Episode #21 – Practical Automation with AI

And welcome back to the Big Cheese AI podcast. We're back, baby. More importantly, Sean's back. My name is DeAndre Herakis, joined by Jacob Wise, Brendan Corbin, and Sean Heiss on the world's second best moderator, self-proclaimed. And today, we're talking about practical automations with AI. We got a lot to say about that. So first of all, Sean, we missed you. Oh, thanks. I thought I got fired, but I get it. They started texting me about 2.05 PM today. Where are you? You know, and I just had to just-- I guess, well, I'm in the parking lot. So I showed up and we're ready to go. Yeah, so in this episode, we dive into how AI is transforming mundane tasks into automated wonders. I figured that was going to say workers. Making life easier and sparking debates over convenience versus privacy. So that was the show notes that you just read, right? Which is fully automated. Right. So our podcast now is pretty much like 95% of it is fully automated. So right away now, we have the ability to go in. We have a script that we can run. And it's be like, hey, what's the topic of this week's show? And you can either be like, here's the topic, or you can be like, I don't know. And then it's going to go and it's going to try to figure-- if you say, I don't have any idea, it's going to go look at all of our past shows and try to come up with what the next show should be about. And then we'll go through and ask, do you have any guests? Do you have anything like this? And then it just-- and then it will go and actually look at all the news and product launches. What did you build that with? So that one's all just running like-- it's like a Bonn JavaScript function or call, I think. I think I did it with Bonn. I don't know. I switch between Bonn Node and Dino all the time. It hits OpenAPI. Yeah, it hits OpenAI. Say that five times fast. Yeah, yeah. Please. It hits OpenAI to basically provide all the information. Then it just writes it all out. We take that, and there's our show note-- or then there's our show-- they're not really the show notes. Show notes are what we actually send out at the end, but are internal show notes, I guess. Then even at the end, we process all of it. So we have this-- we've got a pretty good smooth process going now, right? Where it gets uploaded to-- now, I guess it's Google Drive is where it's going to be stored. I download that, and I run another script. It extracts a transcript. It does the show notes. It does the artwork generation. I mean, so I'd say almost like 95% of this entire process now is fully automated through AI. How much-- so I think some people underestimate how fast and how powerful the transcription stuff is. Yeah. Like, if you have an audio file or a video-- let's say a video file 60 minutes long, and you send it to OpenAI, how much does it cost, and how long does it take? Well, so I actually don't use-- so I use Whisper, which is OpenAI, but it's just running locally on my Mac. And that usually takes about, I don't know, five minutes probably to transcribe the hour-long podcast. And then, I mean, the total time is probably about 10 minutes it probably takes to do the transcription. Then we take the transcript. We use that to generate the show notes-- or yeah, to generate the show notes, to generate the artwork using DALI, generating the YouTube hero using DALI. So it's about 10 minutes probably total for the whole thing to run. And so whenever I think about automations, and just the general public when it comes to AI, I go to, oh, probably GPTs. Wasn't that what GPTs were supposed to do or streamline? Like, how do you see GPTs playing into this idea of automating processes and things like that? And that's what the majority of what mine were. So we all kind of came up with two or three different ideas. And for me, just in terms of thinking through, how am I practically using it, things that people can just use today without having to set up a bunch of Node scripts and bullshit that I'm doing, it's the GPTs. So the blog post, I have-- any time that you want Chat GPT to build you a blog post, you should probably create it as a GPT, because you can go in and specify the way you want it to write. Like, if you just go in and be like, here's a block of text. Write me a blog post. It's just going to go, and it's going to do its best guess. But if you do it as a GPT, you can put in other-- I want you to be a narrative blog post style, writing style. I want you to-- we are this kind of company. We do these types of things. This is the type-- so you can basically just put in all the rules for it. For me, every client that I work with, I have a GPT that's very specific to writing a blog post in their style, in their voice, in their tone, that I could take anything that they send me, paste it in there, and it's going to turn it into a blog post that's actually pretty good. Same thing with the image generation. Like, if you just go to Dolly right now, and you're like, generate me a picture of a chicken running across the road, it's going to be that-- like, the same look that everything that comes out of Dolly is. It's like that kind of hazy, modern, like, AI-looking art generation. Instead, do a GPT where you really ground it on what your style is. So here's our brand colors. So for Prompt Privacy, I've been doing that a lot, because I'm doing a lot of blog posts with them. And it's like, I need to come up with our own unique style. So I wanted, like, 1960s illustration style with a dash of modern. And so everything that I put in there now comes out with these robots that look like they are from, like, 1960 illustrations. It's actually kind of cool. But those are all GPTs now that I just set up that that's where I just go to. And then things with, like, make.com, which is what kind of kicked off this whole thing, because I shared a TikTok to Jacob. It was crazy. Yeah, with make.com, where you can go through, and you can basically set up all of these automations. I mean, we're talking, like, thousands of different integrations. You know, everything from, like, S3 to Notion to Dropbox to anything can be, like, this basic trigger. But this guy built this whole thing that was, like, doing email replies or whatnot. I forget how I even got on that track. Yeah, well, what you reminded me of is, like, for the show notes and the automations we've set up to help with us on the podcast, we had a problem that was recurring that was a good use case for automation using AI. It was a generative problem. We needed to take an asset, a video, transcribe it, then summarize it. That's, like, a bread and butter of automating using