Anybody Can Build Anything Now
Nina migrates Matt's podcast search app from Next.js and Python to her first-ever Java deploy, six years into a software career. Plus the economics of running agents on a subscription versus an API, multi-agent pipelines that pass JSON instead of prose, why AI job interviews feel terrible, and how episode 1 got its title.
Episode two. No longer a fluke, now a coincidence.
What we talked about:
- 0:19 Podcast hardware struggles: the hub that only carries power, and recording off a MacBook Air with two ports.
- 1:35 Nina migrates the podcast search app from Next.js + Python on Vercel to Java, her first JVM deploy after six years in software.
- 2:22 What the app does: turn a podcast into a knowledge base Claude can actually search, timestamps included.
- 10:47 Subscription vs API economics: a chief of staff bot that runs five minutes a day is fine; ten straight hours of transcript crunching is a different story.
- 21:08 Multi-agent patterns: agents handing each other JSON instead of prose, with three follow-ups max.
- 32:21 Nina tries an AI-run job interview and hates it: all you can do is say the right keyword and hope.
- 43:27 How episode 1 got its title: feed the transcript to Claude, keep the line you forgot you said.
- 44:14 “Anybody can build anything now”: JVM memory tuning by asking questions instead of knowing things.
- 52:38 Join or fall away.
Transcript
Automatic transcript, lightly corrected. Also available as WebVTT.
Nina 0:19
Hello, how are you doing, Matt?
Matt 0:22
Doing good. Number two. We've got ourselves no longer a fluke. It's a coincidence, no?
Nina 0:29
Yeah. Yeah, number two. How is your week going with building? What are you working on?
Matt 0:38
It's good. I've been working on my podcast setup and failing miserably. I have one of those hub things to connect everything into, and then that would connect into my MacBook Pro but it's not it's like works only for power it's like not sending any data now i have enough i only have two micro usb ports on my air so i can only do power a screen and the camera
Nina 1:06
Well no i just got a little hub too because i have a macbook air um and.
Matt 1:12
Uh yes i mean link tool on your butt, because mine's not working.
Nina 1:16
Yeah. Uh, it's an INTPW, that's the name of the brand. It works well for me. Like there, there are different separations between like giving powder or power and like transferring information. I don't know. I'll send you a link to a podcast, but, um. Yeah. So this week is programming a lot of side projects and migrated one of your projects this week. It was Next.js and Python. And yeah, it was the podcast search app that you have deployed and migrated it to a JVM. And figured out how to host and deploy my, first JVM or Java app. Had zero experience with deploying Java before, despite working in it for six years, which is very strange.
Matt 2:22
Yes. So the concept of this one was, it basically takes a podcast and it turns it into a knowledge base like Claude can search against. And then like spit out like basically have a nicer search across a podcast more than just like it looks for titles and tells you which one it is and at best looks in the descriptions like this one kind of pulls out timestamps and things and What was the architecture before?
Nina 2:54
It was Next and Python, and, it was hosted on Vercel.
Matt 3:03
Was there even a server, or was it just a database?
Nina 3:07
I think you just had just the... well, we had Vercel. Well, no, you had it set up in Next, I believe.
Matt 3:17
Okay, yeah, so it was the next server on Vercel. Mm-hmm, on Vercel.
Nina 3:22
It's been a bit. Yeah, so the app was really interesting to poke through just how it worked. So like 121 episodes of a specific podcast transcribed with Whisper, and then just the way that the search was set up in the app, Hosted on Vercel, everything would kind of load at once. It would have to boot the model every single time. So migrating things to JVM was probably one of the best architecture improvements for it.
Matt 4:05
Yeah, I remember now. It was like Vercel's Lambda functions, whatever.
Nina 4:08
Yes, it was. It was the internal Lambda functions. And it was totally inefficient for when loading models. Yeah, expensive. Um very expensive um yeah there uh, but you know so what moved mainly was the the serving api so just like the search uh asking questions caching um, streaming answers so there were next js route handlers and that that's what was migrated to to java spring um did it migrate the front end um, with claude it's it's interesting to like research to research up up front and figure out where where is the biggest benefit of migrating different sections of this um and i think like having a persistent application um, inside a serverless environment with 130 megabytes embedding a model is really bad so yeah um yeah uh cold starts are pretty pretty darn gross um, so yeah it it definitely was something that needed like a long running process and progressive streaming, and um in the process of migrating this um it's there's a bit of a bug, on my end and it's not actually streaming it's actually, um sending everything in like one lump sum still so i'm still learning um going through going through with claude but um.
