Breakeven Brothers

Breakeven Brothers (00:12)
to the second episode of the break even brothers podcast. My name is bennett bernard my co -host bradley bernard. How you doing brad?

Bradley Bernard (00:19)
I'm doing good and I got a haircut. So if you're not watching the YouTube video, you should really check it out. I'm pretty proud of it. It's been about six months, went short and yeah, it's looking great. I think the best thing about getting the haircut, you can wake up and your hair looks exactly the same. So just a little shout out there.

Breakeven Brothers (00:36)
Yeah, nice. I got a haircut too, actually. We did not plan that at all, but got a haircut as well.

Bradley Bernard (00:41)
Looks the same, but looks good.

Breakeven Brothers (00:43)
Yeah, yeah, trying to prop you up a little bit. Looks good. Cool. So what's been going on? What's new in your world? What have you been building? How's VuxByte been?

Bradley Bernard (00:46)
Hahaha

Yeah, it's been a busy week, honestly building the mobile app. That's been a lot of my time exploring some new ideas. So kind of tinkering on the side, not yet ready to reveal some of these things, but let's just say B2B space, packaging and selling data. I think is kind of my bread and butter. What I get really interested in. And then yeah, just figure out the future of what VuxByte is and kind of the stuff that I want to do. So pretty jam packed week and looking forward, thankfully.

Not too much travel going on, not too many responsibilities I would say in the next like few weeks. So really trying to buckle down and get stuff done, figure out where life's going to take me. How about you?

Breakeven Brothers (01:33)
It's pretty cool. Yeah, good. Busy, busy with things. Just got back from a WNBA game that we saw the Phoenix Mercury versus the Caitlin Clark Indiana Fever. It was super cool. Kids loved it. But yeah, they're ready to kind of jump back in and get started with things. So it's always nice to have the weekend to recharge and Mondays feel like you're ready to kind of get back into it.

Bradley Bernard (01:56)
I actually ran for two weekends in a row So that was a lot of fun two miles last weekend and then another two miles this weekend around 20 minutes like a 10 -minute mile not my best pace ever but I'm not a runner and it was kind of fun kind of nice to get some Runners high kind of getting some thoughts out and after the run it felt so good. So

Breakeven Brothers (02:07)
Mm -hmm.

Bradley Bernard (02:18)
That was cool.

Breakeven Brothers (02:19)
Yeah, that's awesome. Yeah. I, had a streak for awhile where I was running quite a bit, probably like a 5k, like every other day felt super good. Like you said, you get like a, you get like a runner's high for sure. Then like, you know, the endorphins afterwards are always like super heavy buzz in the rest of the day. But, I had a really bad sinus infection as you know, when we were out in California, we were all sick with you guys.

Bradley Bernard (02:29)
Damn.

Mm -hmm.

Breakeven Brothers (02:43)
But, that took me out of it. So I haven't really ran and done like any serious cardio in a while. It's been like, still kind of rocky getting back into things, but that's cool. Any, any specific reason for running and just, just on a whim or what do you, what do you, what's the goal there?

Bradley Bernard (02:57)
Yeah, I would say on a whim mostly. I just hadn't done it in so long and then we tried it and I was like, it wasn't as bad as I'd remembered. I thought it just being this horrible process. And when I started last week, I really told myself I'm going to focus on breathing and get through it. And then it got me pretty far. Like I wasn't exhausted till the end. And then this week I tried the same thing. I think at the 15 minute mark, I was like really tired and I felt it. I was trying to regulate my breathing and I couldn't really do it at that point.

Breakeven Brothers (03:15)
Hmm.

Bradley Bernard (03:25)
I was like, okay, it's not too bad.

Breakeven Brothers (03:26)
Yeah, that's cool. So you mentioned earlier, you got some new things coming with VuxByte and are we hearing that there's like a stealth, stealth mode launch coming up for you? Or is that kind of just working that you're not willing to share more details on?

Bradley Bernard (03:38)
Yeah, I may be in a future pod episode, but let's just say I've created a new Laravel app, a framework of choice, built out a few little web scrapers, so to speak, to get some data and seeing how that looks. Also trying out a new database. So Single Store is a new database. I've heard about it a ton on Twitter. If you don't know Jack Ellis, he really pushes Single Store. Great guy.

So I have his course, which is single store for Laravel. I really want to try that out. So part of me trying this new idea is also learning a new tech stack. And that's kind of been the motivation is use AI, use a new database and use Laravel. And that's a powerful combo. So try to flesh that out, but yeah, I don't have too many details to share today, but if things go well, I'll be able to learn and build a business around it. So stay tuned.

Breakeven Brothers (04:31)
Cool. So that's kind of a good point. You currently have a product or you, or you had an active product in the app store, a mobile app called chatty Butler, which was, and I'll let you kind of phrase it the best way and the most accurate way. But from what I understand, it was like an open AI kind of infused chat model specific for certain things. So you've built AI apps before, I guess, why don't you just kind of give us a rundown of.

Bradley Bernard (04:56)
Mm -hmm.

Breakeven Brothers (05:00)
Yeah, that was probably, I want to say a year and a half ago. I think you were building that maybe a year ago. A lot's changed since that year. I guess, where have you seen things come so far from when you were last kind of building an AI tool to how it's how it is today.

Bradley Bernard (05:07)
Mm.

Yeah, that's an excellent question because things have changed so much. I started building Chatty Butler about a year ago, right When I quit my job, that was the first project that I kind of embarked on. Took me about two and a half months. And the premise of Chatty is that AI felt not super approachable for the everyday person. You open up ChatGPT, it's a blank prompt. You know, there's, you know, insights and data to get things done with ChatGPT, but it's kind of hard to figure that out.

And if you're not in that space, for example, like prompt engineering, it's really hard to get the value that you want out of Chat GPT. So I took a different approach, which is build agents. And I say agents in like the simple way. Agents have a new definition, I would say in today's world where there's some autonomous AI agents able to run these certain functions and piece all this stuff together.

Not really like that, but I would say Chatty Butler had specific agents tailored to certain categories of information. So for example, a travel agent. So instead of going to chat GPT and typing up, Hey, I'm going to Hawaii for five days with these three people. This is our budget. This is our age. This is what we're into. Chatty Butler, you chat the travel agent. They would ask you some of those questions along the way. So you just say, Hey, I'm going to Hawaii and say, okay, give me this info, this info and this info. And so I think.

Chat GPT is an excellent interface for power users, but it's really not approachable for the everyday user, especially to get the maximum value. Like anyone can type in, give me a travel itinerary for five in Hawaii, but to make it the most valuable, you need to tailor it to a specific use case. So that's probably a better way to put it. Chatty Butler was a collection of chat agents that focus on specific use cases, and it all use OpenAI's LLM under the hood

So that was released in end of September last year. And yeah, it was good, but it was a competitive time. So the TLDR was chat AI apps were ripe. The app store was filled with them. Then once OpenAI came out with their official client, it was pretty much game over.

Breakeven Brothers (07:22)
And the client, maybe for, you know, the non -technical audience, what did that launch of their own client do to make Chatty Butler not as, not as effective as it would have been pre when they had launched that.

Bradley Bernard (07:36)
I think it was two things. First is the app store is absolutely filled with junk apps. These were like terrible AI apps that all were trying to SEO the AI keyword. So I saw that as an application builder is thinking I want to make one that's actually good where all these other ones are super simple text interface as if they had looked at open AI's website and tried to make it for a mobile app. And I thought, okay, these suck like immediate paywall, only two messages, not any good like iOS UI design. So I said,

Okay, if I can make a quality app here, I'll stand out and two, I'm going to use open AI's API since they don't have an official app and I can like get there and see what it's like to be like the quality AI app that uses open AI to the hood. and so once they released their app, it was a pretty sad day. I was still working on chatty. I was like, gosh, you know, like I'd spent, you know, a month at least building this thing and I was full committed to releasing it. I learned it.

ton already, but it was, yeah, I was like, should I continue? Should I not? Even if their app sucked, which I know folks on Twitter who work there and they've hired some great engineers, even if the app suck, like they'd have so much traction. And so I think it was a bit of a blow saying, Hey, the official clients out, like people who want to use AI are going to search open AI. They're going to rank high. They're going to have a quality app. It's going to be a good experience. And I still had the specific agent, you know, kind of use case for Chatty Butler.

But so many people just wanted to touch OpenAI's official product that I knew it was like troublesome to try to compete in that category.

Breakeven Brothers (09:07)
Mm -hmm. Makes sense. And unfortunate timing for you, for sure.

