Generalist models & Iceman's voice

In this “fully connected” episode of the podcast, we catch up on some recent developments in the AI world, including a new model from DeepMind called Gato. This generalist model can play video games, caption images, respond to chat messages, control robot arms, and much more. We also discuss the use of AI in the entertainment industry (e.g., in new Top Gun movie). Featuring: Chris Benson – Website https://chrisbenson.com, GitHub https://github.com/chrisbenson, LinkedIn https://www.linkedin.com/in/chrisbenson, X https://x.com/chrisbensonDaniel Whitenack – Website https://www.datadan.io/, GitHub https://github.com/dwhitena, X https://x.com/dwhitenaShow Notes: DeepMind’s Gato: DeepMind article https://www.deepmind.com/publications/a-generalist-agentTNW article https://thenextweb.com/news/deepminds-astounding-new-gato-ai-makes-fear-humans-will-never-achieve-agiIceman’s voice in Top Gun: Variety article https://variety.com/2022/film/news/val-kilmer-top-gun-maverick-voice-artificial-intelligence-1235281512/Sonantic - the company behind the synthetic voice https://www.sonantic.io/Responsible AI review processes: From a developer’s point of view https://io.google/2022/program/08c90d15-7627-4b7d-ac63-7a2c83fcb88d/ Upcoming Events: Register for upcoming webinars here https://practicalai.fm/webinars!

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Chris Benson?

  It's interesting. I'll often be in meetings where I have a specialist in a particular thing and a bunch of different specialists, and they have a very particular lens that they're looking at things through. And usually my job is to be in the room is to fuse all those things together and generate the big ideas from that fusion. That's where I think there is a lot of untapped magic out there in industry at large. Where these different things you don't expect come together and you combine them and go, wow, there's some amazing opportunities there. There needs to be a form of recognition academically for doing fusion of multiple discipline, I think.

• 0:40Chris Benson?

  And I think that would be very valuable to motivate people to learn how to do fusion from multiple areas.

• 0:56Speaker 01

  Welcome to Practical AI, a weekly podcast making artificial intelligence practical, productive, and accessible to everyone. This is where conversations around AI, machine learning, and data science happen. Join us at practicalai.fm/community and follow the show on Twitter. We're at Practical AI FM. Thank you to our partners at Fastly for shipping our pods super fast all around the world. Check them out at fastly.com.

• 1:30Daniel Whitenack?

  Welcome to another fully connected episode of the Practical AI Podcast. This is where Chris and I keep you fully connected with everything that's happening in the AI community. We'll take some time to discuss the latest AI news and we'll dig into learning resources to help you level up your machine learning game. My name is Daniel Whitek. I'm a data scientist at S I L International and I'm joined as always by my co-host, Chris Benson, who is a tech strategist at Lockheed Martin. How you doing, Chris?

• 2:02Chris Benson?

  Doing fine. There's so much going on these days to talk about and been having deep thoughts about DeepMind lately.

• 2:09Daniel Whitenack?

  Deep thoughts about DeepMind.

• 2:11Chris Benson?

  My name is not Jack Handy, though, for those in the audience old enough to remember that.

• 2:17Daniel Whitenack?

  Yeah, I mean, there seems to be a lot going on. And to be honest, we haven't talked about this for a while, Chris. How do you keep up with some of the things that are kind of coming across your path in terms of news related to AI and such? Like, where do you see that most? Does it come word of mouth? How does that work in your life these days?

• 2:43Chris Benson?

  It's a little bit of everything. I have some, some kind of carefully sculpted Google News for different topics that I use as kind of my primary thing. But at work, we talk a lot about AI issues and stuff. And my media manager is actually really good about it. I often hear things from him before I hear them from anywhere else. Cause he stays right on top of stuff, Brent Sequel, in case he's listening. And having said that, just kind of all over the place, I think the harder thing now isn't where to get it. It's, it's kind of how to keep the noise under control because there's so much coming out, you know, and across and trying to, to focus on the interesting things, or at least the things that are interesting to me and, and the people I work with.

• 3:25Daniel Whitenack?

  Yeah.

• 3:25Chris Benson?

  How about you?

