Suno AI: A Robot... With Soul

This is AI Inside Episode 12 recorded Wednesday, April 10, 2024, a robot with soul. This episode of AI Inside is made possible by our wonderful patrons at patreon.com slash AI Inside show. If you like what you hear, head on over and support us directly. And thank you for making independent podcasting possible. What's going on, everybody? 

My name is Jason Howell. Welcome to another episode of AI Inside. Hopefully you're weak to source for, you know, AI news. Yes, we also have some really great interviews. Today we don't have an interview. Next week we will have two and I'll talk about that at the end of the show. But this week it's going to be Jeff Jarvis and I talking about what's up. How you doing, Jeff? Hey, how are you, boss? 

I'm good. Good, you know, a little technical difficulties, but they keep you on your toes. As hard as technology. 

Yeah, the way of the world, unfortunately. Real quick to everyone watching and listening. Thank you so much for your support, for listening, for sharing. If you have someone that special someone in your life that is really interested in learning more about artificial intelligence, they can learn with us. Just point them to our site, AI inside dot show. And then, of course, those of you supporting us directly via Patreon. We love you to Patreon.com/aiinsideshow, ad ad free versions of the show. Of course, bonus material, bonus content that you get that no one else gets all there, patreon.com/aiinsideshow. And you also get the pleasure of being read out your name being called out at the top of the show, which could be a detraction for some people, I'm sure. 

But Mike Mattatall. Hopefully you enjoy having your name read out because I just did it. So there you go. Thank you for your support. Thank you, everybody. Thank you, Mike. Thank you for your support. We appreciate you. All right. Well, let's let's get into this long row of news, because we do have a lot to talk about. And before we get there, though, you were busy today. You were involved with the Nordic Media and AI conference. What was that all about, Jeff? 

Yeah, I wish I were there in person. I tried desperately to go. It was two days of people from Nordic AI. Sven Stormer Thaolow, who we interviewed in our second episode from Schibsted, introduced me to somebody who was doing the conference. They wanted me to come over, but, hey, when I left my job, I lost my expense account. 

So not so easy. And but I got to speak to them on video today and also to listen to some. And it was really interesting, Jason. 

The session before me was a deep dive into AI audio. These are news people. So it wasn't your world. It wasn't music. Right. You spoke the word. But it was fascinating for a few ways. 

First, it's just their attitude. Yes, they mentioned that there can be misuse. Yes, they mentioned that they need standards. They said that too, but they didn't lead with that. They didn't let that take over the conversation. Instead, they talked about all the interesting things they're doing. One woman from Schibsted also was there. She said she was happy to be around this summer. 

She was the only person around. They said, do you want to be our like the voice of the company and train the AI on your voice? She said, sure. And so she recorded a whole bunch of like seven thousand statements, something ridiculous. Wow. And now it's her voice. And she I couldn't see the video. But evidently her expression the first time she heard herself talk to herself. 

Like, hey, it's got to be weird. It's got to be weird to hear your voice talking back to you. 

It was also interesting. She immediately said, oh, we got to have we got to have a contract about this. Yeah. And so the contract says that if she quits or leaves Schibsted, they stop using her voice because she can't be the voice of ship's. But anymore, she doesn't want to be sure that they can't. They can only use it for editorial. They can't use it for promotion or advertising. And also they can't use it for somebody's bachelor party, as she said. You know, I know it's a toy you want to play with, but no, it's my voice. 

You can't do that. So it's really smart. And then they went through a whole bunch of uses where they're having it read stories, read in different languages, which is just a huge opportunity to increase, especially if you're in Norway to increase pronunciation. I mean, I mean, translation and other languages. And that's a funny stuff about translation, because it would come along and it would get the Norwegian OK. Then if it came into a different language, it would do it differently. 

And they emphasize that there's actually I didn't I've heard this before. There are two Norwegians. There are two languages and I don't fully understand who speaks which one, when, how. And one seems to be a little more high Norwegian than the other. 

Yeah. But it's not like it's not like Switzerland, where you have Schweizerdeutsch, which I can't understand to say by soul and German, which I understand, you know, a tiny bit. So similarly there, I guess. And there's also a rhythm to Norway, to Nordic languages. There's a song aspect to it, which the machine has a hard time getting. 

But they said it's got a lot better. The funny one, if you don't mind me getting close to a bad word here. Was that when they were going along in Norwegian, then they come to a certain English phrase, it mispronounced the English phrase. The phrase was deep fake. 

And it was an imagine. Yeah. Thank you. Deep beats. 

So the head. Hey, she was making an honest mistake. They had to retrain it. 

So they went through. Oh, funny. Fifty minutes of phenomenal examples like that. 

And then my shtick was, and I said this after I testified before the Senate and we talked to Sven and I've been saying, why can't we go like Norway? So I went through the attitude here in the US, where the news industry really acts like a victim of technology, rather than a master of it and a user of it, and rather than seeing the opportunities. And so their reflex is to go for lobbying and regulation or for court cases, as in the case of the New York Times and open a open AI versus Schibsted, which is saying, what can we do with this stuff? 