There were a couple improvements to it. Like, you know, like TypeScripts had two versions of the ask pipeline. So there was like one regular JSON and then another server sent events for streaming. And like there was actually a comment in the code reminding you that if you changed one, you had to remember to change the other. So it has one pipeline now. um and uh, yeah the logic for generating the answer exists once um, and it was really cool to also like like go back and forth on potential features, like given you have like 130 megabytes of data it's like looking at the app and anybody could go go on um i believe the url is, we could link it in the podcast, right? ETA brain.vercel.app.
Um, yeah, it's, you just have one search and then you have the transcripts that you could search directly, and then you have, uh, common questions that people ask, but with 130 meg meg, uh, megabytes of, of data. You could definitely go through and semantically make like a tag cloud of some of the most weighted terms um and make that like a very interesting experience for the user um to kind of, look for look for specific things um in your case matt when you were collecting this data how long did it take you to, I guess convert 121.
Matt 7:30
So did you rerun the conversions or are you using what was already in the database
Nina 7:34
I'm using what was already in the database.
Matt 7:36
Yeah okay that took about 10 hours and then basically what it did was it went through it took the podcast audio downloaded it onto my macbook or not my macbook the other one, the desktop Mac version. And then did a Hugging Face transcription And then on top of that, it did a Claude session to do the semantic understanding of it and then upload that as like a rag with the transcription timestamps. And so I think anything costs like $40 of API tokens and it took like 10 hours.
Nina 8:27
Okay, wow. Are you speaking $44? like with like a like a pro account or do you mean like you went by usage.
Matt 8:38
Or no there's something like i couldn't get it to work with my subscription thing i think because the way we're i was doing the lambda stuff in riverside i needed to do an api token um so i probably could have salt that But I didn't want to waste my time for 40 bucks.
Nina 8:59
Yeah, so that's interesting because that's something that I've encountered in the most recent updates of Claude when developing with it. I have two apps, right? I have one app that actually loads the Claude code terminal and output in the app. And it's directly connected to my Claude code account. And then I recently tried making another app that grills you and gives you tech interview questions with a code editor. I'm in the job market right now for people listening, so I wanted to make something, that could evaluate my answers and grade them out of 10, so there's Claude underneath the app.
And I noticed when I went into the Anthropic console that I had like a $5 spending limit and was actually directly consuming tokens. So it had set it up directly to the API instead of to like-
Matt 10:09
It created an API token without even letting you know.
Nina 10:15
I think I might have missed in the setup what was going on. But you can make a Claude app that just directly connects to your local instance of Claude without going through the API. That's also possible.
Matt 10:34
I am not entirely sure where that stands in the terms of services land. Because that was the big thing when the crap thing came out, OpenClaude. But they kind of just didn't care. And I mean, I have my little chief of staff bot running in my subscription that spends like five minutes a day at most running. And so they're not going to care. I mean, they really shouldn't have that, just kind of small usage. But going through all those transcripts for 10 hours straight, I'm pretty sure my subscription could have got dinged for that.
Nina 11:19
Yeah, probably if you had it directly hooked up to your sub. Yeah. Yeah, but those are the things that you learn when you're, I guess, developing these kinds of applications. Yeah, so I was draining money prepping for interviews to get a job to earn me money.
Matt 11:41
The opposite you want to be doing during that time.
Nina 11:44
Yeah, exactly. But I wonder for the future of AI in apps, I, wonder where that's going to go. Do you think that apps are going to use the Claude API, or do you think there are going to be some that say, sign up with your Claude subscription and and just log in with you oh yeah in the app so you can use your pro subscription within our app because there's going to be ai and everything and the companies are going to have to make like this determination right of.
Matt 12:23
Yeah which is that's tough right now that's like the mcp like discussion like so you build apis right to have people connect to your ai bot or do you have like an ai chat bot that you like custom the agent for um and that's like really hard like i mean jira's a good example of this right jira has a shitty ai bot and it's useless and nobody likes it but someone built a cli around the api for jira and claude is very good at just using it and so like you get your tickets there and you can move things over pretty easily with that Jira or just like give Claude a link to your Jira to read tickets and do updates there instead of actually using the Jira MCP.