• 3:26Daniel Whitenack?

  Well, I see a good bit of the things that I see on Twitter. I think that's, I don't really hang out that much on LinkedIn or Facebook or other places. So Twitter is, is generally where I see things. I've noticed that my feed and Twitter has become increasingly NLP related, which is, is definitely my, my zone. So I noticed that recently and I'm like, I need to kind of search for some more like general sources of, of other things. Cause I definitely, I think that there's a benefit in learning from a variety of, even a variety of other disciplines, kind of what's happening and how innovation is happening.

• 4:11Daniel Whitenack?

  I'm getting way off topic now.

• 4:13Chris Benson?

  No, it's fine.

• 4:13Daniel Whitenack?

  Are there, are there like other disciplines outside of computer science and AI that you feel like have influenced some of your own thinking in, in certain areas?

• 4:26Chris Benson?

  I am very kind of future application focused. Boy, I'm gonna, I'll probably upset a lot of people. Things that I get bored with are just the next implementation of a fairly common, you know, AI paradigm. Yeah. You know, to, and stuff like that. And I'm just like, oh, ho, ho, okay. Seen a lot of those. The things that I really like are when research comes out, some of the science, some of the really interesting scientific papers come out, especially if they're very application focused. And because that's, my job is to think about AI and data in the context of the industry that I'm in for 15, 20 years out. And so it's very, we do a lot of work with DARPA, you know, and, and so I'm very, very future focused.

• 5:11Chris Benson?

  And so I really have a deep interest in finding people who are very forward thinking and less about today's typical, you know, implementations that you see in a lot of commercial space and more about way out there. And so if anybody hears about it, things, things like that in the audience, that I love talking and exploring things that are, that are both practical, as we are practical AI, but also very aspirational in terms of where they might go. And that, that's where I spend a lot of time, but it's not strictly in AI. I've kind of come full circle, going back to some of the things that we've talked about. You can't separate the AI from the software infrastructure around it and all the other data concerns. And so

• 5:56Chris Benson?

  autonomy, robotics, a lot of these applications, but, but less what we're doing now and more where we're trying to get to is the area that I'm keenly interested in. And so hopefully, maybe if some of our listeners have some clues that they can point us toward that might make some for some great episodes going forward.

• 6:17Daniel Whitenack?

  Yeah, I think this question was triggered in my mind. I mean, one of the, so I get some things off of Twitter, but also I, you know, I, I watch YouTube videos related to a number of things, including like, I have a set of channels that I like watching. Uh, I'm, I'm a fan of like, old time, American music of like banjo and, and guitar and all, all that sort of stuff. So I, I like watching those sorts of videos. And so I watched a bunch of random things, but somehow I've sort of got gradually subscribed to generally like computer science and other types of, of videos that are out there. And I just watched

• 6:59Daniel Whitenack?

  one, I'll have to remember what, maybe get the link for our show notes, but it was talking about the importance of kind of getting exposed to various disciplines and how they can influence your, your own thought. Like the example was this guy and he was generally considered in physics, but you know, was making contributions in computer science as well. And his sort of knowledge of certain subjects in physics informed how he like thought about algorithms and, and other things. And he made the point that increasingly there's like a penalty for doing that in our society. Like you, you advance more by

• 7:38Daniel Whitenack?

  being more focused like in your area and it's harder to like be recognized in other disciplines outside of your area. And so it kind of, it's prohibitive for you to kind of enter into those zones and like publish both in computer science journals and in physics journals, for example, but it's probably true in industry too, to operate in different zones, right?

• 8:01Chris Benson?

  So you've really hit, you've probably just derailed everything we were gonna do.

• 8:05Daniel Whitenack?

  Probably. So I'm good at that.

• 8:07Chris Benson?