And access to the technology to work together, which I think is a much smarter, strategic, collaborative, sensible, productive attitude. So next year, I hope to be this time of the year. I'm by God, I want to be in this was in Copenhagen, which I would have loved to have been in great restaurants there. 

But anyway, so I was. Yeah. Yeah. That's super interesting. That's that's like that's that's your that's your breakfast experience. This is how you start your day. You just, you know, go and speak at John to Copenhagen. 

You know, back by lunch. Right on. Well, you also. So just to kind of transition here last week, when we were getting ready to do the show, you had shared with me a PDF on the Meta Community Forum on AI chatbot principles. And we kind of, I think we mentioned it on the show last week and we're like, OK, well, we'll take a look, a closer look at this, at least on next week's show. 

And here we are. So, you know, I kind of it's a it's a it's a large document, by the way. But there was a ton of really interesting information there. 

The community forum actually took place last October. This report details the findings there. And it had fifteen hundred forty five participants from Brazil, Germany, Spain, the US all focused on the principles that should guide how AI chatbots interact with users. And kind of in my estimation, kind of analyzing the support or the feeling that the participants have about certain aspects of this, like should chatbots inform users that they are bots or not? 

Right. Should these be predictable? Like what how do we feel about like romantic kind of interactions with chatbots? 

All that kind of stuff. What was your take on what you read? So this is part of this was done with Meta and the Stanford Deliberative Democracy Lab. There's a lot of movements in journalism. I helped start one called Engagement Journalism. There's Solutions Journalism, Constructive Journalism, Repairative Journalism. 

There's lots of movements. One of them is Deliberative Journalism and Deliberative Democracy. And the and the aim here is to say that if you took people and understand what their attitude is now and then if you could feed them a bunch of information, of course, it depends on what you feed them. But but you you you're paid to actually read stuff. You're paid to get informed. You're given a bunch of material to do that with. Then they look and say how much that changed attitudes afterwards to see whether there was an impact of good information on people. So they wanted to do this to understand what the people's baselines if you were if yous were and then also next what they think afterwards. So that we're going to go to the most changed attitudes and well, first time spent using AI chatbots at worker school. 

So US and this is page 13 Jason US and red means USA, blue means Germany. We're up about 60 percent none. And but up to six or more hours was very little. But if you get in that mid range, it was it was between one and three hours a day where they clustered that people did a fair amount. 

Average time spent using chatbots outside of worker school again, half are none, but around one to three hours a day. So people there's there's a you know, there's a substantial number of people who were starting to use this. Then they looked at at whether that behavior had changed. People who'd used chatbots or not before deliberation in Brazil, it was 62 percent in the US. It was 59 percent and so on. Afterwards, the the increase after they learned went up everywhere from seven to 12 percent increase in chatbot usage. So people became more curious. Mm hmm. 

The belief that that AI has a positive impact. Here, it went up much more modestly, four percent ish in all the countries except Spain, where it went up double that eight percent. So then this part is interesting to me. The the statement that received the most approval in all four countries. AI chatbots capacity to increase efficiency by automating tasks is saving many companies a lot of time and resources. 

And everybody's kind of said, yeah, so around 70 percent is the statement that received the least approval in all four countries. People will feel less lonely with AI chatbots. That was really interesting. It goes to what you said there's a finding later about about romance and chatbots and whether it's right or wrong to use this. And I think there's a there's a creepiness factor there about. 

Perceived, fake human relationships. And that makes me glad. I think people are going to be suspicious of this and they should be suspicious of this. I'm reading an AI book right now called. Now I got to remember the title of it. 

Back to my library listening to a cointelligence by Ethan Mollack. And he's arguing that that we need to treat chatbots as human, not because we think they are human, because it's the best way to interact with them. But you're hitting a real fuzzy line there where that occurs. Sure, sure. The statement that gained the most approval as a result of deliberations that numbers changed. Chatbots replicate biases that exist in the data they were trained on. And so that increased statements, the lost the most approval as a result of deliberation, that is to say, of learning across all four countries. 

The increased use of AI chatbots will lead students to losing their ability to think critically. So people were fearful of that. And then after studying, I became less fearful of that. Yeah, that's great news. 

I think that's yeah, I think it is. So there's all kinds of data in here. Then they went with proposals of what should it do. And I'll end here. Should AI chatbots inform users that they are interacting with a bot? Percentage of participants who support the proposals for this. And every time an AI chatbot responds to a question, most said no. Periodically, up from there. 

First time they registered to use AI chatbot. I'm surprised it wasn't over half, but it was nearly 50 percent. And then the statement unnecessary for AI chatbot to inform user as users assume responses are generated three to five percent. In other words, no, OK, you can't just assume this. Right. 