AI. Now, when I started thinking about what things do I automate in my daily life or in my work life for this show in preparation, I don't really automate that much or anything. I couldn't really come up with a concrete, like, oh, this is a thing I fully automate in my life. And that's because I just don't have or haven't thought of the use cases that cause me enough pain to automate that, right? So I think a lot of people, like, initially, when they think about AI and automation, they're always, like, reaching for a solution to a problem that doesn't exist, right? So I think that's good to keep in mind of, like, you know, it's have the problem before you reach for the solution. But I will share what I do use AI a lot for. And, like, a quasi, like, automation is, like, any time I'm starting a new project or a new venture that I don't really know that much about, trying to build a fence at my house, my fence is a piece of shit, time to build a new one. And I'm like, well, I'm not going to pay some expert to do that. I'm going to learn how to do it myself. So my process is, first, I do some light Redditing or Googling and kind of get the base lay of the land. And then I take that to, like, chat GPT or Gemini. And I say, OK, here's where I'm at. Here's what I'm trying to accomplish. I know I want a cedar fence. I want a top post. I want-- or whatever plate, seal plate. X, Y, and Z. And give me a parts list. Tell me, like, how should I go about planning the project. It gives me a whole outline of the project plan, who I called the lumber company and sounded intelligent on the phone. Like, I know what I'm talking about. Yeah, and that took a process that used to take a lot of time of reading a dozen articles and trying to piece together what was common between them and for my use case and what didn't apply to my use case kind of thing. So that's what I love it for. It's not exactly an automation, but it's sort of automating, I guess, Googling for me. Yeah, no, I think-- so it is an automation. And that's a piece of-- and I struggled with it when we kind of came up with this idea for the show, was I kept thinking of that one that we just sent. There was like 15 different steps that this guy does, right? New email comes in. Go and read the email, process it, send it to a GPT, have it analyzed, figure out what information it needs to pull in, all of these different steps. But no, I mean, again, a single GPT that just takes some text and transforms it into something else is a process. So-- I think it kind of brings up an actually interesting thought. What Jacob mentioned is a lot of people, their job isn't necessarily the same every day. And then there's different problems. So for the folks that have those repeatable problems that they're solving over and over, it's so much easier to say, oh, well, you know, I should probably use AI to do this. And then eventually, those things might take less time, or they might not have to do them at all anymore. But I think for a lot of people, their days are-- it's like a lot of meetings. It's a lot of kind of trying to figure out how to creatively solve problems that weren't on your desk the day before. I do think it's kind of an interesting exercise for people to maybe start actually inventorying how they could make their days more process-driven and more repeatable so that they could utilize AI. Like I was sitting there going, you know what? I probably could create an AI that gives me a report on how many hours each person at my company billed yesterday. Right. And have the three systems talk to each other and send me something. That's a great idea. And so maybe thinking about something that I might be able to do. And I think the second thing I think is tough is people have a hard time going, OK, I'm going to use a large language model to summarize a bunch of data or to convert something into, I'm going to actually take this, connect to a system, get an output, go in as an input to the next process, and another output, and have a three-step, three-system layer. And I think there's such a small percentage of people that could even mentally model that without AI. Yeah. But I think in some of the deals Jacob and I are working on at work right now, the next step with some of this stuff is to use a middle-tier LLM, but actually connect to systems and actually have that. And you know what? It probably isn't going to be something where you go and buy niche SaaS product $30 a month, niche SaaS product $30 a month. It's probably going to be the make.com or the Zapier. I don't know. Yeah. I agree. And if you're a business that's struggling of how to automate, maybe you don't have a developer or an IT person, I think that, first of all, lots of companies that didn't have technical staff are going to start getting them, and they're going to start leveraging these types of tools to automate some of their tasks internally. Because if you haven't tried them out before, the Zapier or Zapier-- I never know how to say that one. It's Zapier or Zapier. It's a Zap. Yeah, I know. It's a Zap. It's a Zap. It's got to be Zap. It's got to be. But I never called it that. Yeah. Yeah, Zapier, and then you've got make, and then we use one that's a little bit more code oriented. But those are two building block tools that you're supposed to say, here's a trigger, a thing that happens. Maybe it's a pipe drive or HubSpot. A contact is created. Now what? Now, based off of that, do something else. Maybe it's a text message to them. Or the possibilities are pretty endless. And those tools, while not completely approachable to a non-technical person, are much more approachable than-- If I was a business owner, and I maybe had, let's say, a decently diverse demographic of folks working for me, I would put my just a fresh out of college person in a rotation, have them learn the business. Usually they would just be like your Excel wizard, right? But I would put them on that stuff. Like, hey, what systems do we have that are AI enabled? What should we do? How can we process these? And have them start setting up those types of things. Because I don't think it can be-- even though it seems easy, it's really not. No, you have to understand API concept. You know what I mean? It's something that-- and from your experience, you're always trying to connect different systems and leverage automation. Yeah. Well, that's what I was going to say, is that like-- OK, so a couple of things on this topic. One, Jacob said both points together, which at first, I was a little bit on the opposite side. I thought, well, now I can see them both. My original thought was you can't build-- from a