Katyella · Episode 5 · October 2, 2026 · 50:39

I Shouldn't Anger the Time Zone Gods

First guest episode: Ben Sun of Meadow joins from Berlin. Pirates vs the Navy on startup life, whether agile survives AI-speed development, review bottlenecks, bigger tickets, multi-agent testing, and what managed infrastructure is still worth. Plus the continuing saga of Nina not charging for her fiction app.

Episode five, and our first guest: Ben Sun, software engineer at Meadow, four startups deep and a Berlin neighbor.

What we talked about:

  • 0:17 Ben’s startup journey: pirates vs the Navy, and why he keeps choosing the pirate ship.
  • 4:20 Where the bottleneck moved: AI writes faster than anyone can review.
  • 11:01 Is agile even working anymore at AI speed?
  • 19:38 Agent search and model costs.
  • 24:06 Multi-agent testing and workflows.
  • 29:51 Bigger tickets and coordinating a team around them.
  • 32:41 Rethinking AI code reviews.
  • 38:46 The continuing saga of Nina’s fiction app and the money she refuses to make.
  • 43:03 Rebuilding media apps in Rust.
  • 45:29 OpenSearch and what managed infrastructure is still worth.
  • 47:17 Personal AI assistants, and why Matt shouldn’t anger the time zone gods.
Transcript

Automatic transcript, lightly corrected. Also available as WebVTT.

Matt 0:17

Hello nina oh matt i have a guest today you do this is ben ben yeah ben's a buddy of mine out in berlin with me he's been saving the world with various startups he's been working now let you introduce yourself

Ben 0:35

some more about uh i'm ben the software engineer i studied compute engineering and then became a software engineer that was about 20 years ago. I've been with four different startups and in between I worked for larger companies and quickly fled those. The CTO that I originally worked for, my first job out of college, she compared it to being a pirate at a startup or working for the Navy. So the pirate lifestyle really appealed to me. So that's why I keep joining startups.

Matt 1:15

Yes. Go ahead.

Nina 1:17

I was going to ask, is it you prefer the startups because there's less red tape and more freedom and of technologies you can choose and stuff like that? Or I guess, why do you prefer startups?

Ben 1:29

I think it's a personal preference, and neither of them, I would say, is wrong, and everybody has their own preference. But I like startups for the ability to bring on larger changes in the product and in the code base. There's much more space for experimentation and exploration. At a startup that I worked at, we really leaned hard into microservices and almost every two weeks we would tear down a microservice and rebuild it, sometimes in a different language. And that is obviously harder in a larger infrastructure, in a larger company where there's significantly more stakeholders, where the product is much more mature and where I would say product fit has already been established. Some of the startups that I've worked for had to pivot, and that I think is also a very interesting process that I was really happy to be part of.

Obviously, it's a bit of a downer to figure out that your business model doesn't quite work, but being able to create something new is very exciting to me and it's very appealing.

Matt 2:44

What size are you guys at now, developer-wise?

Ben 2:47

I think, don't quote me on it, but I think we're 12. Yes, we have some integration engineers and I would like to have them be included as engineers. And I think the threshold for what is an engineer now with AI has definitely changed. But yeah, about a dozen.

Matt 3:11

And how many products would you say? Like just one product line or multiple? We have three products.

Ben 3:18

The first one was a wedge product. The company is called Meadow. It's in a higher ed space. So we take tuition payments and process them for students and really provide more transparency in that process. This is the second higher ed startup that I've joined. The first one is called Simplicity. And that was also my first job straight out of the college. It's definitely a very interesting space. Higher ed, a lot of very smart people with really bad software. I don't know if you are familiar with that space at all, but they're very reluctant to change. Some of these institutions are hundreds of years old, and a lot of them are also public.

So that brings its own kind of hurdles in terms of regulations and rules that need to be followed. The release cadence kind of follows the academic year. So sometimes it's very hard to release something mid-year But yeah, it's been very interesting.

Nina 4:20

So with your experience then in startups where, you know, you kind of have to refactor something or completely redo an entire microservice, how has AI then kind of changed like your day-to-day workflow in a startup?

Ben 4:37

A specific example where we started to cut you off, but that specific example back then we still used to handwrite code. But the olden days, we still had IDEs and all that good stuff. I definitely, I mean, development as a whole has gotten so much faster, right? And rewriting something with well-defined contracts, which is obviously what enabled us to swap these microservices in and out, having well-defined contracts. I would imagine it would be significantly faster now. And if it is Go or JavaScript or TypeScript or whatever language, I think it has become less relevant now. with AI.

