Voices of Video
Explore the inner workings of video technology with Voices of Video: Inside the Tech. This podcast gathers industry experts and innovators to examine every facet of video technology, from decoding and encoding processes to the latest advancements in hardware versus software processing and codecs. Alongside these technical insights, we dive into practical techniques, emerging trends, and industry-shaping facts that define the future of video.
Ideal for engineers, developers, and tech enthusiasts, each episode offers hands-on advice and the in-depth knowledge you need to excel in today’s fast-evolving video landscape. Join us to master the tools, technologies, and trends driving the future of digital video.
Voices of Video
Training Your Own LLM: The Untapped Value of Media Archives
The AI revolution in media production is moving beyond marketing hype into practical applications delivering real value. Ryan Jesperson from Cires21 takes us deep into how their Media Co-Pilot platform is transforming workflows for tier-one broadcasters through thoughtfully implemented artificial intelligence.
Unlike basic AI integrations that simply bolt onto existing systems, Media Co-Pilot addresses the nuanced needs of professional media organizations by training models specific to broadcast content. This approach solves the "garbage in, garbage out" problem that plagues many rudimentary AI implementations, delivering broadcast-quality results for transcription, translation, face detection, and scene analysis.
What makes their approach particularly valuable is the hybrid integration model combining public LLMs with private, custom-trained solutions. National broadcasters with decades of archival footage aren't just protecting their workflows – they're safeguarding valuable intellectual property that could otherwise be scraped and utilized by competing services. This represents a fundamental shift in how media companies view their content libraries, recognizing them as valuable AI training data beyond traditional monetization channels.
The real-world applications are diverse and compelling: reality show producers tracking contestants across multiple cameras with trained face detection; concert promoters generating social media clips from live feeds to drive viewership; news organizations rapidly translating content across languages and platforms; and contextual ad placement based on content sentiment analysis. Each implementation is tailored to specific industry needs rather than forcing one-size-fits-all solutions.
Looking ahead, the next frontier lies in applying AI to video distribution itself – optimizing just-in-time encoding, improving device detection, and predicting viewing patterns to reduce infrastructure costs while enhancing viewer experiences across devices and scenarios.
• Media Co-Pilot brings AI capabilities to complex streaming workflows while maintaining broadcast quality
• Creating private LLMs allows broadcasters to protect valuable content while leveraging AI capabilities
• Face detection and object recognition enable efficient reality show production across multiple cameras
• AI-powered scene analysis improves ad placement by understanding content context and sentiment
• Social media clip generation from live events happens automatically through AI processing
• Transcription, translation and dubbing workflows become more efficient through trained AI models
• Media archives represent valuable training data that companies should protect from unauthorized scraping
• Future applications will focus on using AI to optimize video distribution and just-in-time encoding
• Neural Content Processing servers enable faster training for customer-specific AI use cases
Ready to explore how AI can transform your media workflows? Connect with Ryan Jespersen to see Cires21's Media Co-Pilot demonstrations and discover the practical applications of AI in professional media production.
Stay tuned for more in-depth insights on video technology, trends, and practical applications. Subscribe to Voices of Video: Inside the Tech for exclusive, hands-on knowledge from the experts. For more resources, visit Voices of Video.
voices of video voices of video.
Mark Donnigan:The voices of video voices of video well, welcome back to another very exciting session of voices of video Video. So I'm Mark Donegan. I am here with Ryan Jesperson from Cirrus21. Ryan, you and I go way back. It's good to see you.
Ryan Jespersen:Yeah, it's good to see you again, mark. We tend to see each other NAB and IBC, and then virtually in between.
Mark Donnigan:That's right. That's right, and I think I've had you on I don't know at least three of the various podcasts that I've hosted or co-hosted. So, yeah, it's wonderful to be talking again. So I know today you're doing some really exciting things in AI and we're going to talk about that today, but it's sort of a little bit of a preamble and a setup so that you can tell us exactly what you're doing. Everybody's talking about AI. It's like a requirement. If you go to a website, it has to lead with AI. Go to IBC, every booth's going to have AI on it. Why don't you tell us what you're doing and why it's special and unique and the ground that you're breaking?
Ryan Jespersen:Absolutely, I think.
Ryan Jespersen:Well, I don't think I have to explain why AI is special.
