[00:00:07] Speaker A: All right, thanks for joining the broadcast to post podcast, I'm Jeff Sangpil, CTO here at Keycode Media. Today we're going to dive into artificial intelligence, enhance media management and what it means for the future for creative teams. I'm joined by Jason Perrott, periphery and Bryson Jones at North Shore Automation. Let's get into it. Jason, you've been at the forefront of media technology innovations, including peripheries initiatives and AI driven media management solutions. That's a tongue twister. While AI solutions require you to connect to AI providers like Google or Chat GPT Jason, your team has developed a way to bring AI media management solutions to on prem facilities that are air gapped from the Internet, satisfying a lot over security requirements from major Hollywood studios. So Jason, can you give me an overview of your career and what drives your passion for media and entertainment technology? What brought you to AI?
[00:00:59] Speaker B: You know, I started out actually my career working on film set and doing everything you can imagine from working as a pa up to working as a first ad. And then I transitioned into the production offices and working the production coordinator and then finally a visual effects coordinator where I felt tremendous amounts of pain of broken workflows and workflows of mailing hard drives all over the world and tracking spreadsheets, which is what led me into this world of media asset management event and transitioning into the tech side of things. So I feel like I've been there and felt so many of those pain points of just trying to figure out where the heck the content was, what format was it. Is the right person getting the right thing at the right time in the right place, managing transfer speeds and paying for transfer speeds and trying to get those things done. That's where a lot of this came from. And then I working with media asset management systems basically when final cut server was released, and working with Apple very closely at the time. And then I've worked with just about every major mam system out there since then.
And really with the acquisition of workflow intelligence nexus into data core last year, we've seen so many different great technologies over the years come and go with these different mam systems and automation tools. And with the new kind of era of AI happening, with all the new generative AI and all the new different things happening with transformers and a lot of the technology that's evolved just over the past year and a half to two years, we saw huge opportunities there to finally make a lot of things even simpler than they were before, to be able to really bring this technology down to users so they could see some true benefits. And that's what we've been really focused on for the better part of the past year, is seeing how can we keep leveraging these really powerful new technologies in a way people can use. And one of the first things we saw is, yeah, it's great that you can do all this stuff in the cloud, and they make it sound so inexpensive because a couple cents here is a couple cents there. But when you're processing large volumes of content, or if you want to know a lot about that content, those couple cents add up really fast. And all of a sudden you're averaging a dollar or more per video to get AI intelligence about it. We look at all these customers we've worked with over the years that oftentimes have millions and millions of videos. Spending millions and millions of dollars on your AI just to understand what's in it isn't going to happen. It just doesn't make sense.
And then there's the security side of it, is that many of our customers always want to know. They know exactly where their content is. They know it's in their storage on prem, or it's in their buckets and cloud storage that they own. But with AI, it raises a lot of questions of, is it going to be getting trained on your content? Is it going to remember your content? Where is your content really going for it to be able to process it? And the nice thing about this space is that the open source side of AI is incredibly aggressive and moving forward at a very rapid pace versus the closed source side. And we've been able to do some incredible things in the open source world. And that's where we saw something of an opportunity that we thought made a lot of sense, is let's take these open source tools that tend to be really complicated and changing all the time, and do the same kind of thing we did for many years with win, which is simplify them, bring it down, and make it into a really easy to consume, simple user experience and make it cost effective and highly secure because it can all run 100% in your environment that you control with your content, and it just makes a much better solution. So that's what we're focused on, is really doing that and making sure that you have control over your AI, your content, and your metadata on your storage.
[00:05:25] Speaker A: And speaking of working in your environment, you're actually coming to a slide from my office upstairs. It's just no AI involved there at the moment. So Bryson Bryson's the CIO and founder of North Shore Automation. Bryson, you're an expert in automation and media workflows. He founded North Shore to provide the real solutions to media production challenges.
Share your journey into the media automation space. How did we get here from there?
[00:05:53] Speaker C: Yeah, well, so Jay said that it's really funny, and I apologize for my voice. I'm Covid. So anyway, and I'm not in key code, sadly, but so anyway, with Jason, I very, very funny. At a certain point in our career, our career on a very similar trajectory. So it was really great to hear him tell that story because where he was working on film sets, I was coming from the IT side and then working on editorial systems. So Jeff, you know, I was the old school avid guy and all that. And basically my point where I love that, you know, Jason, like you said, suffer from the bad workflows, right. Well, I got to the point where I was supporting like seven brands of fiber storage and I realized that storage is going to become a commodity. It's awesome how, you know, cheap and powerful storage is now. And I understood that that was going to change. It was not going to be a big deal to install storage anymore. But I also understood that we were making way more media than we could manage. And so our joke about Northshore automation is we founded Northshore automation probably five or six years before anyone wanted to automate anything, but we had customers who wanted to. And so we went from asset management solutions. I was originally focused on one on Cat TV originally, and now that's expanded out into several other options. So we did a lot of data migrations for people. Still do. So we provide data migrations from mans and storage through systems. We're a little bit more cloud centric because of our customer base. So once we had built asset management solutions and really our journey into AI happened around a product, we made some technical decisions three years ago when we were starting to redesign our call me dam. And as we were doing that, to me, we realized that we had built a system that was a pretty good AI metadata platform. And then we were able to start looking at what metadata was available. And one of the things Jason and I share is this idea that right now, v one AI is a wants to tag everything. But we sort of joke. It's like, well, it's probably not as useful that there's a red bicycle in that shot. Sure, that's good to know that that's Jeff or, you know, Tyler or whoever in a shot. That's important. But what's really going to start to happen is when you start to understand the media in a better way with multimodal AI. And then also when you get to where you can actually interact with the media in different ways. Right now, the central problem you have is that asset management systems are not ready for the data that's coming, and then they're also not giving you a user experience that's effective. I think that's what the last thing drives the adoption. One last point, because I'm just kind of following Jason's template on this, is we are periphery is one of the solutions that we think is a great option.
Prim is possible for certain people. The cloud makes sense because they're already there. So no source business is a little bit unusual because we do work for the mom and pops who work for mid sized. We also work for people who have enterprise plans with AWS, who don't really care if it's, you know, if you, if you're like and if you're working in distribution, it's so expensive to buy content. If you own 10,000 feature films or television episodes, which is Gannick organization might only own 10,000, well then when you're talking about a few dollars pertain video in the cloud, it isn't as daunting, right. So if you divide production, I think today, as we talk production from content ownership, content owners have a different metric than people who are producing. Because producing you have hundreds of thousands of hours that you very cheaply were able to capture thanks to all the new technology. Whereas if you're talking about a feature film or an episode of a tv show. So a lot more northwest work right now focused on these are the final assets and we really want to understand that. So its interesting thing. I just want to bring that up because thats the piece. But yeah, we think that. I think this cloud thing is being worked out and sorted. The vote on Prem, I love the fact that there is an on prem solution. There hasnt been for a long time. And I think for people, I think its going to be really cool to see. But I like the fact that. So in your intro video, youre also very sober about the fact that, like I joke, it's like, well, if you've never owned shared storage or if you've never bought us a good server, then the cloud might be first thing you do because turning on an EC two instance is cheap and just takes a minute. Whereas like ordering, you know, a dell server with Nvidia GPU was correctly, that's where key code comes in. So I would say that that's an interesting piece with this. But, you know, I think that long term. Yeah. If you're going to do large content, I want to say that I agree with the economics. Work out the same. Yeah, yeah.
