I Do Miss the Code Editor
Matt and Nina sort the week's tool hype from the useful parts: multi-agent coding in Amp and Codex, model routers, and classifiers that don't need an LLM. Plus live-coding interviews in the AI era, big tech layoffs and golden handcuffs, Nina's collaborative writing app, and the case for finally charging for it.
Episode four. Tool hype, interview war stories, and homework Nina didn’t ask for.
What we talked about:
- 0:17 Sorting the week’s hype from the practical utility.
- 4:46 Where fast classification pipelines actually earn their keep.
- 10:35 Model routers and when micro-optimization stops being worth it.
- 17:50 Nina’s interview rounds and live pair-programming tests. “If someone just shot from the hip a correct regex, I would hire them.”
- 21:41 Algorithmic coding tests and hiring in the AI era.
- 25:34 Big tech layoffs, overvaluation, and golden handcuffs.
- 28:25 Hands-on with Amp and Codex for multi-agent coding, and where “let it take the wheel” ends. The title line lives here.
- 32:04 Combining models for writing and UI design workflows.
- 34:26 Nina’s collaborative writing app for a Discord community, and its concurrency bugs.
- 37:34 Monetizing side projects. Nina: “Because I hate making money.”
- 46:46 Remote vibe-coding tools, and the worst agent Nina’s ever used.
Transcript
Automatic transcript, lightly corrected. Also available as WebVTT.
Matt 0:17
Hello, Nina.
Nina 0:19
Hello, Matt.
Matt 0:20
Well, it's been going good. I got a whole lot of stuff to talk about this time around. A little more prepared.
Nina 0:27
Same. Same. It was a very, very busy week in the tech world, it seems. I could not be on YouTube without seeing something about Jev come up.
Matt 0:38
Yeah.
Nina 0:39
Yeah.
Matt 0:40
Exactly. So I jumped on the hype train for Jev. I jumped on the hype train for Amp. I jumped on the hype train for Codex. and basically two of them seemed very overhyped for me, and one was real. So I guess two truths and a truth, we could play that game.
Nina 0:59
Two hypes and a truth.
Matt 1:00
You could play two truths and a lie?
Nina 1:02
One truth and two lies.
Matt 1:04
Yeah, something like that. Whatever.
Nina 1:06
Yeah.
Matt 1:07
Yeah, so I started off with Jev, because everyone was talking about it, and it's just a classifier. that does some, some JSON outputs, which we kind of already do in our own cloud development.
Nina 1:26
Right. You could technically do what Jev is doing with... Well, no. Well, okay. I'll let you make your case for it first. Why you think that?
Matt 1:38
In theory, it can't. But if you put code, like we do it in ours, we're doing like a data pipeline, you're going to get you can get json because you're going to format it through or you can at least be passing it through code that needs it formatted.
Nina 1:53
Right right but like in for so for okay for eta brain like we have we have those four scrubs right i can see i can see why it might not be necessary for the eta brain which is for viewers who don't know that's the podcast app that we have where there's like four to six Haiku calls, you know, there's a scrub, does the query need rewriting, there's the router, selects one of three of your question types. It's, there's certain things that you can just manually program in there, but Jev could optimize whether or not some of those calls are useful.
Like, especially for the first one, does this query need rewriting? I mean, you're right you're right it's easy to program a yes or no for that but
Matt 2:43
Yeah like why would we introduce a whole nother system to do what save us instead of like a tenth of a penny a hundredth of a penny, would be the cost of the call because we were using Haiku for it like you just like load tested it yesterday and spent like a dollar.
Nina 3:00
Yeah that was after optimizing the ask pipeline and a bunch of other things but yeah
Matt 3:07
Like we would have to be in just area hitting so many queries for that to be a noticeable cost.
Nina 3:16
Right but if there's four to six in and we're both hitting we're hitting a model that spoo's tokens and techs jev can make make those decisions early and decide if like it's even worth making those calls with like less code i don't know
Matt 3:34
There's so how much time would you save like and the whole thing takes like at worst a minute.
Nina 3:41
Right 1.2 seconds per call i mean i'm gonna have to go through the ask pipeline and just use jev because i haven't had a chance to use jev i've just been seeing the hype yeah i
Matt 3:54
Tried i tried to use it with the chief of staff stuff and it just was it's not useful for me i mean i'm not classifying anything other than like prs which you can't really do well it's too much context.