startup perspective, you're not in all the way established company. You're wanting to build new processes. OK, great, you want to automate them. Don't just sprint to automate an entire system. And then if you do have the eye for it, link all these things together. Turn it on and realize it was never going to work, even if it was manual. So why did you just spend all this time automating it? Test it and actually implement something that works, and then come back to it and then automate. You could automate your sales process once you figured out how to sell your thing. But I think that's the key, is that you have to understand the process before you automate. That's right. Yeah, that's right. Because a lot of people will run to it. So a great example-- I don't know if you guys saw it. I think I made a comment on it on LinkedIn. And I think there might be from India, but it was a company who built an AI recruiter to interview folks. And so it was India. They had the headdress on. The guy, whatever. It might have been-- I don't know. It could have been Dubai or whatever. But what was fascinating is it was basically the way that the company who built this made the argument. It was like, we're just making it so recruiters can focus on the things that they're good at and not the things that they're bad at. And all the recruiters came in and be like, actually talking to humans is the one thing that we're good at. Why would you go and try to automate that when there's all of this other stuff that we suck at that should be automated and not the actual interview process? Because again, it was clunky. It was weird. Like, you're talking to an AI, and it's like, I-- [MAKES AI NOISES] [LAUGHTER] And it was just-- it was stupid. But the reality is don't try to automate the things that humans are good at. Automate the things that humans are bad at. Yeah, I think that's what Sean was saying, and I completely agree. And that seems like it was a top-down automation. OK, we're at the C-suite. We're just in a meeting. We say, oh, we should probably automate this whole process. Not once considering the people actually doing the job. Because if they would have probably had, I don't know, one conversation with them, they would have realized that that's probably not the area to automate. But either bringing in a group, a consultant, or your 21-year-old new hire to come in with this view of AI and technology and say, hey, look. I've been now doing this job for three weeks. And I think that these are the three or four areas-- before I flip my computer over. These are three or four areas-- I'm excited. --that we could automate using AI. And it's looking at those high-level processes and connective, aka, tissue of the organization, how data is transferring from Notion to HubSpot down to an email. And then the person didn't show up for the meetings. Then you have to drop them into drip marketing. You have to be able to see it and actually probably be in it to truly implement the AI. But at the same time, that transfers over to this next part of-- Sean, your side on how you guys-- Well, I got a quick story first about automation. So I'm leaving on a road trip right after this. And last week, I always take my wife's car to get looked at before we take the kids on a road trip. You've got three kids, right? You don't want to have a breakdown on the side of the road. And so I go, and I'm like, hmm, how should I do this? OK, I'm busy. I'm kind of like-- I'm just sitting in the parking lot. I was literally sitting in a parking lot, like in between lunches or something. I'm like, I've got to get this scheduled. I go, and I Google my dealership, right? And I get-- there's that website, and I call. And I'm like, OK. And it's like, oh, schedule an appointment. And I was like, I don't want to do that. I just want to call. OK, I called. OK, so I get on there. It rings forever. And then it's like, hey, no one's here. And it's like, would you like to schedule an appointment for service? I was like, OK, sure. I click one. And then this AI person gets on the phone, which is-- you've seen that before. Like, doot, doot, doot, doot, doot. But this was a different AI. It was trying to-- it almost felt like it was a real person. And they were like-- and I was like, this is what I-- this is when I'd like to take it in. And they were like, what service would you like? And I was like, I just need a checkup. And they were like, oh, you'd like to have your vehicle regular maintenance provided. And I was like, yeah. And they were like, do you want anything else? I was like, tire rotation. And they were like-- and it just sat there for like 30 seconds. And then I was like, we don't know how to do that. And I was like, OK, I just want regular service. And they're like, that's not in our system. No. And then I was like-- they were like, would you like to be changed to a person? And so it transferred me to a person. Of course, I sat on hold for like 20 seconds. I got on the phone with the person. And the person was like, oh, would you like the road trip special? We do this, this, this, and this. And I was like, well, hell yeah. I want the road trip special. And I got the discount. You know what I mean? And it took literally 25 seconds. And I got the appointment instead of what the AI was telling me was sometime next week. It got me in the next morning. Yeah. Yeah, that's interesting. Because first of all, that's probably an issue with the data in their system, right? They probably don't have a road trip special in their system. And that's just from that person's experience. Or they're just really good at sales, because they clipped me for like $1,400. [LAUGHTER] Yeah, that's why this is called the road trip special. It's a road trip. This guy's dumb. You're going to need four new tires. He couldn't get past the AI. Double charge him. Are you driving a cross country? Yeah, we've actually got the cross country special. [LAUGHTER] What are you guys doing? We're driving to Hilton Head Island in South Carolina. Nice. You're going to go golfing a little bit? I actually might bring my clubs. I might bring my daughter's clubs, too. One of the nicest. It's one of the-- Oh, yeah. I think the tournament is either next weekend or the Heritage. It's at-- yeah, there's a really nice course there. They should have never let me on there. I created so many divots. Oh, you got golf clubs there? I should have. Oh, yeah, I remember that. Yeah, I should have never been allowed. Was that for the wedding? No, that was for my buddy's-- parents have a place out there or something. I just went out right after college, and I hacked it all to hell. We're driving