Ryan Jespersen:I think it's quite obvious to all of us and I think the amazing thing about AI in the way we're using it is you know, Sirius 21 and then myself included have 20 years of experience in live streaming and in the media and entertainment space broadcast and so on and I think a lot of what we're bringing is a lot of our experience as a company and personal experience to create workflows and use AI in workflows that are specific to our industry.
Ryan Jespersen:So you know, everyone and their brother is doing an AI integration and you can go and do an integration with Whisper or LLMs quite easily and build them into your pipelines. But there's actually quite a bit of additional work that needs to happen to be able to make those products usable, to make it accessible. And then the democratization of the production of video and media has been happening since the 90s, when we went analog, digital, digital to cloud, and now what AI allows us to do is take some really complex streaming workflows and make it accessible to people working in editorial newsrooms or people who are working on social media production from a live event. These kind of additional kind of product gaps is what we're trying to fill with our media co-pilot product.
Mark Donnigan:Media co pilot. Is Media Copilot the implementation of AI, LLMs and all the goodness of what artificial intelligence brings into your existing platform and pipeline? Or does it sit outside of that, or how is it integrated?
Ryan Jespersen:Yeah, so it started with some R&D work that we did quite a few years ago. So Cirrus21 hired a really great engineer called Juan Casal, who used to be at Nokia, used to kind of head the R&D department there for Nokia Video, and a lot of what he's been doing at C21 over the last couple of years is being able to apply a lot of that R&D research around AI specific to these streaming workflows. So, to give you an example, our R&D platform we built it in the cloud. We can workflows. So, to give you an example, our R&D platform we built it in the cloud. We can go integrate with a variety of different LLMs. We can even work with tier one broadcasters, which is kind of where C21 has been.
Ryan Jespersen:I've always described a Cirrus 21 as kind of the elemental of Spain. They work with every major national broadcaster. They work with Olympic broadcast services, mediapro, ebu in Europe and because they work with tier one broadcasters, a lot of the products that we're trying to do is make them tier one friendly. So it's very easy to go and use. You know, google Meet and use Gemini to do a transcription, but is it broadcast quality? Can you actually use it to do real dubbing work that can be used in movies or TV production or for tier one La Liga games.
Ryan Jespersen:You know, these kind of things is where maybe a lot of those AI integrations you need to train those LLM models specific to your content, and that's another piece of this is kind of the privacy side of AI and there's a lot of news coming around about copyright who owns that content? That's right. The value in kind of creating your own LLM. There's a value in doing that and not using just a general purpose, publicly available LLM. Publicly available LLM so if you are a tier one broadcaster in any country and you own the content, you don't want the kind of intelligence you have in analyzing the content to be available to other broadcasters or to other people.
Ryan Jespersen:You own it right. So we're working a lot with these kind of hybrid integrations where we can leverage the power of public LLMs in addition to the value of a private LLM to marry those things together and create much better models to be able to use for a variety of these kind of use cases. And this could be, you know, multimodal analysis for live transcription, speech to text, which is quite a popular subject with things like Whisper and a lot of other technologies that are out there, but the multimodal analysis of trading face and object detection. So if you're doing like a reality show and you can train the model with a hundred pictures of a particular contestant and immediately we can analyze that and be able to help with the editorial production process of doing daily reviews, or you know, the use cases are actually quite varied.
Mark Donnigan:Interesting. Do you find that? You know you obviously have to start with some core set of features, I guess you would say or functionality, but does every broadcaster, every customer have their own unique, either twist or requirements, or is there some universal functions that you can build, you can bring to market? I'm curious what you're into.
Ryan Jespersen:Yeah, I think we're starting to see a lot of these broadcasters like the tier one broadcasters, national broadcasters, globally put out a lot of tenders. Now the broadcasting market tends to be a little bit behind some of the streaming space, but they are starting to put tenders out because they realize the value of this. So the use cases we're seeing are varied. There's some very, very obvious ones and there are other ones that are quite interesting. So I'll talk more about the traditional ones. I think, when it comes to an editorial room, if you're like the Associated Press or Thomson Reuters or Washington Post, you're producing a lot of video content and you need to very quickly take that video content whether it's live or video on demand, and be able to create a draft of an editorial piece that a news professional can then take and go write the blog, to go around it, to go write a script to help with the production of a documentary or a live news report in Gaza or whatever it happens to be, also to translate that, so you can have local language content that's being recorded and immediately create a language version for English, spanish, whatever it happens to be and then give you a starting point. So I think there's the news production side of it which is very interesting from an editorial perspective.
