Nothing Is Intimidating Anymore
Matt and Nina introduce themselves and get into what AI has done to building software: Nina ships a desktop audio tool that would have taken a team, Matt tries the cloud-agent chief of staff trend, and they argue about whether the curve is flattening and whether the silver bullet that kills programmers actually exists.
First episode. Who we are, and what building software actually looks like now that the language you write in stopped mattering.
What we talked about:
- 0:00 Intros. Nina: eight years in software after illustration and music, Node.js, and building desktop tools with Claude. Matt: twenty years from enterprises to startups, independent for the last five.
- 2:30 Nina’s audio app: an Electron + Vue 3 + TypeScript tool for tagging and organizing huge music libraries, with stem separation built in. The kind of feature list that used to need a team.
- 10:48 “It’s ready. I just don’t know how to market it.” Why we’re developers and not startup owners. Yet.
- 11:00 The chief of staff trend: cloud agents with sub-agents, and what changes when the work stops happening on your own machine.
- 20:40 Peaks and plateaus. Is the curve flattening, or does it just feel that way every six months?
- 31:00 Where the bar rises instead: AR face mapping to 3D models, and features that used to be impossible becoming table stakes.
- 31:45 The silver bullet fear. The problem was never writing the for statement. It was translating a normal person’s idea of the world into a computer’s idea of the world. That still needs a special kind of person.
Transcript
Automatic transcript, lightly corrected. Also available as WebVTT.
Matt 0:00
Hey, Nina.
Nina 0:01
Hey, Matt.
Matt 0:02
How are you doing? I'm good. This is our first one. So congratulations to us for starting.
Nina 0:08
Yeah. Yeah, this is exciting. Should we do a little intro or?
Matt 0:12
Yeah, I guess so. People probably want to know who the hell we are at some point.
Nina 0:16
Right. Yeah. So, yeah, my name is Nina. Software engineer. Did illustration and music before getting into software eight years ago. I worked for a privacy company, React Shop throughout my career, a lot of Node.js development, a little bit of AWS. And in my personal time, I do code a lot with Claude for building desktop tools to improve my workflow. So mostly for my creative workflow when it comes to music editing and yeah, just a lot of script automation as well. That's kind of my focus right now with AI-driven development.
Matt 0:55
And met we used to work together Nina and I, so that's why we started this podcast I've been dev for too long, maybe 20 years now, and I've been doing things from large enterprises to some all-startups. I've been running now independent for about four or five years now, doing client work. And so I've been working in Java, Python, Node, a lot of stuff. And it's all changing now with AI, so it doesn't matter which language I'm using anymore.
Nina 1:30
Yeah, right? It doesn't really matter because, I mean, you can learn and pick up anything in an instant at this point. Like, I had flirted with using Electron for desktop apps. And, you know, Claude has made it so much easier to, like, develop a plan right off the bat and do research that it kind of feels like we could code in anything.
Matt 1:55
It doesn't matter. Yeah. I've tried Java. I've tried Python. It's the same.
Nina 2:02
Yeah. Like, nothing is intimidating anymore, which is, you know, or you don't have to invest, like, a hundred hours in something. But especially when it comes to like the research aspect of getting something started, I don't know. Claude has been really great for that. So like recently I made an Electron application with Vue3 and TypeScript that I've wanted to make for a while. And it was kind of a daunting task because all the features I think would take normally a huge dev team to complete. It's like a tag management you tag your audio files it's like a daw system for like music engineers streamers you could make sound boards and it's just for, organizing like huge music libraries on your local machine and it has like outside of like the tagging it has like stem separation with uh like what.
Matt 3:05
What's that mean
Nina 3:06
It means so stem separation is where you have a an audio track that's already finalized, mastered, and mixed. But let's say you want to hear the guitars or you want to hear the vocals isolated. I installed a machine learning model that I researched all of them with Claude, like which ones are free, open source. And Demucs separates the audio files. So I was able to separate like vocals, guitars, drums. And that's really great if you're somebody who likes to mix and remix tracks, so i guess it could be good for djs too but yeah it was just a little bit of python on the back end for separating out those scripts, and i know this yeah go ahead.