to Asheville. We're staying in Asheville tonight. Nice. Asheville's a nice town in North Carolina. Oh, yeah. It's a good place. It's awesome. It's where the Biltmore Mansion is. Highly creative community out there. Yeah. That was the one area that I've always-- I was going to say, you should-- I would like to go out there. Yeah, I think it's got more breweries, more micro breweries than restaurants or something like that. You like mountain biking and drinking beer? You're right. Go there. Hairy armpits, it's perfect. Yeah. Yeah. Womp, womp, womp. So, Sean, when it comes to scaffolding out and the fun and components-- [LAUGHTER] We love you, Asheville. So-- Actually, by the way, in Asheville, I think I've been there. They have this really cool-- it's called a ghost ride, or a haunted bus tour. And I think it's in Asheville, in the little downtown area. If you've got your kids, it's so much fun. Oh, nice. There's actors, and they're joking, and they're actually out on the grounds, and these old-- I think it is the mansion. Yeah, so I think it's Asheville. Anyways, scaffolding, front end-- Oh, yeah. What do you want to talk about? I want to talk about-- they asked me about automation. So there's a couple of things that I use, or have been using. I'm involved in a project right now that is a large-- and I did this two years ago with a different project. And it's one of those things where I've done it so many times that I end up tackling this. And it's basically taking an older front end project that may be two or three years old, in terms of its versioning, and its underlying infrastructure, and fully upgrading it to everything new. And that's upgrading all the underlying software, upgrading the way we do state management data, components, everything. Whoa. I know. And so I did it, and we implemented kind of an atom molecule organism level component architecture in one of the applications. But the other application was the administrative application, and that wasn't upgraded. And one of the things-- I'll skip to the second thing. The first thing that I've found that ChatGPT is really, really good at-- and it's kind of funny. I was wondering if anybody was going to talk about something that they use that isn't ChatGPT. And so far, it's no. Well, kind of no. Yeah, so my note transcriber, which I'll talk about a little bit, I do use Olama for. But we can get into that. And I think the proof is in the product market fit putting there, which is-- Kate was at our office was talking about how she introduced her parents to ChatGPT, and they're world travelers. And now they're using ChatGPT all the time to set up all of their travel plans. It's just good. Right. It's just good. And so I think the moral of this episode, so far, that I was looking at, I was like, man, ChatGPT is still killing it. But anyways, the one thing that it does a really good job of is taking a working piece of code and converting it to something else. So I'm like, OK, this is an old-- so this was a very complex use case. And for the front end developers out there, they might understand or appreciate this. This component was a class component, React, that used Redux and get initial props. And so-- Yeah. Right? So you're taking data from the server. You're passing it down to the client, storing it in Redux, right? And you're using a class component to do it. Right. And I just said, hey, ChatGPT, I'm upgrading this to Next 13. I want you to make this-- I didn't actually do React server components. I'm going to take that. I'm going to do that next. But because-- so it's not the coolest thing ever. But anyways, I said, hey, I want you to use hooks for everything. I want you to upgrade it to be a functional component. And I want you to make sure that you don't miss anything. Yeah. Because it'll just give you a little recipe of how to do it, or it'll convert 20% of the code and be like, and the rest of it goes here. Yeah, dot, dot, dot, dot, dot, dot, dot, dot, dot, dot. And I'm like, oh, OK, not today, lazy asshole. You're doing the whole fucking thing. Yeah. I don't know how many times I'm like, finish it, you ass. I know. So that's one thing that ChatGPT does when it writes code, is it just won't do the complete thing, just because I think it's probably trained, hey, you don't need to spend all this money on this guy. He's only paying $20 a month. But if you tell it just to do the whole thing, it will. Anyways, the results were really, really impressive. And I do the same thing where recently-- and basically, it'll convert it for you. And think about that at scale. You could pretty much have it do a whole project. And people have done that traditionally with code mods, certain things. And those are using very complex code regex and other comparison tools. But another thing that I had to do is write infrastructure as code. So a lot of platforms will write code for GitHub Actions, which is like the very-- it's like, oh, we have to have a GitHub Action. They'll check that box. But Bitbucket Pipelines is-- we use Bitbucket a lot. And so it'll literally one-to-one translate a GitHub Action to a Bitbucket Pipeline with 100% accuracy. That's amazing. I know. And it's like, it doesn't-- so step back. ChatGPT 3.5 wouldn't have had that level of context, right? I mean, right? No, no. 3.5, I mean, it was good, transformational when it came out, but compared to like 4, there's not even a comparison. No. Funny enough, I think-- so Sam on the Lex Friedman podcast was saying that he thinks that the 4 to 5 jump will be just as dramatic as the 3.5 to 4 jump. So for a coder that's using modern, modern stuff, the big reason why ChatGPT 3.5 and 4 is different is just the ability for it to kind of stay up to date with things. So you can even tell it, hey, if you're unsure, go check Bing. And Bing means go check the internet for the most part, right? And I don't know if that's-- I'm just lucky because 4.0 came out at a certain time, or if it's actually staying up to date and updating its model with new information. But I mean, it knows things about very recently released pieces of software. The other thing that I think is cool to do is if you are writing code and you're building out a-- let's say a pretty complex page where you have data flows, you have a UI, and you are taking a design and building the UI. So you're just looking at it, and you're just-- right? Go give chat-- go start two or three prompts, and go give ChatGPT the hard JavaScript front end stuff. Be like, hey, I need this to-- I need a dropdown list that auto filters actually locally, not from the server. And then, hey, I've got this accordion component that I need