Ryan Jespersen:I think from a broadcaster perspective, we're doing a lot of like scene analysis. Try to create you can give it like a 95% margin where we can try to find the best break in a video on demand, in a fast channel or in a live production workflow to be able to be like no, this is where the scene break should be, so it doesn't cut in the middle of content. You even see like youtube doesn't do the best job to actually do a very clean cut when it comes to to scene analysis. So we can do that with ai and then we can also inform the ad server. This is one example that we're doing with broadcasters of what is the sentiment of and the context, context of what's coming before and after the ad break. So if you're in the middle of a, a news report, and they're talking about a conflict of war, you don't want to be selling disney uh, world trips, right? So these things can be do, can be, can be used to create better metadata around informing and improving the ad experience, wow.
Mark Donnigan:Yeah, I can certainly see the value there. So it strikes me, though, that there's a lot of systems integration required, and let's just talk about the last use case around advertising. Who are you integrating with, or are you now being drawn into, building your own ad product? You know, or, or how does that work? Um, so that you can inform the ad decision engine? Uh, you know the DSP and, uh, and and and the other parts of the uh advertising stack. You know to not show that Disney ad when you know it was just showing a natural disaster somewhere, or whatever. You know something that would be maybe not so appropriate.
Ryan Jespersen:Yeah, we're definitely not getting into doing an ad service. That's not our. Our model. We come from the kind of live streaming space what we want to do is add value around improving the live streaming experience and even for video-on-demand streaming as well, we can go, analyze the library and create all this metadata to be in a sidecar file and use in parallel to whatever your streaming pipeline happens to be. A good example of this would be we can do an API call to our media co-pilot service. So we very clearly wanted to build media co-pilot as a cloud service with an API. We do have a web GUI as well, so you can go and use this in the context of a user interface, but the API can do a call to your content, analyze the content and then create the metadata return to your streaming pipeline that you can then use to kind of inform the media server.
Ryan Jespersen:Another good example of this pipeline is it can be used in parallel with streaming workflows to do.
Ryan Jespersen:You know, Apple has their new immersive format that they've been announcing and putting a lot of money around to create just better immersive experiences.
Ryan Jespersen:What we can do is go analyze a piece of content not get involved in the media distribution side, but do create the metadata necessary to go and deliver the media the immersive format for Apple. So this is something where you can use, maybe, legacy streaming media components, whether it's FFmpeg or GStreamer or whatever your packager happens to be, if it's GPAC MP4 box or if it's Shaka, and without having to update all that, you can call to our API and we can add all this additional metadata to help with all these experiences, whether it's improving the video and audio at the distribution level or whether it's adding the value to creating these better ad breaks with your service. So I think that's a part of not having to go and do major updates to streaming ecosystems and pipelines and use it in parallel to add a lot of AI to this value. So that's a lot of what we're doing with these legacy broadcasters because they have very complex streaming ecosystems, yeah, and without having to pull out all the plumbing, we can do this kind of in parallel to their network.
Mark Donnigan:So it sounds like, uh, what they're really asking is for your service to return. You know API calls, metadata et cetera that then they are taking and doing something with that right. You know, either integrating themselves into, maybe, systems they build or then going and working with their other vendors. Is that the relationship, then here as to you know how they're actually actioning?
Ryan Jespersen:media co-pilot yeah I think, what we're looking for kind of in the infancy here. We've had media co-pilot since ibc last year, so it's almost a good year. It's a bit in production as a self-service sas product, whether it's through the api or through the web gui. What we what we really found is broadcasters have very unique needs and this isn't just broadcasters is anyone who's trying to do something with media, whether it's a video or audio, you know? So I'll use voices of video as a good example. You can get this webinar in 30 minutes, 60 minutes, and I know you write a lot of LinkedIn content. You can take the video, throw it at Media Copilot, upload it or point it at an S3 bucket or wherever your storage happens to be for the file. We'll analyze the file. We'll create a first version of a summary that you can then use in LinkedIn. We'll create automatic clips and highlights that you can then reformat for social media in 9x16 or in 4x3 or 3x4 or whatever format you want to be.