I don't I've coded in all of these languages I've also coded in Go and I've done Ruby on Rails and I think Yeah I wish we had AI back then to kind of coach us and kind of create these things faster, allow us to do much more fast prototyping to experiment more.

Matt 5:43

Would you say like in, where's the bottleneck now? Like for the larger corporations, it's the people. They can't keep up at the pace of the developers anymore. They're just cranking out code too fast.

Ben 5:59

It's funny that you say I'm sitting on six PRs right now waiting for reviews and it's a real problem that we've been discussing a lot at Neto. We've enabled product now to submit PRs but they don't they don't review them, so they're just contributing further to the problem. We obviously have Claude look over and provide code reviews and score them and a lot of our reviewers know based on that and I still find myself looking at the file changes to kind of double-check that, but I feel like that's maybe a habit that I have to get rid of, right?

I still want to understand how the code works and to truly fully understand how something is implemented rather than just reading the blurb. That Claude or Codex pops out with the warnings and whatnot. But yeah, it's code reviews right now to have a better understanding of what we're actually creating, which I think we're somewhat losing touch. This has been a comment that I've provided in the past. We're so fast to create things, we're losing touch a little bit with what exactly we're creating.

Nina 7:19

Yeah this is that is a huge issue that i mean i experienced it at my last gig where things were moving too fast, and just using using code like automation or pr reviews pr pr review automation with Claude or whatever tool, i i don't want to lose touch with, like what we're actually merging myself. So I know that a company's project manager would potentially get upset if you're, spending extra time looking at differences in files. But I think every dev I've spoken to has felt like we're losing touch with what is actually being written.

And I feel like that's down the line. We'll always have AI to kind of correct things now, but I don't want to have like a fire drill randomly in my career because I didn't use my own eyes to look at code, if that makes sense. And although like fire drills are like far few between because of AI, it's still something that I miss doing my code reviews a little bit. Matt, you said you're code reviewing six PRs right now at once. Yep.

Matt 8:38

So has Ben. Ben's got nine, maybe? I submitted them. I'm just waiting for somebody to tickle I only got three of my own out there.

Nina 8:48

Yeah. And actually, there was a video that was released this week by the creator of Rails, I believe. And he was going into a whole bunch of topics about just the speed of things and the ability to program in different languages now, but Everybody's coding less. And I always think, how is that going to affect, how is that going to even affect our ability to review code down the line if we're not coding as much? So I always think, like, where's that balance now? He has a line in this video where he says, he asks the audience at the summit, how many of you in the audience actually write a line of code now?

And five people raised their hand in a room of like hundreds and just with that in mind i mean is there anything that you guys do, do you guys actually take the time every once in a while to code something by hand and maybe just use Claude as like an assistant in a new language because, i do that occasionally because i'm always worried about skill regression. When it comes to actually architecting and coding because i learned how to architect, applications through coding and making mistakes along the way so this is a question do any of you code for yourself by hand anymore, no no okay.

Matt 10:19

Even for like a one-line change i'll have claude do it

Nina 10:22

Really all okay.

Ben 10:25

I think the best analogy that i've heard is i know i know how to drive a stick shift but in an automatic car why i think the knowledge is good to to to understand how gears work and how a clutch works and and whatnot and gives you a better appreciation i think it makes you a better driver but like same as in today's cars they have even automatic cars allow you to shift up and down but why would I want to do that? And even worse so, I don't have the time to do that.

Nina 10:57

Right.

Matt 10:58

I think we're hitting a wall with AI and Agile development is not working anymore. We had the same thing happen with Waterfall and Agile when the convention over config started happening and you were able to crank out some more code. And that started to push more into Agile. Now we have AI, we're going faster. And I think really where it's hitting and seems Ben's also hitting it, like PR is where is like the tip of the spear of the breakdown of agile. And I think PR is going to have to go away.

I don't think we can rely on them. Human PRs, it's going to have to be agentic reviews.

Ben 11:46

I mean, it's checkpoints, right? We need more checkpoints. And over the last year, more so than ever, my work style has changed. At first, AI would generate the code and I would still write the unit tests, right? To at least know and guard against, to be able to write down the scenarios that I want to guard against. Then that went away and I was never really a huge fan of writing tests to begin with. but then we relied on end-to-end tests and now there is no time for that.

Matt 12:23

There's no time to write them or there's no time to run the suites?