Matt 3:53
So this is just a tool that existed before are you using like claude to verify like the outputs yeah Are you verifying that?
Nina 4:01
I verify the outputs. So it's something, it's open source, and I installed it on the node backend, and I already had streaming and buffering set up for all the audio files. So it makes a request on the front side, and then on the backend, it does the separation using the model. And the model has its own. Internal functions. It's like a model slash library kind of that you can use and it splits them apart for you and then just stream and buffer them to the front end and it separates audio files and they're right on your operating system.
So the application is also kind of like it's kind of like a file explorer for your operating system but just for audio files. So, One of the things I learned with it was, okay, how do I search audio files on my computer extremely fast? And I researched with Claude, let's come up with the most efficient algorithm in terms of latency and how fast it could search, because that's what matters to the users. And so it helped me implement a parallel recursive walk, and it recursively goes through all the sub directories it's and you can actually set it up to search in batches with multiple workers, and in like the advanced settings i set it up where you could customize the workers so it can search through my entire downloads folder, which is extremely messy i think there's like 4 000 files in there i can't remember the like the gigabytes in there but And it can search through that in like a couple seconds.
Matt 5:51
What do you like even search for? Like cool guitar riff?
Nina 5:55
Well, first you have to point to the directory and it'll go through all of the files in that directory. So you could technically point to your root directory and it could search your whole computer and it'll find every audio file, on your computer, excluding like private audio files used by applications, like, you know, like sound effects for, I don't know. Like your pages app or when you get like a message notification, it'll omit those.
Matt 6:23
So is it just searching? Is it an audio file at this point? Or is it like looking into the audio file to be like, Oh, this is a jazz song. And like, this is a pop song or something.
Nina 6:35
Oh no. So, When it comes to classifying the files, like with tags, in terms of what those files are, that's up to you. There is a search and filter, though, where you can search by file type a string in there, and then once it's a string in the file name, and then once you add a tag, to the file, like you could tag something that's like, download 200. You could tag it, guitar riff. And you could organize those tags and then filter by those tags. So it's for people who have a very, very messy, messy computer, and they didn't digitally archive all their sound bits.
Matt 7:19
So at some point, are you going to try to get Claude to do the tagging for you?
Nina 7:22
Yes, yes. So that's an in-progress feature. I did add Claude. I have added Claude to other apps where I stream the output, in the app for like a document updater. And it was relatively easy. And I was able to hook up multiple models from Hugging Face, whichever ones don't affect my MacBook Air. You have to be very careful with what I install. But yeah, it was a huge learning process. And looking at the podcast that, you know, you were sending to me, what Claude has made the easiest is we're kind of almost like our own product engineers at this point.
Like, we just have to focus on end results and then verify the inner workings as they're being implemented. Like, as long as we research things and make sure things are sustainable and scalable, like, anything is possible.
Matt 8:26
You're talking about that panel podcast that I mentioned? Yeah. Jordan Gall is just talking about not... Do you even need developers? But you do, because they can't do it. But also, you don't want a full-time developer doing this work. It's a weird in-between of where you hire a developer, where you hire a SaaS. Like this could be a SaaS I feel like the Forge Audio thing like there's enough people wanting that
Nina 9:01
Yeah, I think so.
Matt 9:03
Like who would be type of customer, do you think? Would it be like a DJ or?
Nina 9:08
Yeah, I think DJs, Twitch streamers, people who use definitely like game streamers because there is a soundboard aspect. And a lot of Twitch streamers, you have to pay to make soundboards and stuff on Discord and a lot of tools that they use. So, I mean, this might be a nice open source kind of option for them. And audio engineers. I know it would definitely help audio engineers because when they're in the studio and they're mixing bands, they have all these tracks. And I can see them working and I know they're not naming the files.