  I'll arrest it before it gets too far, but you're hitting on a great point. And that is the fact that the real magic for where you can do practical things going forward in an application sense, in all the cases I work in and see, it's at the intersection of multiple disciplines. It's interesting. I'll often be in meetings where I have specialists in a particular thing and a bunch of different specialists and they have a very particular lens that they're looking at things through. And usually my job is to be in the room is to fuse all those things together and generate the big ideas from that fusion. And that's where I

• 8:50Chris Benson?

  think there is a lot of untapped magic out there in industry at large in the world is where these different things you don't expect come together and you combine them and go, wow, there's some amazing opportunities there to dig into. And that's, that's really where I focus, is trying to find those, to mine those ideas and do that. And I think that that's not industry specific. I think that anyone out there can do that. But to your point to finish there is I'm often the only person in the room that doesn't have a PhD in something. And so there needs to be a form of recognition academically for doing fusion of multiple discipline, I think. It's, it's not, it doesn't really exist today

• 9:33Chris Benson?

  in that capacity. And I think that would be very valuable to motivate people to learn how to do fusion from multiple areas.

• 9:40Daniel Whitenack?

  Yeah. And I think, so one example I'm thinking of recently, we had, you know, the series of episodes on AI for Africa and the, the sort of discussion about AI applications in agriculture came up multiple times. And I think it's just by the fact that the applications that they were talking about were not like what popped up immediately in my mind. And it was because they had sort of firsthand experience in the environment, maybe even like growing up in a sort of more agricultural rural setting where they participated in

• 10:16Daniel Whitenack?

  agricultural things. And so like being hands-on in a variety of areas can really spark, like you said, it kind of unlocks where the real value is in a problem and prevents you from kind of building, building these solutions that might be interesting but not Not valuable.

• 10:37Chris Benson?

  Right.

• 10:38Daniel Whitenack?

  So I don't know, I don't have a huge desire to like get into agriculture or gardening. But I think in like, in like my, my own work in, in LP, there's like all of these different fields. I've noticed recently, like this has come up a lot. There's all of these different fields like socio, sociolinguistics. Or like the more traditional forms of linguistics field kind of study. Like we, we recently interviewed Sarah Mo- Moeller about kind of field linguistics and being in the field and being with language communities. And there's so much wisdom to be gained from there. So even though I don't really have the right to, like, or the qualifications, I guess you could say, to operate in those areas.

• 11:21Daniel Whitenack?

  And I should be careful. It, it still behooves me to, I want to read more about those things. I want to kind of participate in those discussions and learn what I can to sort of influence my own thinking. But I think people, I don't know, there's a lot of people that maybe struggle with that confidence to operate in areas where they maybe don't have the qualifications to operate, or they feel like they don't have something to offer. Whereas like you said, it's really when you do kind of put forward ideas in a multidisciplinary environment where really valuable work happens.

• 11:57Chris Benson?

  I think that's a form of, of imposter syndrome, you know, when you feel I'm not qualified and all that. I think, you know, you're only limited, and this goes for everybody, everybody who's listening to this. You're only limited by what you're going to choose to go do and make the effort on and learn. And so it occurs to me from time to time, as I mentioned before, but not most days. I'll be like, I'm the only person without an, an advanced degree in here that's a specialist. And for me, it's because I wanted to do stuff and I wouldn't pursue it. And I know other people, including you, who've done that as well. I've seen you work outside of, of, of your field and have insights and stuff. And I just encourage people, we're in a moment in history, a lot of people focus on the negatives.

• 12:39Chris Benson?

  Because there's a lot of challenges in the world, but there's also more opportunity right now at advancing a number of fields. And there's ever been before and there's not enough people to do all those things. And so you're only limited by what you want to go do, convincing people to give you a chance. And so I hope that not too many people are letting a self perception of not being qualified get in the way. Go get yourself qualified quickly.

• 13:28Daniel Whitenack?

  Well, Chris, I think I led us down the rabbit hole after you initially mentioned DeepMind and maybe some things we've been seeing in the in the news about DeepMind. I'm assuming maybe you're mentioning that with reference to their Gato.

• 13:44Chris Benson?

  Might possibly be. Some big news there.

• 13:46Daniel Whitenack?