Nina 4:09
You're right okay so that that's a limitation of it because there's yes or no but there's there's also isn't there two other types there's there's the score where i can rank something on a metric like see okay
Matt 4:23
All of it is basically a score with categories and confidence levels. The binary one is just a category with two. They're just over-hyping that for no reason. I question how much is their own fault and how much is just the internet's fault. I think a lot more of it's just the internet's fault. This play makes more sense for like, you have a set pipeline already and you're just jamming through terabytes of data, And like, you need a category on it. Like I have a past example that I could see Jev being useful for.
Like I talked about where we were generating pictures for like Amazon catalog items to like show in your hand, like how big it looked in your hand. And every once in a while we'd have an outlier that was like a tuna can the size of your head, even though it wasn't, it was just a normal one. Someone put in centimeters instead of inches or something or meters instead of inches. and so that i could see that because what we had to do is we had to give it to mechanical turk and have people go and manually click it, like if we could just go through all of them super fast and then like a person at the company could spend an hour and through whatever's left or double check the false for the positives that came out, like that makes sense and you would be doing enough with because those are images so the tokens are huge and you you don't want the like you want the speed because you just want yes no on that data.
Or on my unlike pr reviews like you're like oh i see this i want to dive more into this oh i need to actually look at something that's not in the pr like that you want a full model for that.
Nina 6:13
Right right so okay since you worked with it more i have i have an app that i made over the course of like i think since we started this podcast and what it is is it's similar to your chief of staff in which it's evaluating a huge amount of context but it's not a pr it's an answer to a coding exercise question a very complicated like like this isn't like a like evaluating like a like a finite output it's more like abstract like text answers like asking how would you build this if you had these constraints and like just very complicated i use claude api to evaluate the answer like zero out of a hundred now if i were to integrate jev into that based off your experience with chief of staff like what would that what would that look like and would it be no because
Matt 7:09
You're scoring that's not qualification of yes or no you want context a normal model the more general yeah because what you could do is like yes or no like correct or not correct.
Nina 7:24
Yeah but it's more it's more on a percentage you want a gradient sir but but here's the thing like you could you could have You could be looking for specific fragments of words, like fragments of sentences in the answer, and that kind of be your checkpoint of, okay, they mentioned A, B, C, D, E, right? And that could be a part of the evaluation. By just parsing that text and seeing that it has a specific amount of mentions of certain topics, and that could be a part of the grading, and maybe Jev could come in there.
Matt 8:06
You could give it, the critiques a little bit, outline, and be like, did it do this? Did it do B? Did it do C? And then it would give you a yes or no on each of them.
Nina 8:20
Right, and that could be where Jev could be integrated. But is it worth it?
Matt 8:26
No, I don't see how you make it worth it unless you are using gigantic data models.
Nina 8:34
Yeah.
Matt 8:35
To a point where you should start to think about building your own model.
Nina 8:40
Jev being used to determine data that could be relevant.
Matt 8:46
No, at the point you would start using Jev, I feel. You would already have enough data to want to do your own model.
Nina 8:56
Right, that's what I mean. Let's say hypothetically, on a different topic, now I'm jumping to that since you mentioned that. Let's say you want to train a model and you want to have Jev like sifle through like a million entries of data and determine based off a metric, does this qualify as useful to train this model for this specific purpose? And you kind of have that iterative approach or not iterative, but like recursive approach to... Yeah, going through millions of entries to make a more efficient model.
That would be a good use case.
Matt 9:35
Yeah, so you're doing it as making the seed data for the model, right?
Nina 9:40
Right, right.
Matt 9:42
And yeah, that makes some sense. But that's like a one-and-done thing.
Nina 9:47
It is. It is, but...
Matt 9:50
Like, I have so much trouble finding, like, this is a good company idea. It's just... It's so niche in what it does to a point where someone that needs it is good enough to build it themselves.
Nina 10:08
Yeah, I think you're right. But I'm still going to use that API token and I'm going to see what I can do with it. I'm still going to mess around and see.
Matt 10:17
Yeah, give it a shot. I don't know. What would you grade it on, right? It's got a RAG for the ETA brain. It's got a RAG that looks... for like some topics so you can get a confidence.
Nina 10:32
Yeah you could load up
Matt 10:34
Oh go ahead.
Nina 10:35
So here here's where the it could be useful in eta brain so there's like there could be a simple retrieval there could be a deep retrieval i mean that's that's where it could it could make the decision of like how many how many calls to make like determining like i mean outside of like does the query need rewording or something? That's silly. But in the router, there is a which of, I believe the Ask pipeline has like, which of these three types of questions is this?