Matt 9:52
They just like kind of make it and then throw it away and like not even try to organize the stuff and just make it new next time and things like that
Nina 9:59
Yeah they'll have like a they'll have a huge directory for all the different takes right so let's say like like somebody's recording guitar one, they'll have like probably a hundred takes of guitar one and they'll have to write down and remember which was the good take before they mix it in, for anyone on the podcast i i used to do music and go into the studio and see audio engineers work and a lot of the times they'd be like I can't find the take I can't sorry I can't find, the one from this day so if they had Forge Audio they could be editing write an app and.
Matt 10:36
It's now ready selling oh sweet
Nina 10:41
Yeah I don't know how to market though.
Matt 10:44
That's why we're developers and not startup owners yet
Nina 10:47
Yeah, But how has AI helped you with your consulting and migrating legacy code?
Matt 10:56
Lately, I've been doing the Chief of Staff thing. Have you been kind of following this trend that's been popping up?
Nina 11:02
No.
Matt 11:02
So basically, a lot of people are using like Rockbot. Have you heard of that one? Yeah, so it's the XAI thing. And it's like a cloud-based agent with sub-agents easily. So basically, instead of doing all this stuff on your own machine, there is just a bunch of random number of amount of computers you get to use in the cloud and what everyone's doing is like having a main agent that's like the chief of staff and that guy will spin off things to like go to your GitHub, check out your repos see if there's any issues and go to your email, check what emails you got check your calendar and like spins up all the agents then comes back that builds your daily to-do list of what to do and what agents are up to.
What I ended up doing was not using GrapBot at all. I went and I used the Hetzner. It's like a cheaper AWS, but you still get a full box. And so I put my account onto there, and then it can go. And also the best part is it'll go through the PRs that have requests and it kind of scores them for me. And so it's just like, oh, this is like a 7 out of 10 good.
Nina 12:47
On what metric, though? Do you get to dictate the metric in which it's?
Matt 12:50
The metric is, would you approve this PR?
Nina 12:54
But can you customize that?
Matt 12:56
Oh, yeah, you can customize it as much as you want. I mean, it's just Claude talking to Claude. So it's already like agents I had in each box that like have their own special sauce to be like, oh, in this repo, we do it this way and blah, blah, blah. And so would it then go ahead
Nina 13:14
I was going to say that please, Do I got beef? Yeah.
Matt 13:19
That's a pretty deep sigh.
Nina 13:21
I mean, it's like you're being an overseer. I like the analogy of, let's say you're building a building on site, and you have somebody who's pointing, work or go there, work or go there. It's like, how do you keep track of everything at once? What does the interface look like when you see all that? It's like, how do you manage all these individual agents? I mean, it sounds like N8N or like LangGraph or one of those other tools where you're kind of orchestrating agents to do a million things.
Matt 13:58
Yeah, that'd be similar. But this one's a little more wishy-washy. So basically what's on the Hetzner server is just its own Java server. And then inside the Java server, I got like Spring AI going. And it has like pretty strict things that it can do it's like oh i can go to the box and like run this script and the script only can do so much and that way it's like just can't go nuts like there is an escape hatch but it has to be talking to me that it can just like talk directly to its Claude or the Claude on a box but it's good it comes back with good scores or at least like it's like oh this is like six out of ten you should probably look at this one and like there's ones that are just like oh this is fine it's just like a one-liner change don't worry about it but i also got it so it sends me a slack message with all this stuff right but slack has this thing where you can kind of put like in little apps and messages and so the message comes with a button to approve and i can just hit the button from a phone and the phone then goes from the slack to the hetzner from the hetzner to the client box and then hits what does the GitHub approved command for that PR number.
