Breakeven Brothers

[00:00] Bennett: All righty, we are back with a brand-new season: season two of the Breakeven Brothers podcast.

[00:22] Bradley: That’s right. Yeah, we took a little hiatus. I think it was around two months or so. We had some travel planned.

[00:28] Bradley: We had some just busy life stuff going on.

[00:31] Bennett: So while we were gone, lots and lots of AI stuff has, of course, taken place.

[00:36] Bennett: So excited to get back in front of people and start recording again.

[00:40] Bennett: So how’s it going, Bradley?

[00:42] Bradley: It’s going great.

[00:43] Bradley: I came back from a two-week stint in Europe for a friend’s wedding.

[00:47] Bradley: So I did Copenhagen for the first time.

[00:49] Bradley: Had amazing food there.

[00:50] Bradley: Everything was like a tasting menu.

[00:52] Bradley: That was excellent.

[00:54] Bradley: Did a little boat tour as well.

[00:56] Bradley: Colorful buildings and nice people.

[00:57] Bradley: But definitely quite expensive over there.

[01:00] Bradley: And the one thing we did see when we were out there

[01:02] Bradley: was the boat that Marcus went on,

[01:05] Bradley: if you’ve watched The Bear.

[01:07] Bradley: So he does a little stint in Copenhagen

[01:09] Bradley: for, I think, being a pastry chef.

[01:12] Bradley: And along the kind of narrow canals,

[01:15] Bradley: there’s one boat that he was on.

[01:16] Bradley: So I walked past that boat, got a little video clip.

[01:19] Bradley: I was like, “I was here.”

[01:20] Bradley: And then after that, went to Cannes in France,

[01:23] Bradley: saw this amazing castle wedding that I was a part of.

[01:25] Bradley: Honestly, the best wedding I’ve ever been to. It was super crazy, super fun, amazing. And then after that, I did a few days in Barcelona and nearby areas. So I did the whole Europe sprint.

[01:38] Bradley: It was super hot. I jumped through Airbnbs and hotels, but honestly, an excellent time off. I was able to completely disconnect from work, not do anything. It was the longest time I’d been without my laptop in a long time. What I did before I left, which probably wasn’t ideal, was I had set up my computer to be connected to a charger and locked so I could do Codex Remote, because you can do that now.

[01:58] Bradley: But I think it dropped connection after like one day, so for whatever reason it didn’t work out. And I thought, “You know, maybe that’s for the better.” So yeah, I got back last week on Sunday night and then did work the next Monday, and I was just dead and jet-lagged. But it was kind of nice. I’m slowly getting back to the normal schedule where I wake up at like 8 instead of 6 a.m., and I sleep at midnight instead of 10 p.m. So finally back and ready. But yeah, it was a great reset.

[02:27] Bennett: Nice. Nice. That’s awesome. Well, if it’s your first time listening to us, I guess we should do a quick introduction, Bradley. So I’ll go ahead and kick it off. My name is Bennett, of course. I’m traditionally an accountant and CPA, but just have a passion for technology. And so in 2026, that means, of course, AI. And so I love trying to make things with AI and learning what AI can kind of do, not just for accounting but just in terms of knowledge work in general. So yeah, that’s a little bit about me. Bradley, turning it over to you.

[02:52] Bradley: Yeah, I’ve been building apps for about 10 years now. Started at PayPal, went through a bunch of different companies. Always had a passion for building things, even in high school and college. So yeah, software engineer by trade, working on the iOS side, but full-stack developer interested in Laravel on the PHP side. I know it’s kind of a hot take, the gamer. So yeah, just building stuff in public, and I’m on Twitter, on YouTube, have an awesome documentary there, and yeah, just trying to build cool stuff. I think half the time is learning coding, but with AI, it’s kind of less learning the coding part and more shipping good products. So definitely drop a follow on X. I’m trying to post a bit more there, also on LinkedIn and YouTube. So yeah, for first-time listeners, welcome. We have quite a footprint. You have about 40 episodes to catch up on, but we’re happy to have you.

[03:40] Bennett: Cool. Awesome. All righty. So after being gone for two months or so, there are a ton of things we could catch up on, and we can’t do it all in one episode. So where should we start, Bradley? Do you want to start with your item, or should I jump right into mine? What are you thinking?

[03:54] Bradley: I’ll jump into mine. I’ve been having a lot of fun with ChatGPT finances. This has been a feature that’s kind of been in development for a long time. I think it came out probably a month or two ago for ChatGPT Pro users, which, for those who aren’t aware, is the $200-a-month plan.

[04:10] Bradley: Definitely not the cheapest plan and not for everybody, but technically finances is out for them. I think a week or two ago, it came out for Plus users, the $20-a-month subscription.

[04:21] Bradley: So I had waited until it became more broadly available. I tried it out last week, and I loved it.

[04:28] Bradley: So you open up the ChatGPT mobile app, you link your accounts using Plaid, which is kind of a de facto standard for authenticating banks and credit cards and not giving risky control. Plaid does a really good job of isolating your logins and doing official Chase OAuth logins.

[04:46] Bradley: Once you get past that part and you add all your accounts, it’s pretty smooth sailing. So it pulls in your credit card expenses, your bank account transactions, any mutual funds or stocks that you own on Robinhood or other platforms. So it really has a holistic view of all of your investments, your cash, your debt, etc. I let that bake for a day or so to pull in data.

[05:07] Bradley: And then I started asking queries about it.

[05:09] Bradley: And the good part about it is that it pre-prompts you with queries that save money to audit your spending.

[05:14] Bradley: So sometimes it’s a little bit of a blank-slate problem.

[05:16] Bradley: And I’m like, “I don’t really know what I want to do with this feature.”

[05:18] Bradley: It feels nice to have full visibility.

[05:20] Bradley: But I asked, “Hey, how can I reduce my business spending?”

[05:24] Bradley: And it went through and said, “Hey, you spend this much money a month and per year on Buxbyte and business expenses.”

[05:30] Bradley: “Here’s how you could reduce a few with overlap.”

[05:33] Bradley: And I thought, “You know, it identified one or two that I could probably cancel.”

[05:36] Bradley: One of them actually being Whisperflow.

