. Like, we want to use it everywhere because it's been so general-purpose. Although these might be coding models and coding agents, clearly they're applicable in other domains. So to me, it almost feels like it paints a different picture of these layoffs being indicators that business is shifting because we're trying to forecast what this might be.

I imagine for people like myself who spend all this money on AI tools, one, people will check in and figure out, "Do I actually need this?" And two, I think there'll be another wave of broad adoption for 2026. I don't know about you, but I feel at the enterprise level, you know, companies hear about the promise of AI. And it's so proven now from everything that we've seen that to me, 2026 feels like the year of, "How do we get AI in everybody's hands? No ifs, ands, or buts." Where 2025 felt like a testing period, to me, 2026 feels like we know what's valuable. How do we get people educated and onboarded and just integrate it into their flows and change their mindset to say "AI first"? How can we reduce some of these bottlenecks? Because we've talked about this so much in 2025: one, play with the tools because they change, and two, rethink how you work.

I think in 2026, it's not "play with the tools and rethink things," it's "you must do this or else, you know, someone's gonna take your job." I feel like that's a bit of an extreme, but you get the point of, you know, there feels like there's almost a deadline looming in 2026 where the AI expectation is here or being formed as we speak across all these organizations. And it's tough. I mean, I think a lot of people still are against AI, and it's a hard stance to hold in 2026 given the productivity gains. But yeah, interesting year, interesting start. Trying to paint a little bit of a picture of maybe this is a bigger signal for a bit of a change-up in spending and headcount and productivity and almost an expectations-setting heading into 2026.

Yeah. Well, I think, you know, one thing that you hit on was redoing the work that you're doing, but with AI, with agents. And I think probably comparing this year to last year, what I think has been a really big development is the agents SDK. So I know Claude has a really good agents SDK. I haven't used it, but I've heard only good things. OpenAI has one, which I have used here and there. You know, I've talked about LangChain, but I think the agents piece of it... like before, it was just chat and responses API, and now it's like, you have this agents SDK where you can build. Now you have skills, and then now you have MCP. It's been hit-or-miss, I think, but you can see the ecosystem start to come around to, "Okay, we can give this thing knowledge, we can give this thing context, and then we can give it certain actions that we need it to do for us." Like we can delegate to this agent, you know, how we do those actions in the form of skills and tools and stuff like that.

So it's kind of all come together. And I feel like for this year, there isn't really anything, at least in the knowledge work world, I would say, except for rare exceptions... there isn't really anything in that knowledge work world that it can't do. You just need to figure out how to glue it all together, which I think is the challenge, especially for non-technical departments. Speaking for accounting, it's like, you know, how do we get comfortable? We're so used to Excel, so used to being able to put our initials in certain cells and save the file. There's comfort in that, but that's not the best way to do it anymore. I think everyone would agree to that. So like, how do we... you know, I see these agents SDKs are out there. I see that Claude can work with Excel files or Claude can do all this cool stuff for me. How do I take what I do right now and put it into an AI where I'm also comfortable with how it gets the answers?

You know, one example that I was actually working through in my day job was some kind of reconciliation, right? And part of the build, we were discussing... you know, there was one camp that was like, "Let's just give AI the two data sets and let it reconcile, let it just find out the differences." And while I don't doubt that it can, especially as these models get more and more intelligent, it's like, no, no, I want it to be more of a skill where this is the code to run to identify the differences. And there's probably a little bit of nuance to that, but I want, you know, again, as the architect, I guess, of how I want this work to be done by an agent, I want AI in certain spots and I don't want AI in other spots. You know, I want to be selective about where I deploy it. So, but that's what it looks like this year. I think it's more about, you have all the tools, you just need to go out and learn it and also put it into practice because like you said, if you're not, someone else will, and that's the expectation now. You know, that's definitely what everyone knows it's capable of. You've just got to figure out how to put it all together, you know.

Yeah, an idea I just thought of when you were mentioning that was being almost like an AI contractor, coming into businesses, looking at their inefficiencies, knowing actually what Claude Code can do for you given its heaps and heaps of changes over the past few months, even the agents SDK, and kind of packaging this up. So, yeah, if you're really into AI and you are looking to help companies, I feel like that would be the perfect opportunity to distill what feels like a forever-moving stream of information about AI and make that actually actionable. Because there's a lot of overwhelm, especially on Twitter, especially just reading the news, of how fast these things move. If you can put your finger on it and say, "This is what's capable today, this is probably what's capable in three to six months as well," you're going to be in a good spot. And I think companies will be begging to have an AI expert just leading some of the charge here.