I agree with that. Which sources should this is important? Which sources should AI chatbots draw information from? And so peer reviewed scientific information or discussions in major press outlets. People up to 84 percent in Germany said they should get that information. However, at the same time, you have paywalls going up, not just to readers, but also to chat to AI companies around news. It's one of the things I discussed with the Nordics today is saying we have a responsibility to the information ecosystem and to this future technology to make sure it's not all screwed up. 

And we've got to figure that out how to do that. Globally recognized authoritative sources like WHO, again, Germany, 84 percent. Then sources from users, national organizations. 

I'm not sure exactly what the member doesn't down more half percent. So people are saying we want these things to be trained on authoritative sources. But the authoritative sources are bucking and saying, no, we're going to sue you first. So we need a conversation in society and this kind of data can bring that. So there's a lot more in here. But I think that was, you know, some interesting stuff to get us to understand how this is. 

You really want to put what ultimately points to is this general, at least in my mind, is this general idea that the the knee jerk reaction when it comes to the impact of AI where we're at right now is to, you know, once again, be afraid and to give it too much like overwhelming power and and and potential when we understand more about it and when we take the time to interact with it and research and and really take the time to understand it better, then that fear dissipates or at least reduces. That's really ultimately kind of seems to be the overall theme of what we're seeing here over time. They realize, oh, wait a minute, maybe maybe we don't need to be so reluctant just right from the jump. Once we know more, we're willing to accept it for what it is, you know, within. 

But it's not all in that direction, interestingly. So one question here was, should AI chat bots use the user's past conversations to improve user experience? And they have three different variations of this. And the first one was even if the user is not informed and beforehand, 42 percent said no, afterwards, 57 percent said no. So the deliberation also went to understanding reasons for caution and standards that should be set. So it works both ways, right? It's not just all toward toward that. I think it's I think it and what this also does, what I like about this, this movement of deliberative democracy is that it says that if people are actually informed, they are going to make better decisions. I'm not talking about simple misinformation where something stupid appears on Facebook, we think that that corrupts them for life. 

It's not that. It's that when people study, when they learn, when they are educated, they'll make better decisions. And that also is borne out by this, that we can have a better relationship with this technology, a responsible relationship with the technology. If we can inform people well, but too much of media coverage is either. 

I mean, I've tried to read a bunch of AI books lately, and it's either this is cool. This is how you should use it or this is dangerous. Watch out. And and I don't see something as smart like this deliberative work. So thank you for giving me all that time to. 

Yeah, yeah, I think it's fascinating. And I think you just put it very you you were able to put into words better. What I was like, where my mind was at right then as well. How you phrase that? 

Is there anything else interesting that stuck out from you from the report? 

The predictability change too. Yeah, I mean, it's just it's so it is super dense. And I think that was that was the kind of the general theme that I walked away from it from is just that, you know, we should all be doing what we can to understand this a little bit more beyond the kind of alarmist or or the, you know, super, you know, overly positive fanboyish kind of perspectives. There's a lot of value here if we spend the time to get to understand it and to know what what it all means in the context of it. 

And if people are informed and feel informed. Yeah, yeah, absolutely, absolutely. Well, cool stuff. I don't know why I put Johnny Ive and Sam Altman next, except that I do because this picture really cracks me up. I when I look at this picture of Jony Ive staring deep into the soul of this iPad, I just I imagine laser beams coming out of his eyes. But why are we even talking about Jony Ive and Sam Altman? I mean, we've been talking and this has been the rumor for a while now that Jony Ive EX-Apple and then Sam Altman, current open AI were teaming up potentially to build an AI personal device, a personal device of some sort built around AI at its core. Those rumors stoked once again, because the information, which I do not have a subscription to, unfortunately, so you'll just have to see this little image here, has a report that says they are now seeking funding, which I can't imagine is going to be very hard for them. 

This is kind of a star duo attempting to do this. Ive is aiming for up to one billion dollars in funding, but we still don't really have any details as far as what exactly this would be. They might not even know really, it's it's pretty early for what a device is. What is the post smartphone device that, you know, exists solely built around this, you know, this current kind of evolution of artificial intelligence? Is that the humane AI pin which Sam Altman is, you know, a pretty large investor in, or is it something else? 

We don't say says in the chat, Oh, Lord, help us. It will have one button and cost $2,000. Yeah, probably. What are you supposed to get your rabbit, Jason? 

That's a good question. Yes, so I do have the rabbit R1 on order. I want to say the last time I looked into it, I want to say it's something June. So here in a couple of months, I was not first in like I was a little bit later. I jumped when they made the announcement of a perplexity being included. That was when I was like, OK, this this makes sense. I want to try. 

So what's what's the other one? The button with the hand projection. 

Oh, oh, what is the oh, dang. Why am I suddenly blanking blanking? 

As you asked. Sorry, yes, my fault. Totally your fault. 

Projection. No, that's not. Humane. Yeah, that is that is the same thing. Yeah, the human AI pin. 