[05:38] Bradley: So I love Whisperflow.

[05:40] Bradley: However, the Codex app does have good voice mode.

[05:43] Bradley: So I actually canceled that two days ago.

[05:46] Bradley: Big shout-out to them.

[05:47] Bradley: They create great software, but I feel like that might become a commodity.

[05:50] Bradley: I don’t know.

[05:51] Bradley: But other cool stuff with finances: you get great charts, great visuals, and great identification of subscriptions.

[05:56] Bradley: So it shows you what’s upcoming and what you’ve already paid for, and what cards that’s associated with.

[06:02] Bradley: And what I really liked is it has memories.

[06:06] Bradley: So ChatGPT has kind of revolutionized memories.

[06:09] Bradley: There used to be a very simple list of facts.

[06:12] Bradley: I think this is like memories V2.

[06:14] Bradley: It kind of sucked.

[06:15] Bradley: Wasn’t great.

[06:16] Bradley: Memories like V3, I think they coined it like a dreaming mode, where you learn things over time after chats and you kind of forget old things, just like a normal human brain.

[06:26] Bradley: And so ChatGPT finances has its own memory mode.

[06:29] Bradley: So you could chat with it about your goals and finance, and it’ll remember that and take that through other conversations.

[06:35] Bradley: So I was really impressed. I think everyone, if you’re a paying user, should sign up. And I know people kind of don’t love giving all their data, but if you’re okay with that, I think it is really fun to talk shop about how to optimize finances, because there’s a long kind of tail of things you could do there.

[06:52] Bennett: Yeah, I don’t use it, but I did see the announcement, and it kind of made sense to do because a lot of times I was exporting bank details and just in Excel, then basically just giving it to ChatGPT or Codex really, and it would just run a bunch of Python scripts and understand the data, then kind of do exactly what I was asking it to. Kind of like you said, with the pre-prompts and stuff. I use Monarch Money for my budgeting stuff.

[07:19] Bradley: Okay. Yeah, yeah.

[07:19] Bennett: Yeah. And it’s interesting. I’m in the camp right now of, like, I don’t want all my stuff within one vendor on anything. I feel like there’s some benefit to having things spread out a little bit. I don’t know if that’s true or not, but that’s just my personal thought right now. But I do like it a lot, especially for, like, I can see it being super successful with small business owners where instead of having to use QuickBooks, can you just use some kind of ChatGPT finances, but have it be a little bit more catered to business? I’m not sure if that’s part of it. I think right now it seems like it’s more like a personal-finance target audience, but I love QuickBooks, and so I’m happy to see anything that eats into that market share. And I could see something like this for, again, small businesses, simple single-member LLCs, sole proprietorships, if there’s a way to help with that instead of using QuickBooks. Because QuickBooks will try and push its own AI. I think it has some kind of licensing agreement with one of the frontier companies, but it’s QuickBooks. It’s just a legacy dinosaur product that just keeps getting bolted on and bolted on and bolted on, and it’s just not that good. And yeah, I just don’t like it.

[08:27] Bradley: Yeah, I agree. I think it’s quite nice. I had built the Empower CLI MCP and hooked that up within Codex so it could query and do all that, like pull my data from Empower, which was another aggregator that didn’t use Plaid. It was kind of closer to Mint. But I don’t know, I didn’t have these top-tier models hooked up. It didn’t feel as rich. ChatGPT finances can render you charts and widgets and categorize your spending, all these things that, if you have the raw transaction data, it doesn’t do as much there. So yeah, I would highly recommend it. I even set up an automation, which is suggested in the ChatGPT mobile app where you talk about your finances, and then at the end it’s like, “Hey, do this weekly. Take a snapshot of what I’ve spent this week and compare it to last week, and look at the overall trends and suggest ways for me to do better.” I thought that was freaking awesome, because there are certain high-value adds. Once you get your finances data in there, it feels super high value, and the experience is really good. And you kind of build up this repetition over time to check your finances. It’s way easier to have a consolidated view and AI on top and a memory. When I use these other tools like Empower, I have to go manually scan from last time to this time and do all this manual work. But if there’s an intelligent PhD-level reasoner behind the scenes that’s able to say, “I know what you want. Here’s the last snapshot. Here’s the current snapshot,” and give me something reasonable much quicker. So I think that part is a super smooth experience. If you’re a Plus user, definitely try it out. Only in the U.S. for now. Probably expanding; not 100% sure there. But it seems very early for this feature and a quick win to try out if you’re curious.

[10:06] Bennett: Yeah. Well, I think too, the other part of that that’s nice is, and I think this is what people were using—and certainly in my case using ChatGPT-4—is almost like kind of a coach or an advisor when it comes to that kind of stuff. So it’s one thing to see your data and understand where you’re spending money, but then to have someone—and, quote unquote, an AI model; it’s not really a person—but have someone kind of talk you through your goals and tell you if you’re tracking to what you want, or if this is something that is the opposite of what you want. Right? And so I think there’s a helpful accountability loop that you can create with something like this, where you connect your finances, where you’re not just understanding your spending and your income, but then also someone can kind of tell you, “Hey, you’re on the right track for the goals that you made earlier this year,” or something like that.

[10:56] Bennett: So I think that’s a big—because with NBA players too, in sports, it’s like, you know, a coach. Sometimes people say, “Oh, coaches aren’t that necessary because the basketball players know how to play and they know the positions,” and all that kind of stuff. But I feel like a good coach is someone who can take a step back and see big picture and adjust if you need tweaking here or tweaking there. And so I think with finances it’s kind of a similar thing, where it’s like, yeah, everyone knows you should save a lot of money, you should not spend frivolously. Everyone kind of knows these rules and chains of thought with finances. But to have someone, again, coach or be like an accountability partner, I think is valuable. And I think that’s something that hopefully, if you can connect your finances with ChatGPT, you can use it in that way and not just like, “Hey, this is my data. How is it? Tell me how it’s tracking to what I want to do long term with my goals,” using the memories that you mentioned earlier. It’s like, “Hey, you told me back in January that you wanted to buy your first house or whatever. If you’re spending this much money on Starbucks, then you’re not tracking to that goal.” Just kind of having that feedback loop I think is great too.