And so, yeah, I think we've talked about it so many times, but my advice, and I think Ben would agree, would be, you know, definitely try these tools out. 2026 is going to be a big year for AI. The competition was very hot leading into the past three months of 2025. So you can only imagine that as competition grows, these models are going to get better. The harnesses which run the models, the products which the models are embedded in, will only get better. And that's just, at the end of the day, really exciting for the consumer and enterprises trying to derive more productivity and value from it.

Yeah, and I was just, as you were talking about that, because the contractor thing was a good thought. And I was like, let me just go look at LinkedIn or a job board and see what an AI engineer type role is these days. Not like an engineer where you work at a company and you build out products, right? Not like that, but where you are being hired to specifically build out agent workflows at different places. And so there's a job post here, I won't read all the details, but you know, "Have a foundation in Python, LangGraph, Pydantic AI, Google ADK (the agent development kit)." So we're already starting to see this stuff in job postings. You know, "Be familiar with tool-augmented reasoning chains, you know, RAG, all that kind of stuff that we've talked about, MCP." So it's here, it's being sought after now. Whereas before it might have been more of a hobby and so cutting-edge that it wasn't really in enterprise yet. Now it's here. You know, it talks about integrating MCP in different places for this business. So, um, it's super cool.

And I think too, I'm curious, you know, what will... you've probably worked a lot of different places, Brad, and, you know, my experience in almost every place I've worked is there are lots of different systems around, and that in itself makes a huge challenge too. That becomes a problem of maintaining all the different systems, and, "Oh, this was set up here because so-and-so did it that way and it's just been that way ever since." I'm curious, in three or four years down the road, do we have a similar problem where there are just all these agents running rogue and people are afraid to turn them off because they don't know what it really does, you know?

Honestly, probably yes. I feel like that will actually happen because if there's anything about the revolution of AI and AI coding agents, it's that I for one have created so many layers of personal software where I'll create one tool, and the next tool will use that tool, and I'm just kind of building my own personal software stack. And I think at work, I would also employ the same behavior where, "Oh, I have to use this system, which is a really legacy system, not easy to work with. Let me create a small CLI tool to make that easier." And then now that I have that, I can invoke that CLI tool from various entry points. And oftentimes it becomes easier to create the software than it does to go through the official process to ask that team to fix it or find a new vendor.

So to me, I think it'll become even more important to have good documentation and have the ability to create software. But with that, I think you need to not be scared to make changes to other code bases. So I think, yes, there will definitely be agents that are critical that no one knows what the heck is going on, but, you know, it's one of those legacy systems, legacy agents in the future. However, with agents, we do get the power to make changes, even broad changes, pretty quickly. So, if anything, it should be a time where there's a lot more software churn. You know, the cost of software is really cheap now, which provides a weird market dynamic, but having good and reliable software... we're not there yet. You know, Salesforce isn't going out of business, for instance, which is the classic example of, "Let me just vibe-code a Salesforce replacement." So we're not there yet.

Yeah, or a Slack replacement, right? I think people want to do that for Slack because sometimes the Slack bill can be pretty high.

Yeah. Cool. One other thing I wanted to mention too, before we get into kind of Codex stuff, because I know you wanted to hit on that, is actually, it goes into the knowledge work and the AI stuff in the accounting world. And so they released—they, being TaxDome, I should preface that. TaxDome is like a document portal. It is like an organizer, like a CRM almost, for if you're a CPA firm and you do taxes. You know, this is a place where your clients can upload their docs essentially, and you can communicate with them. So it's like an all-in-one CRM document management platform.

They just, I think probably a month or two ago now, I don't know the exact date, but they released a beta, and I think they've been rolling it out more and more where you can send your client docs that clients upload. So say, you know, Brad, you're going to do your taxes with me, hypothetically. "Hey, I request your W-2, I request any 1099s that you have." You'll upload them to TaxDome, and I have these documents here. A lot of times the CPAs will then, either themselves or have a preparer, take the information in those documents and put them into their tax software, right? That's a different thing where you go and file the taxes. Like, you take the data from there and you put it in the system before you hit submit to the IRS.

Well, it basically—I don't want to downplay their tech stack, but I'm just trying to summarize—does document extraction to put it into tax forms. So they basically help you take documents, PDFs, images, you know, handwritten notes, extract the data points from that material, and then they can work and input it into your tax software. So I thought that was pretty cool because tax is one of the things that is... if you look at most tax software, not TaxDome, but the actual filing software, it's pretty ancient. It's like a dinosaur. It is the most, you know, 1995-looking software that you'll find out there. So accounting needs this bad. And I'm curious to see how well it works. So it's a beta launch. I'm going to try and see from folks that I'm connected with and that are going to use it for this tax season, how good is it? Like if I just give you a W-2, that seems pretty straightforward. Anyone can really probably pull information from a W-2 these days with AI. But is it good with handwritten notes? Is it good with receipts that have all kinds of different lines? So I'll be curious to see how it shakes out. We'll report back on the pod, you know, sometime in April or May after tax season to see how it went.