[00:10:52] Speaker A: It's also interesting you're pointing out the difference between content owners and production because content owners used to, we'd only worry about them at the end when there was the box, the box of stuff in the vault, who's taking this well. And no one would stand up. And everyone in production didn't care. They were off to the next so much different set. And the thing is, we're also getting into the idea where that content that's in the box, we can do something else with it. All right. A basic kind of thing. How would you define AI in the context of media management? In a perfect world, give me your elevator pitch on how AI is going to solve all your creative teams problems. What are the basic things customers need to get this going? Bryson, let's go with you first on this one.
[00:11:37] Speaker C: Yeah, I have a really short one, which is what's in the box. Right. To go back to the box. Right. What's in the box? Because we can, I've been saying this for years with asset management. I got into asset management because we can now create and capture more media than we can manage. When tapes were a $100, when it was $50 for an XD cam disk, when you had a cost of a film reel, you didn't shoot much. Right. I worked on a film with vendors and he shot dv for the first time and he shot 350 hours. He would have shot 35 and that, he told me that. He was like, I would have shot 35 hours. Shot 350, and then we had to post that. That's what's happened for the last few years, right? Like, we're all walking around now with like, really high quality cameras and we're capturing media in a way we, we can't go through. So one of the key code customers, right. Foundation Milk Educational foundation, shot 4000 hours of interviews. They have 4000 hours of interviews to go through. So you go, well, cool, let's get. A producer can do that. Eight hour days, two years, right? No, you run transcription on it in a matter of hours. You know, mean days. Right. And then that transcription is searchable.
Beyond that, what else could you do, right? Could you put an LLM on it? Right. All of a sudden you can ask it questions. So for me, the whole thing right now, in the future, what will we get? Oh, my God. Right? We're going to watch the news feeds, and we're going to pull from your archive, you know, oh, my God. A horrible tragedy has happened. We're going to all the footage we have from that location. Oh, someone has passed away. We're going to pull all that footage up for their obituary. That's what's going to happen in AI, media management in the, you know, pretty fast. But right now, what Jason and I are doing is pay all that stuff that it was too expensive to deal with, like you said, that she might repurpose. Right? What would I do? You know? So for me right now, I just want to make the archives that we've been building, all that stuff we've been shoveling, I want to make that visible. What do we have? Yeah.
[00:13:47] Speaker A: So, one thing, just to be clear, LLM is a large language model. It's a. A rubric, basic, that allows you to the framework that allows you to process media a lot faster.
[00:14:00] Speaker C: Chat Cpt. Right? Chat, CPT, Gemini, Claude, those things.
[00:14:04] Speaker A: Yeah, all that fun stuff. Sorry, Jason, what's your elevator pitch, or AI, in media management?
[00:14:10] Speaker B: At the end of the day, the simplest way, I think, to explain it is we've had all this media content that people gather, like Bryson says, expanding and expanding as it's gotten less and less expensive to be able to gather high quality content, and where it used to be about being able to find that one needle in the haystack.
Now, what I find that AI does is with AI, it's like having a gigantic magnet that we're able to run across all your haystacks to find all the needles very, very quickly and be able to have those things all be very organized. And why that's important is that we're seeing more and more content reuse and people even thinking about and planning that and how they do the production. Right. When people are planning productions we've seen and heard about many times, where people go out to a more exotic location for a productionist part of that shoot, they will intentionally shoot additional content they know is going to work well to be able to serve as b roll for future things, or to be able to even end up within a stoppage library that that organization might also manage, because they know the value of being able to capture this content. But that value, again, is only as valuable as being able to find it. And one of the challenges we've seen in the space of mams since the beginning of me asset management systems, is that your ability to find something is only as good as the quality of the metadata that got attached to that content. At the forefront when it got ingested. And we've seen over again that when we deploy a large mam system, what happens? We go in and we configure all the fields, we talk with the customers and figure out the 40 metadata fields that everyone wants and thinks is really important to have on that content. You give everyone a budget training, you come back six months later, maybe a year later, you're lucky if two of those fields are getting filled out.
And so the mam system is capable of searching for things and finding things, but without that metadata, it doesn't matter. You can't find anything. And so with AI, the first we see is being able to transcribe that content and use an LLM to be able to actually query for and find and generate all the metadata to populate those existing fields you already had and just do that job that no other person ever wanted to do. But now that content becomes much more searchable and you actually are fulfilling kind of what the original vision was with the mam systems in the first place. But then where we're going with it is to take a step beyond and do things that we could never do before and, well, not that we couldn't, but it would have taken too much manual labor for anyone to ever do before. And being able to detect really powerful things like we just did an implementation recently for a customer where we're leveraging a combination of object detection and vision to be able to look at things like when people are playing poker games and when we're looking at that content, instead of being able to just identify, well, here's a playing card, and this is this card, or this is that card. Number one, not actually showing playing cards, they have playing card icons, so you can't just do that. But number two, once we can identify those icons, what someone really cares about isn't, oh, there was an ace of spades on the screen. We need to add the intelligence layer to say, okay, we've trained this LLM a little bit to just say, here's how the game of poker works. So now when we look at a piece of content you could search for all the time, someone has a full house. You can search for every instance of a royal flush that's happened. You can actually see who the winner is at any given point in time in the game before they get bluffed out. Right. And then what's the actual, uh, the actual game flow really looking like? And being able to do all of that purely through AI without any human intervention needed once it's been trained is incredibly powerful. And those are the kinds of opportunities that I think the AI is opening up and really changing the way that media management is going to look for people. And then the last bit is going beyond metadata, and that's what we're working on right now with Ice and some new things that we'll be talking about more and more over the next couple months. But we really envision getting to a point where you want to have more of a conversational interface with your library so that people who dont necessarily ever even use EMS today, the people in the organization and the marketing side of things and the sales side of things that need to be able to just ask the question of what content do we have about x? What kind of things did we do with this person in the past? Could just ask those questions in natural language and have answers in the form of content and explanations provided back to them. That's where I think we're really going.
[00:19:18] Speaker A: I'll give you my pitch on it. You shot more and you said fix it in post. We don't have the bucket to fix it in post because you didn't give it to us. AI is our way to do this at a fraction of what it would really cost to make this happen.
[00:19:35] Speaker C: Yeah, dude. No, no, 100%. Let me give you two things. So Jason and I all is really actually one of my favorite calls in the last few years. We had a call one day, we were like going to do the first event. We're like talk and, and Jason, I came up with an analogy. You and I have been making exercise equipment for all these years. And every treadmill, every like, peloton that's sitting in the corner, unused. That's what our ma'ams were. We would build these things and so we would joke about it. Like, like you said, there's no time, there's nobody to budget. People were like, oh, we're going to spend the money to buy a ma'am, but we're not going to do the effort. So part of the reason you hear me talk about content owners is that content owners, there was a budget associated because it was a value. So first thing I want to say that. And then, Jeff, let's talk about budgeting a little bit because brought up something really and Jason, you keep hearing us say one of the things that Jason and I also believe about is that we don't, we're not talking about generative AI to build content. We don't. Neither of us really thinks that's the thing right now. I don't even know if that's even gonna work out legally for a million reasons. Right. So we're not really worried about that. But what we. What we do think that's important and I keep hearing this talked about is that I'll give you a real world example. I'm salivating. There's a key code customer that we work with that they. It's the situation. I'm not gonna mention them specifically but I want this business case to be understood because it actually will drive a budget decision. They make reality shows, they shoot 5000 hours per season. 5000. They obviously don't go through 5000 hours of footage. They had producers writing down what happened and they tell those pieces.