You mean I've been right. So those seem to be the only two user interfaces that are AI specific. My question is, and your doctor Android here. Do you think that it absolutes the phone or do you think the phone is the best vehicle for AI? 

Well, I think that's really the question that people like Johnny Ivan, Sam Altman are probably looking into. I mean, I'm sure that they believe that there is a post smartphone device that the smartphone has been played. We know there's not a whole lot more we can do with it. What what if we take this new technology and we lean into this and this is the future? I'm not entirely convinced that at least where we are right now that a device like this can do much more than a smartphone that also does these things. I don't know what the differentiation is there unless we start getting into and we have with the humane AI pin and, you know, some of these other devices, the rings and all that stuff. 

We can get the different glasses. I forgot the glasses. Right. Like it really kind of seems like where they're headed right now is this kind of like in some way, shape or form a wearable device. 

So not something that we keep in our pocket all the time. Something that's ambient and absorbing the the world around us and, you know, turning all the data that surrounds us out as we walk the world into the data stream that informs the AI potentially. And I suppose that could be interesting, but I think it's it's really early to tell whether it's going to differentiate itself too much from what we have right now with smartphones that can do a lot of this stuff. Benedict Evans, who I quote often, argues that the the Q &A interface is not terribly useful, and I agree with him. I never used my madam A or my Google thing to query it for information. Yeah. Right. 

I did. And when I first got the Google home, I would do that a handful of times, but it was not sticky. It was not it did not become the place that I went to answer questions. You know. Right. 

And. So what's the. Is the format. I'm going to point to something and say, what's that or do something about that, which is kind of the glasses perspective. Right. Right. 

Right. Is it the the voice stuff we had on our desks and some people use, but some people like me didn't because Google and even the chat type interface, I just don't my reflex maybe because I'm old and now I'm addicted to Google. And I think search first. I don't think of things in question form. 

I would fail at jeopardy. Mm hmm. Right. And and so I don't know whether these devices, if you're talking about if it's if it's AI based, which means right now it's LLM based and chat based. 

I I don't know. Taya has another good point in the chat saying that the phone was so successful because it was social, not because of access to information, which is interesting. I think I think that that goes against the fears of some people who think that the phone is anti social. I agree with you that it is essentially social. People are often interacting with other people using the device. And so do you really want to interact with the machine this way? I don't know. Yeah. I mean, I'm the guy who reboxed my iPad and said it'll never take off. So I'm completely wrong. 

I mean, I mean, I think, you know, do we want to do we want to interface or communicate with a machine like this right now? Maybe not because we haven't experienced. 

I mean, we're starting to experience a little bit of it, but we haven't until recently experienced a machine that we could interact with in a human like way and get anything that we felt kind of matched what we're used to. And I think we're getting further. We're going further down that road. 

The more we go down that road, maybe the more comfortable we get in, you know, asking questions of a device and everything. But I agree. Like it's not it's not my preferred method of living life and interacting with technology is always asking my phone. I do sometimes also there's also this idea of like, and this has been a promise for so many years that technology will get to know us so well that it will give us the thing we need before we feel like we have to ask for it. 

And I could see these devices trying to do something along those lines. But again, my experience with systems like that is that it rarely ever gets it right. And even if it knows me really well and it knows what it what it thinks I want right now, that doesn't mean I want it right now. Like I might want that later, but I don't necessarily want it right now. 

Why would you think I want it now? And maybe the systems get to a point to where suddenly it's really good at that. And it's like, OK, now I get it, but I'm still unconvinced as far as I'm concerned. Yeah. 

And I think because it can't know the full context of what you're thinking or what you're doing at the moment. I had a discussion many, many years ago with Marissa Meyer when she was at Google still where I was talking about hyper local and she said, No, I think you're wrong, Jeff. I think what we want is the hyper personal that it does know you that well. 

And certainly that becomes a shortcut to doing certain things. But the other problem is that as we've discussed again and again, these systems aren't good at facts. If they don't know something, they don't know how to say, I don't know. They make something up. 

And that's good. You know, can you give me a, is there a five o'clock flight to Chicago? Well, if it doesn't know, it's going to tell me yes. 

And then I'm not going to be on the flight. And that goes to the second problem of agents. And we have to have a trust the machine before we're going to trust tasks to the machine. And these interfaces strike me as requiring that leap that we haven't made yet. Yeah. 

Yeah. There's a lot of nuance that's at play in there, like an agent, like I want to fly it at five o'clock. I want to fly it at five o'clock, but I'm not willing to fly that airplane. 

And I'm not, you know, willing to sit, you know, the back of the plane. I like to, I guess these are all things that it can learn about you over time and everything. But I think that there's a lot of nuance and a lot of stuff that we as humans probably just instinctively know about ourselves, whether we know to verbalize it or, you know, whether we're conscious of it or not, there are certain things that we know about ourselves and we act upon that, that maybe a machine, at least right now, isn't capable of picking up on. And that leads to an unsatisfactory experience, you know, I don't know. 