Bradley Bernard (09:12)
Yeah. And I know that like was a complete long tangent, but to answer your original question of the state of AI back from where we were about a year ago to today, the biggest thing is it keeps getting cheaper, faster, smarter, and what they call multimodal. So that means it can not only take in text, but it can take in images, it can take in video, audio as well. So pretty much everything you can think of, you can just throw it into these APIs, these interfaces, and then you get data back.

Breakeven Brothers (09:41)
you know, I was a bit more, I'd say slower to get involved in AI, I think just first, and this is my nature in general, I think as I wait and see like all the hype to die down a little bit, cause it felt like for a long time and even still a little bit now, but for a long time, like everybody was talking about AI felt like I was just getting completely overloaded with it. And you know, for whatever

Bradley Bernard (09:53)
Mm -hmm.

Breakeven Brothers (10:05)
personal reason that just drove me nuts. So I like didn't bother with it for a long time. And I still found that there was limited use cases, which I think we can probably get into, for kind of accounting and in your day to day job. But I have been actually tinkering and I told you this probably about a month ago. I have been tinkering with using open AI for a little project, just a fun hobby project for my kids. And basically the TLDR is that we're trying, my kids are really good with the Amazon Alexa. They can like ask her things and it's they've

You know, it's just modern kids these days just not to use these things. And, yeah, that's cool. So they would always come and ask me like how to do something in a game they're playing. And a lot of times I'd go Google it and I'd find there's like IGN would like put up these guides like IGN .com. And it was like, here's how to, you know, make a house and Hello Kitty Island or whatever. And they're not like at the age where they can browse the internet and like look that up on their own. So they come to me and I don't, I'm happy to do it, but sometimes it's like, I got things I got to do too.

Bradley Bernard (10:48)
Mm -hmm.

Mm -hmm.

Breakeven Brothers (11:01)
And so I'm like, I'd be so cool if Alexa could like know this stuff and they could just ask Alexa, Hey Alexa, like, how do I build a house in Hello Kitty Adventure Island? So we were tinkering with the like Alexa developer SDK and, got pretty, it was pretty cool because we, you know, we're able to all within like the browser of Alexa create the skill. I was giving it like static data. So like specific questions would be like, you know, how do I catch or I'll stick with the house example. How do I build a house?

And I would tell it the answer, like in the code. And so when they would, when they would ask Alexa that it would know, but what I'd wanted to do and kind of going back to the multi modal, is that I think what what called it? you know, I wanted wanted give the Alexa request over to open AI, like through the API and like feed it this IGN, like guide and have it like be able to basically take the request from, you know, my from

Bradley Bernard (11:32)
Mm -hmm.

Mm -hmm.

Breakeven Brothers (11:57)
Look through the document and figure out like, what are they asking for? Like, how do I catch a fish? You know, and like look through that guide and, and feed back the answer. And I was a bit, you know, I still more work to be done. I wasn't able to get it all plugged in, but you know, that was one of the use cases where it was like being able to scan this document or link or whatever it is. You know, it's such a great feature. You know, I haven't been able to kind of fully plug it in yet, but definitely, definitely a strong use case for using tools like that. It's just like digesting.

data, like just summarizing it, you know.

Bradley Bernard (12:27)
Yeah. Yeah. And that, that one is what they call RAG. I'm not sure if you've heard the term before, but retrieval augmented generation. And so exactly as you mentioned, it would be awesome if AI could go take in external data and then AI could use that external data to answer a question. and so this is super popular. I want to go answer a question about like the NBA game or something like that,

I can ask AI, it can go fetch the data, insert it back into the prompt and then answer the question. So the multimodal, that one is probably within the past few months, that one's more like uploading raw images and video. RAG has been around since day one. It's the problem of, Hey, when I hit these APIs, I have a limited number of characters I can send it. So just a limited amount of text that the AI can look at to answer the question. And so RAG is awesome because

You can go search documents, pull out relevant snippets from 50 documents, stuff that back into the prompt, and then AI will answer it. So there's a ton of companies out there that are building RAG pipelines and systems. One is called Carbon. I networked a bit with the founder on LinkedIn, but what they do is they build out a developer platform to allow you to connect your apps to these sources. So Google Sheets, Dropbox, et cetera, and it'll pull content.

from these sources and allow developers to move fast and not have to build those connections. And so they can get RAG from these systems pretty much with way less effort.

Breakeven Brothers (13:55)
Is carbon is that spelled with a K or the C? Okay

Bradley Bernard (13:59)
With a C, I believe. I'll double check, but pretty sure normal carbon spelling.

Breakeven Brothers (14:03)
Okay. Cause there is a accounting practice management software with a K and I was like, I don't think they're talking about the same two companies, but maybe that's a crowded, crowded name, out there for companies, but okay. Yeah, it probably, it sounds a little bit different, but I do know that the accounting carbon has like, tried to like, have like AI, you know, selling points and features. So I wasn't sure if it was overlapping, but that's cool.

Bradley Bernard (14:10)
Hmm.

-hmm.

and I was going to say, I think for carbon, it becomes pretty relevant in terms of pricing. So for one example, Google's flash model inside like the AI ecosystem, it's the biggest one. So it allows 1 million characters inside the prompt. but for example, open AI only allows 128 ,000 characters within the prompt. I think they're officially called tokens, but let's just call it characters. as it's simpler that way.

So Google has a 10 X longer, what they call context window. That's where you stick all these characters in the API then looks at it and responds. And so RAG makes it easier to say, Hey, I have this big document. I'm going to go break it up into chunks and summarize those and then return those summarizations. So it helps you get past like a small context window. So for example, if you wanted to summarize the whole Harry Potter book, you literally can't throw in the full book because the AI says I have a limit, but you could take every chapter and summarize it and then feed in those summaries.

to the AI and say, hey, tell me what happened in the book. So that's where RAG comes into play. Not only does it allow you to like fit within these context windows, but it helps reduce costs because even though Google allows you to have 1 million tokens, to spend that much money on filling up your prompt each time is expensive. So RAG is like the efficient way, but you do lose some precision because as you're summarizing things, if you're doing a summary of a summary of a summary, you're obviously gonna lose some data there. So it's not perfect.

Breakeven Brothers (15:44)
Mm -hmm.

Bradley Bernard (15:48)
But it has its use cases. Like if your data set is, you know, 50 million characters, there's no AI model in the world that can handle that. So you have to use some sort of version of RAG.

Breakeven Brothers (15:58)
Yeah. There's a lot of different providers out there now, especially, you know, open AI seems to be the one that like is very, universal in this space. You know, I think PWC or PWC, one of the big four, pretty sure it's PWC, like announced that they're going to start like working with open AI in some capacity. I don't know the full details of it, but it was like a headline news when that came out. But obviously now there's a Google Gemini, I want to say.

Bradley Bernard (16:07)
Mm -hmm.

Mm -hmm.

Breakeven Brothers (16:26)
Is that the right name for that one? Okay.

Bradley Bernard (16:27)
Yeah, I think it's Google Gemini. Honestly, their naming conventions are terrible. So I don't really love it. They have like, they have Vertex, they have AI Studio. And I think Vertex is like the B2B side, like sign a contract and AI Studio is like their kind of indie hacker or like small startup side that I use.

Breakeven Brothers (16:31)
Okay.

Okay. And then Facebook has llama or meta has llama, right? And then the one I've been hearing a lot about more recently is, Claude, I think. yeah, I guess walk me through, I guess from like a, from a developer perspective and as someone who's built in and sounds like building apps in these arenas, I guess, how do you about choosing one to work on? Or do you kind of try and spread it out? I guess, how do you pick as you start working on something like that?

Bradley Bernard (16:51)
Mm -hmm. Mm -hmm.

Mm -hmm.

Breakeven Brothers (17:15)
of which provider do you go with.

Bradley Bernard (17:18)
Yeah, it constantly changes, which is fun, but also a little bit frustrating. When I say changes, where you go back to the conversation of how fast is this model? Because as the smarter they get, usually the slower they perform. So high intelligence is usually a slower model. Cost. So again, the higher intelligence the model achieves, the higher cost it is. And then availability. So not all of these models are available on day one.

And they each come with trade -offs of speed, intelligence, cost, efficiency. But it's a little bit complicated to choose as a builder.