[12:06] Bradley: Yeah, it doesn’t have budgeting built in, but it has categories that you can modify. So that Monarch budgeting bit, I don’t know, maybe will come. I’m not sure what the team’s working on. The other one is the memories from your normal ChatGPT are kind of brought into finances. So I think there’s a subset of memories in finances, and then your overall profile has memory. So it did reference things that I had talked about in other chats in the finances chat. And I was like, “Oh, this is kind of nice.” It felt like it knew me already, which I was like, “That is freaking awesome.” So yeah, super cool software, a great experience. The automations for it make it super high value. And I’ll report back in a few months on what I’ve done and where it’s saving me money. Because I think at the end of the day, yeah, getting someone who is on top of your finances in a digestible way and a good experience is the best thing that I want. That’s the end goal for a lot of the software for me, and I think finances does a pretty good job getting me there.

[12:59] Bennett: Yeah. And I think it’s like, I believe in expertise, and there’s value in expertise. And I think it doesn’t replace financial advisors, I don’t think. And I feel the same way for good CPAs and stuff. But I think it’s a very helpful first layer of review and just bouncing ideas off of and stuff like that. So I think there’s still a place for expertise, but I think, again, it just makes it easier to have a better holistic view of your finances.

[13:27] Bradley: My last anecdote here before we move on is when I did my initial audit, post-linking all of my accounts, it highlighted that I had, I think it was like an $800 American Express fee. And it was like, “Hey, this fee is huge. Are you sure you want this credit card?” And then I told it, “No, no, no. I do my best to use all the American Express benefits. Go add those up,” which some benefits are like $15 a month on Uber, which aren’t really tracked within transactions—that’s within the Uber app. But it did kind of comb through a lot of the refunds that Amex gives you if you spend in certain areas for the credit card benefits, and it tallied those up and was like, “You spend $800 or $900 on the credit card yearly fee, but you use like $800 of the benefits that are refunds. Plus, if you add $300 of Uber credit, you’re already over what the annual fee would be.” And so it was nice for it to flag it, and it was nice for me to challenge it, and it kind of agreed with me. And I think the models have a long history of being agreeable, but this one, I was like, “No, no, I’m pretty sure I’m right here.” And it’s getting better, and I think it’ll probably learn that as a memory of, “Hey, I use the benefits for the card, but also challenge me and make sure I am using it.” So it was nice for it to call that out, because I think you can forget about these things from time to time, and it highlighting it is just like, that is front and center. It is an egregiously expensive card, but if you use it correctly, you can get the benefit out of it. So yeah, shout-out to finances for that.

[14:52] Bennett: Well, and that reminds me of something I did recently where it wasn’t connected to the finance plug-in, but I use Render for some of my websites and apps and stuff. And I had it connected with the Codex plug-in.

[15:04] Bennett: And I asked, “Hey, I think I’m spending way too much money on the way that these sites are set up.”

[15:11] Bennett: Some of them are static pages.

[15:12] Bennett: Some of them are actual web apps.

[15:14] Bennett: And I was like, “Just look at my projects and tell me, am I overpaying, and what is a better alternative?”

[15:19] Bennett: And yeah, it managed to save me probably like $70, which again, it’s probably a skill issue.

[15:24] Bradley: Yeah, it was probably a skill issue.

[15:26] Bennett: Because I just set it up and clicked through some of the Render setup pages, and probably should have slowed down and just been like, “Do I really need this thing for this project?” But yeah, it was like, “Oh yeah, you should move a couple of these websites to just Cloudflare Pages, and that’s free.” And then that cut off like $35. And then there was one that I just wasn’t using anymore. And it was like, “Hey, you don’t have any traffic,” and I also kind of just forgot about it, but I was still paying for it. And so it was like, “Consider sunsetting or just deactivating this site.” And so yeah, it was, again, not through the finances app that we’re talking about, but just something that understands your projects and can advise you on some helpful tips—in this case, in regard to finance. So that’s cool.

[16:08] Bradley: $70 is a lot. I mean, that’s $70 per month, right?

[16:12] Bennett: Yeah, yeah. Something like that. Don’t quote me an exact amount, but it was substantial because there was a web app, and the way that Render works is they kind of push you into a certain tier if you have X amount of apps or something. And so it was like, if I got rid of one of those apps and then moved a couple of them to just a Cloudflare Page, then I would fall back below that tier where I could just go back to the cheaper monthly plan instead of the one up that they had to push me into.

[16:37] Bennett: So yeah, it was nice.

[16:39] Bennett: I was like, “This is good.”

[16:40] Bradley: And that’s kind of the stuff that you don’t care for, and you’re not going to look at yourself. But if you ask AI or other agents to dive into this, and then it finds that, you’re like, “Yeah, why would I not do that?” But I don’t know if I’d spend 30 minutes looking into that myself personally.

[16:54] Bennett: No. Yeah. And also too, it’s one of those things that over time, as you add new projects and add new sites, you don’t again take a step back and look at it holistically. And so I have the ability to do that and have it look at everything. And again, I’m pretty sure I was using the Render plug-in within Codex, and it just worked, which is awesome.

[17:14] Bradley: Yeah, the plug-ins are good now. Codex is great. I know you’re talking about Cloudflare. I saw an article about Cloudflare being a little bit too AI-pilled, but they recently released Cloudflare OS. I have not caught up with it, but I know you listed it as a topic, so please catch me up. What is going on here?

[17:30] Bennett: Yeah, and so I’m still wrapping my head around it, but I think the interesting thing about it is I feel like right now there’s a kind of fork in the road of what the experiences are that we’re building for who’s going to be using the web in the future. Those are all things I’ve seen talked about online, definitely. And I have two things from Cloudflare I wanted to chat about. This announcement is from Cloudflare OS, and basically what it looks like is a workspace where you can chat with agents and kick off a bunch of workflows. So almost like a Slack kind of thing, it looks like. And if you look at the videos, it kind of looks like a chat window where you’re writing a question to the model, but then it has access to your environment. It can run deterministic workflows.

[18:17] Bradley: Very interesting. Is this like a Claude Code/Codex competitor?

[18:25] Bennett: I don’t know.