Tax season. Yeah. Nice. It's cool. I was just looking at their pricing, $700 a year per seat with a three-year commitment. It looks like they have a toggle. For TaxDome's pricing. It looks like they have a toggle at the top, one year, two year, or three year. Wow, that's crazy to lock into a three-year software plan for software that I hope changes, but maybe in that world, you know, things don't change as much.

Yeah. TaxDome's pretty cool. I will give them, I think they've done a lot of cool stuff. There are a lot of different players in that area right now with tax organizers and workflow organizers, but I think they're doing a good job. So, yeah.

Okay. It's worth it.

Well, yeah, I mean, I think the high annual price kind of was like, "Whoa," but I think, you know, if it's a thousand bucks for the year, what is that? Like $83 a month essentially, something like that. So...

That's true, yeah.

You know, like it's and, yeah, it's compared to, again, the context for CPA firms is that they're paper, dinosaur. So TaxDome is... and there are other firms that do it too. There's one called Carbon, there's one called Ignition. So there are other firms that do this, but it's all good. We got to get everyone out of the dinosaur age for accounting and bring them into the modern world. So, yeah, it's a little pricey, but, you know, I think it's on the right track. I think it's working.

Yeah, speaking of AI plus taxes, I saw on Twitter recently someone talking about feeding in their expenses and trying to get some California business forms filed through Claude Code. So, you know, dumping your expenses list from various bank accounts or credit cards and saying, "Claude, go for it." I think it was decently successful. I wonder how far you could get with that because I think I tried this the year prior where I was trying to just dump my info into AI. I think at this time it was Claude, just the normal Claude chat app versus Claude Code, where I think Claude Code has the skills to, you know, modify and create CSVs, probably also to fill PDFs, where I think the chat app didn't really know how to do that. So I want to give it a run-through this season and just see what happens. I'm not going to say I'll submit it because it probably won't get that far, but I'm curious to see how far these computer agents can get with it.

And in the same vein, I'm seeing these new usages for Claude Code. One that I wanted to call out is taking your ancestry data, so the raw DNA data, and bringing that into Claude. And if you heard about the news in the past few months, I think 23andMe was acquired by private equity and shut down, or acquired by private equity and turned into some menace, which people don't like anymore. During that time, there were lots of publications about getting your data out of 23andMe. I didn't really care too much, so I left my data in there, but I saw this post about using AI to talk about your genes and see your DNA. That's kind of cool.

So I exported my data from 23andMe. It takes about two days to get a zip archive of all of your genetic markers that they measure, so not everything, but a decent chunk. And I chucked it into Claude, and I will have to report back because I'll be quite honest, the original poster on Twitter was like, "Hey, upload into Claude Code. It'll do everything for you." When I had talked to Claude about it, it was like, "Do you want to focus on heart health? Do you want to focus on diabetes?"—like all these kind of key areas of risk that you can go look at a subset of your DNA. So I need to finish the conversation, so a little bit of a tease there, but these new use cases are emerging where we talked about Claude Code being the general agent, and I think there's a lot of "overhang," which is the term people are using in AI now, saying that these models are so good, we just don't know how to use them yet to find these new breakthrough advances. And one of those seems to be analyzing your DNA from private companies that collect your DNA. So maybe next pod, I'll give a breakdown of things I found that I'm happy to share.

But, you know, taxes, it's becoming more realistic. DNA, more realistic. Spending habits in general, I think there was a trending post from a user on Twitter saying, "I dumped all my transactions in, I asked Claude to use my computer"—which Claude has good computer-use skills now, like using Chrome—"and it chatted with, I think, AT&T or someone and reduced his subscription amount." So there are startups that do this bill negotiation for you, but now they have Claude as a general agent that can look at your spending, look at Chrome, and see what's typed from an agent and just chat with itself, like, you know, send chat messages. We get to a point where we have powerful tools, it's just how do you use them and unlock new things.

And so I'm getting the sense that, you know, even a while ago we heard about a few use cases, but there's been a big rise in Claude over the holiday break because they had doubled everybody's limits. I think they gave a week free as well in December that anyone could try Opus. And my Twitter feed has been on fire talking about people who have either never picked it up or picked it up again where they had tried a really old model but now picked up the latest one and are loving it. So long story short is I think there are so many capabilities in these models, and I'm glad more people are picking it up because there is so much more to learn from them. Like, if AI stopped right here, right now, on January 7th, 2026, we'd still live a pretty good life, I think. I think there's a lot to eke out in Opus 4.5 and Codex 5.2. So yeah, a pretty exciting world we live in.