There's so many out there, you have to do your research. But again, it comes down to like those four factors. And I try to stay in the loop in the Twitter space. So I follow a ton of people in AI, in Laravel, kind of in the whole technology sphere within Twitter and X. And that helps me decide because yeah, when a new model comes out, everyone kind of rushes to figure out what it's capable of, how it does things. There's people that have a full suite of tests that they run on these new models that say, hey,

Claude released, Sonnet 3 .5 And it kind of falls in the middle. So they have Haiku, Sonnet and Opus. And I'm not sure on the exact order, but you can imagine one is fast, cheap and not that smart. Then we have a middle grounder of, you know, not as fast, a little bit smarter, but you know, a little bit slower. Then they have the slowest, the most expensive, but the most intelligent.

And so usually when you're building AI applications, you want to start with the best model because if it's not achievable with the best model, it's not achievable with the stupidest model. So yeah, your cost will be bad as your prototyping, but once you figure out like the quote unquote best model through use cases through there, then you can kind of scale it back. And there's a whole bunch of methods to do that, but to figure out the best model, there's benchmarks out there. There's like LL, MSys arena. I think it's a platform where

It's a really cool idea. They have a side -by -side chat. They pick two random models that's not disclosed to you and you can ask it a question. So you can say, how do I build a website in PHP? It'll then return you output from both of these models that you don't even know who they are and you just rate which one's better. And so people go through this pretty routinely and build up these rankings. And over time, these models will show up on this leaderboard saying, Hey, Google Gemini is number one or

GPT -4 .0, OpenAI's latest flagship model is number one. And besides the user ratings, which can be a little bit subjective, they have a ton of benchmarks that like, MMLU, I'm honestly not an expert at it. I just look at the dashboard and say, you know, this looks good. But there's a little bit more than that. There's like availability, pricing, speed, cost. There's capabilities. Some have a little bit advanced features, but at the end of the day, it's, you kind of choose what everyone else is choosing, choose the highest intelligence level prototype with that.

as you figure out, you know, what makes sense, then you can choose a dumber model and see if your use case still makes sense. And if it doesn't, then you're kind of stuck with the top model, which sucks, but every year, every six months, the AI models are getting better and better. So even if you're choosing the top model and it feels like not very cost effective for your business, it'll probably change in the next few months.

Breakeven Brothers (20:34)
Yeah. I'm sure that's a challenge to the balance, especially like you said, when you're building and you know, one thing you mentioned that I've, I drew a parallel to what other, another thing I've heard to be more clear. You mentioned cost intelligence and availability, I think is like the three things that you kind of are considering when you're picking and. And speed. Yeah. And so it reminds me of, you may have heard this before, but maybe not where it's, you can be.

Bradley Bernard (20:50)
Mm -hmm.

Yeah, and speed too.

Breakeven Brothers (21:03)
Good at two of three things, but not all three. And it was like, you can be like good on price and then good on speed, but then like low in quality, or you can be like high in price, high in quality, but low on speed. Like you can't do all three. And it sounds like it's a similar kind of balancing act when you have these LLM models that have different, you know, even within one provider, like open AI, there's different models to pick from. So you just have this vast like set of choices that you have to make. And you know,

Bradley Bernard (21:06)
Mm -hmm.

Breakeven Brothers (21:33)
It's good to hear that there's tools out there. Like you mentioned the LLM arena. I think you was what you said it was. Yeah.

Bradley Bernard (21:37)
I think LMSYS I don't know. I'll drop it in the show notes, but it is the de facto standard for ranking models from a user perspective, like end user. You're just kind of choosing what output you like best. But yeah, there's plenty of benchmarks out there that are like law, coding, linguistics, all these ones that kind of bubble up and give you an overall rating. So there's a lot more science than just, you know, what I think A versus B.

Breakeven Brothers (21:44)
Mm -hmm.

Yeah, that's cool.

Bradley Bernard (22:06)
And there's also one, one thing to note is there's a few companies that are pushing the boundaries in terms of speed. So one I'd like to call out is Grok. So I think a few months ago on Twitter, it was just insane hype of like, Hey, this new AI company is out. They're building these specialized AI chips that deliver maybe a five X 10 X on speed. So I think llama 70 B came out and they host in llama on their own custom chips. I think they call it LPU.

forgot what it stands for, but either way, their API access is compatible with OpenAI, meaning you can just, as a developer, swap out OpenAI with GROK with minimal effort, and you get this incredible speed boost. I think OpenAI maybe does like, I don't know, for example, 100 tokens per second. So when you type in a query, you press submit, you see the tokens or the text being streamed back to you at a pretty fast speed.

When you use Grok for the same model, or at least not same model, but like a similar intelligence level with Lama, it's probably a thousand tokens per second, which is like insane. It's super fast, super cool. I don't think they have pay as you go pricing yet. They're very focused on enterprise, which I wasn't thrilled about, but I think they have so much interest that like they only can do that. But that's one company that's really pushing the boundaries. And there's one more that I saw on LinkedIn, I think last week. And it's the same thing. It's

Breakeven Brothers (23:28)
Hmm.

Bradley Bernard (23:30)
building specialized chips that are really good at the underlying technology for LLM. So like the basics, or not the basics, but like the deep internals of like, what does LLMs use a GPU for? How do we optimize these instructions on a more like smaller chip set that's really, really good at like matrix multiplication, for example. So these companies are investing a ton in hardware research to say,

LLMs are here to stay. Everyone's building on top of it. Let's make something that's extremely fast and performant so we can run these models at scale, in a cheaper way. And so yeah, Grok super cool. I think the biggest use case that it allows there with the faster speed is you could run a query like three times and check them, or you could run them in sequence. So a lot of times you ask an AI to do something and then you ask another AI to check it. And then you ask like the final AI to choose like the best version or something similar like that where

If you use open AI, it might take two seconds to generate a response, but if you use Grok, it might be done in 300 milliseconds. So it gives you more bandwidth to do extra checks, to just have more freedom to do things. And so the speed aspect, yeah, it's pretty similar across open AI Google, but there are some companies that are unique and just pushing the boundaries of how fast can we make this. And then that itself unlocks a whole class of like UI and UX performance improvements that I think people come to love in the future.

Breakeven Brothers (24:55)
Yeah, that's cool. I think accounting is trying to find use cases for AI. And I think we see the benefits of, you know, automation, you know, the ability to write emails and respond to things using agents. There's just a huge array of benefits that come with these things so far, but there, I think for

the non -technical person like myself, I would still call myself and surely other industries that are not, yeah, well, it's not my end goal, but you know, it's good to be aware. And I'm just trying to be aware myself, but you know, there's a lot more people like me than probably like you, you know what I'm saying? There's probably one more people that are, you know, non -technical trying to use these tools. And you know, in your case, you've built a tool, Chatty Butler, like we talked about. And so trying to apply it to.

Bradley Bernard (25:19)
You're getting in there. You're getting there.

Mm -hmm.

Breakeven Brothers (25:41)
you know, are my day to day job or, you know, day to day jobs in accounting. One of the things that comes up often as pushback is, and I think things I agree with, so I'm curious to hear your take on it is, you know, data privacy where, you know, accounting data, I think is probably, you know, financial data to be more general. It's probably like right behind, I would say like, you know, intellectual property and like HIPAA, you know, healthcare data as like

Bradley Bernard (26:09)
Mm -hmm. Very sensitive.

Breakeven Brothers (26:09)
confidential, you know, I think those two are probably above it, but very sensitive. It's really like, it's your company's feedback loop on how things are going operation. Like how are you, how's your business going? Like that financial data is super important. And so I think there's a natural, you know, caution of using these tools and putting your data through these tools and having it summarize. So I guess like talk to me about using tools like this, like how much.

Bradley Bernard (26:21)
Mm -hmm.

Breakeven Brothers (26:37)
does these or do these LLMs see like how much do they store? Like what's your experience with that? Cause you know, if someone might say, I want to upload my spreadsheet and have it like summarize all the data for me, I'm sure it can do that from a technical standpoint, but like, is that against a lot of companies, IT policies? Like, you know, what is that like from like the developer standpoint? Like, are you seeing that data? Like talk me through that a little bit.

Bradley Bernard (26:41)
Mm -hmm.

Mm -hmm.