[18:26] Bennett: I don’t think it’s that necessarily because you can plug it in with those other models.

[18:31] Bennett: So I’m looking at the screenshots, and it’s a window, and then it has a thing where you can pick your model. In the screenshot they have, they have like Kimi 2.6 or whatever.

[18:41] Bennett: But I saw another video where they were using an Opus model.

[18:44] Bennett: So I guess it’s kind of like that. I feel like it’s like a desktop app that maybe has some similarities to Codex or the desktop apps that you use with these different frontier models. But they call it an “agent workspace,” and so it has all the context and skills that your company has. They have some language in here about actually having security and governance built in first and foremost with the release of this. So, like, how you can provision agent access really cleanly, so there isn’t any kind of unintended agent access that you really wouldn’t want it to have, and it starts doing stuff.

[19:23] Bradley: Right. I just see that it’s open source too. Interesting.

[19:28] Bennett: Yeah. So Cloudflare—it’s a mouthful. I don’t know why I struggle to say that word. But yeah, to your point about being AI-pilled, I feel like I use them so much for all kinds of stuff. I’ve always used them for DNS stuff. But going to what we were talking about earlier with the Render discussion, Codex was like, “Yeah, you should use Cloudflare,” which has a connector or a plug-in in Codex. And so it was like, “You should just use that to host some static pages, and then you can have a worker.” And I’ll be honest, I’m not a worker expert, but it has—

[20:04] Bradley: People love workers.

[20:04] Bennett: Yeah, it can kind of do much stuff for you. And then there’s a CLI with Wrangler. So on one of my websites, it’s just a static page, and then we set up what Cloudflare calls—I don’t know—but it’s like a database thing. It’s like D1 or something like that.

[20:22] Bradley: Yeah, I think that sounds right. They have D1 and R1. I think R1’s S3, I think.

[20:30] Bennett: Yeah, maybe it’s D1 or D2. I don’t know. We don’t know. Something-one. It’s a letter and a one. But yeah, Codex had some of the context and connected to Cloudflare via the plug-in, and it just was like, “Hey, as part of saving money on the Render bill, let’s set you up on Cloudflare like this,” and it just did all that stuff for me. So yeah, Cloudflare seems to be really shipping a lot of AI stuff. And again, I think a lot of the stuff I’ve seen from them comes from a more agent-driven web world, for lack of a better term. It seems like I saw something recently that the majority of internet traffic now is done by agents, or maybe that was either now or soon. I can’t remember exactly what it was, but—

[21:14] Bradley: Which is crazy. I mean, I have firsthand experience in this because I’ve been using Codex so much for web stuff. But part of it is I ask a query to Codex, and it’ll go web search, which means it’s not using my browser, but it’s just querying stuff almost like Google is. But the other half is I ask it to use my browser, my login credentials, and navigate through pages. And it feels like such low effort to get these tasks done now that it’s just amazing to kick one off and approve a few things here and there for security’s sake. But sensitive actions using my browser feel very doable now. And not even sensitive actions, like a bank or something crazy like that, but even normal things I don’t want to do. One that I had as an example was posting an eBay listing of an old gaming laptop that I wanted to get rid of. I took photos, dumped it into a Codex thread, and was saying, “Hey, I’m logged into eBay. Make the listing for me.” And it researched competitors, found the right price, did the shipping estimate, attached the photos, even blurred out the serial number in the photo that I attached because it was like, “Hey, this is sensitive. You shouldn’t post an eBay photo with your serial number.” And I was like, “Damn, this was incredible.” Part of it’s the harness, part of it’s the model, and the other part is enabling Chrome access with the Chrome plug-in that ChatGPT has now, or Codex has now. But I’m still navigating the web. I’m still on X from time to time, and I’m on YouTube and all that fun stuff, but I have a large part of my traffic being driven by an agent driving my own browser. And it feels like that’s only going to ramp up significantly.

[22:51] Bradley: And Cloudflare is almost the internet front door.

[22:54] Bradley: I think it serves like one in three requests, or maybe serves is the wrong word, but is involved in like one out of three requests across the whole internet.

[23:02] Bradley: So they definitely have a finger on the pulse there, and I’m curious what they’re going to build to make that become a reality.

[23:07] Bradley: Likewise, I don’t know if you saw this release, but it was Cursor or SpaceX or xAI, whatever their name is now.

[23:15] Bradley: They released Bot or something like that, which could be similar to Cloudflare OS, where they have some chat interface.

[23:25] Bradley: It has access to things.

[23:27] Bradley: I’ll be honest, I’m not fully up to date on their positioning here, but people seem to like it. So it feels like there’s an emerging category of enabling AI and workflows in a chat interface that is not Codex or Claude Code, but it’s growing. It’s good to have competition there.

[23:43] Bennett: Yeah. One, I think just variety too, because I think there was one that came out recently called Buzz, which I think was being marketed as the Slack replacement, but more agent—

[23:52] Bradley: Yeah, that was Block maybe. I think they did it. Or Jack Dorsey. I’m not sure. I know it was him, but I don’t know if it’s part of Block or something separate.

[24:01] Bennett: It feels like there’s also an appetite from people to have agents just be where people already are. So in this case, Slack. And I think there’s Claude Tag, I think, which is a Slack-based thing too. Or I know Claude has a Slack integration that you can just Slack it like it’s any other employee. And so I think there are differences of where you can go to an app and chat with AI like that. That’s like the Codex app traditionally. Before that, it was the web app, and you go to ChatGPT.com. And now it’s like, “Hey, these agents can kind of just be in your system,” whether that’s in Slack or Buzz or this Cloudflare OS. I struggle with that word. I don’t know why.

[24:41] Bennett: But yeah, it’s funny because you mentioned the eBay thing, because I had something similar where I was like, “I’m just going to use Codex and see what happens.” So I was going to a networking event in my local area, and I didn’t have any business cards. I don’t have business cards. That’s kind of a relic of the past, in my opinion. And so I was like, “Well, I can probably just make a QR code that will go to my website and auto-add my contact information if people scan it, and it pops up on iOS: add contact,” whatever.