Nina 15:14
Oh, man. I know a lot of devs who are doing that, and... I want to get into doing that. It's like too many. Did you take like a, was this something you just learned on your own? Did you watch, was there anything you watched like a tutorial about how to use GrokBot to like orchestrate all of this?
Matt 15:35
I'm lucky. I'm just like, been doing this long enough. I just know this kind of like pattern. Like all it is is just data pipelines and old fashioned ETL. But like you don't have to be as thorough and mapping out the pipelines you could be wishy-washy with cloud ai between the lines it's like kind of what everyone was selling on ETL like 10 years ago yeah never really paid off
Nina 15:58
Yeah yeah so it's deep agent frameworks seem to be what is going to be i mean if anybody makes an independent an independent app that is kind of like software as service that, like, for devs, it's just going to be deep agent frameworks. And a lot of these, you know, that have MCPs that work with, like, Figma, that work with Jira, Atlassian, it's like, having a framework that kind of maps out, exactly the workflow for for etl or pipeline it's, i always worry about just the verification process and i know it's pretty good at verifying itself but like.
Matt 16:42
What kind of verification are you talking about like
Nina 16:45
Like basic ad hoc testing with unit testing like like actually testing the app not just sending, Claude into like let's say the chrome tab browser to verify things like Like actually like testing it at the end yourself and then, figuring out what, because every time I implement a feature, I only run like two or three agents at once.
Matt 17:07
Yeah.
Nina 17:08
Because I say, okay, once you get to this point, I need you to stop. I'm going to go to the Chrome tab and I'm going to test everything. And then I give it feedback. And that's, I feel like if I don't do that, I feel like everything's going to explode.
Matt 17:23
I've been going back and forth, right? So on the client side, I have to test everything, basically try to get 80% coverage type stuff and end-to-end tests, all that stuff. Because there's a Jenkins or GitHub built around that anyways. But then it's just like on the Hetzner side, I think I have like five unit tests maybe. And that's it. And I didn't really care. And it's just like so bespoke that I don't think I'll ever care. That it's that 100 percent like it'll always just be me using it
Nina 18:00
Right right but let's say it's like a it's something that you do want to you know provide as like a subscription service or some application that you want to.
Matt 18:10
Sell well yeah so then like you got people paying i think once you start getting there's money on top of the code you should probably have some tests
Nina 18:19
Yeah, I think Forge Audio has 950 tests. Whoa.
Matt 18:27
How long does it take to run them all? Are we just unit or are we talking EDU as well?
Nina 18:32
No, these are just unit tests. It doesn't take that long.
Matt 18:36
It may not be too long, just unit.
Nina 18:38
Yeah, they're just unit tests. I don't remember, but I always make sure to run them before I commit, and I commit in batches. I don't say every single thing commit. I say commit a group of these. And I should be better about making work trees and stuff like that.
Matt 18:57
Yeah, nah. I don't do work trees. Work trees are useless. Are you trying to do parallel feature implementation at the same time?
Nina 19:08
I have done it, but, I don't like doing it because I like seeing everything that gets written. And I guess so when it comes to this development, then there's like there's another follow up question I want to ask for for for people who didn't listen to the podcast. I guess you could link them to this podcast. Yeah. Who do you think you're you're more of Jordan or Justin?
Matt 19:38
Oh, neither. So basically on the podcast, there's Justin, Jordan and Brian. So brian's the programmer justin's the marketer and jordan's more in operations it's kind of felt like is the way the trio is working at these days so i'm definitely more towards brian how about
Nina 19:58
You um i'm a little little bit of both i i feel like i'm i mean yeah, i feel like i'm as cautious as as justin and jordan but i mean brian is trigger happy you know it's.
Matt 20:16
I mean his entire business is teaching people how to use ai as coding agents so i feel he's i mean an unintended agenda is definitely there
Nina 20:27
Yeah and i feel like when justin and jordan they were arguing that there were no at-scale success stories for this yet. I think that was one of their angles. I think once we get to a point where these, agentic frameworks are less new and more sophisticated, which will probably be, I don't know, like six months or something. It's moving so fast, I imagine.