So if you think about it, to make us a season of that show they might use 100 hours of that footage. Probably not even that. So it means that there are 4900 hours shot that no one knows what happened because they're in a remote location. There's only a couple of people there. So if you weren't taking notes. 4900 hours are there unlogged. It will never be logged. Except I'm a weirdo. I watch a lot of YouTube before I got my age. And I will watch the weird, long video of very specific data of something that I like. Imagine if you could go back to all those old seasons that they shot of your favorite show and they could pull stories that were missed. Right? Now you go back and go, hi.
Whatever television network you had a hit for ten or 15 years what if we could go back and give you another five seasons of deep dive for your hardcore fans who just want to watch weird side stories. There are some. I watched that show. There were some hilarious stories that they would pull. And I know that there's hundreds of other hilarious, interesting, fascinating stories but nobody could spend the money to go do it. But when you do what Jason's talking about doing here what we're working on where you just go. But one thing we should also mention the beauty of what and why we keep talking about these large language models is that they have conceptual understanding. They can say, I need romantic scenes. You can say, I need conflict. Like you said, you could teach what a good poker hand is. But you don't have to. This is a huge thing. This is why the open source movement and the cloud movement. Because, by the way, a lot of the cloud technology can be deployed on Prem. It's open source. You can pull it and employ it. And so why that's a key is that they taught this thing, not us. But all these people out there researching have taught it what is conflict, what is drama, what is romance, what is humorous? And being able to go in and say, find me humorous situations from 10,000, 50,000 hours of footage is fascinating. And the last piece is you're not just going to throw this into your regular budget, but if you think about it and you come and talk to us and key code and people like this and say, we have this idea, I think like you said, you never would say, I have, let's say, 50,000 hours of footage and I'm going to have producers sit down and go through it. You never would do that. You couldn't. It would be unfeasible.
But there's a middle ground where it's like, well, it is feasible to spend x thousands of dollars to go through this with AI and then develop a show. And I think that's the thing that's really interesting is this shift from that is using AI not to replace humans doing things that they do now, but like Jason said at the end of to replace humans doing things that we were never going to hire a human to do. We're not transcribing 50,000 hours of footage with humans.
[00:24:15] Speaker A: And that goes to one of the comments we're seeing in the chat, Colin saying, that's the issue. I have too much content, no time to manage the manually, the metadata. That's the whole point we're trying to get to.
We've fixed it in post for less and now you actually have to do possibly less production the next time.
[00:24:35] Speaker C: Yeah. And the bar. Oh, yeah, sorry. The dream that Jason and I are talking about, which is rapidly approaching, is this sort of magic conversational understanding AI. But what's weird is in your video, you showed the steps along the way, right? If we understand that this important person has passed away and we need that footage, then we can simply run facial recognition.
[00:25:00] Speaker B: Yeah.
[00:25:00] Speaker C: And you can start to do that with a single focus. If you're a news organization and you are not running facial recognition right now, you've lost a huge opportunity because there's no way. You know how many times Shannon Doherty appeared in your television show? There's no way you know that, right? That's a big one there. And so I want to say that. And another thing, Jason, I said, and I want to set out loud, please just do something. Just start. If you can't buy a periphery rig, let key code get you some credits in the cloud to just try something like do a proof of concept, do anything. Because promise you, because Jason and I had our eyes opened. We both have had this revelation from being like, you know, sort of like script kids and then building development companies and then workflows, and it's more complex. And then we open the toolbox that's AI, and we go, oh, my God. We've been waiting for this for years. The biggest thing that we need is we need people with imagination and a little bit of experience. So let key code run some tests for you, like, get a poc, do something. Because what's happening right now, for the last two years, you've been able to do everything we're talking about, really, for like four or five. Facial recognition has been available since 2017.
No one's been doing it. So if you watch me on social media, if you watch me on LinkedIn, I'm saying, please get a database, any database, and we don't care. Iconic, chaddb, delete, whatever you're going to do, whatever you get, get something that's a database where you can have insert and get that started, then run some AI on it and be, guess what? Be disappointed. Be ready to be disappointed. Because it might not be perfect, but the second or third iteration it is, you know, like he said, hey, they ran that poker thing, and that was probably fairly, you know, a fairly expensive thing. But when you understand a business use case, it's not as expensive as hiring a room full of people to watch that content and find the good hands. And that's why you won't hear us talk about either of us. And I appreciate this, Jason, about you so much as well. And Jeff, you've even done a kiko's on it. We're not here saying, come do some weird science project. We're saying, bring us a business case.
Let us develop a business case with you, because that's what we're like right now. That's the thing. Because what's happening in the rest of the world, they're already doing it all by business docs. Right now, my entire 100 some thousand documents are completely analyzed by an LLM. And I can ask questions of my data. And I want that for every customer, right?
[00:27:41] Speaker A: Yeah, definitely. And I want to know, analyze me playing poker, let me know what my tells are. I really toward that. Jason, there's a lot of hype around AI in media management. You shed some light. Some of the common misconceptions are versus the actual capabilities of AI.
What's it going to take to do those pocs and set up AI for media management effectively? What are teams going to need to do to prepare beforehand to implement.
[00:28:14] Speaker B: Yeah, no, totally. So one of the big, I think, challenges and misconceptions a lot of times with AI is that you have something that out of the box, you're just going to plug it in and press a button and now all of a sudden you're going to know everything about your content.
There's always a couple steps in between there.
One of the big pieces that we find is sometimes challenging to get from people is that you need to figure out what it is you actually want to know about your content.
If we take an example like that poker example again, if you look at a poker game, there's a ton of things you could learn about the content. You might say, well, we just want to know what are the hands? Okay, fine. So we can show you something that with a vision model and an object model, you can put those two things together and about a day or two worth of work of any little bit of training, minor stuff these days, because the vision models make it ten times easier than it used to be. But then, yeah, we can tell you what the hands are, but once you have that now people, well, wait a second, we don't care about that because this hand didn't matter because he folded. Oh, so you want to be able to know what the status of the hand is as well?
Oh, well, yeah, of course. Well, then we also wanted to know in this other game, when there was five players versus two players, how many people are doing what with their chips and how much money does each person have. Oh, so you want to keep track of the money side of what's in. There's a lot of different pieces on the screen. So I think what's important is making sure and spending some time, and I think this is where, from the consulting side, this is where we're really working together with our integration partners, like our friends at Keco here, to work with our customers and say, hey, let's talk about what's really the most valuable information that you could get out of this data, what's going to help your business most.