Ryan Jespersen:Think about that pipeline and trying to do that. Riverside's a great tool, which is what we're using here as an online SaaS tool, but you always have to take it into an editing tool, whether it's a frameio or a Premiere or a Final Cut, something. We do that all for you. So, whether it's creating the highlights, which then you can actually choose the in and out points so we can render those MP4s for wherever you want to put them in your MAM, in your CMS, in your social media, content that can all be done automatically, content that can all be done automatically. And then the other thing you can do is take the same content and be able to then create dubbing for it that you can then dub accurately. So there's just a.
Ryan Jespersen:This is so commoditized now, but with any commodity, there's good versions and the really poor versions. And what I'm seeing with a lot of streaming vendors I don't want to name names, but there's a lot of very, very rudimentary integrations with things like whisper, because it's so easy, but they haven't trained the LLMs to actually create very good translations or very good transcription to begin with. So it ends up being becoming garbage in, garbage out, with a huge margin of error. There's no good speaker detection, there's no good scene analysis. So all that additional data, you know, I think that's the problem with what you see now at NAB and IBC is everyone has an AI logo in their booth, but how many are actually doing it. Well, I'm trying to kind of get through the noise and fight through the noise to actually try to create something that's tier one or that's even better, you know.
Ryan Jespersen:For then your rudimentary integration is really key.
Mark Donnigan:Well, let me give a shout out. Thank you for that, because I totally agree with you. There is a lot of noise. Finding the signal is incredibly difficult, so let me give a little bit of a shout out. You are going to be in our booth, meaning Cirrus 21 will be in our booth at IBC, and we're very excited to be hosting you along with seven other partners, so it's going to be quite a party in the NetEnt VPU ecosystem booth, but it'll be a great opportunity. I assume you'll be showing Media Copilot there. Is that correct?
Ryan Jespersen:Absolutely. We'll be showing our integration with the VPUs and the density that we can get, and a couple of my colleagues, Dennis and Nacho, have been on Voices of Video and they've gone deep into what we can do with NetEnt VPUs. And then we've also launched on Akamai Linode, Akamai Cloud with VPUs and with C21 as a live origin and live control pipeline for orchestration as well, and so the density. We demo that and we show the performance of what we do with very high VMAP scores with also huge power savings, which is really what NetIn offers, right, that's right the hardware that you guys manufacture. But then, similarly, we'll be showing a lot of the media co-pilot demo, just like we were in your booth at NAB.
Mark Donnigan:Yeah, amazing, amazing. Well, let's talk about where um, you know it's, it's always popular, right? Where are things headed? Where are we going to be in three years? And I want to back up a little bit and not so much talk about, you know, crystal ball stuff, because you know that's fun for us but I think not very meaningful, uh, for those who are like, okay, fine, I guess I'll have to wait five years to see if your predictions come true.
Mark Donnigan:But I am curious. You know you are, I like your positioning. You know you are, you know you're tier one for European operators. You have some very important, very large customers who may be operating in different size markets than other parts of the world, but are no less tier one. I think that's very important to note.
Mark Donnigan:Where do you see the joining together of where these AI tools, ai technologies, llms? You know, when you just sort of wrap it all up, you say AI is just going to be, you know I'll call it core features or core functionality in video workflows. Do you see that coming anytime soon? Is there a place for just pure just? You know companies or products you know that are just basically high quality video. That's it, just stripped down, it's like back to the basics, or is it going to be in a short period of time, or some period of time that if you don't have AI, like it doesn't matter how good your video encoding is or you know your core function, like you're not going to be competitive? I'm just curious what, what you're seeing and what your general sense is.
Ryan Jespersen:Yeah, my overall view whether you're using chat, gpt or in a way that's kind of like your your grandfather, or if you're actually going deep and adding AI is it just makes you more efficient as a human. So the whole idea of it replacing jobs, I've never really subscribed to that. It just allows you to do a lot more with less.
Mark Donnigan:Yeah, I'm totally with you, by the way, on that point.