Matt 20:54
It feels like it peaks and plateaus at the same time. And just like, I can't tell if it's just a steady up and to the right, or if we really are starting to see the curve flatten and capabilities.
Nina 21:11
I think it will plateau, but I think once there are more of these frameworks... We'll probably see at-scale success stories with it.
Matt 21:21
So would you say like non-programmers, just like building sapses and spinning them out with AI?
Nina 21:27
It depends on the product, right? I think with those people, there'll be more fires to put out. But I don't know. It's hard to tell at this point because everything moves so fast. But i think that yeah we might we might get there where we don't really need dev teams we just have one or two people spinning agents and you know it, kind of makes me a little sad because it, it does because if one person well a because you're not you're not working with a team anymore which has its, perks and its drawbacks you just have one person with one vision and if that person is tunnel visioned and they make a mistake with their agent, I mean there goes their company if it's a big mistake you know?
Matt 22:24
What really is that kind of mistake? Like they delete the database?
Nina 22:29
Yeah they don't have backups or anything or you know It's something like that Something extreme.
Matt 22:35
So like I feel like those have happened But like also They're like people are learning from it And people are kind of backwards Becoming programmers from this yeah so like i've seen people like i've been helping a friend who's been doing this to like just kind of go through acquisition documentation like she does m&a stuff and she started with just like Claude in the on the browser then went to Claude Cowork and then moved to Claude Code and she's just like become a programmer over the past six months like she didn't know what a database was like three months ago and now she's like asking me how to do like as views and postgres and i'm just like what
Nina 23:20
That's that's a new thing i haven't heard of.
Matt 23:23
Like i mean it's basically like they just like hit a problem and then they dive super deep into just a tiny bit that's relevant to them and they pop back out
Nina 23:31
Right because they're looking at what they want in the end and so they've backwards learned what a database is just looking at the cloud stream is what you're saying yeah.
Matt 23:42
Because like they'll hit a problem they'll be like what's going on here and then Claude be like oh it's something with this this and then they just like try a few things and then it doesn't work and they've learned from those things i mean they're just learning they're just like learning the hard way and but they're able to iterate so fast they don't need like books or courses they can just do things
Nina 24:02
Yeah that's incredible that's really cool but that's also that also gets me nervous because people go to school for so many years to be able to do this and now.
Matt 24:13
I mean nothing's gonna match so one-to-one like with programming and ai so programming is definitely gonna become different a different way to learn i would be like way like in the beginning when like you just be like a script kid and like mess around on the internet, and I learned that way. And then there would be the more scientific stuff you'd learn in school that was more if you just wanted to be a logic-proof guy.
Nina 24:43
Were you a script kid growing up?
Matt 24:45
Yeah, I was beginning just like messing around the browser. And then, then I ended up going to college for it, learning all that stuff. But my college is weird. It kind of was like half, like real computer science and half like reading the rest PhD, that guy's dissertation. I had some weird professors who were very practical about the coursework. So I had a weird upbringing in computer science.
Nina 25:14
Where'd you go to?
Matt 25:16
I was in Loyola, Chicago.
Nina 25:17
Okay. Okay. I did not get a computer science degree. I went to a boot camp, which, that was kind of backwards learning as well. Even though they try to teach you the concepts right up front, I felt like I didn't get those concepts until I dove in and they just gave me the list of stacks and technologies and I had to be programming like 12 hours a day no joke it was yeah that was hours.
Matt 25:44
That was like kind of the big idea like back then to you could just do like six weeks of boot camp and then get a job
Nina 25:53
They're not around anymore really it's it's, I mean They have agents, like, it's AI programming boot camps, which that feels like it's just kind of taking your money.
Matt 26:09
Yeah, so, like, not even LLM training, just, like, using Claude to make...