So maybe a little bit of Jev right there, right? Without making a Claude call. So that would be at the router level. So between scrub, router, re-rank, and then you re-rank per hop, I think. I was researching last night and asking Claude the same thing again. No, I want to specifically know about this. Stop generating documentation. I just want to have a little conversation here with your context. which by the way like outside of plan mode how do you actually like like have like a built how do you how do you how do you actually like plan with Claude in a way where it just doesn't jump the gun and just keep generating documentation upon documentation i'm changing subjects too much but definitely the router level i wanted to say
Matt 12:01
Yeah so i've also seen videos of the router, the model router is what I've heard people call it to decide which level to use and that's okay, because you give it the models you can give and then it'll come out with confidence level and then in the code you can route the data to the correct one, But that's like the whole additional system you're adding to your pipeline just to do that routing where you just have Haiku do it. And it is more expensive, like noticeably more expensive.
But again, we're talking like the difference between a penny and like a hundredth of a penny. So unless you have the huge throughput, which if you're doing human interactions is not likely. like if you're doing like fraud detection through logs of financial data yeah okay but they probably already have their own models that do that.
Nina 13:04
True but maybe there's somebody who's building an app that doesn't have like are you talking about like a proprietary model specifically for
Matt 13:14
Yeah again to the point I just don't know who the hell this thing's for.
Nina 13:18
Well so customer support that could be one does this need a human it could be the choice type should we send this question to like billing technical cancellation or fraud and then it could have a score for like urgency I mean that's one thing
Matt 13:39
I guess I mean you'd have to get such a speed boost to make this worth it.
Nina 13:44
Yeah where is it worth it is the question I see the uses but
Matt 13:49
Like into like, If you're going to do like an agent routing, right?
Nina 13:54
Mm-hmm.
Matt 13:55
Is adding like a high, a kaku-kaku fast like going to add that much when it's going to route to Fable or Astra? Like those two are going to eat so much time that you're saving on the wrong part of the model timeline.
Nina 14:13
Right, but each call does take like about 1.2 seconds and I did read that Jev as like milliseconds. So you are saving a little bit of time, too, on the user side.
Matt 14:26
Seconds, which, I mean, and sometimes, yes, that's super important, like, for, like, media-type stuff. Mm-hmm. But when, like, you're a user in an app, like, you can, a little spinny thing for an extra second's not going to kill you. When you, like, I mean, we talk to AI models, and they spin for, like, two to three minutes.
Nina 14:50
Right, right. But FYI, also, I added a spinner to ETA brain. That's one little thing I did aside from all this, like, the one UI UX thing.
Matt 15:02
Yeah, just add a spinner.
Nina 15:07
I mean, it made me happy to just see a spinner. So maybe you're right. But, like, what about, I don't know, so many times, like, you're working with, like, a chatbot. and you're trying to get like a refund for like an order or something like Jev and and most people like get really impatient and they just like i'd like to talk to somebody i'd like to talk to somebody i don't know if Jev's gonna really change that like i don't i don't know I don't know what I'm going to build this week, but I'm still going to see the improvements.
Because every time I keep messing with ETABrain, I do a probe evaluation where I compare where Master was at before based off how many calls made. The Ask pipeline hasn't changed, but how many calls have been made? How long does it take to go through each step of the calls and go through the pipeline? And I got it from, on the old jar, it was 17 to 19 seconds for that first call to 0.8 seconds, it said, 0.8. And then when listing relevant moments to part of the app where you click relevant searches, and it got from 17 to 19 seconds to 8.6 seconds.
Matt 16:36
Yeah, I mean, you're going from serverless to dedicated server, so that makes sense.
Nina 16:40
Yeah. Yeah, but it was still nice to see. But maybe with Jev, I don't know. Maybe I could try to see if, like, I'll see.
Matt 16:51
You can put me wrong. I will accept it if I see it.
Nina 16:54
I don't know until I actually go in and try, but I need to also go in without Claude in the equation. Like, this is a moment where I have to... What?
Matt 17:04
Why?
Nina 17:07
Because sometimes I don't want Claude to come up with the creativity side of looking at a pipeline and seeing how I can use like a third like another third party tool I want to go through and see what I could think of just looking at the code myself I mean do you ever do that
Matt 17:27
Just no no more why wouldn't you have it just this why would you have it just tell you what's in there diagram it and then you tell it what to do.
Nina 17:34
I did I did diagram it I'm gonna diagram that's that's a much better idea but I'm sorry I've been doing so many interviews this week and code exercises and i'm telling you i do miss i do miss the code editor i do it's yeah no yeah no yeah i i have i in other news i have a lot of interviews this week and an offer so that's good congrats thank you i just got two more to finish up and i really want one of them I really want one of them yeah but yeah in this process I have to go back to coding a little bit because
Matt 18:17
So what's the process they giving you?