[25:10] Bennett: So I literally told Codex, “Make me a QR code that, if they take a picture of it, will just populate my information in their contacts,” or whatever. And it took 15 minutes and did it. It did a branded little B, because my name’s Bennett, of course, and so it branded a little B, and then it was the QR code that worked. It was crazy.

[25:35] Bradley: That’s cool.

[25:35] Bennett: And then I know before I did that, I was looking up, “Do I have to pay for this?” I think there was one that was like myqrcode.com. I was like, “I don’t know.” It was paid.

[25:44] Bradley: Yeah, yeah.

[25:47] Bennett: It was like $50 a year or something like that. I’m like, “Not in today’s day and age.” And so I just had Codex figure it out. So yeah, those apps are incredible.

[25:53] Bennett: But then on the thing with Cloudflare too that I was going to mention is they also released—I think it was a couple weeks ago—this thing called Kite Surf. And it’s actually because you mentioned the Google Chrome web browser thing. So they released something called Kite Surf, which is a more agent-enabled or lighter kind of web browser, because I think what they said in the article was modern browsers like Chrome and others are written or made for humans to digest and understand information. And so there’s a lot of overhead, tech-wise, that comes with that. And so they built a brand-new browser, I guess primarily in Rust, so it’s supposed to be really fast and really lightweight. And so they’re saying this might be the future for if you’re using an agent and it needs to do some kind of web search, instead of opening up Google Chrome—or I think Chromium, right, is the other browser that they can use—it would just use Kite Surf and be much faster and have less—

[26:51] Bradley: That’s very interesting, actually. Huh. Yeah, I totally missed that release. I don’t know when that came out, but I must have been—I’m going to say I was in Europe at that time. So I was just taking the article that you linked, and yeah, it looks very interesting. I think the problem that they outline of a more secure, agent-friendly browser sounds right. And if they’re able to rewrite all this using AI over time, sounds like a good gig. I know having the Chrome plug-in for ChatGPT gets you in the zone so that Codex can control the browser. But we did used to have our own browser, Atlas, that got sunset recently. I wonder where these things exist. We’re seeing this bubble up of browsers, of agentic workspaces with chat interfaces. We’re seeing Slack, like Claude Tag. It feels like we’re trying to bring all the agents closer to you and enable them to use your tools. But this browser is almost like a tool built for an agent. And so there are both methods of bringing it closer, but also building tailor-made solutions for agents. So I don’t know where it’ll end up. I’ve seen both sides of the story, but it seems like Cloudflare is shipping a ton. It’s hard for me to even keep up with what they’re doing. And I’m also a happy user of Cloudflare across all my sites, but I don’t touch a lot of their new stuff. I’ll be very honest. I don’t use workers very often. I’m in the PHP bandwagon, but it feels like they’re doing something right. I do wonder—again, I saw an article recently criticizing them about how much they’re shipping, because in the age of AI, every company is trying to do something related to AI, and they are doing a whole lot of it. So I’m curious how their product portfolio and the perception of Cloudflare, as it became this DNS-centric area and now is moving to this agentic AI solutions for almost a lot of things—I don’t know how it’ll end up, but hats off to them for shipping.

[28:42] Bennett: Yeah, and they’ve been doing a lot. I’ve used a worker, but I couldn’t tell you exactly what it was doing.

[28:53] Bradley: Your agent used a worker.

[28:53] Bennett: Yeah, my agent, Codex, did something with a worker, and it worked, and I was like, “Cool.” And I think that goes to what we’ve talked about a lot on this podcast, about what coding means and how do you engineer with AI and stuff now. It’s been a big mental thing for me. I’d say last year and the year before, I felt like I needed to really still understand the code a lot to feel good about shipping something. And it’s been a struggle, but I’ve gradually been like, it’s going to know it better than me, so why do I need to bottleneck myself and be like, “Oh, explain this to me”? There’s a level to it that you do have to understand, of course, because you need to make sure it interpreted what you’re asking for correctly. But the actual architecture, what am I going to say about what GPT-5.6 Soul extra high did? I’m not going to have anything better. There might be things from a business context that I would understand better, or just information it didn’t convey very well. But if it has a full picture and it has a really good understanding of what you’re trying to do, you’ve got to just trust but verify. But that’s been one of my blocks, just having to feel like I still need to understand every single line of code.

[30:11] Bradley: Yeah, there’s an interesting Twitter thread I saw recently talking about, so we have GPT-5.6 Soul, and extra high takes a while to do things. There’s fast mode, which gets you, I think, 1.5x the speed, but obviously costs more money. You can use that in subscription and API billing. But recently there was announced that OpenAI is working on the ultra-fast mode, which is just way, way faster tokens per second. I think it’s in very, very limited preview. But there was a conversation saying you can do multiple things with agents because it takes time per turn. And if you’re generating a website, you’re saying, “Do this.” It takes a minute, two, three, four minutes, depending on what you ask, and it comes back. But if the AI was able to generate code at 10x the speed, you would be more in the zone and have a live conversation where you didn’t have to go wait and do something else. And at that point, good luck reading any code at the speed that it’s able to write.

[31:06] Bradley: I thought, “That’s actually a very interesting take,” because the history of software engineering was, before AI, it took a long time for you to write something. So you had full understanding of what you wrote when it came to code review processes. People trusted you, but also there wasn’t as much volume, so you could review code and not block other people. Now it takes one hour to generate 10,000 lines of code, and it’s your job to trim that down, have validation, verification, all that. But once we get to the world where we’re able to ship and generate 10x the amount of code in an hour, where do we stop reviewing the code? When do we start trusting it more and more? I don’t have an answer for that, and I don’t think a lot of people do. But maybe it’ll become more front and center. We’re seeing a lot of AI companies build their own chips, or just chip companies are pivoting to do a lot more data-center-focused tasks to make inference more optimized. And it feels like all the arrows are pointing this direction. We’re in AI v2, where agents have picked up, the coding models are really damn good, and you can almost put your hands off of the code and let it drive. But once you’re able to do a 10x volume output of code, at that point, I don’t know. It feels like you’ve got to really let go, but then that’s scary. Security incidents are the end-all be-all for all these vibe-coded slop software, and those hurt. You don’t really recover from that as a business. So how do you move fast as a business and try to review all this AI-generated code, and do you even need to? We should definitely talk more about that, but that one really kind of spawned off. I definitely review the code, but yeah, there is a world in which I think things will change, and I don’t know when that’ll be. I think we’ve said this for a while, but maybe it’s sooner than later. I don’t know.