  Yeah, yeah. So we should say that, you know, by no means we're we're experts on on all of this work that has been happening, but it is kind of worth. I think there is a trend here that's happening and or a couple of trends that are intersecting this work by DeepMind that we could comment on. But first, the main thrust of this is it's a a generalist agent. I'm constantly getting made fun of because I'm trying to come up with catchy acronyms. But Gato is pretty good, a generalist agent. The idea here is that it's a single model that can operate

• 14:27Daniel Whitenack?

  in a multimodal, multitask, multi-embodiment generalization policy. So this is like multiple modes. So you've got like text, video, you know, video game environment type of stuff, multitask, related to like text-based tasks or moving robot arms or all sorts of things. Basically, I think the reason this is getting a lot of attention is Because it can do so such a diverse range of tasks from text and image based things to operating robot arms and that sort of thing. So I don't know what caught your attention about this, but that's sort of what I was seeing.

• 15:12Chris Benson?

  It was that and I it's the fact, the thing that, and this is an area that I think you are much more expert in than I am. It's trying to do so many things, you know, going back. That was the thing that caught me when I first looked at it was multimodal, multitask, multi embodiment, generalist policy, which I I hear that and hear kind of all things to all people and immediately wonder like, what can it do? What's its limitations? What are the practical? And they do have the video on the website, obviously, that kind of shows a whole bunch of tasks lining up. But, you know, how do you fit this in if someone's looking at this? And we've been talking about all these big models with increasing capability over the last few years. Well, how do you think about this when you're looking at it, when you read about that?

• 15:56Chris Benson?

  And is it a jump forward? Is it just an iteration? How do you categorize in your mind?

• 16:02Daniel Whitenack?

  Well, I think that the one of the main things or main lines of inquiry that this model follows that is true of other things that have happened recently is the relationships between this model and recent large language models. So recent large language models have approached the problem of sort of particularly the multitask problem. So in natural language processing, there's a variety of tasks, right? You could have text as input and then detect like a label, like a sentiment label,

• 16:42Daniel Whitenack?

  neutral, positive, negative. You could have text as input and detect certain entities in the text. Like here's an organization.

• 16:52Daniel Whitenack?

  You know, that's mentioned Lockheed Martin is mentioned here. It's an organization that's mentioned in the text. You could have text input in one language and text output in another language. That's translation, right? You could have text input and then you could parse out the semantic structure of the text. There's all sorts of things you can do with text and these are all the tasks of natural language processing. And recently, what has been the trend is that initially with language models, things were structured in terms of a kind of meta task where like maybe you would take out certain words from text and try to fill those in. And that would be the task you would train your language model on.

• 17:37Daniel Whitenack?

  And then you would kind of fine tune it downstream for various tasks. Recently, the trend has been to train language models on a variety of tasks by using what's called prompts. And prompts basically is a text input, but it's specifically formatted text input in the, in the way of like, there is a prompt that is structured in a certain way for question answering. There's a prompt that's structured in a certain way, based on certain tags for machine translation, there's a prompt. So it's all sort of, it's not completely unstructured text input. I guess is one way to think about it.

• 18:18Daniel Whitenack?

  There's actually a prompt engineering or an engineering that goes into constructing this input data. But all of that's fed into the same model and the model learns based on the context that's input, what action to take and what to output. So I see that, you know, obviously there's a big engineering and research step in the scatto model, but the spirit of it is very much in that same, same line of thought with the prompt engineering. So based on what the context that is input, it's determining.

• 18:55Daniel Whitenack?

  Actually, what tasks to complete and what kind of format to output. So whether it's outputting, you know, captions of images or it's outputting actions for a robot arm to complete. It's based on that kind of formatting of the input text in the kind of contextual prompts, I guess, to help it know, help the same model know what tasks to do and what kind of things to output.

• 19:25Chris Benson?

  How different is this in your mind? You know, it's transformer based for training and it references that it's similar to other large language models. What are the differences that seem to be there on the surface? You know, without having dived deeply into it, because I know you haven't had a chance yet. Can we think of it in the same kind of context as other large models in terms of framing it? Or, or do you think it's, have we kind of gone down a different branch is what I'm trying to get at?

• 19:51Daniel Whitenack?

  Yeah, I, I think that this is almost like the, in some ways, it's like the fulfillment of a lot of things that started with large language models. I mean, transformers, you know, were discovered in reference to language, but they've, as it turns out, that sort of has been generalized across a lot of modes of, of data. So vision transformers and other things are really popular right now.