Nina 13:10
Yeah.
Matt 13:11
And I mean, it's puts the cost on the customers, so I'm surprised it's not pushed more. Of actually running these agents and keeping the agents like up to date and useful where like so it's like do what you're good at kind of mentality so just like have whoever like if they want to Claude they want to do um chat gpt like make that the client's problem and then you just host an api like you would good old web 2.0 style and let them figure it out
Nina 13:48
Yeah yeah yeah there's also the side of if you do that you put it on the customer is the customer going to even want to use the product like do they even.
Matt 13:59
You know yeah so jira like you're a tech user right so So it's no problem for me to pick up Claude by Notion, like a Notion thing, integrating. Most of those people aren't technical and have enough problems just using the web browser. So they're going to want a more comfortable UI to use. They tend to use Claude Cowork with MCPs sometimes to have the user interface that they're more familiar with. It's like you just hand them a terminal, but it's here. But it's just, you got to think.
I still just don't want Notion to have their own agent. I want it to be Claude, or I want it to be ChatGPT. That's their job. Their job is to have good agents.
Nina 14:56
Mm-hmm. Um i i there was a point where i was really frustrated with i don't i don't want to talk bad about notion but i was really frustrated about like the the paywall behind uh using their agents and um, like i did not want to pay you know 100 bucks a year and then on top of that extra to use their agents so i made an app that worked around, i made an app that worked around that um I think I called it like Novel Notion or something.
And it was in Markdown and it had like a hierarchy of files, but there was a button to sync with Notion. So it would near the hierarchy of files and check against every single one that was in your Notion and make sure they were perfectly synced and up to date. Um and then just it was literally just a copy of notion if i'm being honest and then like in the inputs like on the pages you could type a backslash to access your local Claude so, local yeah so i i spent a whole day making something just so i could have my own a ai agent inside.
Matt 16:17
Insight yeah i mean 100 bucks is nuts like that's what second to top pricing for anthropic and i i don't i have the top one that's like 200 and i barely hit the limit so there's no way
Nina 16:36
With java last night i hit my limit i had to upgrade um no one i didn't realize java was such a token monster.
Matt 16:45
Yeah, it's all those inputs and proboscity.
Nina 16:49
All those types everywhere.
Matt 16:51
Yeah yeah but it's just that's so like the margins on that hundred bucks gotta be great
Nina 16:59
No it's i i believe i believe the perks are for.
Matt 17:02
The business not for us
Nina 17:04
Well yeah never mind yeah you could just sync across multiple devices that's that's the perk and you get like like two prompts of ai to organize information into a table or something Wait.
Matt 17:18
And they meter the AI you're using after the 100? That's just greedy.
Nina 17:25
Yeah, pretty greedy. But there's an alternative called Obsidian that I heard is.
Matt 17:31
Exactly the same. Yeah, I've heard about Obsidian a lot. Yeah. Obsidian, Notion's aimed less at tech people, and Obsidian's more like, if you kind of know what a graph database is, you can have a better time.
Nina 17:48
So yeah, what's something that you've been developing this week or just playing with AI?
Matt 17:56
So I got my chief of staff thing, right, and it's giving me reports. It's still Mostly useful. The thing is, if it's not completely useful, it feels like a pain in the butt. I want it to give me more terse information about my PRs, but also useful information at the same time. When it's like, oh, that's 7 out of 10, I want to be like, why is it a 7 out of 10? Because most of the time something, it goes, and I'm just like, oh, you don't know the context of the app.
So, like, what you're complaining about is something that's taken care of in another repo. Like, if I could get it to tell me that stuff in, like, the report that's building up, I would prefer that. And what I'm having issues with is, like, the division of what gets done on box on the client side and then what gets done in the chief of staff box. Um so the problem i was doing for a while was it was just trying to like get a diff and then like get us some nazis on the box then like in the chief of staff was trying to then crunch that into a smaller like digestible bit but that digestion needs to be done on the client box where it has all the contexts.
Nina 19:30
Oh, okay. So it's just two different contexts kind of separated when doing evaluations.
Matt 19:38
Yeah, because I'm keeping any code from getting off the box. It can only talk about it in abstract terms. And that's just two levels of abstraction. By the time it gets to me, it's just like, I just don't like this one. It's got some architecture stuff. I'm like, no, no, I need to know what the hell is the architecture problem you're running into.