Nina 18:21
Well I guess this is interesting for the podcast I had to write unit tests in Java it was like a pair programming thing So it was, they drive and I tell them what to write for certain cases. I was expecting algorithms and, I don't know, building a little chess game. Remember what you made me go through? Not made me go through, that was fun. But no, this was very straightforward. And then just a lot of technical verbal questions about systems engineering, more so than the coding.
they want to see how you would think about if you have an app at a small scale, a medium scale, a large scale how would you build it differently?
Matt 19:12
How big is the company?
Nina 19:14
There's four companies one's very small the react i'm just going to call them like react shop small broadband cable company like i think like broadband cable software company they make software for broadband cable companies maybe like 50 employees the one i really want they have like 10 000 employees big big corporate and then the fourth one small like maybe 100 so most they're mostly small but i would get to work in rust in one of them so i'm cool that's pretty cool and it's rust typescript Vue 3 aws like that's that's what i worked in for six years plus minus Java Spring and just swap that with rust but yeah today it's probably going to be algorithmic exercises for for one of them they're going to make me do front end and then i also believe back end which i did not prep for
Matt 20:16
Yeah we'll see.
Nina 20:18
We'll see
Matt 20:19
Yeah i'd be surprised if your large one doesn't include many more steps.
Nina 20:25
We have to build a pokemon game in java which is going to be pretty fun that's the next one like object-oriented programming
Matt 20:33
Tweaked this is why i'm freelance Yeah.
Nina 20:38
I don't know how to go about telling one company, can you wait a week? You know, like, I don't know how to do that,
Matt 20:47
But just tell them. tell me you got other things.
Nina 20:51
I i did i said hey i'll let you know friday and then i got another interview for the one i really want next week oh
Matt 20:58
Yeah well you know okay so you already told them one thing and i can backtrack okay.
Nina 21:04
I could verbally say yes and then well whatever i'll figure it out i want to be respectful and honor everybody's time but
Matt 21:11
I mean the fire at will goes both ways.
Nina 21:14
Yeah. They said they put aside money to bring me on. I don't want to talk about the specifics of it though, because this is a podcast where it goes out. Why are people doing the algorithmic exercises when we have these AI tools anymore?
Matt 21:41
Java just has dot sort. You just let it do what it does. I never understood. Ten years ago, I made people do it. No one's writing sorting algorithms.
Nina 21:51
Right. I learned how to do... I forgot how to... Well, not learn. I forgot how to do a quick sort. And I had it open, and I was like, oh my god, I hate this.
Matt 22:04
And it turns out you never do, a quick search. So it's just like, why?
Nina 22:12
They should let people use AI tools outside of Claude for coding exercises, because that's what the day-to-day is going to be anyway. And they should judge and evaluate you by what you search. i think that would be that would be to interview people because if you're if you're searching like a regex like cool if you're if you're searching like a like like a higher higher level decision or something that you should know then red flag
Matt 22:44
I mean if someone just shot from the hip a correct regex i would hire them no matter what it was.
Nina 22:50
If they shot oh
Matt 22:51
Oh if they were just like straight off the dome like this is the regex you need to do the correct search and it worked i'd be my mind would be blown.
Nina 23:01
I know like two or three i i yeah
Matt 23:04
Exactly no one knows that off the top of their head other than complete insane people.
Nina 23:09
I always i always mess up like just starting a regex like i can remember i can remember passing the regex but i can never remember which way it starts, which way you escape, like actually, oh,
Matt 23:23
If they took the time to learn that, they know everything else. So you just hire them.
Nina 23:29
I'm just going to make flashcards of every regex I can think of and just... Yeah, it seems that companies are having a really hard time coming up with creative ways to interview people because of AI.
Matt 23:41
Yeah. There's a lot of people on the other side of the equation, too. It was not as easy to find people, but now you're swamped with people. Everyone's resume is just AI-generated to be perfect. It wasn't useful before, but it's even now it's a false positive.
Nina 24:05
Yeah i mean i i i used Claude design for mine but i embellished it i embellished it as much as i could to not flag ats readers you know like just so just so it stands out but did have you used Claude for like resumes and invoices and stuff I mean,
Matt 24:27
I do my own invoices, yeah. I've helped other people's resumes. So, done that with Claude.
Nina 24:34
And they're like, wow, this looks great. And do you tell them you used Claude?
Matt 24:37
Yeah, I mean.