Ryan Jespersen:And so one of the use cases we're working on right now is with a very large concert producer. You know they do hundreds of concerts a year from a production facility perspective and they're working with a large right owner that is doing this with our live streaming OTT platform for a very large concert coming up this summer and what we're helping them with is to they want to be able to record kind of run and gun content on the fly with an ENG camera in the field, a high vision Makito, into a decoder and then they want to produce social media clips really quickly to go to social media and push everyone to the live stream where they can.
Mark Donnigan:Which is a great use case, by the way. I mean it totally makes sense to me. It's great for the fan, it's good for the artist, it's good. I mean it's good for everybody.
Ryan Jespersen:So you have an artist coming on stage and they can do a quick behind-the-scenes interview, run and gun, create it In a human workflow. We're having to download the content, be able to edit it, reformat it for 9x16, upload it. All that can be done with AI and we already do with our Media Co-Pilot product. So I think the evolution of where we are in our product there's some really good core tech that we've done. I think we've built a good system architecture and a lot of really good integrations with very good products and have the lens of several broadcasters. What we've been missing is a productization of this. Sometimes in a lab you get a little distracted, you get too focused on the technology and you forget about the use cases. So I think this is the lens that we're bringing right now is working with very specific use cases. I think this is the lens that we're bringing right now is working with very specific use cases. One of the ones I really enjoy right now because you know, I was one of the part of the founding team of Millicast, you know, and Millicast started in the remote production sector where we were helping. We helped to build Evercast and helping with daily reviews and over-the-shoulder reviews for pre-production, production and post-production right With our WebRTC-based browser-native platform. But we're kind of going the same way with one of our use cases with Media Copilot.
Ryan Jespersen:Is there is like a reality show that's being produced with a tier one broadcaster and they're shooting. I think it's like 50 cameras, A lot of them run and gun in the field, a lot of them IP cameras, a lot of them in-studio production, and they want to do face and object detection, they want to do speech to text so that for the post-production which post-production is kind of a term that really doesn't happen in Hollywood these days it's actually pretty much happening in parallel to the footage getting shot. So as people are producing, the daily reviews is really in-the-moment reviews. Where it's coming into the MAM, it's immediately available. And what we're doing with this use case is we're actually adding all this additional metadata so that the post-production team can see the time code sync of all these 60 cameras and be able to know who's in each shot, so who's the winner of this week's reality show, or.
Ryan Jespersen:This is an unscripted use case, but this could easily work for scripted content as well. So this wasn't originally a use case that we had thought about with Media Copilot, but working with this vendor and this tier one broadcaster, we've now opened up a whole other vertical for our AI to be applied to. Very cool, yeah, A lot of Hollywood content is. You know, it's kind of old school, right. It's not just broadcast, it's actually even a step beyond that where they're doing things in a very traditional way for the last 20, 30 years, and here we can add AI to make the actual person doing the daily reviews or the person boots on the ground actually almost be part of the editorial process, instead of just the production assistant or production person.
Mark Donnigan:Amazing. Well, I love it. I, you know I like that, even though the use cases you're talking about are, you know, sort of one step removed, you know, one degree removed from, like the core of, say, video encoding and distribution, which is, you know, as a video encoding vendor. You know, creator, the VPU category. You know that's where we spend 99.5% of our time. You know is talking and working in that area, but it's still very much related, you know, to encoding and distribution, because the use cases you're describing are are the inputs to these massive platforms.
Mark Donnigan:And you know, I think you know there's a lot of discussion around. You know how consumers are consuming media. You know their time spent on streaming services versus broadcast. You know versus pay TV and and of course the trends have been known for years people are sort of fading out from some Some of those more traditional outlets. Pay TV could be one doesn't mean they're necessarily stopping completely. You know Comcast still has, you know, tens of millions I haven't seen their numbers in a while, so I don't know off the top of my head but you know still a lot of subscribers. I haven't seen their numbers in a while, so I don't know off the top of my head, but there's still a lot of subscribers.
Mark Donnigan:But these use cases like going back to the concert, I mean, it's so in the moment and that's what consumers are looking for. That's what viewers, the audiences, are looking for. They want to be at the racetrack, at the F1 race, at NASCAR, at the game, at the concert, at the end. I think sometimes that half of the experience is not what's happening on the track or on the stage, it's the things around it. At least that's my view. It may not be true, but so it's cool that you're enabling this is what I'm hearing.