We can take that and say, okay, great, now that we understand what the value is you want to get, here's the approach of how to leverage AI best to get that value. And I think it makes a big difference because a lot of people's experience with AI, with many systems out there is they say, oh, well, it has object detection. What does that do? Well, a lot of systems, you plug in a video and it says object detection. So all the random objects that are in what's called the coco data set, which is kind of the set of all the standard objects that every object detection system is trained on. And it's going to tell you, oh, yeah, there was a stapler in the shot and there was a cup in the shot and there was a hat in the shot.
[00:30:57] Speaker C: Who cares?
[00:30:59] Speaker B: Great. Your AI found all that. But what does that mean to you? It doesn't mean anything. So I think we're really focused on saying no. You got to understand for your business case what it is you care about. And then make sure you're leveraging AI that has the ability to be tuned and structured and set up in such a way to get you just the data you care about while filtering out the noise of all that excess stuff that you really don't care about. And then the second piece is being able to combine things together. We're doing something right now for a sports team. And they came to us and said, well, originally we had someone who already tried this with our content, but it didn't really work because they told us they could give us facial recognition and we thought we'd be able to recognize all our players in all our shots, but, you know, it's inaccurate and we're not seeing everyone. And an example, one of the first example pieces they gave me was one player facing the camera and the other player with his back to the camera. But of course, his name's huge, written right on his jersey.
And so they're like, well, it only identified the one person who's facing the camera.
Yeah, because you only did facial recognition. There's no face yet, like in that jersey. It's obvious to you and me that we know who that is, but that's not how it works. So that's where we say, no, it's a combination. And what we find is normally a combination of whatever recognition specific thing you want, whether it's object or facial or something else, plus a vision model that could actually look at the screen and understand everything that's happening. And then mixing those two things together with a proper prompt to be able to get you the kind of metadata you care about. And what we've done and what we're focused on in periphery is taking that whole process and making it really, really simple so that it's not a big science project to be able to deploy. It's not a big, complicated thing to manage. It's done in a more intuitive way and done in a way that can be done very quickly to get that value.
So one other thing, real quick, I was going to mention, we were talking before rice. You were mentioning about people with 5000 hours of content. One of the other big places where I think AI can play such a huge role in really changing what's possible is in the area of documentaries and in noobs, because in both those spaces in the world of documentaries, I've edited a number of documentaries a long time ago. And one of the things that you always experience in that world is that you get handed, oh, you know, we went out and did this event and here's 300 hours of content.
[00:33:54] Speaker C: Great.
[00:33:55] Speaker B: Had something to do tomorrow.
[00:33:58] Speaker C: Okay. No, no.
[00:34:00] Speaker B: As an editor, you're scanning through and there's no way you're even going to watch more than percent of that content before making all your decisions of everything that you're going to put in. AI gives us the ability to. Again, this thing was talking about before, find all the needles, you run AI against that content, it's going to automatically search through and find all those moments and tag everything so that now, instead of scanning through it on a timeline where you're never going to watch all of it, you can search through it based off of the things that are important to you, the things that matter to you, telling it to look for these people, telling it to look for these kinds of actions, this kind of stuff. And all of a sudden that 200 hours of content, you can actually feel more confident you're getting the best content out of it without having to watch every single second. And I think that's a real big game changer of before without AI.
[00:34:55] Speaker A: Yeah. You could hire people to do that, but you don't have to do so.
[00:35:00] Speaker C: Yeah, you budget that way. Right. And so, Jeff, from my piece of the hype, you gave the great example. Think about it. You're like, what's the height? Right? Well, the hype is that poker's going to like, look at your tells. Right? Teaching a model to pick up a tell is hard. Right. However you think about Jason style and this, this is my whole thing of this, I would. Everybody here is just. Is one Jason. Yes. People. People need a type approach. If you don't have a clear business case, then do not get into AI. You don't because you're just gonna be disappointed. It's not like buying the red camera. We bought the red camera. Why? Well, cause everyone's shooting red now. Okay, cool. Right. But do you need to shoot something cinema style? Well, no, I'm shooting tv news. Well, they don't buy the red camera because that's not for tv news. Right? So, Jeff, with you, you mentioned that it's like. Yeah. Jason was like, no. On screen in this show shows these icons that show what the hands are. Great. That's easy. That's the thing. People are forgetting what Jason just described of doing a documentary. That's where. That's why if you look at our stuff, our whole Akoli unscripted thing is about that. It's cheap. It's just a few dollars an hour to transcribe. And then if you do know that you're looking for greek immigrants, someone's going to say Greece or Greek. And you can search for it right now with a stupid simple, and the easiest AI to run, which is transcription, but just want to say that, like, you can also lower the bar, because what's happening is people are listening. Like, the hardest thing in the world, Nick, is what? Everyone makes me laugh. It's everyone's example. I want you to analyze the plays in football, bro. Are you kidding me? A bunch of people are doing this and it all looks the same. No, that's going to take up. That's going to be a multi million dollar thing. But, hey, find. Every time Pete says Bob Johnson, I can find Bob Johnson right now in 300 hours of your footage. And if you choose those things based on that, you will get. You won't get caught in the hype. Meanwhile, Jason, our teams are over here working on the weird, crazy sides projects, which will trickle down to you. And so that's the piece. But key code. Being able to consult and say, I need. I need this for this reason. I have this much money. You can have AI today. That will do a lot for you.
[00:37:21] Speaker B: Yeah.
[00:37:22] Speaker A: And in case you were wondering, I'm holding aces and aids, so.
Well, one quick question from the audience.
It's actually more about scripted. Robert Harker was asking overlook source of object information is a script breakdown. Script breakdown identifies character, each prop, each costume, and every relocation and set. So script breakdowns are created by the ad identifying everything. Reference script. These can become the tags for identified objects. Any thoughts on how that would work?
[00:37:52] Speaker C: Hold on, let me take this. No, no, no, no, no. Why no? Because you're talking about scripted. Because you're talking about high budget. You have people, and it's not much content. So can Jason do that for you? Sure he can.
Will it be cost effective compared to a good person doing it? Probably not. Unless you're Netflix and you bought a library from someone of thousands of hours. So that's why I want to. I want to stop. I don't want to be weird about it. Right. But I want to stop that line of thinking because that's like, I shot 90 minutes or a television. I shot 40 minutes and I need to go through this. Okay, but that's a solvable problem. Now, if you do own a big library, then yes, absolutely. Right. I bought the Lorimark every, you know, episode of hundreds of episodes. And I need to boot breakdowns 100%. But even in that, I can already tell you, Jason, some of that stuff. Super easy, right? An actor. Sure. I'll break down every actor right now. Location.
I have a weird ray man superhero power. Y'all should know when I watch an old movie, I can tell you where it was shot. The other night, I watched Hooper, Burt Reynolds, 1978. You should watch it. It's an amazing movie. Hilarious. And it was shot in Stowe park in Burbank. And I know that because I knew that park and my friend was with me and he was like, how do you know that AI does not know Stowe park in the background from looking at the hills? Because it's been there. That's where it's difficult, right? Location is hard. Certain things are. But I would say that in that that's a great thing where you need a consultant to be like, maybe not. Maybe so. Again, you have 500 episodes, 5000 episodes, and you have money that you could make if you get breakdowns done very quickly. 100%. But if you shoot ten episodes, see eight episodes a season, and you're doing that right now, I'm going to tell you. I can already tell you, Jason, as you know, like running multiple mode AI to pull all that. That's a lot to train to do and I would need to really be. Now it's weird news, youtubers, people like that. Yeah. Because they're doing content hours and they might need to. But for film and filmwork especially, or scripted, it's. I think that, in my opinion, will be the last place that AI really steps in because I think that when you're dealing with precious small amounts, you will still curate that.