Ben 12:27

The suites are definitely being run but I'm not generating them anymore. I'm not handwriting them anymore. I'm not instructing Cypress to say, okay, now you click on this button and that button now. It's all done via code, via Claude.

Nina 12:42

Yeah, now that Agile methodology is kind of not as fruitful as it once was, and we require, I guess, I mean, do you guys think we need less checkpoints, right? I just wonder how it's going to change if we're not doing PR review anymore. Let's say PRs hypothetically fall to the side because we put so much trust in agents. Do you think companies are going to adopt maybe having an end of sprint kind of review where you kind of go over everything that's been merged in? And I guess, where do you guys see the next kind of, what do you guys see as like the next steps? Like what would be a good plan other than agile for modern AI development?

Is there anything like in your day-to-day that you wish people implemented or your project manager implemented that kind of accompanied this new fast-paced kind of work environment?

Matt 13:49

It's rough. i mean trying to invent a whole new system off the top of my head

Nina 13:53

Right right but you know while you're working do you ever think man i wish we had a system for this or that or this or that i mean like, i i i think like i've thought with everything moving fast like everybody we were working together matt i thought that, i had the idea like oh i wish we had like yeah an end of sprint meeting where we go over absolutely everything that's been that's been merged or just have an overview view of every PR and everybody look at it together and just have like a big overarching like kind of architecture discussion.

Matt 14:27

You heard of shape-up or Linear cycles

Nina 14:31

No I haven't.

Matt 14:33

So basically the shape-up version shape-up and Linear cycles are basically the same thing where you have like six week sprints and then you have like a three-week cooldown afterwards and that would be like kind of where you fix bugs you do kind of architecture reviews of what's happened That could possibly be a new system that pops out of this. But that's also written up by DHH as well, ShapeUp is, and then Linear kind of took it and ran with it. So Linear is based off of longer cycles, not just two-week sprints.

The idea kind of being that two weeks is time too small to do anything.

Ben 15:17

These days, two weeks is too long. so you get a lot done in two weeks compared to before

Nina 15:25

Yeah yeah it seems like this type of methodology shape up it, six weeks is a long long span is that like six weeks before release like i've never heard of companies i guess using this before.

Matt 15:43

Yeah i've only heard of Basecamp using it. I've never heard of anyone like in a larger company using it. It's just too loose for most larger companies. They want the kind of two-week control over what gets done and not a divergence between like an infrastructure team and an application team.

Ben 16:07

We used it for a second at the startup that I'm currently at and we abandoned

Matt 16:12

it. What happened? What happened?

Ben 16:14

It cycles for too long. I also, by the way, I hate this intense focus, and I think it's American, on finishing something at the end of the quarter. I don't know if you guys ever experienced it, but for a lot of startups, I did a stint at PayPal, and at the end of the quarter, everything had to be done. Something had to be delivered, and if a project didn't fit, it either didn't get tackled or it got broken up. But yeah, we're I mean, we've also been changing things fairly rapidly.

So we used to be two weeks friends, then six weeks cycles, and we're trying new things, which is also something that gets afforded to us by being a startup.

Matt 17:06

Yeah, like the quarterly stuff's coming from the top usually, because you have your quarterly deadlines and that gets pushed down to managers' reviews and the managers have to then hit things on the quarter.

Ben 17:16

The quarterly reports also for public companies are also very intense. I've heard DHH and conventional configuration. Am I amongst Rails people?

Matt 17:28

No, I've done Rails stuff, but also Spring picked it up. So Java does it now, too. Java prefers to keep it to annotations instead, but a lot of the stuff is spilled over. And I've done a lot of Python, like FastAPI and stuff like that, which is just Python Rails.

Ben 17:49

I've done Ruby on Rails for about a year for a first startup out of DC. And we went to a Ruby on Rails conference in Portland, out of all places. And it felt a little culty.

Matt 18:05

It is culty. Have you watched the video? It's great. He basically tells them Rails is stupid and stopped doing Rails.

Nina 18:11

And he made Rails. Yeah.

Matt 18:14

He's just like, we're done. We're done. What are we doing here, guys?

Nina 18:18

Yeah he became a big rust guy yeah because of, like he said i hate he said i used to hate rust when i had to look at it now because i don't have to look at it and an agent can write it i love it it's the best thing ever i hate rails it's really funny yeah i think we linked it in the document.