to build out. And hey, go build this slider. And then give it your base tail-- say I want it in tailwind, give it your base code. It'll go out, and it'll take a second. It'll take a good two minutes per, because it's sitting there writing the code. And you can just come back to it while you're doing your other stuff. Come back to it. Be like, eh, you messed that up, messed that up, messed that up. And then give two more minutes. And the next thing you know, you're like, OK, boom, pop that in, pop that in, pop that in, and you're done. It just works. Yeah. So I think it's more of like a-- give it mid-level coding assistance tasks, not just one-on-one stuff. And I think that also has, for me, fully replaced any of these other tools, like Builder or zv0. It just comes back down to ChatGPT for me. And I think that it's just giving it a text prompt and telling it what you want, even if you're telling it framework-specific stuff, it's still just as good. You know, it's funny. I thought the same thing the other day about Builder. Builder.io is a tool we've used to take a Figma design file and automate that process into actual code with Tailwind and all the markup that you need. Really, really good at doing that if you have really good design that's achievable. And there's just a lot of-- we just don't do that high of a volume of design to code. You know, like that's just a small portion of our business. A lot of it is in the functionality and DevOps, all that kind of stuff. So it's like that specific tool would be better for like a-- if your business is, that's all we do. You give me Figma, I give you HTML, CSS, and then you go on your happy way and have your devs complete the hard stuff. The funny thing is people think that that's all there is to application development. Yeah, yeah. Right. Right? It is a key part of it, but because I think the problem is because it's not useful enough for us. Because I have the same problem where it's like, I use ChatGPT or Gemini not only for coding help, but for help me write a SQL query, help me reply to this email. And I can keep my context of my work day in Gemini or in ChatGPT rather than-- Hold on. Hold on. So you now are-- every time you're talking about it, are saying ChatGPT or Gemini. I use both. So you're leaning into Gemini. Yeah, I bought the premium version. You did. So you got the premium version. Yeah. And you mentioned this before the show. Gemini's key advantage to me is it's faster. It's way faster. I think it's just as good, if not better. And-- Have you used it for content generation yet? I have a little bit. I just-- because I don't do a ton of that. I don't have enough use cases or examples that I really-- That's the one thing that I found with Gemini is that it's very terse. It's very-- Yeah, OK. --very straight to the point. And it doesn't add any of that nice-- I agree. --additional sex that you add to your text to give it a little oomph. Last night-- this may not be the exact same thing. But last night, I'm making my brackets. And I'm like, OK, time to name the brackets, the hardest part. And I asked Gemini. And they were like, here's two examples. They were very dry. I was like-- my team I picked to win, it was Houston. And the Pacers have a player that went to Houston last year, Jaris Walker. And I was like, I want a pun around Jaris Walker. And that's my bracket name. It's about Houston winning it, walking into the finals, maybe, something like that. And they were like, OK. And it was like, Jaris takes Houston to walking into the finals. And I was like, no! [LAUGHTER] That's exactly what I said, but not smart. So I just made up-- my team ended up being very good. But it does seem that Gemini is not as creative-- No, it's not. --as JGBT. Do you think that's because they-- do you think faster means it's modeled on less stuff? Yeah, no idea. Yeah, no idea what kind of black magic-- Yeah. I think we're at the point where we just have to-- literally, that layer of it, where we started 20 episodes ago to now, it's like, OK. I don't feel bad that I don't know jack-fucking-squat about how this stuff works. Totally. Because I'm just going to be a user of this. You want to know the reason I think that Gemini isn't doing better than OpenAI? We already talked about it. It's like the UI. When I went to Gemini, it says there's a button for me to click Chat with Gemini. So I have to click it, and then I have to click Got It. And for some reason, I didn't even actually get into it. Or you have to enable it, if you're on your Google account. It's just not as accessible. Yeah. It's just not as accessible as OpenAI. But have you used it before, though? I have. OK, because it sounded like it was like the intro of, I've never actually used this. And it takes you to the starting guide or whatever. Because when I go to it, it's just a text screen. I'll defend it. You know me. I'm a Google guy. Well, also-- [LAUGHTER] I've got to visit the-- Let's be honest. How much has Google fumbled the whole thing when it comes to so many different things? I mean, Bard, really? Yeah. Yeah. Are you really bard.google.com? Nobody's going to do that. There's no brand around. How many times has a person's name or a human-based persona-- like, you might as well just call it Clippy. Yeah. And just log out. Bard is like a character in Dungeons and Dragons, right? Or some shit. It's like the magical person. I learned this a few weeks ago. I think that's what it is. I was like, oh my gosh, that's so stupid and nerdy. Why did they pick that? Did you guys see that Apple is-- I didn't know how to take this, but Apple and Google are teaming up with AI. Well, so I-- possibly, right? Like, I did see that. So Apple was reaching out to them about some of their potential, using them for some of their generative AI on the iPhone 18-- or iOS 18. But they also have their own model that they've been pushing out there. I read some-- so they came out with a-- did you see that? It came out with a paper. Yeah. And they published a paper. And it was all their engineers-- yes, and it's on multimodal LLMs. And they're-- I don't even know. Did Chachapiti start that term? Multimodal? So they're calling them MLLMs. They're trying to call it something different. Yeah, well, that's classic Apple, right? So they're literally saying, OK, here's an LLM. Here's a large language model. Here's a MLLM, a multimodal large language model. And they're trying to act like whatever they're publishing, I think they're basically saying, we've figured out a more seamless way to stream your inputs that are not necessarily just text and produce-- I think that's just classic Apple. I know. Like, that's