Nina 20:02
Yeah, so this is a terse report specifically for improving existing architecture in your applications. Like, is it, I guess, what's the...
Matt 20:15
So the thing is, like, right, I got PRs on the box that I need to review. Like, I could just easily get, like, a little report saying, yeah, I have these things, click these links, go to it. But that's basically what I get in the email anyways. What I kind of want is the quick Claude review. Give this thing a score out of 10. Give me what the PR is about. Give me why it's not a 10. What's your main objections? So that's basically the whole thing.
And so getting it's I got everything like structured wise because I'm always good about that like I always I never just kind of like send blobs over it's got to be like JSON formatted at the very least so like Claude always does better when like you switch from agent to agent and you give it JSON instead of just text So I do it that way. And so it's there, like it's almost there, just like where it's getting processed doesn't have the context to build useful information in the structure at the moment. Yeah.
Nina 21:28
So when you're developing with multiple agents, you know, kind of, so you're saying that one kind of generates the prompt for the other agent or the format. that yeah yeah so um.
Matt 21:44
I guess it does allow for some back and forth like it's allowed three turns basically like if it gets something it's just like oh what's this and it can ask again for some more information
Nina 21:56
Yeah so like follow it three max follow-ups to each other essentially yeah.
Matt 22:02
Basically.
Nina 22:03
Yeah. So how are you building this? How are you building, this terse report system? How is this being confused?
Matt 22:12
Yeah, it's Java server, right? And then it's the Claude instances on the box. Like we said last week, it's just my Claude subscriptions on the box and they're on GPNs. And then that's all within my company's VPNs. As well so and then that talks to me through the company slack and uh how it works behind the scenes like under the code is java server um using uh spring ai and that has like special components to do different types of talking to Claude back and forth um i rely a lot more on just like kind of open pipeline of it and i just do some like the open pipeline just to give free text to it that's with json inside of it and then i've kind of just wrapped that around some simple java like cron jobs have you done like spring boot batch or anything like that
Nina 23:17
Um i've done external cron jobs that hit a public API before for.
Matt 23:22
Apps. Yeah, it's basically the same stuff, just add scheduled jobs, and then you give it a little cron in there for when it's running. And it's been, it's been good. Like the structure is fine. Cause that's like all like 20 year old technology. Like people have been running jobs forever. I don't have to worry about that.
Nina 23:39
Yeah.
Matt 23:40
And not doing any custom code there, like really laying it on everything but the agents to pull out the data.
Nina 23:50
Yeah. Yeah. Interesting. Yeah. It's the, one of the things that I was encountering when I was migrating the, uh, the podcast search, was I, you know, I was torn between doing multi agentic, you know, development with it because a portion of this migration was, I also wanted to be learning from, you know, working with, um, working with a new Java app. So, um, I was kind of torn between, do I have an agent for the infrastructure, the database, testing, manager, separate them on work trees? Is that going to be like setting a fire at multiple parts of the app if they have to touch the same thing?
But it took me a lot oh go ahead.
Matt 24:43
I've had no success with multi-agent development I've only had success for multi-agents for reviewing things like even being like okay do the front end side with one agent do the back end side with an agent it just melts and then it can't communicate with each other and I feel it just blows through tokens way faster than just taking the time. So I just let her run and go get a coffee instead of trying to parallelize the agents have you had success with it
Nina 25:20
I haven't attempted to do it uh i mean i've i've if the contracts between the front and back end are so clear i mean i think then then i've had success with it but, um like if one is typing up documentation or one is doing tests that's fine because nothing is going to get mutated like right in the middle they're not going to touch the same file. But I was going to with another personal project where I want to migrate it to a more lightweight library like Svelte or framework like Svelte or something.
I was thinking I do want to have multi-agents. And I'm sure there are a lot of people online that have guidelines about how to do that, but, I imagine it's very, very difficult unless you have, Kind of how you programmed with the terse report where the agents can kind of convene and go back and forth for a max of three times before they go off and do their work. Like maybe something like that would make it safer, but no.
Matt 26:34
Yeah, that's like different agents with different domains. That's not development. It's not building software, my multi-agents. They're just reading code.