[32:42] Bennett: Yeah. My review these days is trying to make sure that I understand it didn’t add anything I didn’t want, because that’s the tendency. It’s like, “Oh, let me also do X, Y, and Z.” It’s like, “I didn’t ask for that. I don’t want you to do that.” And so a lot of times it’s kind of steering that or trimming some of that out. But the actual functionality, I’m like, “Hey, run your tests and make sure everything looks good, and if everything looks good, go ahead and commit and push,” and that’s it. So yeah, it’s pretty remarkable. And that’s a long way from just a year ago. I remember when you were talking about that Empower MCP back in probably November or December 2025, I want to say. That wasn’t that long ago, and now here we are with the full-functioning plug-in with your finances and charts and all that kind of stuff. The pace is crazy.

[33:38] Bradley: Yeah. Once that came out, I was like, “You know, I love the software that I wrote, and I always have a connection to it,” but one, it was vibe-coded heavily, but it worked. And two, this experience is way better. So I was like, I don’t think I’ll ever touch that again, but that’s okay. I want a better experience, and however I get that, whether that’s my software or a published full connector within ChatGPT, I don’t really care. So that’s the beauty of it: the experience is great and software is cheap now. I don’t feel like, “Oh my God, I wasted my time,” where if I had worked on a side project like this and it would take me weeks and weeks, and if I got Sherlocked by some official company releasing something, I’d be like, “Oh, that’s kind of a bummer.” But now I actually could care less, and I prefer the better integration. So it’s a very weird world we live in, where software is so cheap, but high-quality software, having judgment, having taste, that stuff is still incredibly valuable. Most people can detect when it’s an AI-generated website and UI and all that stuff. So I think the models still have a long way to go to getting good taste, and that’s hard to put your finger on. I’ve seen startups come out trying to solve that problem, like understanding what a good design is. Is that design tokens, coloring, spacing? How do you grade a good design? Very interesting questions, but I think the models still have some time to go there.

[34:54] Bennett: Yeah, for sure. Crazy world. Crazy world we’re all living in.

[35:02] Bradley: Yeah. And then the last thing I want to talk about, which I think has probably confused a lot of people, maybe even yourself, Bennett, is there’s been a ChatGPT work mode. So if you open up the mobile app or the web interface, essentially, there’s a toggle on the top of your chats, which is essentially chat or work. And the big factor with work mode is it’s kind of closer to Codex. If you’re a developer and you use Codex, Codex can get a lot of things done. It’ll search the web, it’ll create Python scripts, you name it. And it does this all kind of on your computer. ChatGPT work mode is the culmination of that, but this is more a remote server or a virtual machine. You ask a query, and it’s able to use that machine to fetch data, write Python scripts, execute it, build these artifacts for you, and do much longer-running processes.

[35:49] Bradley: So chat is, I don’t know if this is the perfect phrasing, but in my book, it’s more for quicker questions, not as deep mathematical or researchy. And work is for more complicated things. It can use a browser. So there’s a cloud browser now, which is crazy. But as we’ve talked about this whole agentic browsing, now it can open up a browser for you. Chat mode cannot do that. So this is kind of the differentiating factor: do I want to use the browser? Do I want a virtual machine? It’s a little bit fuzzy. Honestly, the reception has been mixed, but I have been leaning to work mode for the majority of things, and only for the simple things have I been going back to chat.

[36:27] Bradley: I think that part is a really big distinction of, “Do I want to use the browser?” One thing I was trying to book was the Odyssey tickets. And I was like, “Hey, this is where I live. I want IMAX, and I want you to go figure out the seats.” So it basically found my zip code, found the movie theaters around me, opened up a website, which I’d never heard of—I can’t remember the name—but essentially had a seat map view and rendered the web page, took a screenshot, analyzed what seats were available to me, and told me, “Hey, nothing is really good this weekend. There are only seats in the front or a single seat in the back.” And my query was two seats that are not in “you’re going to break your neck” territory. It’s really cool to see ChatGPT work go off on that task, which I know Codex would be able to do because it has my browser. I give it control due to that Chrome plug-in. But ChatGPT work is a closer rendition of that, where it’s able to use a browser, and that one feels like a big unlock to me. So if you’re confused about it, this is my hopefully simple explanation. But I would lean toward work for the most part, chat for simple things. Definitely ask it to use the browser too. If it’s something like movie tickets, I think that’s a great example. I think they’re rolling out support for reservations, so you don’t necessarily need a browser for something like a Yelp reservation or I think it was Resy, maybe. I can’t remember the exact connection. But some of these things are being more built in. For things that don’t have official support, the browser is super valuable, and it does really cool things. So I’ve been using it a lot. I’ve loved it. It’s a little bit confusing, but let me know if you’ve had good experiences with the browser and the work mode.

[38:01] Bradley: I think it really can do a lot, and you kind of have to reframe your mind around asking prompts for things that might seem impossible before but are now possible.

[38:12] Bennett: Yeah. So that announcement, I think, came out right around the same time as the 5.6 Soul and Luna and Terra models. And so I think my main focus was how amazing those models were. Like 5.6 Soul is my daily one. I usually just have it on medium if it’s anything, and then if it’s actually complicated, I’ll go extra high. I don’t ever use ultra or fast mode or anything like that. But on the actual integration of the apps, I will say it was still a little bit confusing for me, just because the main thing, at least when I look at mobile now—and maybe I need to update—but when I look at mobile now, there’s chat and there’s work. But then there’s nothing called Codex. It’s just called remote. And that is how to get to Codex on mobile, whereas on the actual ChatGPT app, there’s the dropdown to do ChatGPT versus Codex. So it’s almost like on the mobile side there’s no recognition that Codex is Codex. It’s just called remote. I know what it is, but just from naming and consistency, that’s kind of where I get a little confused. But yeah, I pretty much only use chat for everything because a lot of my stuff is just little questions here or there. But when I’m actually doing work stuff, I tend to still just use Codex. I haven’t used work very much, maybe a couple times here and there. I’ve always either defaulted to Codex or, if it’s quick questions, just chat.