Yeah, well, I think it's because the progress has moved so fast, faster than people can really keep up with it. And so I guess it still feels like we're uncovering, like, "Oh, this is really cool. You know, I didn't know we could do this." And it's like Opus, I don't know exactly when Opus 4.5 came out, but people are still discovering new things that they can do with it, you know? And people are still coming up with original ideas of what you can do. And so yeah, it's pretty cool. I think... shoot, what was I going to say? I lost my train of thought, but that's okay. We can...

Yeah, I was gonna say, my gut reaction to some of these things is one, I'm happy that I feel like I'm right, which is more of a self-congratulatory, you know, "Phew, I loved Opus and other people do too" moment. And then two is like, sometimes I'm like, "Oh, duh, it's really good." But then, like you mentioned, original ideas... I don't have all the ideas in the world. One person can only have so many ideas, and when these people talk about these cool ways to use Claude, I'm like, "Wow, I've been using Claude since May 2025. I didn't even think of that." Like I've been using it for months and months and months and this person picked it up over the holiday and thought, "Oh, let me try that." And that ended up being a winning combo. So then in the back of my head, I'm like, "We need way more people using this because people are creative." And when you have tools that allow you to be creative and build stuff, cool stuff comes out of that. And I think that part is the most important: talking about how AI can be used by anybody. If you have the idea, you can build it with Opus now, quite literally. And so just come up with these cool ideas, try it out. You can get pretty damn far.

Yeah. The part where I lost my train of thought, got it back on the rails, is I'm sure, definitely, I think Claude and Opus and 5.2 will be able to do the bulk of people's taxes, I think at some point, or at least advise on it, you know. If you're a regular W-2 earner and, you know, you have stock or you have retirement contributions, like if you're just in that kind of 80% of the population, I'm sure it can probably do most of your taxes or at least advise you on most of your taxes. I think if you're where you have multiple real estate partnerships and the operating agreements are different, like, yeah, you probably can still use AI, but I wouldn't trust it wholly, you know.

And I think it's interesting because a lot of the really high-performing CPAs and financial planners in my network, you know, they get paid a lot by people that are doing really well. So there are certain financial planners out there and CPAs that work with high-net-worth individuals and high-earning individuals. And they pay them lots and lots of money for advice that they might be able to get out of a ChatGPT, but it's really...

Don't say that too loud.

Well no, but it's true because where I'm going with that though is that their clients, the high-net-worth individuals and the high-earning individuals, they value, I think one, feeling taken care of, that this person's got it. Like they don't have to be the ones to kind of suss out if AI is right or wrong. They can go, "Okay, Bennett, you got this. I'm going to pay you. This is your problem to deal with. Just make sure I'm compliant, make sure I'm legal, and make sure my tax bill is really low." So I still think there's a place for that, I guess is where I'm going with this, even in the age of AI. You know, it's just the same as there are still certain things that you could do personally that you just now pay someone to do. I still think there'll be that for a couple of years at least. You know, beyond that, who knows? But maybe that's copium, but I think there are certain things where you're just going to want the human to do, you know? Like I'm going to a barber shop. I don't want a robot cutting my hair, not yet, but I don't want them with all the blades getting close to my neck. No thanks.

And even if you do pay someone to do your taxes in that high-net-worth scenario, I'm sure there's some expectation that you are using AI to fact-check things or, you know, look up new things in an effective way. So yeah, it's "get the result that you need." However, do it intelligently and be up to date. How you do that might not be their concern, but yeah, I feel like AI being front and center, I don't think there's a big worry about using it now.

And a lot of times too, just to end on this note on the tax thing, CPAs and lawyers can represent you before the IRS. ChatGPT can't. So if you take a position, you know, and the IRS goes, "No, you shouldn't have done that. We're not allowing that deduction" or whatever it is that you tried to do, you have to go seek out a CPA or a lawyer anyway. So right now, there are still regulatory reasons too for a CPA's value. I don't think that's the only value of a CPA is that they can represent you, but that's part of it. You know, if you go to TurboTax or H&R Block, their paid tax preparers can't really represent you in front of the IRS the same way a CPA or an attorney can. So, yeah, it's, you know, we're getting there. And I don't doubt in the future that maybe there'll be some version of AI that can represent you, but we're not there right now. So...