Yeah. And I heard, I forget what I was looking at the day where someone was speculating about, you know, how you can ask the, ask the machine, should I take this drug with that drug? I would not trust these unstructured systems with that today. A structured system that is programmed with exactly that information. 

Okay. That's what every pharmacist does is they look up, but they're looking up at very structured information. When I was listening to the book whose title I'm forgetting already, because I'm old, Co-intelligence, you know, he makes the point that this is one of the things that came out and that the first story we did is that AI enters in some level of randomness. You ask the same question of it 10 times, you will get 10 varying answers. Mm-hmm. Yeah. So you don't get consistency from it because it's weighing things differently. 

And one word changes the next word, the next word, the next word. So I don't know how good LL, though we can talk to LLM so they can talk to us. I don't think they're going to be reliable agents. I don't know that they ever will be. Yeah. 

Yeah. Curious to see for sure. 

I can't wait to get your rabbit though. I can't wait to play with it. Yeah. 

I'm looking forward to it too. I'm also, I'm not going to lie. I'm also kind of preparing myself to be disappointed. Yeah. And I can't help that because like, you know, sometimes with technology like that, like, I'm curious to get it. And I'm not, I'm not saying that I'm going to, you know, get it and immediately be disappointed. I'm just saying like my experience sometimes with new technologies that promise this new thing is that in the end, it doesn't quite deliver what you are thinking. And I don't even know what I'm thinking about this device. 

I'm just curious to see like, what are they promising? And I will, I will say though, now that I'm thinking about it and talking a lot about it, that my experience with perplexity over the last couple of months since I ordered the rabbit really does kind of prime the pump for how I might use the rabbit because now I'm very used to the things that I use perplexity for and essentially to a certain degree, that's like a perplexity appliance. 

Yeah. Which is a lot of peace. Are you still liking perplexity, by the way, as your primary? 

Totally. It's, it's my, yeah, it's absolutely the AI system that I use on a regular basis is also the one that I paid for. So it's kind of like I bought into that camp. But I find myself using it more like every day I find my, I find new uses for it. And yeah, so I'm totally digging it. 

Give me an example of one new use you found the last few weeks. 

One new use that I found in the last two weeks. I mean, I mean, the things that I end up using it, well, actually, I actually just discovered this today. And this is by no means like a new thing, you know, LLMs are great at summarization, but I didn't realize that the perplexity plug-in had a little button that I had been missing up at the top that says summarized depending on whatever webpage you happen to be on. So, so if I'm showing this article up here, you can see this little summarize button. 

My eyes just never saw that. And so, you know, I would, if I wanted to like take an article and get a short synthesis of it, I would often, you know, copy the text or paste the link into the actual website and that sort of thing. And having this here, it was like, oh, wait a minute. So here's cloud next, I can actually just demonstrate. I'll go ahead and hit summary or summarize. 

It automatically sends it to cloud to to cloud two, because that's what I have it set up as. And it gives me, you know, kind of the embedded summary within the context of the browser. And I can copy that text right down here and you know, it gives me my citations, which of course is just this webpage, because that's what I did. But I mean, you know, we're things like that that like when they're integrated and when they're kind of part of the experience. And the more I use it, the more I understand kind of like the, the, I don't want to say guardrails. Sometimes I feel like guardrails feels like a bad word or something like that. But I understand that like when I'm when I get this summary, I'm not taking it at full value, but I, but I do see it as a way that sometimes like, you know what, that's a long article. 

I just want a little short, you know, brief understanding of it before I dive in deeper or maybe I decide to go somewhere else. I don't know. I'm synthesizing a lot of information now kind of running my own business. And, you know, so sometimes these little shortcuts are very useful and it's nice to make it easier, you know. 

So yeah, I am liking it a lot. But speaking of Google Cloud, because I just showed the webpage for those who watch the video version, Google Cloud Next 2024 has been taken place in Las Vegas. Of course, it's Google, it's now. So AI took center stage. Some of the highlights, Google Cloud's revenue grown five times in the last five years with more than 60 percent of funded generative AI startups. 

Nearly 90 percent of generative AI unicorns are using Google Cloud or Google Cloud customers, which is pretty impressive. Yeah, that's a number that you want at the beginning of your keynote. But I think, you know, some of the big news that they were announcing Gemini 1.5 Pro. So that's becoming available to the public for the first time. It's available through their Vertex AI platform, which is its AI deployment platform for the cloud. It has some new capabilities. It can listen to audio as one form of input now, for example. And then, you know, I think the one of the features that's been getting some some, you know, significant coverage is Google Vids, which I don't know how you feel about this, but it's essentially an extension expansion for workspace. So this is working with, you know, your docs, your sheets, those those Google apps to replace a slide deck and instead kind of create like videos and videos that tell the story so that you don't have to have the interactive slide deck. You can have all this information presented in a video format. And Google says it's really easy to do. 