[39:40] Bradley: Yeah, I think the last thing about work mode on mobile is that it does have access to those automations that feel a lot more first-class. So I think the finances automation that I have for setting up recurring weekly snapshots uses the work mode. I might not be 100% correct on that, but it feels like the scheduled tasks, which I think was the old rendition of doing things on a cadence—there’s a new product name for it. I can’t remember what it is. But essentially, I think it uses work mode, since it does have these VMs on the fly that it’s able to do much more robust integrations.

[40:12] Bradley: But exactly as you’re saying, I think the Codex tab, which was on the ChatGPT app, got renamed to Remote. And now there’s Chat and Work. So Chat and Work: simple things on Chat, more complicated things on Work with browser access. Then if you want to go back to your computer, hit that Remote button; that’ll get you there. Then on the macOS side, yeah, a little complicated. We had the ChatGPT app, we had the Codex app. There was the big merge, which maybe some of you have heard about. If you’re not on X in the AI bubble, you probably haven’t heard about it. You basically update your ChatGPT macOS app, it gets kind of replaced into this Codex app, and depending on what mode you choose in this onboarding flow, you can stay in Codex or you can go back to ChatGPT Classic, which is the old-looking ChatGPT mode. But little bumps and bruises. I’ll be honest, even myself, as an employee of OpenAI, using the software at home and doing the upgrade path, maybe a little bit to be desired. But I think once you get through that hurdle and the way things exist today, it’s still not crystal clear on what goes where, but it’s getting closer to this ultimate merge of, you don’t have to think about what mode you choose, and things are local, and you can access your computer. It feels like there’s a good product vision that’s coming into play, and these things are merging, which is great. It’s not in its final end state, but it’s getting close to what feels to be a first-class experience. So if you haven’t used the work mode, definitely try it. If you haven’t used Codex Remote, definitely try it. I use it a lot, basically connecting your mobile app to your computer, and you can just ask Codex things on the go.

[41:46] Bennett: Yeah. I think in that scenario, people who were used to using the ChatGPT desktop app were going to feel aggrieved with Codex, but the people who were used to Codex didn’t want to go back to the ChatGPT app. But it makes total sense why the merge happened. I think it was just people got into their routines and what they were happy using. It’s funny, because from my own personal experience, starting from the very beginning, I used Cursor in an IDE, and so there was my AI in that. And then I remember you were kind of pushing the CLI, and I was like, “Okay, fine.” I finally got on the CLI, and then I remember using the CLI and being happy with it. And then I remember, like we talked about this a couple months ago, you were in town where I’m at here in Arizona, and you’re using the Codex desktop app. I’m like, “Really? You’re using that after you were all about the CLI?” And you’re like, “No, it’s the best.” So now I’m using that. So it just depends on what you enjoy using, and everyone has their own different preference. But yeah, there’s just been so much change in the space. It’s just hard to keep up.

[42:54] Bradley: It is hard to keep up. I mean, yeah, we’ve been gone for three months, and a lot has changed. Obviously, this is a snapshot of what exists today. But yeah, in two weeks, who knows what’s coming out? Who knows what’s going to change? Who knows what companies are forming or disbanding? Honestly, right now, it feels like I’ve said this before: it feels like you can pretty much do anything. I can ask for things in Codex using 5.6 Soul extra high, and it just gets the job done. I have upgraded my PHP versions from 8.3 to 8.5, which, for people who aren’t familiar, it’s not a huge upgrade, but for software things, it’s super boring. And I kicked off a Codex thread, and it’s doing it. It’s monitoring my deployments and doing all that. As we speak right now, I can see it monitoring things, which is great.

[43:35] Bradley: Even, like, I’m trying to sell my Snapchat shares as a former employee. I’m ready to let it go. If you’ve looked at the stock, it’s been a little bumpy, and I’m ready to part ways. And I told it, “Here’s my logged-in bank. Get ready to do the trade and monitor the trade in the morning so that when it lands, the price is good.” It just feels like you can do anything, and that’s a fun, exciting thing. And these things only get better and better. If I can do all this today, what’s it going to look like in one year, in three years? I don’t freaking know, but that’s exciting and also scary too. I think that part’s to be figured out. But yeah, I think right now, if you open up the Codex app, you can do a whole lot. If you open up ChatGPT, you can do a whole lot, but I’m definitely living in Codex, and I have a great time there.

[44:19] Bennett: Yeah, cool. Should we wrap it up with some bookmarks, Bradley?

[44:22] Bradley: Yeah, let’s do it.

[44:23] Bradley: Kick us off.

[44:24] Bennett: All right, so I’m actually going to do one, and it relates to what you were just talking about. So there was a white paper, I think is what you call them, that was published recently about Codex usage.

[44:36] Bennett: And so I’ll include a link in there.

[44:38] Bennett: But what I found really interesting about it was it talked about how users are using Codex.

[44:44] Bennett: And it ranged from, I think, people that actually work at those frontier labs versus people that are pretty good users of AI versus just completely new users.

[44:53] Bennett: And so they segmented it into three different buckets of people.

[44:55] Bennett: And basically the premise was, what can the highest users tell us about how Codex is being used across organizations? What can non-engineering—I think that was the bucketing. I think it was engineers at the frontier labs, maybe engineers elsewhere, then non-engineers or something like that. Basically, it was like Codex was showing huge growth in non-engineers. So think finance, legal, marketing, and what are some of the traits that were being noticed and observed as why that was?

[45:28] Bennett: And it was super interesting because one of the key themes was obviously the desktop app and being able to work alongside files on your computer and your workspace, which is super useful. Taking it out of the web app, which a lot of people were still stuck in, especially non-engineering people who maybe don’t think and live in the X AI world all the time. And so being able to realize that you can do a lot more things on your computer with browser use and all that kind of stuff.

[45:53] Bradley: Yeah.