Nina 26:47
Right, they're just doing review. for sure but but they do have the ability to like you do set set it up where they can follow up ask each other questions so like maybe, maybe there's a way to program with agents and i was going to test this out where you can like, um set them set both off at the same time and have like a five minute window where they both plan together and then executes and make sure the contract is separated while they're building like, and then they have to reconvene and stop coding i think, if i can figure out a way to get multiple agents to go between plan mode and development mode at the same time and they they wait for each other to reach certain milestones and like, i've never used an application that kind of automates um agent workflow if you have multiple workers i'm sure i'm sure those agentic workflow uh applications do that but.
Matt 27:51
Have you tried ultra code and
Nina 27:54
I i have not.
Matt 27:56
I have not so it's it's just a effort level in Claude like you have high max and then ultra code uh i forgot it used to be something that was different or they needed a rainbow but anyways um It likes to do agent pipelines like that. And it does a little bit of that. It doesn't do completely like you're talking about with like two agents as peers planning. It more does like an orchestrator and then like kicks one off next, kicks one off next more to keep context clean.
I don't even know if that's as important as it used to be. It was back when like context dirtying was a big problem. And i think that's the problem that's trying to fix or another tool that i used to use was this like orchestrator same pattern but it was you'd have the orchestrator right that's always running and you that's what you communicate with it was basically like the pm for your project and then you it would go and then if you like came up with a feature it would then kick off like a project manager and then the project manager would spin off like an architect a programmer and a reviewer and a qa and like each of those would each do its own thing report back up and then like slowly bubble back up to the orchestrator and you'd have something done that would just destroy your tokens i mean you just burn tokens left and right doing that
Nina 29:36
Yeah that's what i was gonna bring up that sounds like you're burning money at that point.
Matt 29:42
And you didn't get much out of it it just looked cool
Nina 29:46
Yeah right like uh one one uh workflow automation uh app i really wanted to try what was the name of that one again that one was called.
Matt 29:55
Uh i forgot the guy like who built it changed the name twice and now i think he has something that's not open source anymore. So I'll look for something in the description. I can now put links in the description. YouTube is keeping me from last episode doing that.
Nina 30:13
Oh, oh, weird.
Matt 30:15
Gotta give them my picture. Some AI, like, verified it was human.
Nina 30:22
Wow. I haven't set up a new YouTube account in many years. Didn't know that was now a thing. N8N is the workflow out of me.
Matt 30:33
Yeah, that's the pipeline one. It's kind of like Zapier. Are you reading Zapier? Are they even older ones, Yahoo Pipelines?
Nina 30:46
No i never see i never i never uh i wasn't doing a lot of pipeline engineering throughout my i was i was mostly doing just mean, well i guess i guess um yeah so the the n8n is yeah, it's it i like it because it has like this nice visual node canvas so i mean i was debating getting that set up to automate like, i don't know clearing of contacts um and also like all the mcps that you have in there for like checking checking email and i just want to automate my job search, yeah that's that's what i want to do i want to i just want to web scrape um everything from indeed and linkedin and, It seems like that's actually really affecting, AI is really affecting the job market right now in terms of competitiveness, and I can't imagine how many recruiters are looking at the same AI-generated cover letters.
Matt 31:52
I mean, they're just having AI do it, right? Is it doing AI interviews?
Nina 31:57
Yeah. I did one AI interview.
Matt 32:00
Oh, yeah, you already had two ones?
Nina 32:02
Yeah it was it was like for some like tiny annotation gig i was just curious because like on linkedin they had a bunch of data annotation gigs and like, that's a hole that you don't want to get into right but um i was just curious it's and it said ai interview i just wanted to experience it um, i hate it i i don't like it um, Because all you can really do is just hope to say the right keyword and hope that you're not rejected. I don't know.
That's a little too weird for me.
Matt 32:50
It's like experience you expect a person. It's like white chatbots aren't great. Kind of okay. but after a while you even like club terminal it's just like you notice it's a bot's writing it's got a certain style to it and feel no matter what you do it's really hard to get rid of that flavor from now on
Nina 33:15
Yeah yeah so it's yeah i like i like talking with a person i don't like customer agents and chatbots like that personally, if i had a company or something i would definitely have to have them but i don't want to deal with them myself um, so yeah i'm i'm looking i'm looking at like uh how many people are just auto applying with agents to the same jobs in a span of two minutes.
Matt 33:49
At least I mean there weren't before there used to be the old just scripts they would run through like the LinkedIn easy apply and click the buttons because you would already have it all set up
Nina 34:00
Yeah I have I have done that before but, I have for the easy apply for sure but not anything if there's the option to add a cover letter you know I'm going to actually write a cover letter, but um yeah everything everything is is bots now and um, Yeah, how do you feel about that?