I think PowerPoint did horrible things to human reasoning and interaction. It oversimplifies everything and makes things seem really simple. That's that's that's a get off kids get off my yard kind of statement. But I kind of despise PowerPoint. Whenever I go give a talk somewhere and I say I don't have PowerPoint slides, people applaud, so I think they're sick of PowerPoint. But I'm wondering whether videos can be better or worse. 

Yeah, right. Yeah, that's a good question. 

Because the other thing is you see those videos to where you see the hand that supposedly drawing and it's drawing that thing or I go to a conference where somebody's off to the side and they're drawing this ridiculous thing on a big board, which just has some buzzwords and it really doesn't tell you anything what happened. It just tells you, oh, you could have just outlined stuff and it would have been better. The AI could have done a better job of outlining what happened. So I don't I ought to see it before I start grousing like an old grouch. But maybe just talking about your ideas and having a conversation might still be better. I know I sell. Yeah. 

You know, people people learn differently, though. There are visual learners and, you know, people who can synthesize by just listening intently versus seeing with their eyes. I kind of I'm kind of like halfway in between both camps. 

It just depends on the context. But this will be public beta this June. Oh, and another thing that they announced is Google Meet offering meeting notes, summaries, you know, like the others already do. Ten dollars per user per month. So useful. 

Yeah. Yeah, a little bit of keeping up with everyone. What's interesting, I put a story in here that open AI, Google and Mistral all released new models almost simultaneously. There's there's a huge arms race going on still. 

It's not like things have calmed down. They're all trying to release new models. They're trying to show new applications of them. 

They're trying to offer them in new ways. And obviously, that's OK. It's what they do for a living. But I think we're still in a position where people. Oftentimes, it is a cool gadget without us, without a problem to solve. So we'll see. 

Yeah, absolutely. We'll see. Well, what what if the problem to solve was was, you know, applications like applying AI to advertising into the marketing industry. Apparently, WPP is the world's largest ad group, and they're doing that. They announced a partnership with Google's Gemini AI becoming at least according to this search engine land article, a pretty integral integral part of its ad process. 

So Coca Cola, L'Oreal, Nestle, other brands leaning into Gemini AI for things like ad narration of voice over script generation. So I. Oh, yeah. OK, I'll get to it in a second. 

Product image creation aligned with the brand guidelines, enhancing creativity, understanding of the brands so they get a better understanding of the brands, optimizing the content and predicting campaign effectiveness. Yeah, I don't I don't know. But at the same time, like if we're if we're, you know, generating ad narration, generating voice over script all with AI at a certain point, doesn't it all kind of start to sound sound like the same kind of language? You know what I mean? Like, yeah, where is where is the generated stuff? It sounds more of the same. 

It's more it is literally more of the same. Yeah. Yeah. And where's the I mean, we all hate ads to some extent or another, but we all recognize that some ads are really creative. The DDB Volkswagen ads back in the day of the little Volkswagen bug. 

You're probably too young for that. They were really creative ads. The absolute vodka with the shape of the bottle as art. You know, I don't know that AI is going to come up with that kind of stuff. On the other hand, you know, I tend to have MSNBC on much of the day, just as background noise. And on Fios, some of the ads are sold by Fios. Some of the ads are sold by MSNBC. Neither company really wants to sell them, apparently, because the audience is too small or not provable. And so we get the same ads for a month over and over and over again. 

And when it's the Loomy lady putting something in butt cheeks, I just want to turn off the sound every time. And so one small advantage here is at least two different creatives. You're going to show me the same damned ad for the same stupid pills. At least show me five different versions of it. I don't know. Yeah. 

Yeah. That's true. Oh, God, that is the worst getting the same ad 20 times. 

There used to be a philosophy of rate caps. So you wouldn't repeat the ad. An advertiser didn't want to pay to have the same ad over and over again. But now it's just cable TV time is so cheap. Yeah, I'll buy out the month. Sure. Sure. Yeah. Right. Yeah. 

I think time will tell whether the quality will be of a point. We'll get to a point to where that's better than the alternative, whatever. These AI systems are actually helping to create. But yeah, interesting stuff. 

All right. We've got more coming up here in a second. Activist groups calling on big tech to take stronger action against AI generated misinformation. 

And of course, things like deep fakes. This is all, you know, ahead of the fact that there's 60 national elections happening around the world this year. These activist groups are calling for more aggressive policies for preventing, you know, the political propaganda. They consider dangerous, prohibit deep fakes, label AI generated content and political ads be more transparent, which I can get behind the transparency. 

Absolutely. About the data also that's powering the AI models. It's a whole laundry list of stuff here. Experts are warning that AI generated misinformation is already causing confusion in elections worldwide and that that's going to ramp up. 

So in my contrarian way, I had a few reactions to this on Twitter and elsewhere. We're at number one, they're concentrating on the technology. And if they're going to go after misinformation related technology, then should they go after Microsoft Word and Photoshop and certainly radio and TV and Smith Corona type writers or whatever, right? 

