We’re back with another episode of Mission: Generate. This time though… with a special re:Invent twist. Meet the real Dr. Ryan Ries and Casey Samulski! They’ll be chatting about how the podcast was created, the future of Gen AI, patterns with customer Gen AI productions, and more.
Ryan
Welcome to Mission: Generate, the podcast where we explore the power of AWS and the cutting edge techniques we use with real customers to build generative AI solutions for them. Join us each episode as we dig into the technical details, cut through the hype and uncover the business possibilities. I'm Ryan Ries, our generative AI Practice Lead and your host. And today, I'm here with Casey and we were both real this time and not generated. So we look forward to talking to you about how Mission Generate came to be the brainchild of Casey as he had to listen to way too many of my webinars and decided, hey, what would be a great idea would be to make this into a podcast. So Casey, how many times did you have to listen to the talk before you decided I want to be able to create Ryan's voice all the time so I can sleep at night?
Casey
A lot. It was many minutes. But you know, it was great. I had a lot of fun doing it. And by the way, you did a good job considering that's your first time actually saying the intro out loud. I was impressed. Yeah, we actually, you know, what I took was the footage that Olivia recorded of you at the go to market Summit. That was what we ended up using for the final audio. Because in between the takes you were like making jokes and stuff. And it like made the voice a little bit more lively and dynamic.
Ryan
So how do you produce each episode? What's your technique?
Casey
So I take content that we've already developed, often for like a use case page or a blog or something like that. I actually feed that to an LLM as a first step. And I say, Hey, give me about a five minute episode. And then it gives me something back. It's usually a little bit too jokey and not really you. But then I basically pull that whole thing apart and then rebuild it and try and put our put our own spin on it. I guess.
Ryan
Have you tried fine tuning the model yet?
Casey
Not really. I mean…
Ryan
We would have Ryan GPT. No longer have to go to a beat in case you're feeling okay, feeling the need here.
Casey
Yeah, 2024, that's the goal
Ryan
You played with the LM, you know, you said it was too jokey. And sometimes they get very verbose. Have you changed any of the features in it, because you can change you know how long they are. Or you can change kind of just the timer and things of that nature?
Casey
I tried a couple of different prompts to sort of navigate it and see if I could get more of the feedback that I was looking for. But what I realized is that, in the end, the way that it writes things, it doesn't write them as they should sound audibly. So actually, one of the biggest elements we had to change was putting back in like, not “Umms” but those kind of like filler words that you actually use when you're talking to people out loud. And without those, it felt much too like scripted and, and sort of straight ahead instead of live and organic.
Ryan
I do say “Umms” way too many times.
Casey
Do you? Actually, I don't know that about you.
Ryan
I get told that every once in a while: “you needed a better filler word” or better yet, a “filler word”.
Casey
You should just listen to the podcast more.
Ryan
And then I'll stop saying “Umms”
Casey
Yeah, you can find better filler words.
Ryan
Yeah. I have to ask me a better filler word for “Umm” or “you know”.
Casey
Yeah, you know, is good. Yeah.
Ryan
You're not supposed to use filler words and you sound less professional.
Casey
Well, I like being a little less professional in the podcasts. So people kind of feel like we're joking around. Which is also hard. The humor is really hard to get across, right?
Ryan
That is the big, the big hard piece. Yeah, we're working with another customer that will probably come to an episode soon that has that exact piece. And we're trying to recreate a celebrity's voice and their mannerisms. And it's exactly that piece, right? How do you do a better job with tambor and everything. And, you know, excitement and emotion. Because, you know, obviously, on our podcast, it's kind of, kind of just monotone to a certain extent. There's, there's a little bit of wobble.
Casey
There's a little, there's a little, yeah.
Ryan
But not a huge amount. Well, you know, if you're gonna do this for real, you've got to really get that.
Casey
Yeah, yeah. So all of the humor is deadpan at this point, because that's really the only thing it can nail in terms of delivery. How are you gonna solve that problem, Ryan?
Ryan
That is a work in progress. With Ben Max trying to figure out that there's some cool models that come out constantly to explore. There's a system out there that we're kind of looking at, as well called out from elven labs that seems to be able to do a similar thing of creating somebody's voice from, from previous voice, but, you know, we haven't played with it enough. We have talked to some customers, right, you know, content localization is a big deal. And so with Magellan TV, obviously, we're doing content localization on, you know, all of their documentaries. And in a documentary, it doesn't matter, right? Because in a documentary, you're just trying to relay content, you're just trying, you know, you can use summarization and all that, sort of to relay the content. But once you start going to things, you know, like Netflix, where they do localization and all that they're in a whole different boat, because now you have to add the emotion right now, you have to look at how can I keep the comedic timing right, how can I rewrite a joke into something that's going to be another joke differently?