• 20:16Chris Benson?

  Yeah.

• 20:17Daniel Whitenack?

  And so I think that this is a natural thing, that this sort of idea of prompt engineering and multitask sort of things have filtered in. It's sort of gone full circle to, to catch up to like all the data and all the types of prompts that we might want to, might want to include. So I think it, it is a really bold kind of step to kind of say, well, how far can we push this sort of idea and the things that are inspired by these large language models? And certainly there's a lot of modifications that need to happen in terms of like how sequences of things are tokenized and organized as input to the model.

• 20:58Daniel Whitenack?

  But I think a lot of the threads or the foundational blocks that this is built on. Are really kind of building on those things from large language models, which is, you know, really how science works, right? It took a lot of pieces. It took the, you know, prompt engineering pieces and the reinforcement learning and action type pieces and the transformers type pieces and the like computational engineering pieces, all to sort of like follow their own paths and development to eventually combine into this thing. At least that would be how I would look at it, maybe.

• 21:37Chris Benson?

  So the thing that I can't help wondering, I don't know if this is a tangent or if this is kind of on the main line of talk, is if you keep extending this, this is where so much research is going, so much attention is going into the continued

• 21:50Daniel Whitenack?

  No pun intended, self attention, maybe?

• 21:52Chris Benson?

  Self attention, yeah. Wasn't intending that, but you know, is this, given the amount of effort that is going into this and the money and the resources, is this possibly, you know, using the words like generalist and using words like multimodal, multitask, possibly a path toward AGI in the sense of you're kind of combining all the things together to get to something that is is more generalized, that kind of that, that's always AGI, which is artificial generalized intelligence, has always been fairly aspirational, you know, that kind of, that kind of holy grail that we're moving toward. But this is where the focus, this is where the work is happening.

• 22:35Chris Benson?

  So is there any, have you heard any speculation? I mean, is this where people think would be a productive pathway to get that way?

• 22:44Daniel Whitenack?

  Yeah, I mean, I guess it-

• 22:46Chris Benson?

  I knew I threw you a curve ball on that.

• 22:47Daniel Whitenack?

  Yeah, there's a whole variety of thoughts here, I think. I think one thing to note is like typically how things have happened is generally people have set a goal, like when we're able to do this with AI, then we'll call it intelligent, or when we're able to do this, you know? And you sort of keep hitting those and then you, the goal line keeps moving out. There's certainly like, you could look at this model and I'm sure that there'll be a variety of studies that come out that show how it's limited in certain ways and doesn't scale in the ways that we think it maybe could. But it's, you know, it's a stepping stone to a variety of, of-

• 23:27Chris Benson?

  As have all.

• 23:28Daniel Whitenack?

  Forward.

• 23:28Chris Benson?

  Yeah.

• 23:28Daniel Whitenack?

  Yeah.

• 23:29Chris Benson?

  I mean, as have all the models that we've talked about on this show in recent years as we've moved through them.

• 23:34Daniel Whitenack?

  Yeah, so I, I mean, I think it's certainly more general, definitely in terms of task and type of data. I think kind of circling back to what we started the conversation with today around like multidisciplinary things. I think that this is general in the sense of the AI tasks that AI researchers have defined over time. Right? Like we've defined certain arbitrary tasks that are well defined in terms of a computer. And like, you know, like sentiment analysis is like a totally arbitrarily defined task for a, you know,

• 24:16Daniel Whitenack?

  it corresponds to reality in some way, right? Like we have a concept maybe of how it corresponds to reality, but it's not really like sentiment is so much more complex. Emotion is so much more complex than like positive, neutral, negative, right? So I think that like, that's the way I look at it, I guess, is this is general in terms of the types of problems that AI researchers have dug into over time, but that's only a very small amount of the complexity that's involved in the world, right?

• 24:49Chris Benson?

  It's general in a sense that AI practitioners and researchers have kind of created the world in which we're developing this stuff. And so it's general to those set of tasks.

• 25:00Daniel Whitenack?