Nina 26:15
Yes, a lot of them are just, like, learn how to use AI to program. They do have machine learning and LLM courses, too, but it's still more... They're just... You're paying for them to just point you to use Claude. Yeah, exactly. And I don't know how to tell a friend who wants to get into programming, that this is probably, it might be the best time, but it also might be the worst time for you to get a job because nobody's... It seems like nobody's hiring junior devs because AI can just do their job for them.
And that's what bums me out about this the most. It's like us who are mid to senior, it's like, yeah, we can do anything. We could, we could like found a consulting company. We can go work for a company that needs AI implemented in it. But the juniors are the ones, I think, that are suffering from this. And that's what bugs me, I think, a little bit. Like, how do we open up the door for people to get into tech now and actually get a job as a junior dev somewhere, you know?
Matt 27:37
Yeah, I mean, but are we just oversaturated? like do we just have too many juniors i mean that was like the big cell to like push more and more people in computers push more and more people into the boot camps but now they're out of the boot camps now it's too many like it's just a real a real not even distribution of people who've come into the field
Nina 28:02
I i feel grateful that i got my boot camp certification almost nine years ago, because it wouldn't have been possible now. But I don't know. I've known people, a lot of people haven't stuck with it. You know, people that got pushed into it or felt like they had to get into it because they just wanted to get into it for the money.
Matt 28:25
Yeah, it was for a while. It was just like the CPA, like what CPA used to be. People just got it because the parents made them do it.
Nina 28:33
That too.
Matt 28:34
And then after a few years, they're just like, I don't want to do this anymore. why this is miserable.
Nina 28:39
I'm going to miss those days when we go, oh, this is miserable. Just a little bit.
Matt 28:47
Just a little bit.
Nina 28:48
You know what I mean? Don't you miss banging your head over a keyboard for hours just a little bit because it feels like you earned whatever you made?
Matt 28:57
I mean, I still get that point with Claude. More now, I just don't bang my own head. I want to punch Claude in the face. It's like, what are you doing?
Nina 29:05
Yeah. Yeah. Yeah. But yeah. So we went a little bit off topic.
Matt 29:12
But yeah, it's our podcast. We do what we want.
Nina 29:16
Yeah. Yeah. The future, future of AI systems engineering is, I feel like we, the people who, who can do it, I mean, we're just going to need less steps and you know, that's, that's just the, the harsh reality of, of things. So people getting into tech, they, they have to learn how to, I guess, The bar is raised. Something that makes me really optimistic about the future is not only is the bar raised for developers' skill sets because of AI, but the bar of what makes a good application is going to be pushed to new levels.
Matt 29:59
Does it, though? I have not seen an increase in quality.
Nina 30:05
Well, I feel like we're going to have innovation and breakthroughs. like. I'm trying to think of an example of like something that could be elevated or pushed like things that would take forever to develop and add into an app that now we could do very quickly, like creativity when it comes to i don't know, for whatever reason let's say you have a very feature heavy app and it's a desktop app but for whatever reason you need to add ar technology, like why would you do that maybe there's a specific case where you have to add ar which is as you know where you, you do some sort of mapping of your face for whatever reason like i'll give an example game development, let's say game engines every there's there's hundreds of them and there's the big ones like unity or unreal engine, let's say they have additional like technology inside where, we're using an LLM under the hood in one of these to do an AR mapping of your face to generate a 3D model that is then optimized by machine learning, to have the vertices be smooth and run great in the app.
So that's an example of a feature that could be elevating. One part of the industry where you can you can just use ar technology to create 3d models from video so that would be like a raised standard so like creative things like that, But I can't think of another example. Exactly.