It's the actors and the behaviors that you want to go after. And to expect the technology company to clean the world up for us is, I think, a rather foolhardy and be impossible. And the other issue is, and this is real heresy for me in journalism, because journalism is in the information business. So we think we're the ones to solve disinformation. I don't think information is the problem. I don't think disinformation is the problem. I don't think that when exposed to disinformation, people who are otherwise sane and sensible suddenly said, oh, OK, I'm going to be insane. 

Now that goes against what I said earlier about the liberty of democracy, but it depends on the information, depends on how serious you are, being educated. The problems we have are deeper than that. It's time to read, Hannah, around about people's lack of sense of belonging in society. And I don't think machines can do anything about that, good or bad, which is also interesting for the first story about whether or not people think that machines ought to make us less lonely or not. Maybe that could help. I don't know. But I think that we're still in a moral panic mode of blaming the technology and thinking that's going to solve all the problems if we can somehow turn it off or fix it. 

And I think that's a fool's errand. So yeah, some more people went and they yelled at Big Tech, and Big Tech should be doing more. They should have more transparency. 

Facebook just announced a whole bunch of things about how they're going to try to label things that come out of AI. That's fine, but good stuff can come out of a tool and bad stuff can come out of a tool, whether that's Photoshop or whether that's Dolly. So I don't know that yelling at the technology companies is the right path here, but it's what they're going to do. 

Yeah, and I don't know that there's any way to really stop it, because there will always be people who jump right to the big, the scary. And there's just a lot of support for that viewpoint, especially with things that people don't understand too deeply. So yeah, interesting stuff. 

But yeah, we've had this conversation quite a bit. We're going to keep having it. Because it keeps happening. That's the thing. Because it keeps happening, does that then mean that it's almost inevitable that that there's going to be success found on the other side of this continuous battle? 

No, I think it's the opposite. I think we start to, well, all right, now, okay, I was too cynical. 

Maybe that's the excuse that I have. 

I think there's success, but I think it comes from the side door. That is to say that it doesn't come from an obvious, if we just pass this one regulation, if we just fire this one executive, if we just change this one piece of software, everything will be okay. Instead, as we adapt to technologies, it's always about norms and standards and understanding. I just got my galleys for my next book. There it is. Oh, look at that. 

Hey, that's awesome. 

It's great. In there and in the Gutenberg parenthesis, the other book, still on sale, this counts. What I see is that after a time with the new technology, the technologies fade in the background. I've talked about this before. We figure out what's good and bad about it. We figure out as a group in a society what works and what doesn't. We know where to hold standards. We invent the institutions needed to enforce that. And the example always is, for me, that when print started, it was not seen as reliable. Anybody could make this thing, like anybody could make a Facebook post or a Twitter post. 

And you don't know where it came from. And authority was a social system before print. And then print came along and we invented these institutions of publishers and editors and so on and so forth. 

And then print took on more authority than what we can now call rumor. Well, now we're switching around the other side of this, where the institutions aren't working anymore and we have to rely on social systems. And in this scale, that's difficult. But I think we will. So yes, I think Jason will come out on the other end better off. I think we will figure it out. But not because we did some simplistic little thing. 

It's because we'll do the work to figure it out. And cut it off at the supposed source. Yes, I think so. Yeah, yeah, interesting. I thought this story was pretty fascinating. The New York Times, data collection, data. I don't want to say data harvesting, but finding data to feed the AI systems that we have, building up everywhere, Meta, OpenAI, Google, all these companies are saying, we need more data to feed into our systems. 

And there's just not quite enough. The New York Times wrote about the race that's taking place right now to get a hold of as many sources of data as possible. So we're talking customer voice recordings, potentially device sensors, YouTube videos, which a large part of the article actually does focus on OpenAI actually created Whisper to transcribe text from YouTube videos. Did so with more than one million YouTube videos, something YouTube prohibits in their terms. OpenAI believed it was fair use. 

So there's that question again. We as humans have the right to watch and even to notate. And all this stuff around the videos that we watch on YouTube as one example, does an AI system also have that right? 

Or is it for some reason different for that system? But according to this report, Google also did this with its own YouTube catalog. It also did this with docs and sheets and everything, training its systems on publicly available data from those avenues. And yeah, it's just kind of interesting how the requirement, the need, the thirst in AI right now is more and more and more data. And at a certain point, at a certain point, have we run out of data? 

Have we run out of quality data to feed into it? And then what happens? Then you start having one system creating for another. And then what does that lead to? It's just interesting. 

Yeah, there's a lot of layers to this topic. I mean, I think the first to me is that we have to have the discussion at the level of principles. As you just said, Jason, if it's okay for me to watch a video or read a news story or watch, read a book and learn from it and use that knowledge, that's the essence of enlightened society is we should be able to do that. 

So at a level of principle, I don't have any problem with this. So where is it that they think that open AI went over the boundary? And I don't know. That's a discussion to be had. Is it scale? Is it they did too much of it? 