That was my next startup idea. Those Boston Dynamics robots are dancing like crazy on stage for these insane videos that are really choreographed. Maybe the next startup is an AI CPA that will actually walk itself to court if, you know, things go wrong, and it'll come on the microphone and say, "Hey, I did everything right." So, you know, if you're in that space and you're doing robotics and taxes, could be a nice hybrid. You never know.

Yeah. And it probably won't have to be very bulletproof. You could probably be pretty, you know, you don't have to invest a bunch. Like, you're not going to hopefully get attacked in court, you know. So, yeah, it should be something you could work on. You know, might be a market out there for that.

Interesting idea for the listeners out there, you're welcome. You heard it here first. We'll let you take it this time only, but yeah.

Well, one thing that I've been working on recently has been a lot more usage of Codex. Peter Steinberger has been a very loud voice on Twitter. He's the man who has been tinkering with things 24/7, 365. And his recent blog post was about how he uses Codex now instead of Claude Code. So there was a meetup in San Francisco back in October where I attended and he was there talking about the era of Claude Code and how Claude Code is so good. A few months after that, kind of at the end of last year, Peter has officially switched to Codex.

And his reasoning for that is basically Codex can one-shot things. So with Claude Code, you do the same up-front planning. For a coding change, you say, "I want to change this, build this user feature, here are my requirements." Claude will go look at some of your existing code to see how to build it, present you a plan, and you say, "Go." I think the difference that Peter outlined, which I've come to agree with, is that Codex does a lot more up-front searching. So if I say, "Hey, go build a feature on Split My Expenses where I want to allow users to search expenses. This means on various screens where we show a list of expenses, there should be a search bar. If you tap into that search bar and you type, you know, 'In-N-Out,' it should filter the expense list by what you typed in. And it needs to be private, safe, and fast." You know, I don't want other people searching "In-N-Out" and getting someone else's expenses in the list.

So I have used Codex to build expense search as one of my top-requested features on my user feedback board for Split My Expenses. And it works so well. I was blown away by choosing Codex 5.2 as a model. There's also a thinking budget, so I think it's low, medium, high, or extra high. And from Peter's blog article, which I'll link in the show notes, he mentions using "high" as a default. Again, there is "extra high," but that is almost too much, is how he puts it. And so "high" is a good in-between of "medium," which thinks a little bit and gets most things done faster. But how he describes it is, "I'd rather spend more time up front and have the task take longer and be done, versus it being faster and there might be room for error."

And so this is my first trial run giving Codex the reins of my web app, which is SplitMyExpenses.com, and the React Native app. So it was looking at both repositories, making code changes on the web app to power the mobile app, and did all that quite seamlessly. And so I would give a huge thumbs up to Codex and say, I think they've come a long way. Their interface has gotten better, the model harness has gotten a ton better, and using the right model and the right settings feels like it's the new unlock.

Again, it could just be I didn't use Codex too much, I'll be quite honest, but using it today in January 2026 with the high 5.2 model really, really feels good. Again, it takes longer to get things done. It's not significantly longer though. Opus still takes some time. So I think there used to be a divide where GPT-5 just took forever, like 25 minutes for something small. Codex gets a lot done, reads a lot more code. Yeah, it's just a different way. I feel like Claude Code is still there, and I still make probably 50% of my changes with Claude Code. But the harder ones, I feel like I'm naturally going to Codex now. The harder coding problems feel like the meat of my work. You know, go fix the UI, go do that—those are easy things to fix. But for it to be the important model for the harder problems is a pretty good achievement. So, you know, hats off to OpenAI. I think Peter put me on Codex, and there are so many tools out there that I have to try to believe it. Once I tried it and saw the power of it, I was like, "I personally do not think Opus would have gotten me this far and one-shotted it in the same vein that Codex did." So if you are looking to do a coding agent and you have picked up Claude Code recently and you love it, please try Codex and do GPT 5.2 high. Give it the same or a similar task. I think you'll be very surprised and impressed at how well it looks at the code and formats code and produces solid results with no back and forth.

Yeah, I would, I think two things, I would second the notion that I'd rather spend time up front getting the plan right versus not spending much time on the plan and then having to constantly check in or course-correct. Because I feel like the course-correcting, in my experience, is where things kind of go off the rails. Because it's like, "No, no, don't go change that. Just change this." And then you start having to basically undo, undo, undo. That's just my perception of it, where it's like if you... I remember being really impressed with the Cursor plan mode because it made the whole document and then it just went, and it was like 98% of what I wanted, which was great. Um, I guess I was curious because you've been such a Claude Code proponent, with Codex, are you using the CLI or how are you using Codex? I was just curious.