  Yeah, yeah, it's, it's general to those set of tasks and it's certainly more general than other things. But I think in relation to the complexity and subtlety of so many things in our, in our world, you know, there's, there's still a lot of, you know, it's not like this is like covering all the general complexities of our, you know, universe.

• 25:24Chris Benson?

  You realize that 20 years out, we're going to be on episode 5000 and we're going to have long gray beards. And I'm going to say, is this the path to AGI, Daniel?

• 25:35Daniel Whitenack?

  Yeah.

• 25:36Chris Benson?

  As we've iterated many times through there.

• 25:39Daniel Whitenack?

  That's probably true, Chris.

• 26:03Daniel Whitenack?

  Well, Chris, have you seen the new Top Gun?

• 26:06Chris Benson?

  I have. Everyone but me has seen it in the last few days, I think. As we're recording this, this is the Wednesday after Memorial Day weekend 2022. So it just came out, I believe, this last Friday, five days ago. And I think I may be the only human being in metropolitan Atlanta who has not seen it. But no, I haven't seen it yet. But I've seen all the trailers.

• 26:32Daniel Whitenack?

  So maybe for, well, I don't know. I mean, Top Gun seems fairly internationally known. But for anyone that isn't aware, Top Gun is a famous fighter pilot movie with Tom Cruise. And a sequel just hit theaters or theaters and streaming. I'm not sure if it's on streaming.

• 26:54Chris Benson?

  I don't think it is, but I could be wrong.

• 26:56Daniel Whitenack?

  Okay, yeah. So again, you know, fighter pilot stuff, which is right up your alley, Chris, being a pilot. I picture you in those sort of Top Gun, yeah, Maverick scenes as you're flying your Cessna.

• 27:11Chris Benson?

  I do have the fantasy when I'm flying. But yeah, not only am I a pilot, but I work at the largest aerospace company in the world that produces fighter planes, obviously. So it is up the alley.

• 27:22Daniel Whitenack?

  Yeah. Well, I mean, despite the, the general interest here, maybe there is a, there is a point to why I brought this up, which is apparently that at least in one place, and I, again, I haven't seen the movie, so I've yet to hear this. But Val Kilmer, who played one of the original characters in the movie, in the original movie, they had a scene where they needed his voice for something. And I'm not giving any spoilers, particularly because I haven't seen it and I can't. But apparently they used AI models to recreate Val Kilmer.

• 28:02Daniel Whitenack?

  Voice because he is, you know, is and has been undergoing throat cancer treatment, or I'm not sure if he's still undergoing it, but at least he's in a place where he cannot speak and has, has that sort of constraint. So they used a variety of AI models to reconstruct his voice and actually produce a voice for Val Kilmer in the Top Gun movie, which is quite interesting. What's your, what's your reaction to that?

• 28:32Chris Benson?

  I think this is the first of this happening many times. I know that they, they used a lot of his old audio from movies to train the model. And not only that, but that the model is very specific to his style of talking. So it's not just generating the audio, it's generating audio the way a younger Val Kilmer, before he was in this particular medical condition, would have, would have addressed it. So it's, it's supposedly very much right on. And I think this fits in, you know, we've, we've had shows on deep fakes and, you know, on the visual side. And I think we're going to enter a time where AI

• 29:13Chris Benson?

  to produce all sorts of audio, video to consume and other, other medium will be pervasive. So we may look back at this and say that kind of kicked off a whole era of movie making.

• 29:26Daniel Whitenack?

  And I know that, of course, there's like video based things that have been done as well to like de-age actors or like change physical appearance and that sort of thing. This one is, is quite interesting to me, especially if, you know, if you think about, like, we've been talking about multitask, multimodal things, and there's certainly things that can be done on the video side. There's, this is an example of something that can be done on the. Sound side, but it, you know, makes me think, well, like in the future, are they just gonna keep having Val Kilmer act in movies with the voice that they've created for him?

• 30:08Daniel Whitenack?

  And like, you know, a realistic avatar that they can create for him, you know, like what, what's stopping that sort of thing from being, you know, widespread, I guess?

• 30:21Chris Benson?