Ryan Jespersen:I think one of the predictions is that I have about AI and how it relates to kind of streaming and kind of. Like you, I spent the bulk of my career actually on the vendor side from a. You know I worked at Wowza from the media server side and the infrastructure side, and even Millicast was more of kind of an architecture and a CDN, you know, and an origin From that side. I haven't seen AI used very much yet. So most of AI has been around. Let's add metadata, because there's a lot of money around metadata or workflows that can help. You know the ones I've mentioned previously.
Ryan Jespersen:Where I think AI can be used and I think it will come is is the actual distribution side. So there's a lot of just-in-time encoding, research and products coming out. Where you're trying to, you know there's a commoditization of the infrastructure and the systems side of delivering streaming packets and streaming content. But what about creating? You know you have a particular person who's asking for a video on-demand file. You're not only doing a just-in-time encode, you're also detecting much more information about what they're detecting, what device they're on. You know, did they pause and go away and then come back on a different device, which happens quite commonly. How can ai be used to actually predict some of this and save a lot of the architecture deployment which right now is quite cumbersome in a lot of these legacy streaming ecosystems to do that?
Ryan Jespersen:So I I think AI, applied to streaming workflows, whether it's live or video on demand, is a place I haven't seen a lot of innovation yet and I think that's probably in the next year or two you're going to start to see a lot. I think you already get a lot of the big kind of, you know, walled garden platforms, be it Netflix, be it Prime Video, who are maybe applying a lot of that, because we're talking about just huge cost savings, but I think a lot of the other streaming OTT services that are in our industry, I think they haven't done a lot of this yet because they really haven't had the budget, I think, to go and invest.
Mark Donnigan:Well correct. And it strikes me one of our taglines, or a phrase that we often use in our marketing and when we're out in the market, is it's well understood now, because we've actually been a part of educating the market, that YouTube runs largely on ASICs or VPU they call it something a little bit different, but it's an ASIC based encoding architecture. It's how they can be so insanely good and profitable at the same time. Meta has their chip. You know there's, there's other large hyperscalers in China who have built their own chips. So so the fact that dedicated hardware you know ASICs are coming into the network is not like, oh, that's unusual. You know the largest platforms in the world are doing it, but unfortunately and that's the opportunity for NetEnt, was the opportunity, is the opportunity None of those platforms have opened up that technology to anyone outside of their egos, anyone other than, like, google, it's only YouTube, it's not on GCP, you know.
Mark Donnigan:And so our tag is we built this for the rest of you. So it's great If you're meta, if you're Google, fine, you know you've got the uh, the balance sheet to go finance a $500 million chip project. Um, but you, you know, for everyone else, they, they don't have that, and so it sounds like that's, uh, very similar to to what you're doing you know?
Ryan Jespersen:yeah, I think so. I think with vpus and what you guys doing with asics, you know, early in my career actually was it was in eda space and worked on fpga, asic design and all that kind of stuff. And it's true, like I look at the efficiency numbers of of what NetInt, uh, asics can do with their VPUs, and it's a 10th of what you can do with general purpose, whether it's a GPU or an FPGA or or or even CPU resources. Yeah, cause it's, it's a obviously application specific. It's built just for this um task, I think. With AI, I think you're going to see an evolution there. On the hardware side as well, I think there are some general purpose. You know hardware that's getting built, but I think it's specific to video. You might see an evolution there as well. I don't know enough about it to comment educatedly on it.
Mark Donnigan:Yeah, well, you know we have, and you know we have a small computational block on the VPU, on the second generation, the Quadra. It's 15 tops. If you have our two-chip PCI version, the T2A, it's actually 18 tops per chip, so it's a total of 36 tops, which is you know. If you're not familiar with, you know with what that means is. That is that you know low power, medium power or high power in terms of computational power. You know it's maybe more towards the lower side, but there are some more basic use cases that can provide a lot of value. You know things around like scene detection. You know it's interesting that you you mentioned that um that's actually. You know there's models that can be run directly on the VPU. That um will reduce the throughput, but not so far that it's not practical.