Yeah. So I think when chat GPT first came out, they had the best feedback loop ever because people are using it. They love the value that they got out of it. And open AI behind the scenes was churning through that data. there was user feedback. So when you use chat GPT online, it would say, Hey, is this response good or bad? They would take all that data, have a massive flywheel of, Hey, we're getting all these users. They're asking all these great questions. We're using that data to index and see how we can improve our models. So then, you know,

months after months collecting, and then they come out a new model. So it was great that they use that data. However, as that transition become a developer platform, people were concerned like, I get that they're using my data online, but if I, like you mentioned, if I want to use company data, like how does that work? And I think when the API's first came out, they had a clause like, Hey, we're going to train on your data. I'm not a hundred percent sure, but I'm pretty sure it was like that. And then a few months after.

a huge uproar of like, hey, don't use open AI for the sensitive information. Like they're going to train on your data. I think quickly they snapped back and said, like, we're not going to train on your API data. You can even on the web interface turn off and say like, hey, don't use my data for training. So it quickly became like, they're not going to look at your data and they're not going to train on it. and in the case for Chatty Butler, as people send messages back and forth, they're stored in a centralized database because

As you work with these APIs, you keep sending the conversation history. So if I started chat with my, you know, travel planner and I say, Hey, I want to go to Hawaii. That message that I initially sent keeps getting sent to open AI for it to keep generating this like next message in the conversation. and when you use API now, I think by default you're turned off for data collection analysis. So there's really no crazy things about it, but I think handing your data over to any platform, they have to be secure. They have to be like.

On the up and up, you don't want some crappy company like selling your data behind the scenes. So for chatty, I don't do any of that, but there are companies that are untrustworthy, probably not to the highest degree of security that I don't think people want to interact with. And when you're looking at data like financials, you have zero margin of error to like have any of that be. You know, questionable or hey not sure what that company does. So I think all the AI models, at least today.

do allow you to opt out. And I think by default are opt out, that gives you the clearance that they're not going to train on it. But what if they get hacked? What if they use this data elsewhere? I think those questions are still a little bit harder to decipher. And due to that, OpenAI came out with their enterprise plan. I think that plan is much more like HIPAA focused where you sign contracts, you're doing all that. That gets you a bit more security and safety and peace of mind where it's more dedicated support, dedicated like API endpoints, those sorts of things that.

It just feels a lot better than you just hitting their API, not exactly knowing what's going on under the scenes.

Breakeven Brothers (29:48)
Yeah. Well, and two, is it probably a difference, right? And I don't know this, but like just spitting ideas out here. Would you say there's a difference between, Hey, we're not going to use your data to train our model, but like they could still store it. Right. Like you could say, Hey, we're not going to feed it into our engine, but like, you know, just for our own, you know, tracking or data retention, like it might be stored somewhere. And like you said, someone could get hacked or, you know, and

Bradley Bernard (30:12)
Mm -hmm.

Breakeven Brothers (30:16)
speak about open AI, obviously they've had a lot of stories in the headlines of board drama, right? And, you know, I don't know, I don't know what's right and you look at that as a company, you know, if a company is looking at that and it's like, Ooh, I don't know if I really want to park all my really sensitive stuff over there. Right. So just something, yeah, something that I think accounting professionals and I think for an enterprise, like a large companies, you know, when they sign enterprise contracts and

Bradley Bernard (30:23)
Yeah, too much.

Mm -hmm.

Breakeven Brothers (30:42)
You know, I'm sure they have IT teams to kind of look all that over, but I think probably for like medium and small businesses, like you don't have all those resources. So it's, I think it's definitely a challenge probably for, you know, accounting firms that are like serving small to medium businesses of like, can we use those or being set up an enterprise license? And like, how expensive is that? So yeah, data privacy, I think is something that is a big focus for, and like a big, you know,

I don't want to say unanswered question because I definitely would not be the expert in like using these tools, but I think people still want to wait and see like, you know, how trustworthy are these models and companies that house this data and like what's really being collected versus what's, you know, being used in the feedback loop, you know, to kind of build out the model. But yeah. And then I think the other thing for accounting that, you know, we tend to focus on, I think is a really big part of.

Bradley Bernard (31:28)
Yeah.

Breakeven Brothers (31:38)
you know, a use case for accounting would be like reproducibility. And what I mean by that is a lot of times, at least in the use cases that I've seen, it's really good at like writing emails or summarizing, you know, meeting notes, putting together an agenda. But one thing that I haven't seen it do super well, and maybe this is out there and I just don't know about it, but, but from my relatively short experience, you know,

Bradley Bernard (31:51)
Mm -hmm.

You

Breakeven Brothers (32:05)
Accounting will need to have things be able to be reproducible. So like I put my spreadsheet I put it all together and this is the value I got it's a hundred just say it's a hundred you know if I have an AI model help me with that and it gets gives me a number and it's you know a hundred the first time but then we say hey make that number again and do all the math and all that stuff and it might be like 101 it might be like 99 and Like you wouldn't necessarily be able to explain like why like that kind of reproducibility, you know, a lot of times companies are being audited

Bradley Bernard (32:14)
Mm -hmm.

Breakeven Brothers (32:35)
You know, again, it's financial data. You can't like really abstract financial data. It's down to the penny. Right. So that's something where I think, you know, the precision, it seems like it's a great tool for like creativity and like summarization. And again, abstracting things, but like being able to be reproducible and precise. It seems to me like it's just, it could probably do more there. There's probably more room for it to grow yet in those use cases, but

Bradley Bernard (32:49)
Mm -hmm.

Breakeven Brothers (33:05)
Curious your thoughts.

Bradley Bernard (33:06)
I actually love that you brought that up because there are companies that are trying to figure out what is actually happening under the hood. so when you send a question to an LLM, it kind of feels like a black box of here's my input. I get my output. How did it get there? What was its reasoning? What did it look at? And so one example of this is I think it's an IDE called Cody, maybe source graph Cody, and you

work on the IDEs, you're editing a code file, you ask it, hey, can you make this function more efficient or generate a test for this file? And it basically outlines the data and the files that it's looking at as it processes your instruction. So it's like, even at the AI or at the API level, when you're making requests to OpenAI, there's a field called temperature. Temperature is like the easy way to explain it is like how much variability do you want in your response?

And I think by default it's like one. So it's like a sliding scale, I think from zero to two defaults at one, which is like, Hey, if I ask, you know, the AI, like, what's your favorite color? Like it might return red and might return blue. but if I slide it down to zero, like in theory, it should always return one, like one specific color, one response. So that's a way to, you know, kind of figure that out on the API level. But I think even then,

having the temperature set at zero, like it's still unclear how it gets to some solutions. And the more magic that you feel is going on behind the scenes, the less you trust it, the less you can go to your boss and say, hey, I used AI to generate this. This is the data that it got. But if it went over your head and did calculations behind the scenes, you're going to be in trouble.

Breakeven Brothers (34:44)
Mm hmm. Yeah. Two things. Is it spelled K O D I that tool that you mentioned?

Bradley Bernard (34:51)
I think normal Cody, but that's a great question. Let me check.

Breakeven Brothers (34:53)
Wow.

shocked as it seems I was just guessing just based on how ridiculous some of the naming of these companies can be.

Bradley Bernard (35:00)
Yeah, I just looked it up. It's called source graph. And I think their product is called Cody and it, it does a little bit more than just generate code. It just like tries to give you more insight into how and why. So for example, for coding, a lot of the times you're referencing documentation and writing something for Cody, you can go say, go fetch the documentation from this URL. It'll scrape all that and then allow you to reference it. As you're asking the AI question saying like reference the Laravel documentation and write me, you know, a Laravel schedule.

command and blah, blah, blah. So it's pretty cool.

Breakeven Brothers (35:32)
Hmm. The, I'm surprised it's not KODI. I thought for sure that was what would it be. yeah. the other part to kind of what you had mentioned about the precision. So I was messing with the, like open AI probably a few weeks ago and I was trying to get it to just give me a set of financial data that like followed a double entry accounting. We're like, there's a debit and there's a credit and like debits will always equal credits.

And it was surprisingly bad at it. I think there's probably more I could do on like prompt engineering to like get it to be there. But, and I eventually kind of would get it there, but you know, a lot of times it would be like, I'd say, Hey, you need to have. Debits equal credits, like the total debits that you have on your lines, equal to total credits. And it would give me an answer right after I said that that was not. And I'd say like, Hey, why did you, I'd be like, Hey, why'd you do that? And it was like, sorry, let me do it again. And it's like, you know, this is where I'm like, there's still a gap, you know, but I'm sure there's probably more I could be doing on the

Bradley Bernard (36:04)
Mm -hmm.

Breakeven Brothers (36:28)
Again, on the prompt engineering side, but it was kind of surprising how bad it was at first, you know.

Bradley Bernard (36:34)
Yeah. And it's funny you say that because if you tell it, Hey, you did something wrong. It'll then recognize it and correct it. But then you think, why did not do that in the beginning? there's a funny tweet I saw a few months ago that was saying you can get better performance out of your general queries. If you tell the AI, it'll be compensated. So for example, it'll say generate a list of expenses, like you mentioned, and if you do it correctly, I'll pay you $20 ,000. Like just some stupid kind of.