[45:53] Bennett: But then, two, the focus was really on delegation versus automation, and I think that’s something that really stuck with me. In my career, being in accounting and being a leader of teams, sometimes people think about AI as automation. And it can automate, certainly. It can write Python scripts or Bash scripts and do a bunch of cool stuff for you. But the way that you need to think about it is more like delegation. And we’ve talked on this podcast a lot about how it’s kind of like an intern. You’ve got to give it some direction, and you’ve got to monitor it a little bit. And it’s getting smarter as these models get smarter, but the overall point was that white paper made it very clear that it’s really about how you can delegate and kick off tasks that free you up to go do something else. And so the productivity gains are really measurable once you adopt that, “Hey, I can tell Codex to go do this.” It doesn’t have to be Codex, but this is what the paper was written on. But any AI that’s living in a desktop app-type scenario, I can go ask it to do these two or three things, and then I can go do something else. And that just frees up your time, makes you more productive, and the intelligence is there where you can trust the results that you’re going to get better than you probably could back when it was o3 or o4.

[47:17] Bennett: So that was a really interesting thing. I’ll include the link, but yeah, that was my bookmark.

[47:23] Bradley: Yeah, that’s interesting. I had one tidbit on that. I was trying to unsubscribe from emails to reclaim my email inbox because I am in spam hell, and there’s no official Apple Mail connector. So this weekend—yesterday, actually—I had asked Codex, “Go through the past 90 days of my emails and unsubscribe from things that are purely marketing. Keep transactional, and if there’s anything you’re not sure about, ask me.” And it made it through all this, used my computer to page through all my email results, and unsubscribed from a ton of them. But one of them, it tried unsubscribing and said the link was bad. And I looked at it, and it was this crypto-looking link, and I go, “How about we don’t click on that one next time?” So I guess that one, it almost made it. Not that I got hacked; everything went well. But I looked at it and thought, “Yeah, we probably shouldn’t click on that anymore.”

[48:10] Bradley: And then right after that, I turned it into a monthly automation in the Codex app. So one is do something you don’t want to do, and you give permissions and you work on something else, which I was doing. Two, you come back and monitor and check the results. And then three, if it works out, package that into a skill, which you don’t have to do anything anymore. You just say, “Set up a recurring automation to run this,” and it knows what this is since it’s in the chat, every month. And so I have an automation to unsubscribe from emails that I don’t care about in Apple Mail on my Mac computer running every month. So yeah, there’s a whole plethora of speedups right now that are quite insane, which I love.

[48:50] Bradley: But yeah, jumping to my bookmark: if you’re using AI tools and Claude, Codex, any AI model is explaining code to you and you’re scratching your head thinking, “There’s got to be a better way to understand this,” there’s a tweet recently by Dex Horthy. Hopefully I said that right. He works at HumanLayer, which is a company that I believe is trying to build an IDE for teams, like a multiplayer AI IDE. He gives a ton of talks at conferences like AI Engineer, Claude Code meetups, etc. So he’s definitely in the know.

[49:21] Bradley: Basically, he created a skill called slash show me that you can invoke that will just show me a visual representation of how the code changed, not necessarily a giant blob of text.

[49:31] Bradley: There are different artifacts. There’s an HTML artifact, which will do more visuals. There’s a call stack diagram, which kind of shows the functions that it calls before and after. Things that you don’t really have to know and dig through this explanation, which can be very technical and jargony, but more: give me something I can understand at a high level. And as we talked about reviewing less code, I feel like this comes hand in hand, where I might not know what this function does internally. I trust Codex to figure that out or Claude to figure that out. But this is the general structure, this is how it feels in my codebase, these are the things that matter to me. I still need to use this skill. When I saw it, I thought, “That is a great idea,” because there are times where I am reading explanations from the latest 5.6 Soul model, and I get why it’s saying this, but man, the mental load on reading some of these things can be quite high, even for someone who has been doing software for a long time. I look at it and I’m like, “I’ve got to read that two or three times.” It’s important stuff, especially working on important code, but there’s probably a better way to package it. And so I’ll try this skill out soon. But if you haven’t heard of it and you run into the same problem, definitely try it out: slash show me. And it’s like a simple npx install. It’s `npx skills add humanlayer/skills --skill show-me`. I’m not going to type that out, but you get how easy that is. And then you can invoke it from your favorite agent.

[50:53] Bradley: But very cool.

[50:54] Bradley: I think we need more of this.

[50:56] Bradley: There are de-slopify skills that are coming out, and show me skills, kind of the next evolution of how you can use AI coding tools a bit better given their high intelligence and understanding these days.

[51:06] Bennett: Yeah, no, I think that’s super useful because, I mean, I agree. Sometimes it’s like all the code that’s written, like we were saying, I don’t really feel like I need to go and read all that, but it’d be nice to get a high-level understanding. And it’s interesting. HTML has really picked up as a good way to visualize things again, because in the past you had to write it, and HTML is very lengthy and strict on the rules and stuff like that. And you had to know the tags and CSS and all that. But in an agent world, it can spin that up in two seconds and give you a beautiful—like, I just had something when I was at work. I was trying to understand a high-level picture of my team’s work and the tasks that they were doing. And so I just gave it to Claude, and it went through and made this really good Sankey diagram. It wasn’t exactly a Sankey diagram, but just kind of a flow. And I was like, that was so much better than an Excel pivot table, which is what I would have defaulted to two years ago probably. And I’m like, that’s the way. So no, that’s cool.

[52:07] Bradley: Yeah, I mean, even sometimes I ask it to make an artifact that I want, but other times I’m hands off the wheel and just say, “Create me a visual.” That’s all I’ll say, and I see what it comes up with. Sometimes it’s not perfect, but we’re getting to that point where I think there are some discussions that giving more context almost hurts the model because it knows more than you. We’re not fully there yet, but interesting conversation. So yeah, definitely check out that skill. And if you use it, let me know and let us know in the comments.

[52:34] Bennett: Cool. All righty. Well, let’s wrap it up. We’re back with season two and getting back in the swing of things. So good stuff to catch up on.

[52:40] Bradley: Yeah, and we’ll come back at you with the next episode here in a little bit. But cool. Good chat, Bennett.

[52:46] Bennett: Sounds good. See you.

[52:46] Bradley: Awesome. See you, guys.

[53:10] Speaker: My dad.