[00:40:17] Speaker B: Yeah. I think the biggest place where the AI is helpful in the film side is when you're looking at something where you're, you know, three films in, or if you're in the fast series, ten films in, and you will be able to find out, you know, when was all the different times that, you know, this car was featured or we did this kind of action or this person said something or these two people were together. Those kinds of things can make sense, right? We can get those kinds of insights by scanning that content and having it.
[00:40:53] Speaker C: But also, Jason, I think the end there in that situation, the bar is lower.
[00:40:59] Speaker B: Yeah.
[00:40:59] Speaker C: Like, we still. So everybody on this call is clear. You still cannot get FCC compliant captions from AI.
[00:41:08] Speaker B: Yeah.
[00:41:08] Speaker C: If you buy FCc compliant captions from rev, an AI company. They run AI and then they have a human do the captioning part. Door slams spooky music so I want to be things that are critical, like a breakdown, like for music and scripting it where there's money and legal involved. But what Jason's talking about, where it's like promo, right? Oh, I got 26 season simpsons and I want to find romantic scenes. Bam. Done. Because that's a low bar that.
No, I need a breakdown that's going to actually pay actors.
[00:41:41] Speaker B: Yeah, yeah. It's a different story. It's like Rice said, it's a good starting point a lot of times, but you still are going to want to have that human element in there to check if it's going to have any kind of legal ramification, financial specific ramifications and things like that.
[00:42:00] Speaker C: I want to give you, all three of us credit for that because there are people right now bsing that AI is going to make their world work. And it's like, I want to appreciate Jeff and the whole key coaching. You have a really sober view of all this. You know, keeping you ahead of technology doesn't mean keeping you five years in the future where you can't use it. Right? Yeah, it's like, you know, keeping you on the cutting edge, but not the bleeding edge. And that. That's something I just want to say thank you, Jason. Because that's another thing that I really appreciate about periphery. Periphery is like, yeah, we know we can't. That's.
It's also a difference of, like, how. How you're funded, how much height do you have to explain?
[00:42:38] Speaker A: But, um, the interesting points about promo, though, is a lot of times you watch a promo or you see a trailer for a film, then you see the episode or the film and you're like, wait, that stuff wasn't in there. So that's the sort of thing where, okay, if you're promo is being fed dailies and there's material that may or may not be used in the episode, being able to kind of equal those.
[00:43:03] Speaker C: Out to the trailer business where they shoot for the trailer as you. As you know, the dirty secret, any of you are not in LA, they shoot scenes that are not for the movie, they're for the train. And that that's happened. Yeah. You know, that's that common. Yeah. But. But I think that that's a great example, though, like, pro overuse news. We're doing a 1.3 petabyte migration to the cloud for a news archive right now. And I can't wait to see what they find because it was just made for tv broadcast daily. They. They took notes and they really don't know. You know, I was. I was thinking about it, like, you know, how many times are famous people not tagged but in a shot because they weren't the focus. Right. Or maybe they weren't that famous, you know, so it's like, you know, we had a president that was a real estate guy. It wasn't a big deal when he was a real estate guy. Now it's a big deal because he's president. So that's a fascinating idea that you find that. And that's where I think it's. When you see Jason, I get really excited because we understand that there's value in that content as opposed to, like I said, the red bicycle. Yeah, I like your next question, but I'm excited also.
[00:44:10] Speaker A: It is a question of who finds value in the content. So I know there are services that record local live news all the time because they have corporations that come in and say, hey, when was our company name mentioned? And they search across the entire archive and grab it. That's the sort of thing where the benefit would pop in. That wouldn't be a benefit to the news organization that produced it in the first place.
I think that leads into the next question I've got here, which is, who do you say benefits the most from implementing AI driven systems? Is it just going to be for enterprise? I know we've got major sports teams, large marketing teams.
Can it make sense for the indie documentary teams?
Can you double types of markets, the size of teams, and the types of content that are best suited for this tech?
[00:45:05] Speaker B: Yeah, no, it's all. I mean, I think really, at the end of the day, and I think Bryson and I both say this, it's the content owners more than anything else, that are the people that can fit most and normally that have the financial incentive to be able to actually invest into AI and components around it. Right. Because it's all about knowing what have across that entire library so that you have better insight. Find what a customer is looking for, a partner's looking for, you know, someone gonna create promos, someone that's doing historical looks into the content. It's that searchability that is so important. And then also on the production side, it's kind of what we're talking about before the news and the documentary. And faith based to me as well, is also a big sector that can really benefit from a budget perspective.
If you're going to do things on prem, you need to have a decent beefy server, you need to have a really good graphics card and those things can cost some money.
But we do offer even some of the basic utilities like the ability to do the transcribe and summarize. If you have a a mam like iconic, we could do some of those things in the cloud. A lot of times what we see is the cloud in many ways is the starting point. Again, like Bryson mentioned earlier, that you can check it out in the cloud and youll start to see benefits and then youll start to see the cost as you do more and more and more, where youll start looking at it. And eventually, depending on how much content you have, you'll probably start to see that, oh, this isn't going to make sense for me to run my entire library through this. If I'm running things as come I could run them through the cloud. Normally that's not too bad, but it's once you're going to get to that, oh, I've got this really big library of millions of assets. That's where it's going to make more sense in our opinion normally to always run it on Prem. And one other thing to that too, because we always talk about so many companies, especially over Covid, have moved so much of their libraries up into the cloud or sometimes into multiple clouds for redundancy. And we're actually having conversations now with customers about implementing more object storage on prem for them to back up their cloud content.
Because it's funny, people always think about oh well, the cloud was that place we were backing it up to and that's our super secure because it's distributed and all these other things. Yet almost every year for the past three years, theres been a moment in time where one of the major clouds was down, sometimes for a day and a half or two days.
And the amount of businesses that completely just lost so much revenue because they had no other way of their content or accessing their content to be able to do other things is tremendous. So as much as the cloud has a lot of redundancy and resiliency, its still not perfect and its still can be valuable to have that backup on site and then especially can be more valuable if that backup on site, even if it's not the entirety of the same size of all your content in the cloud. If that backup can be used to also enrich all that content with metadata, and that metadata can be pushed back, that cloud content, so it can serve as even a temporary space, can be really valuable sound.
[00:48:29] Speaker A: So, Bryson, what are some of the real world challenges you've encountered when you're implementing AI into a media environment? And how have you overcome them?