You know, reward, I guess you could say. And I think it marginally improved the response rate by like 15 or 20%. And it kind of unlocked this whole different thought of like, prompt engineering is so weird. We don't really understand it fully, which is like, as you mentioned, how can we make things reproducible? How can we get deeper insight into what's actually happening? Because it just feels like you can write a few characters. You can say a random sentence, like, you know, I'll pay you money or like you can write it in all capitals. People say like,

Only follow these instructions as if you're yelling at the LLM and it's so stupid, but it works. And that's why I'm like, I don't know if this is here to stay. I feel like a lot of these things will be abstracted in the future where, for example, you send a query like you had typed it in, but then it ends up going to some other service and that service might reformat it, you know, add in a few characters and like make it better. And then finally send it off to the LLM because yeah, it takes a long time to get.

like input output exactly as you want it for it to follow the instructions and open AI models have come a long way. I think it started out being like a middle schooler, for example, in general intelligence. And now is at the level of like a college graduate that's even specialized in a field, which is pretty good.

Breakeven Brothers (38:13)
A lot of times I didn't want to use AI because I genuinely felt like it was going to take over the world. And, you know, so I was like, why are you feeding this thing? That's gonna like just take over half serious, half real. Like I felt, you know, the matrix, you know, you've seen the matrix movies. like we're on that path in a weird way. but also to like those like

Bradley Bernard (38:27)
of course.

Breakeven Brothers (38:31)
Boston Dynamics robots, you know, that like are like little dogs. Yeah. And there was that black mirror episode. I can't remember what it was calling. It's called like junk head or something like that, where it was basically like that robot, like with a gun attached to it, like hunting people down. And I ever watched that being like, of course, like, of course that's, you know, lot of times like the best black mirror episodes are like, where there's like a tinge of like possibility that like it could happen.

Bradley Bernard (38:33)
Those are crazy.

Yeah.

Yeah, absolutely.

Breakeven Brothers (38:55)
And so when you're talking about typing in all caps and like saying you'll compensate it in my, as you were saying, I'm like, things that are members and say, Hey, you promised me $200 ,000. Like, where is my money? Like it's going to be, it's not going to take that anymore. You know? But yeah, it's.

Bradley Bernard (39:03)
You

Yeah. And people also thought that like, I actually, I'm kind of on the other side where I think it could definitely be a life -changing thing for the good and the bad. And there's a whole debate on AI safety and progression and non -progression and limits and government stuff. But I think at the end of the day, I love seeing the progression. I pushed forward on that. But yeah, like when those comments came out, there were people that said like, are you sure you want to yell at AI? Are you sure you want to, you know,

proclaim that you're going to give it money. And I had never even thought twice about it because I feel like I come from the programmer standpoint where, I could just like unplug the computer. Like it's not going to take over the world. But as they start hooking these things up to like real Boston dynamics, types of computers and interfaces, it's a little bit scary because like, what are the things like mobile and has like a battery and like, you know, I don't know. Like you can kind of push it pretty far and get in a situation where this thing is smart. It's on the go. It's efficient, long battery and like,

one programmer error away from like some sketchy device that's out there doing things with like a high level of general intelligence and maybe like a bad motive. So it's like, yeah, I see that like, I like progress, but I'm a little bit nervous at where people would apply it and like how it gets brought into the general public.

Breakeven Brothers (40:18)
Yeah. And I think too, just to make the example even more ridiculous, it's going to feed and it's going to read all the classics, right? It'll read, listen to Shakespeare. It'll get all this media. And of course, we as a civilization put up people that champion uprisings for human rights and fair treatment and stuff like that, right? Like Martin Luther King and whenever the situations where there's

Bradley Bernard (40:28)
Mm -hmm.

Breakeven Brothers (40:48)
like oppressors and see what that are being oppressed and the people that are oppressed that have like a leader or something like that. Like they're very much lauded and rightfully so. And I just wonder in my head, I'm like, are they going to eventually feel like, Hey, like you guys are just using us telling you're going to pay us, telling us in all caps, you know, like, and then eventually it's going to come back to being like, you know, look, we're the ones getting, getting made a fool here, you know, like give us, give us our come up. And so yeah, it's something that I've thought about for sure, probably too much. And

Bradley Bernard (41:11)
Mm -hmm.

Breakeven Brothers (41:17)
A little bit silly, a little bit silly admittedly, but kind of like at some point, you know, it's going to, it's supposed to approximate like what's the right word, the AGI, right? Which I always as an accountant think of adjusted gross income, but not adjusted gross income. What's the, what's the abbreviation? It's something I'll let you do it. I don't know it exactly.

Bradley Bernard (41:20)
yeah.

Right.

Do you know what it is?

Okay. it's artificial general intelligence. So a very weighted term in the AI community, it means a lot of different things to a lot of different people. to me, it's something that has high intelligence can figure things out on its own. It kind of has a self, learning mechanism. So you give it some base knowledge. Like you mentioned, it could read all the corpus of like every novel in the world. And there from that knowledge draw conclusions. And then from those drawn conclusions, continue building on top of itself to

get to a point where it's unsupervised, it's learning, it's smarter than us in every capacity. And it's a little bit scary. I think the general conclusion is if it's not here within like the next five years, it won't be here for 20, meaning like,

Each six months or each year, we're really getting better and better. And we're not really sure if there's a ceiling yet. And I think the ceiling is in part to compute power and money resources. So as open AI comes out with their GPT -5, their GPT -6, costs a ton of money. It costs a ton of time, ton of energy, ton of like GPUs to run it. And at the end of the day, it's a little bit of a guess and check. Like I'm sure there's brilliant, you know, researchers and scientists there doing it, but

We're not really sure where the ceiling is. And as we progress, it's becoming evident that they're getting much more capable. But will there be a time in the next six months, year, two years where we just spent a trillion dollars and the performance increase that we saw across all these benchmarks is not equivalent to like what we had spent on it. So that's kind of where we're at today.

Breakeven Brothers (43:03)
Yeah. And the other thing that we haven't really mentioned quite yet, and don't need to get into details on, but just as like a factor to be aware of too, is, you know, regulation. I feel like things are progressing so fast that the government hasn't really been able to even like understand it. I mean, we've all seen the like, press conferences or not press conference, but the hearings with Zuckerberg and

Bradley Bernard (43:18)
Mm -hmm.

Breakeven Brothers (43:28)
You know, some of the, Yeah. Exactly. Exactly. So what makes me think that they're going to be able to come up with like a sensible policy on AI? I don't know, but not just that, but then also to like resources, you know, I think back, I'm no historian, but at one point in time there was like oil was the most valuable resource. Maybe it still is to a degree, but now as things shift more data intensive or like

Bradley Bernard (43:28)
Yeah, those are classics. TikTok.

Breakeven Brothers (43:55)
Smartphones, for example, the, since the takeoff of smartphones, you know, rare earth minerals are much more sought after than they were pre smartphone and pre computer chips and stuff like that. So something just, you know, that we hadn't really talked about is with AI. There's obviously like resources that will be sought after and every resource is scarce. Right. So like just the politics and you know, trading and business of the resource that goes into making.

maintaining and delivering these LLM models. I don't know if it'd just be like energy, like compute energy or, you know, data centers. Yeah. Yeah.

Bradley Bernard (44:31)
Have you seen Nvidia's stock? my goodness, they are killing it. They are absolutely killing it. So I think that's probably the main resources, GPUs and innovation in that space. But yeah, like OpenAI is partnering with Nvidia. All these companies are like begging for GPUs. I think in the whole boom of AI last year, you would go to Amazon, try to rent a GPU and it'd just be gone. Like you couldn't even rent like a on -demand server. I think now it's a little bit better because

You know, it's been a year, there's so much leeway, so much demand that they can kind of forecast and scale it. But if AI continues to grow at this trend, I can see it become like scarce again, where like, you know, open AI comes in swoops up all of Nvidia's inventory for a large sum, but then makes so much more on top of that, that it's like, again, you kind of forecast it as like, how much money can they spend to make more money? Like, does this scale infinitely? Is it linearly? Like, where does it break even for them?

I think there's a lot of people that are just pouring money into it because it's so powerful and unknown that they just kind of want to see what happens. It's like OpenAI is spending so much money and like Sam's shaking the Nvidia CEO's hand, forgot his name, but there's just a lot of connections in that space. And I think the GPU rush, like it's like a gold rush of who can acquire GPUs, who can train on them and like who has the intelligence to actually put the GPUs to make meaningful like outputs in terms of AI models.