  I don't think anything is stopping that. I think that there is, I'm gonna reach out on something I know very little about and probably folks in the audience can, can educate us on this more, but I know that the band ABBA has a show that's been out there for a while where they use some sort of avatar and I don't know the details of that and I apologize. So it is a young ABBA back in their prime that are performing the show. And I'm, that may, it probably has nothing to do with AI, but what I'm getting at is that idea of being able to, to use technology to present entertainment or, or other things outside of entertainment in the way that you would prefer, I think is here. And we're

• 31:03Chris Benson?

  seeing that, we saw, you know, whatever they happen to do for the ABBA show, we're seeing that with Val Kilmer. And we're also seeing the choices of not using technology and not using AI, not using CGI. And that, going back to the Top Gun Maverick movie, for the flying scenes, they decided not to use CGI. They could clearly have done a lot of the flying stuff with CGI, but they chose maybe for the last time ever to do something strictly without CGI in terms of the flight scenes. And yet they made the, the choice to bring Val, Val Kilmer back into it with, with the AI model. It's, it's interesting. AI models are becoming a tool of creative artists going forward and, and

• 31:46Chris Benson?

  you have a whole palette of tools. So it's interesting to see how the choices are being made now.

• 31:50Daniel Whitenack?

  Yeah, I'm also wondering about the sort of rights related to this. Like, I don't know, you know, I don't know the details and I didn't see, at least in the articles that I read, I didn't see a reference to this, although maybe. It is public information. I'm not sure, but like, did they pay? Like, what is the payment structure around like, well, Val Kilmer, we created your voice, right? You got paid for the previous things, right? And I, maybe the studio owns those audio clips and the video clips and they created the voice out of those, right? So do they owe Val Kilmer any money for like? I mean, they're certainly representing his likeness and his voice, right?

• 32:35Chris Benson?

  I do not know the answer. My guess would be yes, that they've compensated him.

• 32:39Daniel Whitenack?

  Yeah, I would sort of assume so.

• 32:43Chris Benson?

  But?

• 32:43Daniel Whitenack?

  But it brings up the question in my mind, like, well, I assume so, but I don't know. And that sort of like, makes me wonder, this is certainly a much, much lower level. But, you know, we now have 180 episodes or something paired with, with voice and, and text, you know? How would I sort of go about it if someone recreated my voice and was posting things on YouTube? For example, it was clearly my voice. Maybe they had my name, maybe they had my name, maybe they didn't have my name, but it was sort of obvious what was going on, you know? What would I do? I, I, I don't know how I would exactly respond.

• 33:28Chris Benson?

  If you're thinking in particular, as I do in the industry I'm in, about nefarious intents.

• 33:33Daniel Whitenack?

  Yeah.

• 33:33Chris Benson?

  You know, we've had many conversations about AI ethics.

• 33:36Daniel Whitenack?

  And sort of deep fakes and such.

• 33:38Chris Benson?

  Absolutely. It certainly raises a lot of questions about the future and where this might go. As you were describing that, the thing that was in my head is I'm a big Jared Tolkien fan and, uh,

• 33:50Daniel Whitenack?

  Yeah.

• 33:50Chris Benson?

  And have all the books and, you know, I'm not truly a super fan, but I'm bordering on it. And so, you know, that is an estate that's happened and not only has he passed, but his son Christopher passed. And, and I'm wondering, is that the future? If you have a very marketable personality or personality assets, like a voice, does everything become kind of an estate based? Path forward in terms of how you have those assets because it would not surprise me if we hear from Tom Cruise, you know, 50 years from now, when, when I'm expecting, he probably will not be alive at that point. Could be. He's very good at staying young and he has the, the budget to keep himself young, but, but I'm assuming he won't be

• 34:35Chris Benson?

  around in 50 years. But I will bet he will be on screen even on new things. A little speculation there, but I don't know. The Cruz estate may have, uh, may have a few centuries left of bankable filming.

• 34:48Daniel Whitenack?