Matt 31:43
It feels like there's this fear of this silver bullet that's going to kill programmers, and it's just not there. There's like, the problem wasn't writing the for statement. The problem was translating a normal person's idea of the world into a computer's idea of the world. That still needs a special kind of person. Do that right it's like where our creativity exists that's like what we do and it's always been what we do but we just like had to waste a lot of time writing that into a computer language where now we can translate it more into a computer style of talking and then talk to the Claude and it can do the computer language it's just like a new layer if like a virtual engine like we're no longer doing assembly or just writing java
Nina 32:42
Yeah i think one of well you still work with them i don't want to should i mention his name i don't know no no.
Matt 32:49
Okay no names okay
Nina 32:51
Okay no names okay yeah one of the one of the people that we that we worked with when i would talk about ai with him very often, He said, and he's been a developer for like 20 plus years. He said this is like akin to when JavaScript frameworks came out or when Google came out. Like, we're always going to need developers and engineers. It's just that we have new tooling to make us do our jobs faster. And now with the agentic frameworks, even more so. And if I hope we don't reach a plateau I hope we keep pushing and.
Matt 33:32
Finding and engines I don't know it feels like it's slowing down
Nina 33:36
Are we talking about game engines in particular or.
Matt 33:39
Just AI LLM like we kind of hit the big surprise and then a few more now it's like like Fable 5.1 just came out I don't notice a difference
Nina 33:57
I do notice a difference.
Matt 33:58
With fable they do yeah like from opus to fable yeah for sure
Nina 34:03
Oh okay that's that's where i jumped but.
Matt 34:06
No before fable 5 just released a 5.1 oh okay uh so that just came out like this week and then but like opus of fable was like okay i noticed a little bit like sonnet to opus like was a big jump i and then and then i forget whatever the hell was before sonnet basically went from like unusable to usable mm-hmm I think it was just a few versions of Sonnet, but whatever. So maybe it's still just the same amount of increase, but just not usable-to-usable is such a huge feeling increase that the capabilities, even though they're still just incrementing the same amount, aren't as noticeable.
Nina 34:54
Yeah, I think I've got to get better hardware.
Matt 34:58
Oh, I have to run it locally, yeah.
Nina 35:00
Yeah, so I can start tuning models, because, that's something that's having specific models that are tuned for specific use cases, like spatial reasoning for, I don't know, things outside of what we do, web engineering. Like that that's really interesting to me robotics and ai i just want to i feel like a kid in a candy shop and i just want to be experimenting constantly even outside of web engineering but.
Matt 35:35
Yeah i always wonder what specialization gets you like if it really does pay off in the long run yeah
Nina 35:41
I i think it does like i'm sure i'm sure mechanical engineers they're gonna have to be you know are using models for, you know, like, agentic, like, what's the word? Just robotics. I'll just say robotics. Smart robotics. But, yeah, are there any hugging face models that you in particular have experimented with for, like, very special use cases for apps you wanted to make?
Matt 36:13
Yeah, I've used the Whisper one to transcribe podcasts. And that one's, I mean, it's good. And it's also, I mean, I use Super Whisper to like talk to Claude from time to time. And so I guess, yeah, for like audio, it works very well. It works better than whatever the hell YouTube's doing for transcriptions. That's why I use it because if I want to have Claude search a podcast, I have it transcribed with Super Whisper locally and then look at it. It does a way better job than YouTube's transcription. And YouTube's transcriptions doesn't have who's talking where if Super Whisper will tag who's talking about what, even if it doesn't know, they'll just put unknown, bark It's just like someone's dog in the background.
Nina 37:03
Yeah. And that app that you made is really, really useful. Maybe, oh, we got 15 minutes. Well, maybe next time we can talk more in depth about it. Yeah. Yeah, Whisper. Oh, go ahead.
Matt 37:20
Yeah, Whisper's good. I got no beef with it. But I'm never going to make a model. I don't think I ever would. So you let me know how that goes.
Nina 37:28
I got a MacBook Air, Matt. I don't think it's going to happen anytime soon. It's not going to happen.