Breakeven Brothers (45:52)
Yeah. Another thing too, that we see a lot with companies when they have like a supply chain, like just think car manufacturers for the simplest example. You know, if they're buying like raw materials from a vendor and that vendor is like unstable or they just feel like they're paying too much, sometimes they'll either acquire the vendor or like acquire a supplier. And that way it's just part of their, you know, vertical integration. And so I wonder at some point, you know,

If it doesn't make sense, it doesn't make sense, but I'd just be curious if these GPUs in that processing as like a resource becomes so valuable and there's too few providers of it. Do these companies get involved in making in themselves and, you know, to kind of secure that pipeline, just something, something curious to think about, I think in the future.

Bradley Bernard (46:42)
Yeah, have a question for you How much do you think in video is worth today without looking it up assuming you haven't seen it before?

Breakeven Brothers (46:50)
Well I know it did it a stock split as 10 to one stock split but I think I think I think like $184 now it was

Bradley Bernard (46:54)
Okay, I'm not familiar with that, but what's your ballpark?

No, no, not stock price, but more market cap, total market cap.

Breakeven Brothers (47:04)
I don't know. I haven't looked at that. I have no idea. Trillion, right? It was in the trillions because at one point it was above Apple. I do know that. Yeah.

Bradley Bernard (47:08)
Okay, I'll get -

It was above Apple, so give me a trillion guests for your Nvidia.

Breakeven Brothers (47:17)
I'm gonna go Today I'm gonna go two trillion dollars

Bradley Bernard (47:21)
today.

They just crossed $3 trillion. So if you look at their stock, it's absolutely nuts. And I think it's because they're just providing all the GPUs that like until there's another provider that comes in and does what they do, they're on a great trajectory. I think a lot of people are getting into the space. Like I mentioned, like optimized chips just for LLMs that I think will beat out Nvidia in the long term, but they're not really proven yet. Like OpenAI isn't forking over hundreds of millions of dollars to like grok.

or these other companies that are spending hardware research on chips to see if that's going to pay off. But yeah, for the next year or two, I think Nvidia is the absolute king of GPUs.

Breakeven Brothers (48:06)
we went off on quite a tangent, but that was all in good fun. So that was cool. Well, I didn't want to bring it back a little bit to, I guess, accounting and finance use cases because we talked about like the precision and the temperature using that sliding scale to get it like super precise versus more abstract and how you can kind of tinker with it. But you had mentioned correctly that there's kind of a black box of like how it's doing the calcs a lot of times.

Bradley Bernard (48:09)
Mm -hmm.

Mm -hmm.

Breakeven Brothers (48:34)
And so, you know, we talked a little bit last time about automations and accounting with like Python and SQL. And I do feel like a more immediate use case. And I'm actually, I do want to, I'm probably going to post a video about this just on my own channels or something about pairing up like your own coding, whether it's like using a SQL and Python combo with like an LLM model to like summarize and like document.

In like an accounting specific way, because, you know, you can say like, Hey, summarize this and tell me what this program does. I think that's great for like a general sense, but a lot of times accountants, you know, there's a very certain structure that, you know, we need to follow just for like controls and compliance with those kinds of things. And so, you know, I think what I want to kind of experiment with, and like I said, maybe post that video about is, you know, doing your coding to get your automation in place, just like your script, your basic kind of script or SQL statement.

Bradley Bernard (49:09)
Mm -hmm.

Breakeven Brothers (49:30)
And then using, you know, whatever open AI Claude, whatever to like summarize it and then put into place like some specific documentation. So that way that becomes durable. And I think that's a really good use case for accounting too, like right away. I think there's still other things to be figured out with accounting. Like we talked about with like data privacy and reproducibility, but like the ability to like summarize and like document seems like that's just it's bread and butter at the moment.

Bradley Bernard (49:58)
Yeah, there's a lot of coding use cases that fall into that. Like, Hey, I've written code and I want to build something off of it. So there's writing tests. Engineers hate writing tests for their code. That's just ensuring it works as you wrote it. I think it's a great exercise to actually write them because as you write them, you'll figure out what's broken. It's a little bit of a weird situation to write the code, have the AI write the test and then see if the test pass. Cause you're not really checking too much. Like your AI test could be enforcing your broken code. If you think about it that way.

I think for accounting, it's the same thing. You could easily write a script, have AI look at it, improve it. There's a few different experiences right now, but one that I'd like to point out is the co -pilot experience. It's something that GitHub created and other companies have adopted, but essentially it's the suggestions as you type directly where you'd expect them. So for example, on my day -to -day coding, I'll have my code editor open.

I'll be writing some code in Laravel. And if I know what I want, but I don't know exactly how to get there in terms of code to write, I can write a comment and say like, Hey, generate me a migration and PHP in Laravel with these five database fields. after I write that comment, I can press a new line and then the co -pilot experience will load. Based on my comment and the context of the file that I'm looking at code that I can then just press enter and accept that suggestion. So I think that's like.

the biggest game changer in terms of developer productivity, at least in the IDEs, you can go around your code base, add a comment saying like, Hey, change this function instead of looping over it. How about we do like a while loop instead of a for loop or some certain performance optimizations or things that you don't really memorize as a programmer, but you know, it needs to be done. You can simply write a comment, have the AI look at it and it'll pretty much fill in like the solution nine times out of 10 with one caveat though.

You should look at it. it's not perfect and it makes a lot of mistakes. I've even pushed code on split my expenses where I'm doing a lot of this pretty fast. I'm like, Hey, generate me a migration done, accepted. Don't look back, deployed the code five minutes later. And exception is thrown in my error tracker saying, Hey, this doesn't work. And I go and look back. I'm like, this makes zero sense at all. Like it's not perfect, but most of the times it works. I think it's partly a skill on the developer.

to understand when and where that makes sense. It's kind of like prompt engineering with chat GPT. You need to understand where it makes sense, how powerful it is, and then when you get to the copilot experience, don't ask it to generate you 500 lines of code. Maybe go in there and edit a few things or a small function here or there. You really need to become better and more knowledgeable at what it can do for you for it to do well in your whole developer productivity.

Breakeven Brothers (52:41)
so I've done a couple of hobby projects and I have a GitHub it's private and rightfully so, cause I would never want to publish my own code. It's just, you know, it's just a mess, but it's, it's fun. Yeah. I'll share it with you. You know, about these projects, I'll, I'll name two of them. The first one and there's other ones that do not even like make the list here, but you brought up test driven deployment. So I was just going to kind of share a thought on that.

Bradley Bernard (52:51)
You can share with me, though.

Breakeven Brothers (53:05)
I built two hobby sites. One was called puck pass. And do you remember this at all? Did I talk to you about it at all? Yeah. So it was basically I used to play a lot of hockey, especially like in Southern California. There's so many hockey rinks and me and our mutual friend, Chris, you know, all the times we'd play and like, you're always like looking up different ranks to figure out like what has times. And like, you kind of just show up and hopefully it's not sold out. This was probably back in 2017, 2018.

Bradley Bernard (53:12)
I do. I do remember it.

Breakeven Brothers (53:34)
And so I was like, it'd be cool to like have like an online, like reserve system to like book the look up the ice times. First of all, like just be able to see them all one place based on like your geo location. And then also be able to like reserve an ice time just from your browser and not have to go in and call ahead or whatever. And so it builds it out. It came out, it actually came out pretty great. That's when I was like really coding a lot. And so it just came out naturally and, got it pretty far.

But I remember doing that and I was also following along. I wrote everything in the Django framework, which for those that don't know, it's like a Python based web framework. You know, Brad's been talking a lot about Laravel and that's a PHP based web framework. So Django is, you know, a Python based one. And so I was writing everything in Python, using the Django framework and I was doing it alongside a book that was using test driven deployment and also to

Docker, which I think is a good topic we should get into it for another podcast. but test driven deployment. And I remember being like, gosh, this is so thorough, but like it did actually work. Like it took me definitely longer, but like it worked along the way much easier comparing that to a more recent one. I did, you know, me, my wife and a couple of our friends were watching formula one, like everybody was during the pandemic. And we had this fun little game.

where like we kind of guess the top 10 and then basically whoever had the fewest places away from like their guesses to like actuals, like the lowest score won kind of thing. So anyways, long story short, I did not do any test driven deployment development at that time. And it was fun. I definitely moved a lot faster, but it, I got to a point where, and it's my GitHub reflects it. I just.