Matt 34:28
For the interview stuff? It's not great.
Nina 34:32
Really, across all different platforms, all different. How do you feel about that?
Matt 34:40
Yeah. I'm using bots a lot myself.
Nina 34:43
Right.
Matt 34:45
But those are my own personal bots. One being forced upon me. I don't even like doing the talk to ChatGPT or top-to-clawed ones. We figured those would be the best ones. I haven't had much interaction with customer service ones or call center ones. I hear those are becoming more and more popular. So mine are pretty custom and built by me, and I like mine, and mine are good for me. That's where I'm really in.
Nina 35:19
Right, yeah. Yeah, personal development bots for efficiency are definitely the, definitely the ones that you know i'm okay with that's just like, when i'm going you know when i'm calling somebody or i'm trying to talk to a person for a specific thing i do not like them um, but yeah i i needed to update my skills on linkedin And I was like, I am not going to retype the same skills like over again. So I set off an agent to transfer my resume skills right in the browser. Um, And, you know, it's made things so much easier to do that.
Matt 36:09
Yeah, that's just work, right? You're not doing anything creative.
Nina 36:13
No, no.
Matt 36:14
It's like when creative comes in, it's like it doesn't work out. I've been trying to write an article about this chief of staff, and it's just garbage. I don't like it at all. It just sounds like a bot wrote it. So I got to get it to just write an outline of what we did, and then I go in and write everything myself.
Nina 36:39
So you mean you've been trying to write a blog article, essentially?
Matt 36:44
Yeah, explaining what I did and stuff. It was good at kind of structuring it and then generating images to show the architecture. But the actual text just it sounds too ai to me i can't publish it right for some reason like i have older ones that are just like no this is just a new thing so i like ai Claude did it but those like Sonnet 3 and those don't feel as ai to me as the ones that come out now maybe i'm just more experienced maybe i need to go back and read those
Nina 37:17
Yeah i mean there there's if you're looking for you know models for create, like creative writing or something there are plenty on a hugging face that are attuned specifically for like writing blog articles or sounding more human um but, yeah i've i've also tried doing that for a blog post or, any type of post and it they they have the same they say the same things like every single time you know you're trying to talk about like, okay I'm just going to say for work experience because I wanted a blurb you know for my resume that was formatted in a certain way um, Yeah, they'll use the same keywords every time, and the sentence structure is just very convoluted and almost run-on.
Matt 38:10
It's supposed to be probabilities, right? So it's just using the biggest distribution.
Nina 38:16
But you're right. When it comes to structuring, let's say an essay, and you have five paragraphs, and the way it segues into each one and kind of structures all the themes, like bullet points, bam, bam, bam. Um it's very good at that.
Matt 38:31
For sure yeah which is good for me because i like staring at a black blank page i'll just run away where just like fill in this outline i'll be able to work a little bit better
Nina 38:43
And change it afterwards right.
Matt 38:45
Yeah like then then have the i like go through grammar just like fix things up in spelling and punctuation whenever But it's just so bad at writing.
Nina 38:59
It is very bad at writing. However, it is really good at, you know, I think I mentioned this app last time, and I mentioned this app this time. I, like, called this app the right help, right, where you have AI built into, like, a document editor. Um one of the things that it can do is um it has you know it can generate like, you know synonyms and stuff like right right off the bat like and it'll evaluate for tone um and make sure everything you know could you can modify the tone of a paragraph and that's kind of nice like, it It's nice when you're doing very, like, specific things or subtle, yeah, just very, very subtle things like changing, like, sentence structure when you give it explicit instructions.
But, yeah, I found, like, this synonym tool that it has is pretty useful. Like, just, it's good when you're completely blocked. You have a blank page um you know a blank code editing you just need the structure of something to get going so whether that's code or writing um it's good for boilerplating, that's what i'm trying to say.
Matt 40:24
Yeah yeah i mean i just like telling someone how to fix something more than actually making something new myself so
Nina 40:33
Well i mean i did have i i did enjoy improving the the eta brain app i do like improving stuff um but i don't feel like the improvements this round were as, where i wanted them to be i still have more i think i want to do to the migration um like it would be nice to get rid of Vercel altogether by the way, and just have things.