Matt 37:35
I did one, make a model, like, what was going to be, like, 10 years ago, back then when, like, you could, like, have a telepicture was, like, a hot dog or not a hot dog was the trend at the time from Silicon Valley?
Nina 37:47
That was the biggest thing?
Matt 37:49
Yeah, it still took like 10 minutes of training on my laptop.
Nina 37:54
Oh my God. You see, that's where you need such insane hardware that if you're just one person, it's nearly impossible to train something that could keep up with any of these big name models. So it was 10 minutes for just determining is this a hot dog?
Matt 38:16
Yeah. It went through like 100 tagged images. It was like part of a workshop. It had 100 images of hot dogs and dogs. And I had to choose between if it was a hot dog or a dog.
Nina 38:28
What if you squished up the hot dog and then you presented it to you?
Matt 38:33
It had a hot dog that was being eaten by a dog to trick it. And it had a dog in a hot dog costume to trick it. Just to mess with it over.
Nina 38:45
How did it do?
Matt 38:46
Basically like there was enough of us like 100 people each making our own and like when we ran it like about 10 people would get mixed up on each one
Nina 38:56
Was this in what this was in school like a school project.
Matt 38:59
No this is i got lucky and like old company at the time sent me to like an training program oh I was doing a project for, so like basically their idea was they'd be a middleman between companies and like Amazon and Walmart and stuff. So like say Kraft would come in and be like, all right, we want our macaroni and cheese box on all the e-commerce sites, but we don't want to individually put it. So they were just a middleman and they divvied it out, basically retail stuff.
And so what we wanted to do with llms was you know amazon they like give you the image like in your hand so you know how big or small it is well we knew how big it was right and we can measure the hand we took someone's photo of a hand we measured it and so we like had both of them so we could resize the image to fit in the hand without like having to do a whole bunch of like graphic design on top of it to like or like actually take the photo in somebody's hand and and then gray out the hand and stuff.
So that was the idea. We could get LLMs to verify it because the data wasn't 100% good in our database. So sometimes you'd put something in someone's hand and it would just be like a can of tuna the size of your head. Someone used meters instead of inches or something like that.
Nina 40:18
So this was like, how big was the database again?
Matt 40:21
For those images, we had to have had hundreds of thousands, if not more. We never got to actually building it the idea was to build one what we ended up doing is just familiar with Mechanical Turk
Nina 40:37
It's, if we refresh my memory.
Matt 40:39
What time is it? So the idea is it's an Amazon product called Mechanical Turk. And basically you can, it just like pays an engine, pays like people 10 cents to do one tiny task. And our tiny task was to like, be like, is this image correct? Yes or no. And we like show them the image and then like, they'd be like, no, can size of your head. And so we just did that, but we did it internally too. We didn't even pay Mechanical Turk. We just had people inside our own company whenever they had free time.
We built an app to just be like, yes, no, yes, no, yes, no.
Nina 41:14
Yeah. You know, it seems that that's like a huge thing now. I don't know if you've seen on LinkedIn where they're hiring people to just train AI, doing stuff like that like a hundred times.
Matt 41:28
Yeah, yeah.
Nina 41:30
But they're paying like 150 an hour for doing something like that, which is insane.
Matt 41:36
Get paid to have your job destroyed
Nina 41:38
Yeah essentially yeah get get paid to i mean who knows what you have to go through i mean imagine getting trained on like, imagine you're trying to train a model for ethics or something that that would that would stink like like images for what is ethical versus what's not i've heard horror stories about it oh.
Matt 42:00
Yeah i mean yeah breaks people, right? They have to go to therapy and stuff. I mean, you probably end up with AI that has to go through therapy at some point.
Nina 42:07
Yeah. Well, yeah.
Matt 42:12
On that note, shall we end the podcast?
Nina 42:15
Yeah, it's kind of a depressing ending note. Yeah. Good talk.
Matt 42:20
Thank you all. Alright, I'm going to end this. I'll see you next week.
Nina 42:24
See you next week.