Yeah, using the CLI. So it feels very similar. Claude still wins in terms of user experience. It feels better for whatever reason. Text input feels a little bit better, the diff view when it adds and deletes code and it shows that visually just feels better on Claude Code. Some of these things are hard to put your finger on, but that experience feels better, but the logic and the code quality I get feel higher, and significantly so, from Codex. It's not miles ahead, but it's enough that, you know, again, if I do have a hard problem, I've been going to Codex now.

And one last note about Codex that I think is actually an awesome addition is the $20 a month, which you get with ChatGPT Plus, and you get Codex, it gives you a ton of coding within the Codex CLI. So I pay for the $200-a-month Opus plan. I haven't hit my limits in a long time; probably, you know, I could go down to the $100-a-month plan. But the $20 Codex, I feel like I get a ton of value out of building hard stuff with only $20. So it's almost inspired me to think, is Anthropic also shipping this many tokens for $20? Like I could save $180 by rolling back my $200 plan to $20. Because back in May 2025, Opus consumed so many tokens because it was so good that you would hit the limit, so I had to go to $200. I had taken a bite of the coding intelligence and I was like, "I need more." At this point, the models have gotten so good that they use fewer tokens.

So there have been conversations on Twitter of, "Should I be using Haiku or Sonnet for a model?" And the creator of Claude Code, Boris, who's now joined Twitter, he mentions, "No, only use Opus 4.5 because it's so smart that it uses fewer tokens." A dumb model will use more tokens to kind of think its way through things, but for something that knows what it's doing, it uses fewer tokens. So it's this weird dynamic of, "It's smarter, so it's more efficient." And I think in the same vein, Codex has done such a good job that I think for folks who haven't used it, definitely try it out. And that $20 package, although it seems too cheap to be true, for me personally has worked out quite well. Again, I still pay $200 a month to Anthropic for Claude Code and I love it. But I don't know, maybe in a few months, this will change. Oftentimes these models go back and forth and we'll have to see. I think the CLI wars are pretty heated up right now and I'm excited to see the competition. But yeah, for folks out there who haven't used Codex, definitely give it a try. Read Peter's article, which I'll link in the show notes. It helps you understand why Codex is useful and kind of get you on his workflow, and he's spent millions and millions and millions of tokens every day. So pretty impressive.

Yeah. And whenever I look at his page or maybe it's his X profile, there are like seven or eight websites. He's just got all kinds of stuff going on. I'm like, "Wow, that's impressive to be able to even maintain." I know he's using AI a lot, which is not to take anything away from him, but it's like, "Damn, that's impressive to have that much going on," that many, you know, what do they say, irons in the fire or whatever. So yeah, pretty cool. Good for him.

Yeah, I think his latest title was, "I ship code I don't read." So it was kind of an excerpt from his blog article, "Shipping at Inference Speed," published December 28th. But the headline is like, "Why I stopped reading code and started watching it stream by." So essentially, I think his theme is the model is so smart that I'm able to ask it to do something. I have high trust that it does it well, like architecturally in the code, and then I'll just test it and move on. Whereas I think before, the models weren't as good, and so you felt like you needed to be more hands-on with the code. I think he is now transitioning almost full-force to, "I look at some parts of the code, but not a lot of it. And due to that, I'm able to ship faster." And he already shipped fast before, so his developer pace is actually insane.

I feel like... I've never met Peter and I hope this comes off as a compliment and not a diss or whatever. But that movie, um, *Ex Machina*, where it's... I don't know if you remember that movie, it's like the robot. Yeah, it's like there's the guy who has the house and everything, invites the main character over. I feel like he's kind of trending towards being like that... I'm trying to think of the right way to phrase... he's trending towards being like the guy that has the house. He's just so advanced on the usage of AI, I honestly imagine him in his house barely even touching his computer, but then just has all these websites. Like just Tony Stark, maybe Tony Stark is a better example, not the *Ex Machina* guy, but like Tony Stark and just he's in the lab with Jarvis just doing all this cool stuff. So, yeah, it's cool. I'm sure, you know, he probably figured that out through lots and lots of trial and error and just spending time with it and all that kind of stuff. But yeah, it's pretty impressive.

I'll have to ask him for a photo of his setup or something. That's a good question. I feel like all I see is his code output, but like, you know, what does it look like at home? Is he living in four bunkers?

Yeah, he's got to have a lab. He's got to have a lab, you know, like just a giant room, like in *Batman: The Dark Knight*, he's got the... this is getting so nuanced or so specific, but like in *The Dark Knight*, Morgan Freeman's looking at the wall and it's like 50 TVs. I'm like, "Is that his setup, you know?" And they're all Claude Code, one per TV.