Nina 24:38
Okay. I've seen friends that recently had a few friends in IT that were all laid off at the same time. And they were asking me to help proofread their resume and everything. and they there are people who are extremely anti-ai but looking at like their resumes it was very i think it was a case where i think they needed to use ai for like layout stuff i don't know
Matt 25:04
Yeah that's where your resumes used to be useful in helping see how good the person is at human communication, they would like basically they have to describe their life in one sheet of paper like what do they pick that's important what do they leave out how do they present it what's the layout yeah, um but like i usually did that more than like did they work at google or they work at some small mom and pop shop like that was less important to me yeah yeah.
Nina 25:34
I want to move on to the orbing with Amp, but I also want to talk about, there was a video about the tech industry where I think it was like, it was like one of those like kind of vice format, like documentary style, like short, short things where they were interviewing three people who had been laid off from meta, Intel, and then the other one, Microsoft, I think. these were all people with like anywhere from eight to 15 years of experience and they're all complaining and how they couldn't find a job like one was a product product engineer one was a software engineer at meta who claims that she submitted 8 000 applications and she can't find a job and it's been 15 months and then another one was i think like a cloud engineer of some kind But yeah, they were all complaining that they couldn't find jobs and they all came from these big companies.
Matt 26:35
Yeah. I mean, they're overpriced and they're underskilled.
Nina 26:38
Okay. You know what? That answer is... It was bugging me, but if you do stay at a company for many, many years and you just work in maybe a department or a team that still uses archaic things and they haven't migrated and they earn like $180,000, you're right.
Matt 27:00
Yeah, I mean, there's a sweet spot between the evergreen beginner, someone who switches jobs every year, and the giant corporation 15 years. They've gotten very good at handling the beasts of the Fortune 500 company they're in, but they can't learn a new system. They spent 15 years learning one system.
Nina 27:22
Well, Matt, you just said that.
Matt 27:24
And now they want half a mil to join a new job where they're not going to do well because they're used to getting their stock options that were like $400K because they've been there for 10 years.
Nina 27:34
So wait, what's that term? what's the term with like a silver spoon what's the term golden handcuff golden golden handcuffs silver spoon okay so they're spoiled yeah can't find a job if they've
Matt 27:46
Been at facebook for 15 years even eight years i'm pretty sure they got a lot of money.
Nina 27:51
Yeah i kind of like that they're going through like you know there's like these sad montages of like them looking out a window and and meanwhile they're in like these huge houses and walking like there's three dogs and it's like okay anyway you answered my question pretty quickly i was complaining about this particular this particular doc to like two or three friends and they were like they didn't have anything like finite or any opinion to say on it but you just nailed it
Matt 28:20
Yeah probably don't want to be mean okay, Anyways, back to Orbin. Okay, so I tried Orbs, and I signed up for Codex to do it this week. So basically what it lets you do is it's got a Visual Studio interface, basically. And then you got your projects on the side, and then you'd be like, okay, this project is about this repo, so you connect the repos. And then it's like, okay, I want to start a tab. So you start a tab, and that's like now your chat interface with the LLM.
And then you go, okay, I want to do a facelift on the UI. And it's like, okay, I'll go and spin up an Oracle, and then I'll spin up to learn about it, and then I'll spin up some workers to do it all on a different machine. And it goes and does that. And then you can spin up another tab. It's like, okay, make the backend more efficient. And then it spins up another one. So basically, you're spinning up VMs for each of these tests that then have...
code that have ai on it and can run and code and then i'll bring comes back to you with prs and things like that it kind of actually doesn't want you to do prs it wants you to be a vibe coder and not do that stuff but you can do prs if you really make it.
Nina 29:43
Okay so it's it's not built off of just one single ai model right it's it's um it switches between various models for like specific tasks like intelligently
Matt 29:58
It has its own models which i'm pretty sure are Anthropic but they say they didn't i don't know if they said 100 it is and then you have your own ChatGPT, account and then basically it has like a knob which for some reason i don't know why to like turn up the intensity with different models and stuff and it's a lot of configuration there that i did not find enjoyable, and but then like it takes forever just for it to write some code i had to go and like update a blog post and that took like 10 minutes, and like before i even started writing anything just to like spin up the box and load up the model.
Nina 30:40
Was it just the first call because
Matt 30:42
No i did i did about five before i gave up so i did like five within one because i thought that was slow and then i was I was like, oh, maybe it's just this box is iffy. And then I started another model and tried a few times, and then another, and another, and another. And then I was like, yes. I was like, is this Codex or is this Amp? And so I downloaded Codex. And it rewrote the blog post awesomely in like two minutes. And that was the first go. And I think most of that time was me logging in.
And so I like Codex. Amazing.