[00:48:40] Speaker C: Okay, yeah. And let me give you a low, medium high, because I like to give actionable stuff like today, but I will fit that into that. So first off, just so you know, starting, there are weird jumps, okay? There are weird cliffs and canyons you have to get across. So right now, if you're. Because I'm a huge fan of people, DIY, indie filmmakers, if you're doing an indie documentary right now, go get something like Otter AI, which is like three or $30 a month, and upload your audio into that. And then you can get transcriptions. It'll be linked, and it won't be like we're talking about. You won't have them, but you can search transcription, and that is hugely valuable. I do that myself. That's how I prepare all my talks. There's your bottom line. Get that done. Next level up from that. If somebody that wants to, like, for instance, one of our customers is PBR pro bull riders, my nephew, Garrett Jones. Follow him on all socials. He's a pro bull rider and super big thing. But we, they have an economy, and we ran some testing for them, and they're sponsored by Pendleton. And you search for the word Pendleton and find it. And you can do that. Keco, you up with that? You can do a Poc with that. If you own a mam, you can do. Add that for a few hundred bucks, you can literally throw a couple of grand. Get cloud transcription integrated with your mam. If you're running iconic, you can get it for even less. And do that, which me to the point that loves this part when I do it. There are two things that you need to do if you want to implement AI. You either need to have a mail or you need to have intelligent storage, which means object storage. That domic storage can be in the cloud. It could be on Prem. But if you do not have intelligent storage or a mam, it's really hard. You can run AI because Adobe will let you run transcription for free now. Can't really do anything with it because it's just on your system until you can share and retain that metadata. And so I just want to say, if you leave anything here and you're running a facility mid size, at least be like, okay, I need to figure out what object storage is and need to actually finally get a ma'am. And you can make that leap very gently. Challenges are incorrect. Expectations, no idea, no business. Come. Come to us with an outcome, come to us with a project and then you can have success. But I literally just started, gave you a $30 a month solution to deal with doing a small documentary. So it's here now. I think that's a big. That's a big challenge. And the other challenge is the fact that people have not done a good job in the past of preparing their data. And that's what's interesting about periphery and why. Why I'm pitching object stores. You kind of can just dump it in now and figure it out. It's clean and groovy and it's probably a little bit more expensive, but if you had to know what you should do all this time that me and Jason were like, hey, I should get a map. And you're like, oh, we know where everything is. You didn't understand. We weren't trying to get you just to hire us for no reason. We knew this was coming. We knew that you would need semi structured data accessible to some kind of API, whether that's on prem with periphery or in the cloud with somebody else. I just want to say that, like, get it into that. So, like, just to, you know, to be open about other things, wasabi has a solution. You put your storage in the cloud, give you AI. Great. Like, that's the idea of it. That's a. That, you know, that's a. That's a mid sized solution or smaller solution. But I think getting there and then, Jason, you're exactly right for me, I keep pitching. The reason I'm pitching the cloud is that for a few hundred or few thousand dollars, depending on your position, you can actually start doing this.
[00:52:16] Speaker B: Yeah.
[00:52:16] Speaker C: So you mentioned the solution, buying this advertiser. Great. I'll tell you the advertiser. Every time in every sports broadcast that you've done for the last two years, and I can do that for very little money. Then you get an idea, then someone goes, oh, my God, great. Oh, my God, I want to run 20,000 hours. And then you go, yeah, that's going to be. And then you go, you know what? Let's talk to Jeff and Jason. Hey, periphery, what can we do? But what everyone's been doing with AI is literally being like, I want a magic robot in my house that cleans my dishes. And you're like, no, no, no. But we do have really good dishwashers that if you just put the dishes in them, they clean your dishes. No, there's not the Jetsons robot that takes your plate and does it. And I think that that's the thing that I want to get people out of is that quit waiting for that. Just be like, oh, I can do it. Be shocked.
All of this came. All of the north Shore technology really came because I had a transcription that I needed to do, and I found a service on the Internet. I uploaded the video, it transcribed it, I liked it, I saw it had value. I started using it, and then I literally built that into these products. And then once I had turned means I could summarize. Once I could summarize, I could search, and then it just all expands. So that's the piece. But, yeah, I think that's it. I think that. But the challenges are bad. The fact that the data's not in a place that's accessible to this magic stuff. Yeah, it's the biggest one. Let's jump.
[00:53:42] Speaker A: That kind of leads into a question came from Michael Kamas, our old friend.
[00:53:46] Speaker C: Oh, man. Oh.
[00:53:49] Speaker A: Can definitely AI plus run local without some sort of mam and dam.
I can answer that. In our environment, we should have run this on object storage.
However, due to limitations on air conditioning, I couldn't my object storage up, so we were running it on block.
If it was object storage, I could have put all the. All the metadata next to the data, you know, sidecars. I would have all the answers. I believe I still have to have a platform to ask the questions because all the answers next to all of the data still, I need to know.
That's a whole lot of answers. That is equal to watching all my media again.
[00:54:30] Speaker C: So you're separating user experience, right? Yeah, go ahead.
Picture thing. But I want to say that just high level. That's Uihdenhe.
[00:54:40] Speaker B: Yeah.
[00:54:40] Speaker C: UI and UX is a challenge, and that's the two of us aren't worried about.
Let me just wrap this. And Jason, I want. I want you to speak specific conceptually. Jason and I are not worried about AI technology because hundreds of organizations are making the AI better while we sit here talking to y'all. Right. What we are worried about is making it so humans can use it and understand it. That's what he and I are working on right now. The AI is just going to show up, but can you use it? And Thomas and I, it's really funny laughing that because I talk about this all the time. You have it, what can you do? It's in JSON. It's sitting next to my asset. Jason, how do you make that action?
[00:55:21] Speaker B: Yeah, and there's a couple different pieces to that. So within the periphery storage world, one of the things that we have the ability to do with swarm storage is we actually have a concept of annotations within objects. And annotations basically take the ability of adding metadata to an object to a whole nother level. So you can store tremendous amounts of metadata. So all that metadata can actually get indexed into something that is visibly just the same single object on the s three store. So you don't necessarily have to have a sidecar. If you're using other storage, then there would be a sidecar file associated with it. But then as far as, how do you get access to that if you don't have a Mamdez? There are certain things you can do with things like again, in the periphery object storage world, we have a premier panel that goes directly to the object storage. If you have editors who want to be able to search against something, they can actually search against all that metadata. Because in the periphery world we actually leverage elasticsearch, which indexes. And basically think of it as something that is keeping track of all that metadata across all those assets in a single very high speed database that then can be accessed through something like that premier panel. So it gives you complete searchability across your entire library without any mam needing to be involved. And that's great for, again, smaller facilities in different places that might not need a whole mam system. Right. They don't want to invest in a full iconic system or any other kind of system. They don't want to manage the mam system. They just want to be able to get benefit of this metadata and be able to know that they're creating it because maybe they plan to do a mam in the future, but at least want their creatives to get access. And we find the Adobe panel structure is a great way of doing that in a really simple interface that people can access and then also just through the storage interfaces itself. We're working on making that better on the swarm side, but then we also have Vision, which is like a lightweight man. That is something that is just part of the periphery product line, something that you have if you have that you have access to if you have periphery store. So right now, Vision has been very strong with our object matrix storage for a long time. And we have new versions of Vision coming out later this year that have more and more capabilities on the swarm side. And so again, it's not. Vision's not meant to be a direct competitor to some of the major mam systems out there. It's not going to have all those layers of functionality for building out powerful distribution pipeline workflows and doing a lot of those things. But to have an interface to search across all your content and find what it is you need across everything to find those needles in the haystack, it's fantastic for that. And it can leverage all the AI and get benefits from it. And there's one last thing I mentioned too, is that everything we build inside of AI plus is designed to work with any man that's out there. So we show things a lot of times with iconic and we talk a lot about iconic because quite frankly, built over 400 workflows on top of it. So I'm pretty familiar with it.