Matt 18:36

Yeah we'll put in the show notes too yes you should definitely watch it yeah he just like he just rails no pun intended into other developers and it's basically he says 37Signals doesn't write code anymore all the AI writes the code he reviews some of it and so do other people but mostly it's getting like checked automatically with AI and merged in

Ben 19:05

there's still a sunload progression in my YouTube consumption I used to listen to things at 1.25 and then I cranked it out to 1.5. Now I just put it into Gemini and tell it to summarize it for me.

Matt 19:19

I think that's exactly what Nina did for the video.

Nina 19:21

Yeah, that is exactly what I did. I have it open right now on my computer. It transcribed it with Whisper, and then I pulled out keynotes because I just didn't want to listen to him talk, and I could just read it faster. And that's actually one of the projects. I don't know if you showed Ben ETA brain. No, I haven't.

Matt 19:45

We were talking about it earlier. here but yeah yeah but he also goes into that too he's into our idea of agents as search basically the idea is like Elasticsearch is garbage and so unless you're like google able to like do the customization needed to get what you need out of it it's a lot better just to have with an agent that could go through it and do the search for you. If, especially for us, like we're just transcribing podcasts, so the data size is not large.

So you can grep through the whole thing fast enough with enough agents just ripping through it. As long as you can spin up those reps fast enough, it works well.

Ben 20:35

I just have done.

Nina 20:36

Oh, sorry.

Ben 20:38

Sorry, go ahead.

Nina 20:39

I was going to say we're thinking about how to add Jev into it as well. That's like next task. If you think it's something that you want to do, Matt, I'm kind of curious for that.

Matt 20:51

Yeah, I was wondering where you got to it. Jev is just a classifier. Basically, it would help you pick if results returned are good or not, and then it would be a filter step in the process of the agent's aggregating data.

Ben 21:06

I did during the last episode you guys talked about Jev and I mostly agree with it, it's a classifier I do think the one legitimate use case that I've seen so far is picking a cheaper model or analyzing a prompt and then directing it towards a model to have a little bit more cost control

Matt 21:27

I think since that episode an open source model has come out so Jev was was proprietary and the only thing out there for I think three or four days which is amazing by the way the open source one's called Jeff

Nina 21:44

Oh my gosh, Wow, that's four dates. Things are moving so quickly. I don't even know how to keep up anymore. I need an agent to summarize all the new things coming out and determine the best one. Yeah, so Jeff is completely open source. So is it that you're saying that we don't, there's no, I mean, it was cheap before with Jev. Jev wasn't that costly.

Matt 22:13

Yeah, and I don't know what the system requirements are for it. It may eat up our tiny Hetzner facts.

Nina 22:22

Yeah.

Ben 22:23

I also think cost controls are coming, right? The days of token matching are long gone. I haven't received instructions specifically, but I don't send everything to Fable anymore, right? Sonnet or Opus can do the same task. Hey, create a feature flag or do this one or two-line PR. I don't need Fable to do it. But prices are going to dramatically increase. And I think that is the interesting part about Jev, right? It's supposed to reduce costs. It's a cheap alternative, which I don't, at least from my point of view and from what I've read, cost controls haven't been in the super forefront.

It's been more about creating better quality and better models.

Matt 23:14

I mean, even on the user side, we're still in POC phase for almost everything, implementing AI agents. Like, you're not really going to care about that up front until you have something working and then you're going to fine-tune the costs.

Nina 23:32

I personally have never hit my max token usage on like a $200 Claude subscription. But I know like if you work with Java or if you work with different types of languages that eat up tokens, I know you can hit it pretty quick. But even like with making my own harness, I have not hit the token usage. I see a lot of people complaining a lot of the time that they hit that token usage, but maybe that's where Jev fits in for them. Have you guys ever hit your token usage on like a $200 subscription?

Matt 24:06

Yeah, only when I'm doing multiple agent work. And that's usually when I'm doing UI work. The backend rarely uses that many tokens. Backend works pretty straightforward. Where UI work will require images and Cypress tests going back and forth over the agent, which will eat up most of the tokens.

Ben 24:28

I think something that I've done that definitely I mostly use the plan provided to work but if I try to one-shot something I've definitely asked it, spin up five agents come up with a plan, align them all against each other or sometimes I even tell them to build the entire feature and then compare them against each other that is expensive that is expensive and we're churning through a ton of tokens

Matt 24:57

are you talking about like two different like layouts of uis and show them to you or or differently

Nina 25:03

Be building it this way versus that way and doing direct tests performance with two different versions i've done that and that's definitely eight up tokens, but i don't do it often but there's value to it for sure, Is that what you mean?