Mark Donnigan:You know to implement there's things around face detection, not necessarily that you would. You know use it to see oh, is that Mark Donegan or Ryan Jesperson, but it's a human and some information around that that then could be fed back into. You know into encoding optimization. You know into rate control. So there's some interesting things coming. You know that we're certainly looking at Our third generation chip will be taped out early next year and you certainly can expect that we're going to be investing a lot more real estate on the chip for AI. We're being a little bit tight-lipped to say exactly what that's going to look like, but, believe me, we're watching very closely what the use cases are that are emerging, how people want to use AI in video encoding and video processing workflows, and we think we can bring some more added value there.
Ryan Jespersen:Yeah, I think it's an interesting space. I think another prediction you made me think about is it's not even a prediction, it's already happening. So Cloudflare quite famously blocked the, you know LLMs from going and scraping their content right, and I think that's the next evolution. It's already happening where you know, broadcasters are finding a hard time to create value, Like I saw, the Skydance merger acquisition you know went through last week finally got approved. But you know, these huge volumes of information that broadcasters have or that content producers have, that's worth a lot, that's a lot of IP, that's a lot of copyright. I'm not going to mention the politician but very famously last week said that's not our problem and it's such the wrong thing to say. I think you're going to see a lot of acquisition happening around, not trying to monetize content, but scraping content and legally scraping content and licensing people to go and look at your entire library Because there's an argument to all the knowledge in the world that's in books and there's HarperCollins, there's publishers who own that content, plus the owners themselves or the authors themselves, is there's an enormous value and that's an unrealized value for LLMs to become more intelligently trained.
Ryan Jespersen:So I think the thing that we're already seeing is the tenders that we're winning from an AI perspective are very focused on. We want to retain privacy and control of our LLM. We can still use the cloud to go deploy this, but we want it to be in our own private cloud, not to be accessible to anyone outside of that. Those privacy concerns are why, yes, we have a self-service SaaS that runs on public cloud and uses public LLMs, but really a lot of what we're seeing in our POCs that we're working on and some of the tenders we've already started are actually based on.
Ryan Jespersen:Let's go create, start creating our own LLM, because that has a huge intrinsic IP value. When it comes to being the national broadcaster of any country, you have sometimes 75 to 100 years of content. That is worth a lot to a a lot of people, and I think that is going to change. Instead of us going and making money as an ott provider, which is hard to do, it's kind of a race to zero. Yeah, the netflix's and the prime videos of the world here we can go and actually sell this content to to LLMs or people looking to-.
Mark Donnigan:Well, there's a model there, right?
Mark Donnigan:I mean Reddit famously, two and a half, three years ago now, really challenged OpenAI, based on at the time anyway, and I haven't followed all the details, but I read enough headlines to know that you know they're like hey, you're not just going to be scraping Reddit, you know, and getting it all for free.
Mark Donnigan:And then other publishers followed suit and in some cases, again off the top of my head I don't know the numbers, but you know they're able to get my head. I don't know the numbers, but you know they're able to get, as the saying goes, real money because there's real value in that training data and you know. So Reddit or any other publisher can say, no, you're not going to A scrape for free, and yes, it's going to cost you and it's going to cost you, you know, like a significant amount. So that's a new revenue opportunity as well, which strikes me as um is what every media publisher and vertically integrated you know content owner media house is is looking at and certainly is, uh, actively trying to figure out how to unlock the value there and, I think, retaining it as a content producer.
Ryan Jespersen:when you look at YouTube, which is the largest streaming service in the world, you sign off your rights right For your content to go and get scraped and you know Google's been training their models based on that data. Yeah, it's an interesting time. Yeah, I think you know, if you've been in the industry for like 20 years and you're consulting, that consulting.
Mark Donnigan:Yeah, cool, exactly I. It's going to be interesting, but I think no one disagrees. You know that, that the, the, the creators, you know those individuals, those companies. You know those organizations that put in the immense amount of time, energy and dollars. You know capital to creating those content pools, that they need to be rightfully compensated. You know for it. So I want to end here, ryan, because I'm I'm I'm curious. I meant to ask this question earlier, but let's end on this one. So are there any specific LLMs, foundational models that you're finding are more desired? I'm trying to avoid the work better, because everything for model developers to you know to overcome. What can you share around that?