Yeah. Or I guess Codex now. Yeah. And anytime that people are complaining about Opus being bad is probably because he's throttling it himself, you know.

Yeah, really nice guy. I'll probably get some time with him next week, but yeah, really a leader in the AI space and publishes great content. So if you haven't read Peter's blog post, definitely start with the most recent one, but great insights and honestly, fantastic software that he's been working on over the past six months.

Yeah. Cool. Alright. What should we hit next, Brad?

Uh, talk a little bit about agent skills. So I have seen an influx recently from my iOS Twitter crew where they are publishing more open-source agent skills. So I think for iOS specifically, they become more important only because Apple hasn't done the best job at documentation. Oftentimes in recent years, Apple has shipped new APIs. One key example is what they call concurrency. So Swift concurrency is a new language feature within the Swift programming language where there's a different mechanism to do multiple things at once on your computer. So you're coding this up using different APIs to have multiple things run. However, the evolution of concurrency has changed significantly over the past few years. Best practices have changed, APIs have changed. It just feels like a moving target for developers to hit.

And now developers have to spend a lot of time to understand what is good and what is bad with these new APIs and how to migrate a project from pre-Swift concurrency to Swift concurrency. So long story short, Apple has a clear gap in best practices, and skills are the perfect way to summarize how to do this effectively given a real-world codebase. So I've seen so many interesting skills published. A main one that I'm seeing is Swift concurrency, and it makes it dead simple to ideate knowledge, distill it into a markdown file, and share it with the world.

And it's surprisingly effective. Like, we talked about MCPs, which are hard to set up. Skills should be easier. Once you get it in, your agent should look at it, be able to pick it up dynamically. So that kind of context discovery works well. I haven't used the iOS ones yet, but I did clone the repo and take a look at it. So it's nice seeing skills pop up as a first-class citizen within both Claude and ChatGPT, and it's nice that people are actually contributing because one kind of neat fact is it almost feels like you're distilling and outsourcing yourself. Because oftentimes as a developer, you have specific guidelines or taste, and people are distilling that taste factor into a skill, saying like, "This is how I would do it. This is how to make the code better." Some kind of undocumented strategies and guidelines, but they're publishing that for free, just like Adam Wathan published Tailwind.

So much excellent open-source software and information is out there these days that it almost becomes like you're outsourcing yourself. So a very interesting divide where you want to help others, but it's almost secret, extremely useful knowledge. Like, what do you do with that? It's kind of like Peter Steinberger where he creates these really fascinating software packages that he could probably sell, but he just puts them out there for free because he just loves doing it. So we're getting to this world where I think AI is enabling a lot more people, and to make AI effective, we take these really smart people and get knowledge from them. And it feels like skills are becoming that repository of information where we ask smart people, "How do you pull this off? Okay, write that down and share it with the rest of us." And that's where I see it headed. Maybe there'll be paid skills, but I'm happy that people are picking it up and starting to contribute to that.

Yeah. No, skills, I think it's something that I, when I first saw them, I was like, "Oh, those are just tool calls." And that's kind of how I viewed it. Which I think is still right, like they kind of seek to accomplish the same thing, but I think from my understanding of them—I haven't really built out any Claude skills—it's almost like there are tool calls, but it's a bit more like AI where it's text and markdown, because they're markdown files at the end of the day, typically. So, but yeah, I mean, I think it's cool to see and again, it gives you, like you said, that kind of taste or how you want to delegate things. It's not just open-ended, you know. So, yeah, it would be cool to see.

Cool. Awesome. Well, should we wrap it up with that, Brad?

Yeah, let's do some bookmarks. We talked about so much. So I'll kick it off with bookmarks. So I've already talked about it a little bit, but the creator of Claude Code, Boris Cherny, I think that's how you say it, is on Twitter now. I think he joined in December. And Andrej Karpathy tweeted December 26th, the day after Christmas, saying, "I've never felt this far behind as a programmer. This profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse, in between." Uh, really long Twitter post after that, but the TLDR is things are shifting and it's clear, and what do you do about it?

Boris, the creator of Claude Code, then responds to that Karpathy tweet and says, "I feel this way most weeks, to be honest. Sometimes I start approaching a problem manually and have to remind myself, 'Claude can probably do this.'" Um, which is a kind of nice revelation. At the end, Boris closes with, "The last month was the first month in my career as an engineer that I didn't open an IDE at all. Opus 4.0 wrote around 200 GitHub pull requests, every single line within each of those pull requests. Software engineering is radically changing and the hardest part even for our early adopters and practitioners like us is to continue to readjust our expectations. And this is still just the beginning."