Ben 25:21

Yeah, so in Cursor, there's a multitask mode. So spin up five different models, use five different models, use Grok, Fable, and whatever else is out there, and then just perform that same task and then come to a consensus decision, measure the performance of each solution, essentially.

Nina 25:41

That is really, that's, yeah.

Matt 25:43

How do you digest it when it comes back? Do you have it make a separate rapport? How do you judge what comes out of it?

Ben 25:54

I will ask you to summarize the findings. I'm having a really hard time reading AI Slop, and I've been trying to add things to the agent to make it easily digestible. Definitely one late night, I got a response from an agent, and I was just like, explain it like I'm five. And then I still didn't get it. And I jumped to explain it to me like I'm a golden retriever. I just kept on going down. But yeah, it's difficult to search through that. But yeah, for that specific process where I use multitask mode, I have it, summarize it in the end and bring me a consensus decision.

Nina 26:42

I did that with ETA Brain. And that was the first time I've done that. I asked it to look at the past two branches that had been merged where, we updated the Elasticsearch. We migrated from having our server in front-end-based migrated into the Java code. And then I said, okay, let's do a comparison in speed when asking questions. And so it checked out master and, you know, it rebuilt everything. It did Claude and Chrome. And yeah, it took like an hour and a half. And I did eat up a bunch of tokens, but it came back with a nice table where the speed improvements were in terms of how it went through every single pipeline.

And that was my first time using AI to do an extreme test like that. And it was something that didn't even dawn on me. Like, oh yeah, I can do these huge performance tests really quick now with AI, whereas that would take... Long time to the point where i wouldn't even cross my mind to something feasible to do because it would just be such a time time suck but yeah, i uh but i still didn't hit my max token so i still didn't i still didn't hit the end which was weird, but yeah do you come up with um i'm curious in your workflows ben if you do you come up with i guess like consistent performance performance testing like in md files like are there i, guess how do you how do you log your your your different prompts for doing robust testing i guess i'm asking broadly like what, how you automate things like is it is it that you do you use like a harness for your ai, workflow automation or is it more like you you have your own that you made or is it that you just you just rely on like a hierarchy of Claude files and, go-tos like what's your setup like when you work day to day it's.

Ben 28:49

The last one so we have a number of md files in in our monorepo so we switch everything to a monorepo like when we started really started using ai to to provide better context and these md files then become PR requests that we discuss at a bi-weekly engineering meeting to Create better outputs and set certain standards. Not necessarily coding standards, but I think lately we had some problems with the data model and AI pushing really hard for us to change it. But we were quite confident in it. So this is something that we discussed and persist and agents on defile and then tell it that this is the way we want it.

So it's something that kind of grows. It's a, for lack of a better phrase, a living document, and we discuss these things as a group, and it goes through the PR process.

Matt 29:51

How big are your tickets, you would say, these days? Are they like full features now, or are they still broken up into individual pieces?

Ben 30:00

I was thankful that recently, that they were too granular. I've thought about this, and I like granular tickets it's because I have done some self-reflection. Smaller PR reviews gain more feedback. And that is not true anymore, right? This just creates more PRs that nobody wants to look at. So they should be getting bigger. And that was my aha moment. So I would get handed a project and I would instruct to AI, hey, come up with a plan, break this into Linear tickets. We use Linear. We used to use JIRA. I'm not a huge JIRA fan, maybe bordering analogy, but yeah, they're supposed to be getting bigger.

And as part of that, it's going to be less PRs, late, which PR reviews are hard to combine. And if it's a thousand lines of code, I used to shy away from like changes, be like, whoa, this is like a thousand line change. Can you break that down? Nowadays, it's the same effort. If I don't look at the files, I just need to read to what the AI prompted, what the rating of it is. And yeah.

Matt 31:13

What about team size? Do you think, because now you've got bigger PRs, people working on one monorepo, you're going to end up with an unmanageable amount of conflicts?

Ben 31:29

Yes. I mean, from the time that I publish a PR, it will get rebased at least five times because people are just pushing so much more code, right? Team size, I'm a big supporter of having at least two people on a project, even if your secondary is just a PR reviewer or a sounding board or maybe even a silent partner or whatever you may call it because people get hit by the bus. It does happen. And you want to break that silo as fast as possible Because with us creating so much code and having things be in silos and nobody understanding what is actually going on is so much more commonplace now.