just the way that Apple does. Yeah, exactly. It's a watch. Here's a patented diagram. It's this thing, and it has the time on it, right? And it's got a-- Didn't they famously try to patent the rounded corners on the iPad or the iPhone or something? I'm sure. With their specific-- because it's not just rounded, right? Like, it's their squirgle? Squiggle? Oh, it's because it has two radius points, not just one. And now Figma does support it, where you can go in and you can make your rounded corners about that same algorithm. But no, I think it's just-- But they're extremely precise, and they're just so unique. Yeah, whatever, right? But no, I do think Apple is kind of just doing that thing. Because the reality is that I think the majority of these are multimodal, right? And when we say multimodal, it's not like a single large language model. It's actually like a bunch of large language models that are glued together through a different set of processes. That's why Chat GPT-4 is so great right now. Exactly. Because you can just have the same chat, and you're really talking to Dolly and you're talking to Chat GPT-4. Yeah, and I agree. That is why it's so great. It's like you used to have to be like, I now want to work within the context of Dolly. And you switch and you told it your intent. Now it just understands your intent. I want a picture of a cat that's doing X, Y, and Z. And it says, oh, you actually want to use Dolly. And actually, I'm going to take your prompt. I'm going to make it better and more suited. How many-- how much-- I mean, how much of a-- I mean, I could just say it. If Siri was even close to as fast as Chat GPT, and they came out with something that was even as good as Gemini, you would probably-- they would probably dominate the market pretty quickly when it comes to just personal AI use. I think that's what's going to happen, though. I think, yeah, I think Siri's going to roll out. And now Siri won't be the idiot that it is. I won't have to be screaming at it constantly. I'm sure I probably will, because I scream at Dolly all the time. Because we've proven the fact that-- I mean, we've looked at different models. We've looked at models that are faster. And we know that although Chat GPT has-- we can't stop using it, their first mover, their competitive advantage isn't necessarily on the model side alone. It's on the accessibility. It's definitely not alone, right. It's on the UI. It's on the stuff they've layered on top of it. Exactly. But it's not necessarily something that they just can't-- that can't be replaced. Sure. And I think if anyone's going to replace it at a mainstream level, it's Apple. It's because they compete at the same kind of level already. Yeah. Whatever Sam has been able to bring to life, he's just got that similar aura where the things that he's completely creating just have this edge to them, just like these edges. It's just like, whatever it is about it, I like this one more than the normal one. And I think when it comes to Apple, they've got that same sauce. So now they're just coming into the arena with the same kind of sauce. I'll be very curious to see how well it turns out. Because Siri, to me, at this point, is an embarrassment. Like where I'll ask, well, I can show it on your iPhone if you open up your iPhone. I'm just like, just tell me the fucking answer. Last night I had my phone locked in the shower, and I was like, play Spotify. And it was like, I can't access that unless you give me access. I was like, I'll kill you. I will kill you. I had to click a button. My finger was wet, and it was like, no. So it makes me so angry.

• So literally yesterday, I am screaming at Dolly. And I'm just like, all caps, I'm like, quit fucking putting borders on these images. Like four of my rules for the images are no borders, no frames, no vignette. Just fill the whole fucking image. And every image that it was generating was coming out with some framed border. And I'm finally like, all right, stop. Like, I need you to tell me why you keep putting borders in here when I have three rules that very clearly define no borders, no frame, no vignette. It's like, oh, I'm sorry for the mistake. If you put this in your prompt, maybe I'll be able to follow it better. And it was like the same thing, so I copied and I pasted. So finally, my prompt literally said, fucking no borders, never add anything, but fill the whole fucking screen because you refuse to listen to me. Then all of a sudden, it starts doing better. It's literally, you have to abuse these things to get it to a point where it's like, stops doing it. And then, of course, this morning, I go and I generate an image, and there's this god, a frame.
• Hey, so have you seen the Leonardo image generation tool? We talked about it once.
• I remember us talking about it. It's really good, and they came out this week with transparencies.
• Oh, so you can do PNG transparency, that's cool.
• I know, and I was like, you know what? That's a really, really, really good feature.
• That's a very novel feature, yeah.
• It was the only one that ever came close to generating an AI image of me based off a picture. And you know, you got one that looked great, you little asshole, but I didn't. Mine sucked.
• Mine's like making me go to the gym.
• Yours is like four days a week.
• So I think next week we should revisit the image generation thing again.
• We should, all right.
• We'll check back in on that tech.
• Did you guys see the Mid Journey came out with a character tool? So they've introduced--
• Consistently?
• Consistent character, yeah.
• That's key, yes, I did see that.
• I haven't used that one yet, but I know we struggled with that early on. It's almost impossible before now, I guess, to go to Dolly 3 or to Mid Journey and create a character and then say, okay, now I'll have that person in a different scene, and they would just look different. I'm like, how?
• Do you have access to--
• I just canceled my subscription.
• So I've requested access to the Mid Journey, the web interface.
• Yeah, I was like, oh, they came out with it, and then I went, and it's not even real.
• I had been on the wait list for like a month now.
• Right, right. So I would love to try it, but I mean, I abuse Dolly, like, all the time.
• That sounds like it.
• Yeah, like--
• Literally. Because it is, like, again, really great. It can do some amazing things, but man, when it doesn't follow the instructions, I just-- - It's infuriating.
• Infuriating.
• One thing that I had as a bonus for automations, I think this is something that people should definitely be doing, is use ChatGBT to write outlines.
• Oh, yeah.