So fantastic quote from Boris. I think I truly have felt this and I've been singing Claude Code's praises, although I do love Codex now too, ever since May 2025. But yeah, it's becoming weird, you know. I think I have gotten over this thought of "software engineering is different." Like to me, I don't feel as impacted as Karpathy, but I don't know, I'm not sure why I feel that way. I just feel like I go with the times and I'm like, "Hey, this is how things are now. I'm happy about it. I can still build stuff." I don't sit and dwell and think, "Damn, I used to write code by hand," but when people bring this up, it does bring up that feeling in me where I'm like, "Oh, you know, I did used to do that." So it is a little weird, but here we are. I'm happy about it. We have great tools. So if you aren't following Boris or Karpathy, definitely follow them on Twitter. I'll link the Boris response, but very interesting. You know, a very principled engineer, and not opening your IDE for the first time in your career... I don't know about you, but seems like a seismic shift in software engineering.

Yeah, I mean, I think even the Stack Overflow thing to me was big too, because that's... when I was learning Python and all this stuff back in 2016, 2017, I was there every day I was using it, every day I was coding. You know, I didn't code every day, but every day I was coding, I was on there. So yeah, it's stunning. But yeah, I mean, I don't know how you couldn't read those tweets and not get excited too, because it's like, here we are. This is our moment in time. Let's get after it. So yeah, it's pretty cool.

Uh, my bookmark is from Levels, who if you don't know who Levels is, that's your fault because he's one of the better follows on X. I won't go into his background, but just follow him. But basically the most famous indie hacker. I'll leave it at that. But my bookmark from him is where he basically—and I'm sure he doesn't claim to be the person that found out how to do this, but he's making it popular and it came into my feed because of him—you can basically use Claude Code on mobile through a VPS. And so I don't know the exact... it's a bookmark, so I want to look and see exactly how he does it, but I think he has a VPS and he's writing requests into Claude Code. And I think even at one point he was using like Whisper flow or something like that to just dictate into his phone, which was talking to Claude Code, which was writing code for him and doing pull requests or whatever, push requests, whatever the right phrase is.

So, then of course people were replying and being like, "Oh, I'm doing this too. Blah, blah, blah." So I want to check it out. It seems pretty cool because at that point, if you're on your computer and you got Claude Code and you're doing stuff and you're like, "Oh, I got to go to Chipotle, I'll be right back," and you start doing it on your phone, you're just talking to it like, "Oh, hey, actually tweak this." You could just be "cracked," as the kids say, right? Just constantly, constantly "Clauding out," you know. I think I made a joke to you in text like, you know, "Saint Claude," right? It's just like, that's where it's at right now in terms of just productivity and people doing stuff. You know, maybe it'll change in three to six months, but yeah. So anyway, people just can't get enough of it, even setting up a VPS to remote in and be able to do it on mobile. It's crazy.

Yeah. Crazy times we're in, but we love it. You know, not every day can you say that.

For sure. And if you're not learning, then that's boring, right? I even just... look, I might be weird, but I like learning even tax stuff. So if you're not learning, it's boring, right? Otherwise, what's the point? So...

And if you do love learning, AI is the perfect spot to be in because man, there is a new thing every five minutes. So, definitely something to get used to.

And also too, and we'll wrap up with this. I know we're going long on time. You know, learning will be so interesting because I was talking about this with my wife and my buddy, like, what will college look like when my kids are of college age? And if you end up having kids, like what will college look like? Will it look like it does today or will it be like, you got an AI professor? I couldn't stomach paying $80,000. You know what I mean? Like something's... I don't know, but maybe it's not that. Maybe it's where you meet friends and meet your partners or something. I don't know, right? But it becomes like a really expensive dating service. Who knows? But...

Wow, you can't end on that. That's a thought-provoker. Wow.

Yeah, yeah. Like just what would college look like in, you know, my kids... my oldest is like eight. So in 10 years, what would that look like? I don't know.

Yeah, I can't even imagine five years. Ten years feels like, you know, we'll have at-home robots to feed me all meals every day or something. That'll be crazy.

And there's no way it can be as expensive as it is right now. There's no way. But yeah, I digress. We'll see.

Well, if you're listening, maybe that's the comment we'd love to hear from you. What will college be like in 10 years? Let's hear some theories. I for one am hard-pressed to put my finger on anything specific, but it'll be different. That's for darn sure.

Yeah, yeah, for sure.

Cool. All right, Ben, good stuff. We'll wrap it up there.

Cool. Sounds good. See you next time.

Yep. See you everybody.

[Outro Music]

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.

[Outro Music]