But we also want to create a bunch of features. So everybody's kind of like lone wolfing it with AI. AI as their co-pilot and then just punching out code. Which, again, I have absolutely no context. If there's no more PRD, no more engineering document, because there's just a plan, I don't know what you're trying to do. I'm just looking at the code and looking at the review and trying to decide if it's safe or not.

Nina 32:41

Should have something that i would imagine or like to have in the future is like when you're in jira or you're in you're in, you're looking at a diff if there was an ai summary of the diff in the context of the code base just at the top just off to the side like yeah you know up at the top, just you know for thousands of lines just have an ai summary for PRs rather than the actual summary that's, you know, at the top. Like, where you could click and it would automatically scroll down to each line that's relevant to that key point. I don't know.

Matt 33:20

I'm getting closer to that with my chief of staff, Bob.

Nina 33:23

Oh, that's right. I didn't think, yeah. Yeah. You technically have something.

Matt 33:29

I've kind of gotten fed up with GitHub and Bitbucket and slowly making my own. It's a lot more like this should do i even need to look at this is kind of what the purpose of it is now and it and then it has a button for me to just approve or it has in the or i have to go into the box and then look at the actual code and it's getting it's getting better i got it just gives like me a score out of 10 and then now it gives me which i added is helpful a table with like the top three reason why it's not a 10.

And usually, maybe not usually, at least a third of the time, I'm able to see why a bad score really isn't a bad score. And then I still have to fail out and go to the actual PR for the other two thirds. I really think that's like kind of, we're trying to get closer to like, what do you actually need from a PR? Give it a thumbs up.

Nina 34:29

Wish there was a way we could showcase it somehow.

Matt 34:32

I've got way too much client. I'd have to make a fake one. Maybe next time.

Nina 34:37

Make one with fake. It would take a second to make one with...

Matt 34:40

Yeah, I have it in the blog that I wrote about it, but an actual working one, I can make fake soon.

Ben 34:46

I've tried to do something similar once to get more of a feel for what the code actually does to give me each file and pseudocode out very concisely what a function does and what the change actually encapsulates. I don't need to, if it is a new endpoint, I don't need to look at the controller or I really want to just see what the service does, like strip all the boilerplate out and just give me the essence of it, but relate it to a specific line of code so I can actually take a look at it.

It felt very satisfying to do it, but again, I think it takes too long.

Matt 35:28

Reviewed takes too long? Looking at the actual code to actually trace that huge

Ben 35:34

change to like, hey, AI says the service now does something slightly different, and this is how it was implemented. This is the service file. Here are the three lines of code. There's an if statement there and whatnot. It just takes too long.

Matt 35:50

Yeah, like is there a shift to be like, now you need to kind of prove not to greenlight the PR? Like I felt before the idea was you need to see if this is good enough to merge and now I feel you need to see if this is bad enough to actually warrant comments

Ben 36:12

I still call grades it out of the five if it is a if it is a three or below I've got more work to do right and if I when this happened before and I and I argued against it and I got the the score up to a four. And a five out of five is generally a single line change or whatnot. I barely, I don't even worry about the files anymore. At a four out of five, I do want to see the comments. I want to see how they were resolved.

At a three out of five, those comments don't only need to be resolved. They need to be argued against. And there needs to be some argumentation on the developers to have on why AI is wrong and how it was resolved.

Matt 36:53

I've gotten my reviews to now when I actually do dive in. I mean, I'm still diving in with AI. It's going to give me failure cases that it thought of, but either they're not covered in tests. And so it has to create that test first before it's a valid comment to add to the PR. And then that code needs to go into the comment as well. And then if it's an architectural one, it needs to diagram the previous, the current architecture within the PR. And why thinks it's bad and then a correct one so that like there's something for i mean really the other person's agent to come in and and pick up and see and like actually bite into and try to fix with it

Ben 37:40

we have a skill it's called real or paranoid and i think with certain models there's definitely, I mean, Some of the warnings are really paranoid. Time zones, like what it worries about. I shouldn't talk about time zones. Time zone bugs still do exist, and I shouldn't anger the time zone gods. But there's definitely some things where it is over-worried, and so we developed kind of a skill to kind of ground it a little bit.

Matt 38:12

Yeah, it's definitely paranoid. It depends Depends on the model you're using, though. I feel Opus 5 now is a little more loose and just tries to get the job done than Fable was. I've been using Opus 5 exclusively now. I don't go back to Fable. I go to Codex if I need a second opinion now.