Casey
That sounds impossible.
Ryan
You know, I know they're working on it. I mean, we're not, you know, but documentaries are easy.
Casey
Yeah! Maybe the lower lower bar of all those categories?
Ryan
Yeah, I wouldn't say it's easy, because it's not easy. We have work to get it done. Easier than, than that problem. But it's definitely, definitely something we're looking at and going to be trying to build out as we go. Because, that content localization is a key feature. And when you look at it, there aren't a huge number of models out there that will do somebody else's voice, right? When you look at Amazon, they have a service called Polly and Polly just has X number of speakers. And often, for many of the languages, it's just male and female, some languages have more, some have less. And so you end up you know, what are we going to do for companies that want to transcribe or want to translate into other languages? Right, right. You know, there's two voices, but you have four people talking in a documentary, that's not gonna work very well. So it is a, it is a project we're working on, you know, to try to solve that issue.
Casey
So to season two, are we going to be able to tell better jokes to each other?
Ryan
Maybe, it depends on people's humor, and whether they're better jokes, but I'm waiting to see when you translate everything for global reach versus US reach. So Casey, how many more episodes in this season? Are you going to do seasons, or are you just going to toughen it out and just make one every week?
Casey
I'm definitely not going to tough it out. We'll probably wrap up sometime in January, I really want us to at least get some of the bigger topics through before we finish. Like I want to do Ragnarok for sure. At this point. I think that would be really interesting for everybody to hear. So…
Ryan
Everyone needs to know about Ragnarok.
Casey
By the way, could you tell us about Ragnarok?
Ryan
Yeah, so as we've been working on Gen AI projects, you know, you kind of start seeing patterns. And for several of our customers, we've developed, you know, an architecture pattern that we call Ragnarok. You know, we liked the name, obviously, big Marvel fans. And so having, having that name is cool. The RAG and RAG, right, retrieval augmented generation than the NA is, you know, kind of the the complex piece where you're using neoteric agents, right. So that's a thinking agent, where it's able to start doing a little bit more than just, hey, I'm gonna prompt the system. And then the rock part, you know, is the back end of bedrock. Right. So this, this is, you know, using RAG with agents on bedrock. So we're pretty excited by that solution for, for a bunch of customers.
Casey
Yeah, it seems like that is an approach that we found is going to have a lot of flexibility for the customers that we're starting to have walk in the door in terms of what they need for capabilities.
Ryan
Yeah, absolutely. I mean, when you look at it, you know, a lot of people are still on the fence of what's best, right? Do I use RAG? Or do I use try to find right and fine-tuning becomes complicated because a lot of people just don't have a lot of documents, right? So you can't get a super great fine tuned system going well, if you use RAG and you only have a few documents, you know, it's only searching in that documents space and so you can get a lot better answer and reduce hallucinations versus just asking the model itself so it is definitely a first step for everyone on that path into Interjet AI.
Casey
That's awesome. I think you just gave me like a starting prompt for an episode. I think I can get a lot of mileage out of that.
Ryan
Well there's a lot of other cool prompts that can use the models just themselves right like we have a company that works with mission you know on our resell and Cost Ops who's an IT company and they just use, you know, Titan straight out of the box. I said they use Cloud v two out of the box where it is they're asking because they do IT support on-site support and people are like hey my HP blah blah blah printer has a blinking light right? So they call in that guy just typed into, into this into bedrock, hey, I've got a blinking light, how do I fix it? And you know, it'll pop out you know, really good steps because you know, all these different models essentially took all the different books and all the manuals and everything else, you know, they're out are striving for any content in order to train their model.
Casey
Who would have thought in 2024 we'd never have to read a manual again. We're truly living in the future now,
Ryan
Who even read the manual to begin?
Casey
I did not! Ryan, I just want to say it's a pleasure to actually finally talk to you, the human being Ryan, on our podcast for the first time.
Ryan
Okay, so I appreciate you having me live on the podcast instead of generating me this time. Thanks, everyone for tuning into this episode of Missing Generate the Podcast where we explore the power of AWS and the cutting edge techniques that we use for generative AI. I hope you found this informative on what we do to make this podcast and if you really want to do it for yourself, or you want us to do it here at Mission, feel free to reach out and we can help you build your own podcast. Have a good night, good day, wherever you are. Bye.