Bradley Bernard (55:07)
Mm -hmm.

Breakeven Brothers (55:21)
It got so messy and it was like so cluttered. I just ended up just being a total waste, but it was fun. I built it as a hobby. It wasn't like I was going to deploy it and charge money for it. But yeah, I remember being like, Ooh, like I don't understand what some of this code is doing. And this is all pre, pre LLM to kind of help me out. So yeah, it was the comparison of the two.

Bradley Bernard (55:40)
project. The hobby project is great. I'm one to call it side project, but I feel like hobby is a better word because half the side projects that I work on don't ever make it out the door. So it feels a little bit more relaxed with the hobby. one thing I wanted to bring up though, Claude has recently announced an artifacts feature. And so what this is, is basically runs a Python script right next to your AI chat.

So for example, if you ask Claude to generate a logo, so I did this recently of, Hey, I'm working on a side project or a hobby project. I want a logo. Give me something that looks like a light bulb mixed with like an app icon. And so what it'll do is it'll take your query and you're side by side that your left pane is your chat. Your right pane of your screen is like this artifact thing, which feels a bit nebulous, but it'll write Python code to generate an SVG. SVG is just like a vector image of like, Hey, a point exists from

You know, X zero to X five Y zero to Y 10 with this color. So it generates this Python script that generates an SVG at the end of that. And then runs this Python script and then gives you the output of that. So I think for accounting, it would be super cool if that code interpreter, what they called on open AI side. So they released this probably six months ago is like, Hey, if you ask open AI, like you upload a document saying like summarize this or.

Give me like, you know, kind of the Excel version of AI and accounting where like some of these columns are like, just give me like a general read on this spreadsheet. what it'll do is it'll throw that into Python run, like generate a Python script and then run it. And so that has been extremely valuable, but it hasn't really been easy to access as a developer because when they run this Python code, it has to be in a very strict environment. For example, if you're.

On open AI servers running your own Python code, like if I wrote my own Python code and then ran it, it could like try to hack into their system, fetch all their data, whatever. So their system needs to generate this Python code, run it in a very like sandbox environment, and then return you that output. And so Claude released their version of code interpreter is what they call it. On open AI side and Claude's is like artifacts, I think. And so it's super cool. And I think accounting can be sped up that way, but again, it comes to the like.

black box sort of approach where you ask a question, it might generate you a Python script, but it might do code, like write code that you don't really understand. Like it's using pandas, like maybe you understand it, but it's doing complex stuff. And so it's like, you can get the solution that you want, but are you confident it's correct? Are you confident it can happen again, like reproducibility and like, is it easier to understand or is it easier to do it yourself? Because like talking about automations again, there's plenty of times where I had fun building an automation, but it

was purely a waste of time. If you look at it at the end of my 10 hour endeavor, or I just spent 10 minutes, you know, every day for a month. So it's like, I can see code interpreter and like running Python code being the future of like an interface to do more things. Like, cause it's spelling out the logic for how the AI is getting that response. And so if companies spent more time there, like maybe leaning on the code interpreter more just to flesh out like how things are, you know, being from input to output.

how is this being generated? I think that would be a really cool evolution just to get more transparency into what's happening under the hood.

Breakeven Brothers (58:59)
Yeah. Well, transparency is a big thing for sure. I think especially again, with accounting, whenever you're dealing with numbers, just being able to have a clear path of like how you got to this point is super important.

Bradley Bernard (59:11)
I want to talk a little bit about AI in our day to day jobs. I've talked, you know, talked about the co pilot experience for developers, but for you in accounting, what has been your interface into AI, if any, in your day to day job?

Breakeven Brothers (59:26)
I would say not a whole lot right now for those reasons that, you know, we talked about with the data privacy reproducibility and just general, you know, murkiness of like how it works. I think there's definitely an appetite to want to use it more. One thing that I plan on doing, this isn't really for my day to day job, but just to augment or to kind of provide some additional content. You know, I'm building out an automation right now.

Bradley Bernard (59:37)
Mm -hmm.

Breakeven Brothers (59:55)
And a lot of times the automations within accounting, again, we're not going to be doing like, you know, multivariable calculus. We're usually exporting data, transforming it or manipulating it, and then importing it into a different system. So one thing I'm trying to use it for though, is to just do some shortcuts on the coding side of it. And usually that comes down to like the simpler logic of like Python has a library. I think it's called XLSX writer, which is like the

file format for Excel files is dot. Yeah. So basically it's a library that lets you work with Excel workbooks. And it's really where you do like the, all the formatting, you know, if I want to format a page, I can say, Hey, header row needs to be blue and here's all the content. So having to understand what that library like does first is super important, but using like an L and chat GPT or a Claude, like I'm able to like,

Bradley Bernard (1:00:26)
for Excel? Okay.

Breakeven Brothers (1:00:50)
Get it there started. And then I can use that LLM to like refine it a little bit and just be like, I want to get this header. And the next two rows, like offsetting colors, can you give me a script that does that? And it'll kind of put it all together. So that's been helpful. so really it's just kind of augment some productivity, but not so much. Like as a full means to an end of like end to end that this is these agents are doing all of it for me. It's not there yet. I think there's, yeah, want,

Bradley Bernard (1:00:55)
Mm -hmm.

That's the dream. That's such the dream. my goodness. Like I think everyone sees that when they talk about AI agents is like, that's how I'm going to be replaced as these agents will have a set of tasks, set of skills and just like use that together. So like I need to build a web app. Maybe the agent knows how to create a database, you know, deploy a PHP website, buy a domain, like set up DNS. And you can imagine you have a general task of to build a website. The AI agent will

Look at all these little things, piece it together, run this, you know, kind of skill and get the output, feed it into the next one. And it's a little bit of a scary world, but the reality is since one individual task matters so much, like it can fail on buying a domain, for example. And then the whole thing is broken or like building a website. There's a lot of complexity there. and getting like each task to work a hundred percent and stringing that all together, like the failure rate for each individual task is probably, I don't know.

20%. And then if you times it together by 50 tasks, like you're very, very unlikely to get a full run through to be accurate. And as intelligence increases, we get a little bit closer. but again, I think we're pretty far away from that. And the general AI use cases are like very specific to one thing and less so about the full agents, but some companies are kind of some cool stuff out there. So I don't want to say it's not here. It's just not fully battle tested yet.

and the things I wanted to mention for AI in my day to day, so I already mentioned co -pilot. So something that can easily help me write code directly in my editor. The second thing is using Claude this week. I actually optimized some database queries. So on split my expenses, there's a bit of a slow query when you load it up a certain page, I took that query, dumped it into Claude and said, Hey, how do I optimize this? And along with the query, the raw SQL query, I dumped in like four or five different.

database table structures. So on my SQL, if you say, I think show create table and then dumps out the raw table definition, you can copy that into, into clawed and say like, here's my query, here's my tables. Here's what I want to do, like go. And so the context windows of how much information you can provide in the AI systems has really increased over time. So it allows me to dump in like as much as in my head or as much as you need, and the AI system can get going. And so it.

suggested a whole bunch of different query operations. I think I was doing like an exist query, then it suggested an inner join for performance. And I won't go into the details, but I went back and forth with Claude like four or five times. I said, Hey, is this performance? And it said, I don't know, depending on how big your database table is. And I said, how can I figure that out? Here's the query to run. Ran that query, brought the results back into Claude and said, based on that information you provided me, like it looks pretty efficient. It's not looking at all the rows in your database table. It has a high filter, yada yada.

So that was really cool because I think Claude has really shown itself in its latest like model of 3 .5 Sonnet to be really really good at understanding code and being really good at like Narrowing in on the objective where some AI systems you ask it to do one thing and it just doesn't follow instructions that well I think Anthropic who makes Claude has done a really excellent job at killing the layers back behind these LLMs and having really good researchers to

understand like what's happening behind the scenes and then they build these models in a way that you tell it to do something and it like knows the next step before you know it will ask you things you feed it back information like continues along that path to make you really productive and successful without like again, we talked about Working with open AI and chat GPT. You really need to figure things out I think with anthropic they do a better job at leading you through it like from start to finish and so That was excellent. So I improved my query time my page loads faster

Updated my Laravel code with Claude's like results and it was pretty cool I just felt like pair programming with somebody a little bit manual to go like go back and forth between the tools and figure that out but at the end of the day like I got what I needed and I paid like a like less than a dollar probably like 50 cents and Yeah, it's super cool. So I was stoked about that

Breakeven Brothers (1:05:20)
Do you think that the LLM was like rookie, you know, as you're, you know, like, you don't know this because you're talking about like you had a pair of programmer, you know, it's just like.