Nina 31:19
I've never programmed with Codex.
Matt 31:21
Yeah, I finally bit the bullet, tried it out. Well, it wasn't coding. The coding, I find, is okay. It's the same as Claude to me. But the human language writing is like head and shoulders above Fable. And I didn't even try Astra. I was on a lower model.
Nina 31:40
That's the only reason I would like to try Codex. And it doesn't surprise me that the human language coming from Codex is much better, because ChatGPT in general is much better than... human language and just writing natural flowing sentences claude is horrible at it yeah awful at it what's the pricing point on that out of curiosity
Matt 32:03
Because so i used the 200 bucks a month one because i wanted to mess around with Astra later um i think you could do the 20 bucks one and probably be fine.
Nina 32:17
I'm on a 200 Claude subscription
Matt 32:20
Yeah i am too and i'm gonna keep that or i think I'll do coding in Claude and writing in Codex. That's kind of where I'm going to go.
Nina 32:30
When you try Astra, what are you going to use it for in particular?
Matt 32:34
I might redo some UIs. Let's see, it's better at the UI design than Claude.
Nina 32:43
I do use ChatGPT for generating UI mocks over... I mean, I'll use both. Okay, so I've used Claude design before, but typically it works better for UI mocks if you use ChatGPT, generate the mocks, and then feed that to Claude, develop a design system with some inputs. Just bringing them back and forth, typically results get some good results. Like, one of these apps that I maintained that I ran into this week, my first production bug, because I actually have users on this app, Thank you. I had to like come up with threads on comments and I was like, man, this UI is not working for anything.
And I had to, you know, go, I went back and forth developing like what it would look like and coming up with a design system. But yeah, starting in ChatGPT and then switching to Claude and back and forth. And sometimes even like just feeding the HTML and CSS that Claude generates to ChatGPT with an image with some input is really interesting. The latest model on Codex called, is it, I'm sorry, is it Astra or is it, it's Astra, right?
Matt 34:03
Yeah, Astra's the new one.
Nina 34:05
That's not available in like browser ChatGPT. no
Matt 34:10
You gotta like yeah pay the hundred or more for it.
Nina 34:14
Yeah I've seen it generate 3D models and like full games and I'd be really interested in seeing how it does there
Matt 34:24
Because I I've.
Nina 34:26
Used I've used a lot of AI for like for game engine stuff but I'm finally like I'm finally like actually making like I'm making like a like a socket a socket based game with like multiple players for this this website that i have with multiple users we i have like 15 to 20 but like they're on it like every day and yeah i ran into i ran into a bug where i accidentally coupled comments on documents because i was just not paying attention like it should have been a separate table and so we had like one one person who had a document open and then another person who was also in the document commenting on the document and the reads and writes were concurrent and it was duh so I had to decouple decouple them And yeah, it was easy with Claude and doing some research, but I should have never let Claude make the decision to couple them in the first place.
I was like, oh yeah, that'll be fine. And then not thinking that documents could have multiple people viewing them and writing to them.
Matt 35:36
What the hell is this game that has 20 people constantly using it?
Nina 35:39
So what it is is it's for a discord community and it's a round robin writing game where people can write fiction together online so it's for a specific fandom i'm not going to mention the fandom but yeah so how it works is you sign up you get it you you make an account you mark whether you're a part of a certain discord community and then you get certain access to certain features like sharing round robin writing games to that discord and to specific channels And you have a profile and you have you have like an about section where people can link like their their websites and it's also a solo document editor for for fiction too so people can share their stories with with others in this fandom add beta readers to it mark it done have line by line annotations but the round robin writing part is the fun part so every week people get on it you have a host of a game and that host you can have up to like i think it's five or six writers in one one session and a prompt generator and so the prompt generation generates writing prompts based off like characters, location, tags, call to action.
Like it's, it's a pretty, it's a pretty rope, like big, big prompt generator. And so then everybody then gets like four prompts. And in the session, they can all vote, vote on the prompts. So the one that wins is the one people write and you can, you have timed rounds. So every person gets like 60 seconds to write a line or two minutes to write a line. And when the timer starts counting down, there's like music that stresses them out. There's also gimmicks to mess up their writing.
Matt 37:34
And this is free?
Nina 37:35
Yeah, this is something I made.
Matt 37:37
Why are you charging people for this?
Nina 37:39
Because I hate making money. Because it was just for a Discord community, for people who wanted to write essentially fanfiction. and I just wanted to like give some friends like a nice little space for people to write because everybody was paying for like ellipses which is like a writing platform that you have to pay.
Matt 38:02
How much is that?