  Yeah, that's, that's true. I guess. Yeah. I mean, I, I guess this is probably true even for text in that if there's enough text from a certain, and I know there have been attempts to do this with varied levels of success, but like you have enough text from Shakespeare, right? You can write some new, you know, some new plays or, or whatever in the same style. And I know that that's, you know, kind of, there's some toys out there like that. And it's been kind of fiddled around with, with varied levels of success. But the idea I think would be similar as sort of whatever creative arts are out there,

• 35:28Daniel Whitenack?

  including music, you know, writing, uh, movies, acting. It's just interesting to see this technology infiltrate all of those areas. Uh, we, you know, we've had obviously a lot on text generation. We've talked about on this show a lot. We've had even the, the show on, um, it may be a couple of shows now on how AI is influencing music creation and production. We've had, you know, we're now talking about movies and I think have talked about video at other points. So yeah, it's just sort of another example of how pervasive this technology is. And it's bringing up new sorts of questions that people haven't

• 36:12Daniel Whitenack?

  had to, to wrestle with yet.

• 36:14Chris Benson?

  It was kind of when we looked back and AI ethics was first coming about, there's a lot of people that rolled their eyes at the notion of AI ethics. And we were talking about it very early on because it matters a great deal to us. I know that in my personal conversations, I got a lot of pushback about it in a variety of venues. And so I think this points out this conversation that it's not only crucially important to the present, but will be increasingly so for the future and that intellectual property issues are going to not only be present but change, the very nature of them will change going forward because we're moving into a world where we and our AI will become inseparable. And I'm not saying that in a big mushy aspirational way. I'm saying that we're having that now.

• 36:58Chris Benson?

  You know, child's toys, you know, being able to, whether the presence of a person, I'm using the word presence, whether that person is alive or dead, they're still involved in those interactions. So that idea, the term we use in my industry is called mum tea, which is manned unmanned teaming. But in all industries, in all aspects of life, manned unmanned teaming will be present. You will be interacting with your AI in just about everything. So maybe there's somebody listening out there that will, that will kind of take us to the next level on some of these issues because we, we're going to need some help to navigate this.

• 37:35Daniel Whitenack?

  Yeah, for sure. We need that creativity. We need multidisciplinary thinking, all those things that we've talked about in this, in this episode. And that brings us in a sort of natural way to a learning resource that I was going to bring up in this fully connected episode, which is talk and set of, you know, linked tools from the TensorFlow team, which I ran across recently, which is around the responsible AI and machine learning type of things. So this includes, you know, lessons learned from the TensorFlow and Google Teams around these things, but also some relevant tooling around, you know, creating things

• 38:17Daniel Whitenack?

  like model cards and data set cards, even Merva from Hugging Face brought this up in our, in our last, one of our last episodes about model cards and maybe even integrating some of that tooling with, with Keras. But this is more of a sort of in that vein from the TensorFlow. Flow team. And so if you're thinking about this sort of responsible AI and ML things and want it maybe more from the developer perspective and less from the sort of philosophy and kind of social science perspective, this seems like maybe a good, you know, a good starting place, not covering everything, maybe, but a good starting place. So I'll link that, that video

• 39:02Daniel Whitenack?

  and set of tools in our, in our show notes. So people can, can go check it out.

• 39:07Chris Benson?

  That sounds good. I'll finish up by saying you, I'm kind of creeped out by the idea that we have all that audio out there a little bit. I started thinking, oh my gosh, that's, think about that.

• 39:17Daniel Whitenack?

  All of our listeners will now be learning about responsible AI and machine learning practices. So maybe we're good there.

• 39:25Chris Benson?

  I hope, I hope we're good. We're asking, we're asking the audience to save us from this.

• 39:29Daniel Whitenack?

  Yeah, yeah, for sure. Well, yeah, it's been, it's been fun, Chris. Appreciate catching up with you and look forward to talking in our next episode.

• 39:39Chris Benson?

  Sounds good, Daniel.

• 39:49Speaker 01

  All right. That is Practical AI for this week. If this is your first time listening, subscribe now at practicalai.fm or just search for Practical AI in your favorite podcast app. We're in there. And if you're a long time listener, please do share the show with your friends. It is the best way you can help Practical AI succeed. Thanks again to Fastly for shipping our shows super fast all around the world to Break Master Cylinder for the Beats and to you for listening. We appreciate you. That's all for this week. We'll talk to you again next time.