Ben 38:34

We found the same thing. Opus 5 is dumb enough to accept some of these things. Opus, Fable gets a little paranoid.

Matt 38:46

Do you want to change topics to Nina's fictional writing app? I want to know if you did, added the payment stuff.

Nina 38:54

No, I did not. No. Come on, Nina! Because the people who write the fiction, Matt, they're writing fiction for free. Therefore, my demographic would not want to pay to write.

Matt 39:10

Couldn't add ads?

Nina 39:12

I was thinking about ads, but it would ruin it. The beautiful ui it will it would i have like an animated ui such an artist different theme well okay okay, i'm gonna there are there will be plenty of apps to make money with plenty like the forge forge video forge audio those can work, like i know this is an app where i actually have users who use it and they they literally keep spamming hey can you add this hey can you add that? And yes, I'm doing it for free.

I'll figure it out. Give me some time.

Ben 39:50

What about a donation button?

Nina 39:52

There is a donation button. There's a Buy Me a Coffee button and that's good enough for me.

Ben 39:57

Maybe you can have a pop a little bit more.

Nina 39:59

I'm going to make it bigger. I'm going to make it fly across the screen and right in their face. It actually is pretty inconspicuous right now. I should probably make it bigger. It's kind of tiny.

Matt 40:14

You probably have AI figured out. Tell it to be a marketer and figure out how to get people to click on this donate button.

Nina 40:19

How do we psychologically convince the user to...

Matt 40:22

I mean, it's what everyone else is doing.

Nina 40:24

Yeah, I didn't even think to ask a lot of questions like that. Pretend you are a person who is skilled in marketing and branding and want to trick the user into clicking this. Yeah. I know you think it's foolish, but they won't use it if I start charging. I mean... Have you guys added, when you guys add AdSense to a website, is there a way to add AdSense in a way that you could make it like a little transparent or were really inconspicuous? Or are there rules and regulations against doing that? Like, can I have it behind a div that, and so they only see it like sometimes.

Matt 41:07

I mean. Yeah, Google will know that it's like Google's.

Nina 41:11

Okay, see, that's what I thought.

Matt 41:13

You have like a million people focused on that specific problem. But like you could just have it on the side or something or you're just like, well, it's a discord, right? So like it could just be like a message that's no, I don't know. I don't know how you would do it in your discord app.

Nina 41:28

I'll think about it. I'll think about it, do it.

Ben 41:32

Doesn't that defeat the purpose? Aren't you being paid by impression and by click-through? So if you hide it well enough so nobody clicks through it, it

Nina 41:44

Just going to stick with the donation button. I want to stick with that. It's, yeah.

Ben 41:50

I mean, ramp it up a little bit. Don't go full on Wikipedia. Yeah. But if Wikipedia is a 10, maybe try a 4 or a 3. Maybe a B tested.

Nina 42:07

Maybe what tested?

Ben 42:09

Do some A B testing.

Nina 42:10

Okay. Let me make a note. I'm just going to make a note to remember that.

Ben 42:16

The only problem is you need to come up you need to have enough traffic to have a key value rate but with the donation button how many people click through how many people how many people stay on the site and just click it away there's a bunch of metrics that you can pull

Nina 42:35

There's an active now so all the users on the website they have like an old like you know the old myspace where it would say like online now they have that on there so i can see there's maybe like four or five people on at once it's it's a baby app it's in its infancy.

Matt 42:51

But users so your other ones don't have users

Nina 42:56

That's because i haven't shared them well i could yeah right right i'll start sharing i'll start sharing apps but i need to redo them in rust all of them.

Matt 43:08

Yeah and then redo them again and again and again

Nina 43:12

Just rust I'm pretty sure rust's the end all for desktop applications yeah.

Ben 43:22

I think in the first episode you were talking about electron is that what it is

Nina 43:27

Yeah and now i've realized that electron is way too slow it for because it was for a media app it was for an app in which you're you're editing audio files it was like a daw so, that was horrible on the cpu because everything with electron is through chrome, it's it wasn't the best choice electron on it's very very slow in comparison they're like there's so much lag in it so i need to redo it, it was called forge audio and now i made something called forge video to help matt and i out with our podcast editing, and i did it in rust with Tauri and metal and it's so much faster there's there's no hammering up the cpu at all so, yeah it's yeah it was it's better for like real-time performance so i'm probably going to redo all my applications that's all my applications that are media.