• And I mean, I know it sounds silly, but like, for your problem, the fence thing, like a detailed, an outline, then detailing from there. But like, if you want to create, like, one thing that, it actually came to, this whole thing came to my mind when I watched Apollo 13 on vacation. I talked to you about this. (laughing) Is that I could not believe how technical that movie was. It had no plot, it had no character development or plot. It was a very, it almost was like a historical documentary. But the one thing that happens is that they get, when they're testing the software that they built, they throw in all these variables, things break, you know, and they literally coded to the nth degree on how to build a resilient system that could have it. What happens if the cooling system goes down to 10% or goes out completely? How do we land this thing? Okay, there's a procedure for that. So they wrote, they have all this software, but then they have all these twisty knobs, right? And they have buttons. And then they have a procedure for everything, every single problem that happens. And the procedure is literally, they go, we need a procedure for that, we need a procedure. We have a procedure for that. And it's literally, you just turn to the book and the page and it goes, okay, the procedure is, flip this button, click this latch, undo this hood. And then you have fulfilled the procedure. With business and with yourself, you're trying to like do something that's extremely detailed, right? You're trying to cover all your bases. And so like, what's one of the hardest things you could do is like, I need to write literally the de facto architecture or the processes, the standards, and they need to be complete, right? And so in order for something to be truly complete, it has to have a complete outline. And the other thing that happens because of all this is because it's so overwhelming, you're never gonna get it done, right? So I think one thing that's really cool to do is create an outline, give that outline structure, like it needs to have a paragraph, it needs to have some bullet points, it needs to have a reference table and get that all just like scaffolded out, start filling in the gaps. And then as you're going, okay, give me, you're just going through and you're, hey, chat, give me a bullet point on what a no sequel is. Give me a bullet point on what, you know, abusive HR, you know, languages, you know, like, and just start boop, boop, boop, boop, filling in the gaps and just turn yourself into like a total human AI hybrid when it comes to like writing really long, complex things.
• Totally, it's like, that's how is the counter point to my original point on you need to have the system that already exists. We don't necessarily have to, the other side of that is exactly what you just said. So like, let's say that you wanna create entirely new sales part of your organization.
• Yeah, you've done a good job doing this, by the way. You've done this a lot.
• Yes, and so it's a situation where you say, okay, I need to build a sales organization. What are the things I need to do? I don't know the first, you're a vetted 10 year VP of sales for SaaS startups. Okay, great, and then there's a really long list and you might've known three of them, but there's seven you have no clue you should have had. And then you say, then you go through each one of those and you say, elaborate, elaborate, elaborate, elaborate. And by the end of it, let's say that you've gone through all of this. Now you're all the way from top level, like hiring and compensation structures and outlines to real pitch and all the details in between and all the platforms you should be using. And maybe that you should be integrating these. And like, it's just, that is totally a great point. If you're looking to use AI to automate a newer system, make sure that if you're gonna be automating a new system, you don't miss any basis. And it's a really good way to help you not miss any basis.
• I think that's a problem a lot of small businesses have is just being able to behave like a well functioning, like larger corporate entity that has lots of these processes and procedures in place.
• Yeah, I think the key is for the future of AI within enterprises, I have no doubt that it's gonna be, they're all gonna be grounded on pre-existing processes. So that APQC, the American Productivity Quality Center, they go through and they identify every single possible process that all companies use. So like we're talking 13 different top level categories, and it goes down to like, this point, this point, this point and they literally write out every single process that the companies who are doing the best work, this is how they do it. That's the way that these AI should be grounded. They're on pre-existing processes, pre-existing things. And a lot of them can just run in the background. So one of my clients is an insurance company and they're sitting there and they're looking at what's called subrogation. And all subrogation really is, and this really pisses me off because for three years, I thought subrogation was way more complicated than this. Subrogation is literally just the behavior of going and collecting money from the at-will party or the at-fault party. So like if I get into an accident and it's their fault and my insurance provider needs to go after them for the money, that's all subrogation is. But we're talking like millions of dollars that insurance companies lose on not going and pursuing something that might be subrogated. This is a perfect example of, and that's what I've been working on is helping them basically go through, take all these notes on a claim and be able to say, should this be subrogated or not? These AIs are great at it, right? It goes through, it pulls it down. It says, yep, yeah, there's a high likelihood of subrogation because one, we've got the documents that say that they're at fault and blah, blah, blah, blah, blah, blah. Those types of things I think are gonna be where the future of AI within enterprises is gonna be. It's just gonna be these autonomous agents that are just running behind the scenes, just following the process.
• And that could even create more jobs because--
• 100%.
• It's like, well, now you need the agents, the people to go out and serve and do the thing. And that's what they're better at is--
• Exactly, yeah.
• They're good at going and finding the person, talking to the person, whatever they need to do to accomplish that task. That's why they don't do it.
• But what about the, no, I'm just kidding, the bias. (laughing) It does differ to, no, but the agent piece is extremely important. And I think there's one thing, people that I wanted to clarify, and I kind of put it in here, but the future of, I think, AI code contribution is not what people think it is today. So especially for the enterprise. So right now you go, okay, I have a coding assistant and as I code, I'm gonna use AI. Sure, that's great. But the real application for AI is for AI to be inside of your source control system. Okay, so they're gonna review, let's say, for example, you have a process where you have a source control and every time somebody wants to put code in the system, they have to do what's called a pull request. Well, those pull requests basically say, hey, this guy's trying to put this code in here. Okay, and usually a human goes, oh, I'm gonna review that pull request. And then they do what's called rubber stamping. (laughing) Looks great, Johnny. And then they get--
• And then all the ship breaks.
• And then they're like, oh, that build failed. I don't know why. It must've been Johnny's pull request. I don't know who approved that. Anyways, but the AIs are just sitting there waiting. And now they're looking at it and they're going, oh, what's the, does it meet the, does it look good from a security perspective? Does it meet the functional requirements? Does it, is there merge conflicts? Can I go ahead and fix some of those? And the next thing you do is you connect that with your ticketing system. You're gonna start recreating smaller achievable tickets in your sprint management, and if anybody knows anything about Agile, you just basically create tasks and people do them every day and they can't do it, do it, do it. You just work, it's called working. (laughing) It's the most overblown bullshit you'll ever hear about in your life. (laughing) But basically you create a task and then the AI might elect to do it. Does that, you know what I mean?