Nina 38:04
I'm going to open it up for multiple fandoms. So this would be for like, you know, the Tumblr cretins. It's $6 per month but an annual plan of $60 per year.
Matt 38:18
Yeah. And how many uses do you have?
Nina 38:22
I have... It's growing. Because people outside of the Discord are like, Hey, my friend wants to join.
Matt 38:28
God, Nina! You're just sitting on this thing? Yeah, it's okay. So how many uses do you have? A hundred?
Nina 38:35
I can send it to you and you tell me if I should start charging. So I'm a little embarrassed about it because it's for fanfiction. It's for weirdos who like writing fanfiction,
Matt 38:45
You know? Yeah, but you could just make it about marketing and then you could do B2B. and have like b2b ad making scram sessions.
Nina 38:54
Oh i could have ad revenue on it can i
Matt 38:57
Well you can add ad revenue as well like you do the free version for the fandoms and they could be like getting ads and then you can do a paid version for businesses who want to like have writing bashes and their marketing team, and they could pay like 100 bucks a month.
Nina 39:14
Writing bashes for marketing Thank you. I guess it's it's style. There's 20 different themes for the specific fandom. And like, if I show it to you, you'll probably be like, oh, my God, how much time did you spend on this?
Matt 39:29
Yeah, I'm already thinking that. So it can't get worse.
Nina 39:33
Yeah, that's a good idea. So this.
Matt 39:36
So what do you what do you have? You had 50. You had 100. You had 10.
Nina 39:41
50, 100, 10.
Matt 39:42
What do you mean?
Nina 39:43
No i've kept it just to a single discord because every time i tell somebody they want to jump on it and yeah i capped it at a hundred just in case so i can handle traffic so if the hundredth user joins they will or tries to sign up or the hundred and first they get a nice little animated banner that says hello i'm one person join the waiting list
Matt 40:06
Oh my god Nina so you have 600 a month in revenue that you're not even trying?
Nina 40:12
Show me how to do, I don't know how to market. I don't know. This is why, this is why I'm, I'm, I'm a contractor with your consulting. Maybe you could show me how to market myself here because I, I'm very good at marking, marketing, creative, like things that I just make, but I'm not good at like coming up with a plan to make money at them. So yeah, I know. What would you do if I, because I'm starting to open it up and I'm starting to like post it on, like, I'm going to start posting.
Matt 40:46
You got to start charging. You got to start charging like now. At least throw ads on there and then have a free no ads. Pay me six bucks a month.
Nina 40:57
Okay. I put a lot of time into caching and snapshots and seeding like content data for prop prompts. If something goes wrong, I have version history and rollbacks now.
Matt 41:14
Yeah. So there's this thing called Stripe. You should spend like a day on it and add Stripe.
Nina 41:20
Stripe for payment?
Matt 41:22
Yeah. And then you get Google Ads, and then you add that into it as well.
Nina 41:27
But Google Ads are so ugly.
Matt 41:31
Facebook Ads? Which ads do you prefer? Maybe you can get some authors to sell their books or something. I don't know.
Nina 41:39
It could compete with AO3.
Matt 41:42
I don't know what that is. What's AO3?
Nina 41:43
Okay. It's a fan fiction writing website. But here, after this call, I'll send it to you. It's on my resume because it looks really nice.
Matt 41:53
Yeah, you have a working business. Of course that sounds nice.
Nina 41:57
I am answering to bugs in the Discord because somebody was like, the other day, this is hilarious to you, but no, someone the other day, and this is something I learned coding with AI, by the way. Somebody the other day was like, help, my story keeps having amp and amp ampersand, and it kept appending. And I was like, immediately, I was like, sanitizer. The sanitizer screwed up. Like when you switch from HTML to rich text. And it kept saving and doing that on their document. Which, it's corrupt data, but it kind of is like a memory leak.
And if it's spam saving because there's changes every 30 seconds on the page,
Matt 42:42
That could... And they can pay you $6 and you will fix it.
Nina 42:47
I fixed it. I did fix it. It was because I, instead of having a third party like dom sanitizer, I coated my own with Claude because I was like, why not? And there was one bug in it. So I'll charge. I'll charge. I will. I'll figure it out. That's the thing about, like, these fan spaces is that, like, they're typically, like, anywhere from, like, college kids to, like, a middle-aged mom who doesn't have a job, like, that are in, like, these groups, like, you know? So it's kind of, I do have a...
Matt 43:27
And they pay for, like, Wordle stuff, or they'll do ads, like, for those little games they play. Like, you're not breaking the mold, too.