And we do like that system. But we've done integrations with CapTV. We've done integrations and looked at integrations with probably at least six or seven other major managed systems. And those integrations are very, very straightforward because all we're doing is saying, okay, you want us to be able to add metadata there. Normally two or three API calls. It's less than a day or two of time to be able to integrate that. So the metadata is generated locally, it can be stored directly in your storage, but it's also designed to easily be pushed into whatever mam you happen to be using.
[00:59:15] Speaker C: Yeah, and I think that will be. There'll be less and less. Ma'am, we have a thing called a call me. It's the minimum viable thing to search something. And I'm trying to make less and less and less, less and less. I, you know, Jason, are mam people that don't love the way the man world evolved. So we're trying to change that, I think to speak to commas. I can tell you that there's a certain class of customer at North Shore that will just go to a slack channel or something and get video results. You know, you just go to a chat and be like, what about this? And get it. And then if you want to go in there and look. But there's no reason that that. So I think the evolution will be that this data will be available everywhere. And then, you know, you do that, you know. So at this point, I really. I really believe that a lot of people are going to move into like teams and Slack and Skype, you know, things like that where you can be like, I'm talking to my data. Why do you need to go then if you need to click in? Same idea with the adobe thing. Yeah, I love the panel structure. It's like, why do I need a ma'am? Editors, you know, we joke like, editors don't need ma'am. You know, they, they might ask a couple questions or look or something. But yeah, I think that's a big change too. What we talk about needing. When we say the ma'am, a lot of times we just mean a database somewhere. There's a database that holds this data for you. And periphery is building that in their platform for us. We have several versions of it.
[01:00:33] Speaker B: Yeah. And that's one other thing I mentioned too, is Arduino organizations. One of the things we've seen is that it's not uncommon that they might have multiple mams across that larger organization. And so we're actually doing a project right now where we're putting together kind of a proof of concept system where we have storage in multiple locations, multiple mam systems. And we take all the metadata that we're generating and it's stored directly against the assets, directly on the storage and the object storage. But then we also will leverage a central utility to host all that metadata and become the traffic cop of that metadata. And there's a couple different products out there that can do that. But then you have that idea of a central metadata layer that those mans can pull from and that can be directly from periphery storage. It can be through another system that's designed just to be a central metadata layer. But the idea is that when you have AI generating that metadata, one of the big differences versus having people do it is consistency. You can know that you have that metadata generated out of every single asset in the library and every single asset as it comes in. And that it's always going to be created with a consistent kind of prompting and a consistent process which makes the content so much more findable. And that's what we start to talk about more is it's about content findability. Right.
[01:02:06] Speaker C: Is how content discovery. That's, that's our whole eight b pits content gov. Right. That's what we're, that's what this whole series has been about. Yeah, I. Jeff, how we doing? How we do? Oh, you're muted. You had your plug there. Come on. McClain got mad at you. He turned you off.
[01:02:24] Speaker A: Edit. It happens.
[01:02:25] Speaker C: Let's get into a quick live thing round.
[01:02:29] Speaker A: We're in the relatively early days of AI for media. What parts of the solution today are clunky? How do you think it's going to improve in the next one to two years? Give me your 32nd blurb. Jason, you.
[01:02:40] Speaker B: I think the part that's clunky is the fact that even some of the best models that we have right now still have room for improvement. And we're seeing new models every two to three months. They can have exponential improvements over the previous ones. And I think the quality, the level of that difference is going to shrink and shrink and shrink as we're getting better and better so much.
[01:03:00] Speaker A: Bryson, 30 seconds.
[01:03:03] Speaker C: Yeah, the clunky part is that is the upgrade of it, the transport of the data. Do not let people run AI for you and into a proprietary system that you can't take away because if you're going to spend ten grand or 100 grand or whatever on it, you need to be able to take it away. That's a big one. The other clunkiness for me right now is just, yeah, the inconsistency of results that certain types of AI are better than others. And so, you know, make sure that you test your POC and that you don't go buy something off the shelf that is supposed. There is no one size fits all right now.
[01:03:36] Speaker A: Definitely. So, Bryson, what's the next AI thing future that's going to significantly impact media management? And how should people prepare for that.
[01:03:47] Speaker C: Better integrated multimodal AI? Jason gave you the best example with the facial recognition and the jersey right now. That's got to be built together as a kid. So when you hear that term, I think that's going to be the big deal. When you have a thing that's a pretty good, I think, I mean, I faith that periphery will get there. Like, hey, this is a blend that people see what like people like twelve labs are doing. This is a blend that most people want. So I think multimodal becoming a thing where it's like transcription, facial recognition, object detection, OCR, and then we go, yeah, great division all. Now here's the result. That's going to be the biggest change and it's coming back.
[01:04:24] Speaker A: Biggest change. Jason, what do you think?
[01:04:26] Speaker B: I think it's AI workflow, automation. I think what we're trying to do in our focus is doing the same thing for AI that we did for all the other kind of media tools for years is that we've had all these tools that are great little pieces on their own, but no one needs that little piece on their own. No one needs just facial recognition, just object russian, just vision. You need a combination of all three. And being able to build out those workflows and build out automation between those workflows is what's really going to bring the value and really be able to actually make it worth something. That and the fact of the chips getting lower and lower in cost, your GPU's are going to get lower and lower in cost, become more available. And then there's a whole world of AsiCs that we don't get into on here. But there's a lot of people out there designing chips that are very purpose built for very specific parts of this that could potentially have 30 40 x performance at a much lower cost. And I think we'll start to see those in the next couple of years.
[01:05:27] Speaker A: Awesome. So learning audience today has got a wide range of technical set skills and technical and creative roles. If folks want to get training on either use AI solutions or how to deploy them, what kind of skill sets do you recommend they train themselves on and where to look for those? Jason, where do you think that would be the best thing to start?
[01:05:47] Speaker C: Jason?
[01:05:47] Speaker B: Oh, right now I feel like, kind of like Bryson would say a lot of this. Get your hands on it. Get into the actual different AI systems, learn about prompting. YouTube is an amazing resource. I've probably learned a 60%, 70% of my education of AI comes from YouTube. And it's kind of the place you can learn about all the bleeding edge and then attending events like this and working with the integration partners and asking questions, you know, asking and getting your hands on stuff as much as possible. That's the best way to learn.