Media editing apps like at the time i was just very stubborn about using electron because it was what i knew and it being all javascript and that being what i was most comfortable coding in like maybe five years ago, yeah ai has like in terms of like helping me with research and figuring out like what's the best best way to build something, yeah i would not have had the ability to get out of my comfort zone if i if i didn't have ai to help like you know Claude to help research and code it for me, because the the learning what is it called the uh.

Rust is very difficult to pick up quickly and I would never have experimented with it if I didn't have an agent coding in it for me. Especially coming from a background of preferring dynamic languages. So, yeah.

Ben 45:29

I have a question. Both you guys brought up a specific technology and this is for my own selfish reasons. I recently, for work, we stood up at we set up OpenSearch, which is AWS's version of Elastic. Are you guys familiar with the offering? And how does it assist? Are the performance implications noticeable? And if you guys are familiar with how they fort, do you guys foresee OpenSearch deviating much further from Elastic?

Matt 46:06

Or if it follows DocumentDB compared to MongoDB yeah it'll immediately slow down and not keep up performance wise unless you have super fine-tuned your mongos or your Elasticsearch i'm just guessing i don't know for real but unless you have something special in Elasticsearch you're not going to notice that much of a difference it'll just going to be easier to maintain, if that's become a problem, that infrastructure, to move it off to something that Amazon worries about is worth it. We did, I've done the OpenSearch migration, as well as a DocDB migration.

And both the reasons is specifically to not have an IT person worry about the box.

Ben 46:55

Hosted in AWS, so that was essentially the winning argument.

Matt 47:01

I mean, it's a real argument. You save 150k on a person, at the very least. There's no way OpenSearch can cost that much more over Elasticsearch.

Ben 47:12

It is expensive. It is surprisingly expensive.

Matt 47:17

I've also been playing with Instinct, completely unrelated. Have you guys heard of Instinct?

Nina 47:22

No, it's not.

Matt 47:23

It's basically what OpenClaw wanted to be and so it's also the same as facebook's muse that just came out

Nina 47:31

Oh it's uh it's oh it's it's like a it's like dot dots.

Matt 47:36

I don't know i don't know if it is like dots or not it's basically the interface is whatsapp at least for me you can do it so that's like your messenger app and it's really easy to connect it to things unlike OpenClaw and you don't have to host it yourself so i'm giving it way too much data but i've been enjoying what i've been doing with it been doing it to like generate leads and kind of go out and research like other companies to like offer services to as well as find conferences that have cfps open and then write up a cfp that fits the conference and drafts and things like that it's been good i like it

Nina 48:15

It's it's not like dots where it works like a wrong around the clock right like.

Matt 48:19

You can make it work around the clock it is very opaque what is happening behind the app where i feel if dots is anything like amp you're kind of like seeing the sausage get made and for some reason like watching a thing in the terminal blink i have a lot less patience for than watching my whatsapp someone's typing go forward

Nina 48:43

Many different so i know we're towards the end of time but how many different ai tools do each of you use because i'm curious because you have instinct you have grok ai.

Matt 48:54

I don't have great i have am

Nina 48:56

Uh okay and.

Matt 48:57

Then i have uh i do codex and Claude such between the two codex

Nina 49:03

Okay yeah i Just have Claude and ChatGPT.

Ben 49:09

Predominantly used person, although I've recently been using Claude Code exclusively. And then for whatever reason, I have a subscription to all three of the biggest ones. I have a ChatGPT personal one, Gemini and Claude Code just for ChatGPT, mostly for writing, Gemini for general research, and Claude Code for random site projects that I want to do.

Nina 49:35

Yeah. Yeah, same. I have, I recently canceled Gemini, but yeah, I just have ChatGPT for, for like writing or sometimes research outside of coding. And then yeah, Claude, Claude code for all of my side projects. But I am very curious about Dots that just came out.

Ben 49:57

What's that supposed to do?

Nina 49:59

It's an assistant that kind of works like around the clock.

Matt 50:02

Okay. So it is the same thing. It's the same thing.

Nina 50:04

Yeah, it's a 24-7 AI agent, and it uses the latest ChatGPT model, and supposedly is extremely fast, and the highest subscription tier is $500. So I don't know what the lowest subscription tier is, but I'm curious to try it, at least. Not the $500 one, though.

Matt 50:30

It should be good.

Nina 50:30

That's, yeah.

Matt 50:32

Right. I think we got to end it there. So that's that sweet.

Ben 50:38

Thank you so much for having.