Matt 40:59
Yeah, but you got it all in the Hetzner box now. I don't see why I need to keep Vercel.
Nina 41:07
Yeah.
Matt 41:08
Other than I got some, I got the domains like set up and things like that. Yeah. Let me just change that over.
Nina 41:17
Yeah, I think pulling everything the UI or the user experience out of Vercel is like the next step in a new box and stuff.
Matt 41:27
So, It shouldn't be too rough. And that's pretty structured. Yeah, I can do that very simply. Yeah, I don't do any infrastructure stuff anymore. Also, so little new stuff to do in infrastructure work. I'm just like, just set up this box and go figure it out. Give it my API token.
Nina 41:49
Mm hmm. Yeah, so. Yeah, I, that, that, uh, yeah, I wish I, I definitely learned a lot from migrating that. And I think up front, there's so many things I wish, wish I could have done. I wish I could have said, we need the VPN up front or we need, you know.
Matt 42:12
Oh, yeah. I mean, that's like regular development, though. You didn't know that stuff until you hit the wall. You're just like, why is this not working? It's like, ah, yeah, I forgot to tell you about the VPN that exists.
Nina 42:23
Yeah.
Matt 42:24
Yeah. Yeah, it's normal software. That's a people problem.
Nina 42:29
You're right. You're right. Yeah. That's what I would run into that anyway.
Matt 42:35
Yeah, you can't do that all up front. If that worked, then people would still be doing waterfall.
Nina 42:40
Yeah. And the proxy and the changing the environment files, all that stuff. Dealing with a bug in caddy so everything's returning a lump sum. Yeah, those are things that you don't run into until you're in it. So, yeah.
Matt 43:04
Well, what else do you see? See your notes?
Nina 43:10
My notes are all over the board. I put together the notes last night, and I think next time I'm going to be earlier.
Matt 43:16
Nah. You got to keep it fresh. People want the news. Yes. Um yeah yeah we got everything mom yeah what else oh so one thing it did yeah i did do well is i put um put the podcast transcript from last week that's how i got the title and it was good title i liked it
Nina 43:41
Nothing is intimidating anymore.
Matt 43:43
Right yeah that was sweet and sweet title i was like damn i don't even remember Nina saying that
Nina 43:48
I think i yeah i think i did say that i i was i I was having a sky's the limit moment. Yeah.
Matt 43:58
It's like so profound.
Nina 44:03
It's very it's no i feel like i feel like an optimistic bob the builder or something no it's like uh, but no it's it is true though like anybody can build anything now and that's what's really nice like when i was going back and forth with um, again not never had to think about like memory and ram when it comes to hosting a jvm i could just ask it questions like hey is this enough with 130 megabyte models is this, what are we going to run into and, it actually it spit out ways to optimize um, running like on a cheaper cheaper uh instance you know a cheaper um container, with less power to it and it was like oh well if you do a a swap of two gigabytes of data they will be fine and it gave me all the different options so i would have never been able to, do that before unless i googled for hours when i used.
Matt 45:11
To be my job is just optimizing JVMs it's like my first joke it was miserable
Nina 45:19
You should use the the tool that you have to evaluate the code that I put together the report tool that, i feel like i feel like it's it's gonna have some things to say um, because i was i didn't have much input when it came to like my own input about, optimizing like jvm so what was that what was that like like before before claude.
Matt 45:50
You just have uh uh you had this like thing called jcat i believe which like listened on to the garbage collection and like jvm metrics, and you'd run the app, and you'd load test it, and you'd see what broke, and then when something broke or didn't break, you'd just adjust the levels. You're like, oh, this one needs more heap. This one needs more minimum memory. This one needs less maximum memory. It needs more CPU. And then you'd add that to the startup script for the app next day.
And do it all over again you go to a place and you'd spend a week doing it and then they give you a new place to go to the next week to do it all over again
Nina 46:38
Wow wow, that's that's what you did full-time before we worked together.
Matt 46:46
No this is like my first job out of college it was a consultant ever had uh you just like it wasn't every project some projects was like actually app development But when you're between projects, then you need to keep you busy until they found a bigger one. You do small engagements like that.
Nina 47:06
Yeah. And now that wouldn't even be necessary.
Matt 47:10
Now you can do like 100 of them in a day. If someone didn't already just do it themselves.