• Yeah, totally.
• Like it might not be, hey, the AI, this is their job. No, the AI might see that, oh, you know what? This is just updating a dependency and I'm gonna go ahead and do that.
• Yeah. - Right?
• I already have, on some projects, we have automatic dependency check and then once something does bump, it creates a pull request. So it will do it, it'll bump the version and then it'll create a pull request and be like, email me and be like, hey, yo, this went from 10.4 to 10.5, check it out and then approve it if you think that looks good.
• Are you guys doing that with GitHub?
• Yeah. - Okay.
• Yeah, and then we've used, I've started the process of trying to use a new tool. What was it called?
• Bear. - Bear CLI.
• Yeah, yeah, yeah. - So, Bear CLI is a open source-- - B-E, how do you spell it?
• B-E, like bear, - Like bear token.
• Rer token. - B-E-A-R, okay, bear.
• B-E-A-R-E-R, yeah, and that is, it basically will review all your code. It'll scan your code or review your pull request for security problems. - Nice.
• And it can be attached. It's a tool that you can put into your DevOps system and read from an open source Docker image. So you don't have to pay a licensing fee to put it inside your environment. It is an open source tool that you can Dockerize.
• That's awesome. - You can put it into Bitbucket. That was the one where I converted it from GitHub to Bitbucket pipelines. So those types of things are really good because people right now don't have good answers on security when it comes to code level security. That actually integrates well with a developer's productive process.
• I just heard, that reminds me, I just heard an example of a nightmare scenario where a person put into production, I'll leave all names out, a redirect that was like sending a person from their email or something. Anyway, it was collecting credit card information and then sending them somewhere else. Well, it was making a get request. So it had that information in the URL and they are logging all HTTP requests as most servers do. So there was a log of someone's unencrypted, just redacted credit card. And something like Bear probably would have found that at the code level and flagged that. They found it months later because they're a good company and that does security audits. And they were like, "What the hell is this? We're seeing credit cards in our database." But to your point, you can mitigate those risks. It's not perfect, but you can find those things faster and help mitigate that risk.
• So a little sidetrack that people don't, that a lot of developers don't think of is exactly that problem of using get parameters for all sorts of things. Because they don't think that like Google Analytics is gonna read every single thing that goes into that URL. And it is, man. It's like-- - I didn't even think about Google Analytics. I was thinking about the server level, but I'm like, "Well, shit, you're right."
• Yeah, just right there, Google Analytics, anything that is in that URL.
• I'm an absolute HTTP method freakazoid. So that's a delete method, not a put.
• What are you doing? - Pull request denied.
• Delete.
• Yeah. - It is. It blows me away how few people actually think about that in terms of the amount of information that you can leak out from a privacy standpoint just in the URL parameters is crazy.
• I'd say that all in all, from an automation perspective and from everything that we've talked about and really going off of Sean's last point or one of his points is that if you're a small startup, smaller business, medium-sized business, what these AI automations allow you to do is simulate as if you are a larger organization and let you implement systems that larger organizations currently have. And if you aren't growing as an organization, the only reason you're not is because you're not using systems that organizations just like yours use to grow and operate as. And so if you wanna operate like a larger organization, the only way to get there is to act like a larger organization. And so that's why I think all in all is the biggest takeaway, at least from my perspective.
• Boom. - Drop the mic.
• Drop the mic, Dre. - That's it.
• Great, great part, I agree. I agree. You know, learn how to automate stuff.
• Yeah. - Do it. And enable it in your business processes.
• My show notes, real quick, this is the last one. So not show notes, meeting notes. So every meeting note, every meeting that I have, I now record just locally. I then run that through Olama, so it's all local. There's no, I'm not leaking any data to VHFGBT. And then I use, I've kind of hopped in between GEMMA and Open Hermes, I think is the model that I'm using, to basically take those notes, totally build it into like a very clean outline of like, here's your discussion points, your takeaways, who was involved, who's this. And then I just copy those and I send them to the people that were in the meetings, and it's great. I mean, like I use it.
• People love that stuff.
• Oh, and for someone like me, where again, I get all amped up about the current moment, then I forget everything that I talked about or whatever, 'cause I just was like, you know, in the moment. For me, it's just absolutely critical now that I have all these notes, I can go be like, all right, here's the key takeaways that I need to go do. It's just been a lifesaver. Now there's a product called Plaid or Plaud, P-L-A-U-D, I think. And it's a little like card shaped thing that records your calls, your conversations, your different things, and then uses chat GBT to do the transcription. I'm thinking about buying, I might buy that because it's doing pretty much the same exact thing that I'm doing, but mine's fully automated, just on my computer. But this thing's just on your phone, you just hit a button and it just does its thing. So that'd be kind of cool.
• That's awesome.
• Well, guys, look, thank you so much for watching the Big Cheese AI Podcast.
• If you got all the way here, you're a winner.
• Yeah, if you got all the way here.
• Let us know, actually. So this is something that I did on my old podcast. If you make it to the end, let us know because then we know people are actually making it to the end, that'd be awesome.
• If you made it to the end, mention how much you like me in the comments. (laughing)
• How Sean's your favorite.
• Yes, Jacob Wise, Brandon Corbin, Sean Hines, and DeAndre here.