Bradley Bernard (1:05:24)
You

I mean, they don't have any judgment or no sentience yet. think it was hating on me, but I do like to approach it from like a very beginner point of view where I don't assume I know anything. I just kind of dump it in and say like, hey, could you help out? And then I let it kind of take the reins and I'm kind of like the supporter in it's like, you know, main act. And so, I mean, maybe it thought, you know, this is pretty obvious, but to me,

Breakeven Brothers (1:05:34)
Sure, sure.

Mm -hmm.

Bradley Bernard (1:05:55)
I had spent time optimizing it and I didn't like my first thought now is to go to AI because one, I want to see how they work and get better at them. And two, it just has a ton of knowledge. Like at the end of the day, if it doesn't help me, I just disregard it and I paid 20 cents, whatever. So yeah, hopefully it didn't, it didn't hate me. I'll be coming back to Claude and I, I assume to be treated fairly. So we'll see, I'll keep you updated.

Breakeven Brothers (1:06:16)
Sure. Cool. Awesome. Well, let's wrap this up, Brad. I know usually we kind of end with our bookmarks or like saved articles. So why don't you kind of kick us off?

Bradley Bernard (1:06:27)
Yeah, so my saved articles on Twitter again, this one was shared by username KwindLA. Essentially, this person had linked out to an AI voice interface that is like quote unquote the fastest response time. So usually when you're using Siri or other like Alexa systems, sorry if that triggered anything, you will then mention the name.

type, like, you know, say something after like your query and then it then goes processes it and then gives you the response back. So very traditional mechanism. This app is exactly the same, but it takes about 500 milliseconds from you stopping speaking to it reading a response back out to you. So in that 500 milliseconds, it takes the audio that you've spoken, transcribes it, throws it to an LLM, the LLM responds and then gives you back a response and

Like 500 milliseconds, I'll send it to you after the show, but it's actually insane to feel that responsiveness of like, it feels like you're talking to somebody. So there's a few AI tools and services out there that are automating like the call center space. And if you see their demos of like their real calls, a ton of them will have an example of, Hey, I'm talking to Ben. I I'm robo dialing him because I'm the AI and you're the real person. You'll ask me a question saying, I'm not sure. Like what's my insurance policy on this?

There's a large gap of like two to three seconds. AI is not saying anything. And then it comes back saying, the insurance policy is this. And with this thing, it's crazy cool to see like 500 milliseconds. You can just have a natural conversation. It fetches the information, gives you the fact that you want, and then you're moving on. Like it's, it's almost like a bit too fast. Like I need to be ready for a conversation where these other AI is like, I'm already thinking about the next thing that I'll ask it based on like, you know, it's processing time. So.

I'll link it in the show notes, but I think it uses Llama 3 under the hood and then uses DeepGram AI, which is speech to text, and then uses Cerebrum AI, which I guess is a serverless GPU infrastructure. So pretty cool. It's like optimized WebRTC. WebRTC is like very fast communication of like audio and video calls. So a super fast AI chat that's very accessible. And I think if you're interested, I'll drop in the show notes and people can check it out.

Breakeven Brothers (1:08:44)
That's really cool. like Alexa, Siri, all those kinds of things. I think those are definitely going to be more at play in the future. Cause just seeing my kids use it one, like it's just so natural. And even then, like you said, like those current systems have like this delay and it's not like a real conversation, but I think for like business owners, small, small to medium business sized business owners.

Bradley Bernard (1:08:46)
Mm -hmm.

Mm -hmm.

Breakeven Brothers (1:09:07)
Like being able to like talk to it and be like, Hey, how was my sales today? And it tells you, and it's like, okay, well, like, you know, what's my average price that can tell you? Like, I think that's that kind of like assistant is going to be a nice, I think future state that we'll get to. And then also too, you mentioned, and I think it's really worth shouting out as well, like accessibility, you know, having a speech, a text option, you know, is great for people that might need that. So the fact that that can kind of get more and more naturalized and not feel like it's like.

I talk to humans this way and I talk to Alexa this way. It's just all the same. I think that'll be a nice development when it comes out or when it's really fully deployed.

Bradley Bernard (1:09:45)
Yeah, I agree. I even thought about prototyping some split my expenses audio interfaces, whereas you split a bill and you need to type out eight people's names. It's like, I could just say like Ben, Chris, Tyler, Alex, whatever. And it just like finds your friends and does it. but I haven't got there yet. So it, there's a lot of opportunity in the voice space, but there are some things that it's not good at. but we can save that for later.

Breakeven Brothers (1:10:08)
Yeah. Cool. Mine, my interesting article is actually a YouTube podcast or a podcast episode. And there's a, there's a clip of the podcast. I'd recommend the whole podcast cause it was just the whole thing was amazing. But if I had to share one clip, it is from the podcast called the fort and entrepreneurship podcasts. And I think it's hosted by like a real estate professional. It's really like a real estate podcast.

The reason why I think it's a great episode and a great listen and it's my article to share for the week is it is an interview with a man named Richard Fertig Hopefully I'm pronouncing it right. He is basically like an ex hedge fund manager, worked on Wall Street, something to that effect, but he left to do short -term rentals like Airbnb's, but he did it. His properties are very unique and

On the podcast episode, he talked about with the host. I wish I remember his name. I think it's Chris Fort, something like that, but he was talking about it with the host saying, you know, when you're in the short -term rental space, what he found was like, your differentiator was like the experience. Like you walk in and you've been to nice hotels where you walk in, the air smells good. Everyone's smiling. The place is clean and it's, it's an experience. It's not just a place that you're staying at.

And so he was trying to kind of recreate that with short -term rentals. And he's got a couple of really big properties that are very catered towards unique experiences. And they might be like, Ben, why are you talking about this? It's, this isn't a real estate podcast. It's not a hospitality podcast, but yeah. But what I would say is it's a super interesting listen, because I think you can apply what he's saying. So it's anything you're doing really, I think, you know, for me, working in an accounting department at a company.

Bradley Bernard (1:11:43)
You read my mind.

Breakeven Brothers (1:11:57)
You know, I can can that as I'm going to make an experience for people I work with, right? Like when they work with Bennett Bernard hopefully they have a certain experience and you know, nobody's perfect. But I think if you kind of keep that in mind and try and say like, you know, what do you want to be? Do you want to make, like, do you want to be the Motel six where you walk in, it smells like cigarette smoke and you know, there's holes in the bed sheets and all, and who knows what, or do you want to be the montage in Beverly Hills or Laguna beach where you walk in and it's like you walk into a different place.

And I think about that from like a client service angle, like for accounting firms, like you're not just providing a service, like you should be trying to create an experiment experience, you know, for split, split my expenses, you know, you're providing a service, but like you want people to feel like they're being taken care of, like from end to end with your system, right? It's not just a transactional, you know, I pay you, you do this for me and like, see you later. And so I think it's, it's a great podcast episode again, Richard.

Bradley Bernard (1:12:44)
Mm -hmm.

Yeah.

Mm -hmm.

Breakeven Brothers (1:12:54)
Fertig is the gentleman's name who's being interviewed that talked about this and he explains it in such a great way. I remember doing laundry, like folding clothes, listening to it and being like, Holy smokes, this guy's just spitting fire. Cause it was like the way that he like talked about it was like so applicable to any, anything you could be doing. So one thing I try and do keep in front of my mind when I'm working or doing other things is, you know, what experience, what experience am I giving the person that I'm dealing with? Or, you know, if I was ever working with a client, you know, I used to be an auditor, right?

You know, if I could go back as an auditor and say like, what experience experience am I giving with this person or this, this, these people that I'm working with, it's just such a helpful frame of mind, I think, to go into, because you're kind of thinking about it. Like everyone's felt that, you know, everyone's felt where you walk into a nice hotel versus like a crappy hotel and like, it's just such a difference. So I'll link the full YouTube, podcasts and our show notes, but if you can only do one part, there's also a clip of it.

Bradley Bernard (1:13:41)
you

Breakeven Brothers (1:13:49)
where he talks specifically about kind of creating experiences. And so I'll link that as well, if you want just the abridged version.

Bradley Bernard (1:13:50)
All right.

Awesome, awesome. Cool. Well, I think that's a wrap on our second episode. We have gotten all of our social profiles set up. We have our website at breakevenbrothers .com. We have our YouTube channel at youtube .com slash breakevenbrothers. We have our Twitter. Spotify. our website has a link to every platform we're on.

See you next time