[01:06:18] Speaker C: Yeah, well, for me, short term right now, yeah. Start using an LLM, right? Use chats. One of the coolest things if Google customer is if you have Google workspace, you can turn Gemini on, on your Google workspace and you can actually use that. And you get for your gmail, you can summarize email. Once you learn that, why is because it's just like learning to look like it. Like when you're learning an edit tool. What could I do? Because all those things that do with those LLMs and that level of understanding or lack thereof is what constrains periphery, north shore, these companies that are using that. So those tools as they get better. But like I would say, you know, I don't know. Jason, what do you think? So chat, GPT, Gemini and Claude, is that kind of a start for people there? If you start there and start playing with it and start solving problems in text data right now, you will start to understand because it's kind of weird. Everything we've talked about kind of comes back to converting things to text and then doing something based on the text. So everything Jason's talking about, you're getting text output metadata, and then that metadata is parsed. So if you understand what large language are doing and can do, then you can come to key code, Jason and I, and say, I have an idea and I do this. Could we do that? That's the thing right now, everyone. If it's all theoretical, you don't understand it. That, and also make your boss happy because an LLM will save you so much time. I don't even like reformatting spreadsheets. That kind of stuff all done with me. The other thing is, like you said also, you know, we mentioned, like, start getting and looking at. You can go to Google's computer vision page and you can upload an image and just look. You get back, there's all kinds of goofing off and then, yeah, man, YouTube. Be very careful, though. Focus on functional. If you're a content producer, don't go off and watch a four hour, like, oh, Jason's watching a four hour thing on, like, you know, the lava models.
But how do I use this in my day to day? Cause as you do that, I promise you that's one of things, like pitch out. Cause he's watching Collis. Michael Collis messes around with things he can do and then some bigger idea and extrapolates. And I've watched him for years do it. And I know Jason, you do the same thing. We play with something, and then we're like, oh, we could do that. I think our biggest problem right now is our customers aren't playing with this stuff. So they're not bringing us these crazy ideas. They're bringing us ideas that might be too far out. Right? So if you play with it, you're like, oh, it can do this. And you will be shocked. You will, you would. Pleasantly surprised in certain areas.
[01:08:54] Speaker A: One of the things, a couple things I'm seeing on the 30,000 foot view, there was a mailing list called TLDR, which is, which is kind of commercialized. So it's kind of fallen off for me. And then I've got an app on my iPhone called substack, and it's just all these different channels about all sorts of different tech things. And there's a ton of AI stuff in there. Gives you the quick blurbs of, okay, what's coming out and going about Gemini weeks before you see it all right. So one thing that one audience that we want to get clarified from Matt, does it cost $1 per video or $1 per process? AI in the cloud. So what if the video is 30 seconds versus 3 hours? What's the good rule of thumb in there so people can figure out budgets?
[01:09:46] Speaker C: Wait, there is, it's, what's a hin way? There is no. Right. So like, I guess in Jason, you work with this, but in cloud services actually, by the way, here's the deal right now. You can go to Google Azure or Aws. There's a calculator. You go to the AWS calculator and you can be like, I want to run transcription and I want to do it on a thousand hours. If it's video content, it's done typically by per minute, is normally how you could calculate it. If it's images, actually go, go price images. Because it's insane. Like, it's really bizarre. It's like, like we have customers that like, gave us, like, I remember we laughed. They were like, oh, we're going to do a few thousand a month. And it was like, I don't think it actually, I think it was a zero and didn't even read. So there's a free tier, by the way, of all those services. So what does it cost? Transcription costs anywhere from a couple of dollars an hour to the more expensive services. Might be ten or 15 depending on how it's integrated into stuff. But typically in, in two to $5 an hour, you can get transcription done for you on the cloud. Jason, I don't know, like cloud wise, I object detection is the most expensive, right? LLMs are very expensive. Yeah, but they said, I can't even begin to explain how LLM is build. I don't think we can do that on here.
[01:11:07] Speaker B: No. I think that the key thing that I was driving at though, too, when I talk about how we did some testing, we were seeing like a dollar per video. It's because there's so many different processes. If you want to really know and understand a video, the first thing you want to do is extract the transcript and run that through an LLM. So you got a feed there, then you have to run it through the LLM for every prompt that you want. So let's say you have 15 metadata fields. That's 15 times you're running that thing through the LLM. Then you take the visual side of it. And if you want to understand what's happening visually now, you have to run that through a vision model. And vision models are much more expensive than text models, especially online. If you want to get a good understanding of that video, you probably want to sample it at least one frame per second. If it's high action, if it's low action, maybe five or ten frames per minute. Every single one of those frames is another call to the LLM, is another charge.
All of that quickly adds up that even in a ten minute or 15 minutes video, you'll be over a dollar a video just to be able to process that. So the thing is, don't be caught up in the fact that these LLMs will say, oh, well, you know, a single API call only cost $0.10.
You don't do much with a single API call.
API calls, one question, one answer in a certain amount of tokens.
And tokens are a whole nother complicated thing to get into. Yeah, that's why it adds up, Mary.
[01:12:35] Speaker C: Yeah, and that's why I pitch talking to key code, something like that. Talk to an integrator and do a proof of concept. Because you can say, I need to do this. How do this for me. You can come to us, you can come and talk to Jeff and say, jeff, I have 50 hours of stuff. We go, great, let's do this. And we run it on the periphery system, we run it in the cloud, whatever, and then say, okay, what does it cost to set up a periphery to do this with system? What does it cost to do this in a best we may. What you really going to see is we're going to talk you down from your amazing big idea down to something simple. I keep the transcription because people tell me, transcription. If you just did that concern so much, if you like, if you just did logo recognition, oh my God, if you're in sports doing logo recognition, things like that. So that's the thing is like it's where you need a caddy. You need to come to someone like key code and say, I need you. This is what I need. I need to hit that green from here. And they go, great, that's a night iron. Do that. And that's what you really, right now, that's the biggest thing is that this is not commoditized. And I think that's one of my biggest worries about the people that are like,
[email protected]. yeah. And you can sign up and it's cheap and it's easy.
I think that's really going to disappoint people. And so I have respect for Jason because I keep pointing at him because he's Brady bunch down here, because he doesn't over promise, he codes the same idea. It's a very sober view. I keep using that. It's a very sane, real worldview of what we can do. But there is no such thing right now as you run this other than transcription. Right now you can go get a price for your hourly transcription online and get that.
[01:14:11] Speaker A: Actually, I already got that for you. That's the wonderful thing about having a team. Someone ran the transcribe calculator for AWS. Forty cents a minute. There's nothing there. Yeah, no, just transcribe. And if you're watching YouTube, there should probably be a link here right now for the AWS calculator.
[01:14:27] Speaker C: But that sounds. Yeah, but wait, that sounds high. Just so you know, because I know that on iconic you can buy for a couple of bucks an hour and I know that in a coli you can do it for. So just know that you should look at that because that may not be per minute, but anyway, there's ways to.
[01:14:42] Speaker A: Figure out your budgetary pricing and then there's ways to get economies of scale by utilizing other folks. So we're out of time. Big thanks always periphery Bryson and North Shore for joining us today. Azure reminder key code media is the systems integrator of choice for all things live production post production av. If you're wanting to schedule a free whiteboard consultation on AI workflows, please hit us up here at Key code Media. Our apologies to Matt Damon. Thanks everybody for joining us today.
[01:15:10] Speaker B: Thank you.
[01:15:12] Speaker A: Thanks for watching broadcast. To follow us, please make sure to subscribe to the podcast to receive future episodes. Follow keycode Media on LinkedIn, Twitter, Facebook or Instagram to receive news on additional av broadcast and post production technology content. See you next time, folks.