Nina 43:35
I was going to add little games, too.
Matt 43:37
It already is a game. They're playing these things. You could have leaderboards. They're sharing data.
Nina 43:44
Leaderboards, but the only way that, because there's only 25, I host a round-robin writing once a week on a Saturday for two and a half hours. But it is going to have a calendar where people can actually see who's writing at which times. because there is an active now and you can see who's online.
Matt 44:09
But first, it will have ad revenue. Alright, I'll accept ad revenue.
Nina 44:16
Maybe I should just ask Claude, how do I make this make me money? Yeah. Just do that.
Matt 44:22
I mean, so many people are using it and there's a waitlist too? Yeah.
Nina 44:27
Only if we hit 100 users, it's pre-programmed.
Matt 44:30
Okay, so you haven't hit 100 yet. Okay.
Nina 44:31
No, it's pre-programmed because I'm trying not to share it too much. But I'll implement the ad revenue, and then I'll start sharing it to other discords and see if they want to use it. So I will. I will, I promise. Is there a way to have Google Ads where it's not, it's not like in your face every two seconds and you have to like play a little game where you're hitting X every, you know?
Matt 44:57
It's a discord app right so i don't know you know i'll figure that one out sorry.
Nina 45:02
Okay okay i do have somebody subscribing like a few like a couple people subscribing like ten dollars on like a personal like kind of patreon page That's something, but yeah. Yeah, help me with marketing, because I don't know how to market.
Matt 45:18
I don't know- Dude, what do you mean to help you market? You already have marketing. People are trying to share this thing already for you. You have viral marketing.
Nina 45:25
But I need to know how to make, outside of ad revenue, a subscription plan might turn people off. I don't know, it might. I have a support, support me at the bottom. I do have that.
Matt 45:39
You just have it do ads for free and pay for not ads. And then you can add more stuff later for a higher pay point.
Nina 45:48
And maybe if they watch ads, they could unlock cool things and that's more incentive to...
Matt 45:54
I want to get that. It's not illegal, but it's just a pain in the butt. Get some ads in there and then it just keeps popping up and you get a few cents for every hour or something they play. it's better than nothing and then see how many people move over to the subscription, or if you have new features if it'll make you feel better you can put them behind the paywall and then be like oh I got some cool new stuff, and so they're paying for the new stuff not the free stuff you've been giving to them for god knows how long.
Nina 46:28
Yeah there's even like a like a quick command for writer's reference too
Matt 46:32
You don't need to tell me about hotkey stuff you got in there just make them pay for this.
Nina 46:38
I'm sorry, it's painful. It's painful. Yeah. Okay. I didn't even get to talking about how orbing technically sucks with Replit, because...
Matt 46:52
Oh yeah, what are they doing?
Nina 46:54
The agent takes forever. Just like you said with Amp. There's no stream of what it's doing. Like, you know, you have with Claude or anything in your terminal.
Matt 47:06
Yeah, same with Amp. You can get a little bit of it, but it's definitely meant to be vibed coded. you're just supposed to let Amp take the wheel and hope for the best.
Nina 47:18
It's broken if I've pushed up from my local machine to Replit and I've tried to use the Replit agent it's broken my app before so I just stay away from it
Matt 47:30
Replit was the first one on the train, you would think it would be the best one.
Nina 47:35
It's not good I don't even let it deploy I deploy from my command line because I just don't I don't want to touch it. It's the worst agent I've ever used in my life.
Matt 47:47
I feel like if you're at the point, right? The same thing with Jev. Like, if you're at the point, you need this stuff. Like, I want to be vibe coding at, like, a multi-agent level on multiple boxes. You might be technical enough to spin up your own box and have your local agent SSH into these and kick off, like, what I do for the chief of staff stuff.
Nina 48:11
I did that with Hetzner. That was something that I never had, I mean, I knew I could do it, but I just, if I have to write all that out by hand, like, or not by hand, but myself, I'm just, I'm so happy I have Claude to SSH. Yeah. Yeah.
Matt 48:29
Makes sense.
Nina 48:30
Yeah. True. But do we have time for your other point about small tickets?
Matt 48:37
I haven't even gotten to that. I will do it next week. Okay. I'm spent. I've ranted too much today.
Nina 48:44
Yeah. Well, this was a good one. Lots to talk about.
Matt 48:48
Yeah. So, I want you to have ads or payments by next week.
Nina 48:53
Okay.
Matt 48:55
Make some money.
Nina 48:56
I'll write it down. I swear. I promise. it'll just be one one little little banner in the corner that's fine all right
Matt 49:08
All right that's the pod I guess.