Ryan Jespersen:I think it depends. I mean, we obviously use several when it comes to our service, depending on what we're doing, because there are better models that have just better experience, that have been trained for specific purposes. One really interesting one that I've been doing some work on is Cloud Code, or Claude Code, I should say, from Google is anyone who's been coding. It's far and away the best LLM to be using for doing any kind of coding. Yeah, and because they've been purposely training it to actually do that particular purpose right. I think the LLMs are interesting because we've had a variety of use cases.
Ryan Jespersen:Everything that we did with Media Copilot was try to build this in the cloud. Yet we've had a lot of demand, not a lot big. Pocs have been almost local or, you know, wide area hybrid deployments and you're limited by the size of the language model that you can use on a local deployment or at the edge. So you know that we've had to adapt it, some of our use cases or our infrastructure, to be able to train a model specific for a particular purpose. Like I was saying for the remote production use case and the post-production use case, we trained it for this unscripted reality show. We're going to train it with pictures of the contestants and content from the previous season so it actually knows what we're trying to do. So, similarly, with the ones you have to look at, depending on what you're trying to build, go and do some research on where they have and use AI to do this. Go find research on which is the best model for X purpose, whatever X happens to be.
Ryan Jespersen:That's right, and I don't mean X within the context of Gronk?
Mark Donnigan:Yeah, I mean, I mean X in the context of yeah, I think I'm leaning incredibly heavy into AI usage across literally everything I do and I feel like I'm quite advanced. And then I go on X or LinkedIn or whatever, and I see these vibe coders, these vibe marketers, and I'm just like, wow, I'm not even scratching the surface. They're like here's my whole company of one with my 35 AI agents who are handling everything from accounting, to like, wow, okay, I'm not even scratching the surface there, but but it is good. Guidance is that you just have to. You have to actually use test. It is so application specific and you know there is no sort of. You know, use this model, it's the best.
Ryan Jespersen:Yeah, so I think the other piece I'll throw on there is is using an NCP servers.
Ryan Jespersen:That's the big trend that I'm seeing now is you can use them to actually train it for a particular use case, and I think we're already adapting media copa to be able to to train that against someone's NCP server for their specific use case, because there there are workflows that we haven't even thought of. So what an MCP server allows you to do is very quickly train it for your use case but leveraging someone's technology at a kind of deeper level right.
Mark Donnigan:Well, would you guys even turn Media Copilot into an MCP server?
Ryan Jespersen:That's what that's our CTO is actually working on that right now. So, yes, so that that's watch here, watch the space. But, yeah, this has come out of once again, yeah, the idea of selling an agent right To do that, and I think in the short term here by IBC we're going to actually have some demos here of different.
Mark Donnigan:Very cool.
Ryan Jespersen:For real use cases.
Mark Donnigan:All right. Well, I can't wait to see it. Well, ryan, this was a great conversation. I'm really happy we were able to talk about this. You know it's a hot space, but it's. You know. The thing that I like is now we're far enough down the road.
Mark Donnigan:A year and a half ago, you know I think all of us there was a little bit of rolling of the eyes with AI. You know and I'm speaking in the media, entertainment, in the streaming space, because everybody had bolted on AI to their marketing. You know it's like wherever you could insert AI, there it was. You know AI was there. Now we're a year and a half down the road and there's still plenty of companies who, I think, are still sort of bolting it on.
Mark Donnigan:But now there's companies like Cirrus 21, uh and many others, um, who are building real um, you know applications and use cases and bringing real value, and it's only growing and expanding every day. So it's wonderful to hear what you're doing and I would encourage for all the listeners, if you're going to be at IBC this is a shameless plug to stop by the NetEnt VPU ecosystem booth and have a look at what Cirrus 21 is doing, get a demo and hang out, enjoy a coffee on us, and I think we're going to have a cocktail party and there's a whole bunch of stuff going on. So, yeah, it's going to be a fun place. So it was good talking with you, ryan.
Ryan Jespersen:Yeah, great to see you. As always, mark, and we'll share a beer together in Amsterdam.
Mark Donnigan:Yes, we will, absolutely All right. Well, thank you again for listening to Voices of Video and until next time, happy encoding with AI. This episode of Voices of Video is brought to you by NetInt Technologies. If you are looking for cutting-edge video encoding solutions, check out NetInt's products at netintcom.