Nina 47:18
How do you feel about that, though? As somebody who had to put in the time to do that.
Matt 47:26
Oh, yeah. Like the slow way. No. I've never been one of those gray beards that are just like, Back in my day, you used to have to manage the JVM by hand.
Nina 47:38
Garbage collection. And yeah, it's, yeah, that's, I feel like I'm going to be one of those though. I feel like, I feel like I am set up to be one of those because, yeah, I'm kind of envious. I'm kind of envious of AI sometimes. I kind of get mad at it. like oh i put in all this time and this son of a bitch can do it in x amount of minutes seconds you know but um no it's, new frontier.
Matt 48:15
Yeah it just gotta be product focused and that's about the code and more about the product yeah just a lot better
Nina 48:21
Yeah for sure.
Matt 48:23
Build software you don't write code
Nina 48:28
I used to love writing code.
Matt 48:31
Yeah, I mean, that's the people who have qualms about it are the ones who like the code.
Nina 48:37
Yeah. Yeah, I saw, I know we're getting towards the end. Maybe I could table this for next time, but I saw a video, a vlog of a guy saying that he was going to refuse to use AI for coding it.
Matt 48:51
No, that's a bad idea.
Nina 48:53
Even if it gets him fired.
Matt 48:55
Well, yeah, I mean, that is going to, It's like, it's, well, some higher ops who doesn't code that's in charge of hiring, firing is going to see they're not using the tokens and they get upset.
Nina 49:09
Yes, and he brought that up. He brought up that, like, his project manager was upset that he wasn't consuming enough tokens and he thought he wasn't working or something.
Matt 49:21
That's exactly what happened when people switched to Agile. like it was and people didn't like points and what points don't what points don't mean time but then they totally mean time and then people like higher up that didn't know what's going on why aren't you getting more points then right and
Nina 49:40
So that something recent like when i when i did have a job that happened to me i wasn't using enough tokens at a previous place of employment.
Matt 49:50
Jesus.
Nina 49:52
I don't know if I ever told you that. Did I ever tell you that?
Matt 49:55
No. It doesn't make sense if they would track that.
Nina 49:58
Right, right. So I was asked, are you using the product? Are you using Claude? Are you using? And I was like, yes, I am. It's just that the work that I've been doing, I feel like, don't need to use it as much. They were just like a bunch of bug changes. I know where it is. I'm not going to send an agent off to do it if I know where it is.
Matt 50:23
Yeah, it's the metric du jour, so whatever it becomes, you always got to do it.
Nina 50:28
Right right but no he he did the the guy in the vlog brought up like skill regression and um, you know he he felt like he was forgetting how to actually engineer and then therefore if claude's building something he's losing the ability to, code review from not from that skill regression of not writing code and that does, that does make sense but now Claude reviews code so like you know what's what's the what's the train off there like what's the what's the point bun like.
Matt 51:09
I mean i don't agree with this guy at all i think it's a very bad idea both on the superficial side of it and the practical side of things like if everyone starts doing Claude Code like he's not going to be able to keep up reviewing manually
Nina 51:25
Right i think he's treating it kind of like join me in this revolution this this revolution like and no one's gonna join him that's that's the thing no one is, the the way he was phrasing things first they came from my code editor and so he was describing it like and it was auto-completing for me then it came for it like, it was really it was kind of funny it was kind of.
Matt 51:48
I bet this dude writes haskell as well just like does all the high-minded programming things you can do
Nina 51:56
No he was actually he's actually just a regular javascript developer surprisingly no it even makes it even funnier to me that that makes it really funny to me um.
Matt 52:08
Well he better at least be an anti-framework guy that be a vanilla js guy
Nina 52:14
Right. If he likes to do it the hard way, right?
Matt 52:18
Yeah.
Nina 52:22
But, yeah. Yeah, maybe I could pull up this video, like, another time. We could talk about it and why it's a bad idea. Debate it.
Matt 52:34
He better be in a very small company where he has a lot of way. Because, I mean, there's a game to this. it's not just do what's right all the time you gotta do what people want done sometimes
Nina 52:51
Yeah it's this is this is not the place to make make a statement about this like we're not going to be able to stop how things are changing you have to you have to join or you have to like you know, fall fall away but um yeah.
Matt 53:14
I think I'm good I
Nina 53:16
Guess that's.
Matt 53:17
The vibe that's
Nina 53:19
The vibe.