I think scale is up a lot. But again, it is scale. But at the level of principle, I don't know. 

I mean, if you could, I could spend my whole life doing nothing but reading books as fast as possible in the library. I would read more than I do now or you do because we're busy doing other things, trying to make a living. 

I will have advanced that scale. Does it make it any worse? Or is it better? Am I more educated? 

Is that better? And what do we think as a society? Do we want to educate these machines? So that's one layer is what's the principle involved? 

Where is the line? And I don't think I haven't heard any clear expression of that, except the second point is how it was acquired. And if you read a million books and you didn't pay for them, then books three or however many books there are in that database, that's an open question. If you did go behind the paywall of the New York Times and didn't pay for it, that's an open question. However, if you had one subscription in the New York Times and read it and it's not scaled at that level, it's just I'm using it for a different purpose, which is to train my machine. 

What's so wrong with that New York Times? I paid for the use of it. So what's the limits of terms of service at that level of principle? 

Third is, as I already hinted at, what's the responsibility to society? If these machines are going to be out there speaking to us, do we want them to be stupid and wrong all the time? Or do we want them to be better educated? Is it our interest to do that and to figure out how to do that? My argument to the news industry, and what I saw when I talked to the Nordics this morning, because they're trying to be more constructive about it and saying, okay, I know you're going to reach for our stuff. Now give us access to the technology so we can get benefit from it and work together. That's a smarter way to do it. Or I argue that the news industry should create an API for new. They should say, okay, okay, you want our news, you can have it. Here's the deal. 

Here's how you earn a key. You give us certain standards and you pay us and we negotiate that. And fine, there's a lot of ways that we can negotiate all of this. The New York Times, in this case, is in a beautiful conflict of interest considering that they're suing open AI. So there is a corporate view of this, of taking data is bad. Okay, in some cases, one can argue that. 

But in other cases, I can argue the opposite. The last angle to the story to me does come back to scale. And as ever, I think it's going to become a law that every three episodes, I have to quote the stochastic parents paper. But this is where they argue that building really large models gets us into this fix. 

And do we really need really, really large models? I don't think I've seen really good reporting on that. We've seen that they get bigger and bigger and bigger. They have ever more tokens. They have ever more relationships mapped. 

They do learn better, but they learn at a very high cost and a very low ability to monitor what they're doing to audit it. And they're running out of text. So when you get into this world of synthetic text, which is one of the things the Times wrote, it's about, that feels really hinky because it loses ground truth. 

It loses that basis. And I'm not smart enough about this stuff to understand what the uses of synthetic text are. When I went to the last World Economic Forum event about this, there was a lot of talk about synthetic text. And obviously, they're going to try to do synthetic data, I should say. They're going to try to do it. 

But I wonder if it goes to my friend Matthew Kirshenbaum's story, the text apocalypse, where it just ends up with this gray goo of remade stuff. It's like the kind of a cow becomes our culture. Yeah. 

Right. That's what comes to mind is something along those lines, like the game of telephone. It's like it started off as one thing. And in bouncing back and forth between the systems a million times, what do you even end up with at the end? 

It's the most basic, boring. I don't know what that looks like. Is it even English at the end of it? 

Who the heck knows from all the translating and understanding and then putting back into the same type of grammatical structure that the other system does anyways? And does that just end up, like you said, into a goo at the end of the day that's even usable? I don't know. 

Great. It's nice to have a bunch of a solid block of data to train your systems on. But there's also something about integrity with the data that you have there and high quality data. And there's something about synthetic output and training your systems on that that doesn't scream high quality to me. Again, who the heck knows long term? 

Who knows? It would be lovely to get somebody who's a proponent of synthetic data to argue with it and explain how it works and how you can guarantee its quality and why you need it. Or maybe at some point you just run out of. Again, the other point that I was going to make here is that, yes, you run out of data and the response to that is not to make it up. 

The response to that should be to say who's not there. What perspectives? What expertise? What history? What groups of society? What languages are not there and should be represented? And one way to do this is to put a whole bunch of money into Wikipedia to get them to get more diverse contributors, Wikipedians from around the world, from different perspectives, and to increase the amount of knowledge there and to support academics to write more papers. That's what we should be talking about, I think, is how do we add in more quality information into this rather than, any data is good enough. It's not. 

No, exactly. There needs to be some sort of level of quality, I think. And just real quick, before we move on to the last story, it was kind of along this lines, but Reuters had a story that focused on some early 2000s properties and services. You remember PhotoBucket? I remember PhotoBucket big in the MySpace and Friendster days. And now, at their height, they had 780 million users. Now, they're a fraction of that, 2 million users, which I'm just kind of surprised that they're still around. But what they do have is 13 billion photos and videos that they can now monetize by licensing out to AI training. And there's other services like FreePick and others that find themselves in a similar position. It's like, oh, we've been sitting on this data. We're not quite as hot as we used to be, but now here's a really great opportunity to start monetizing in a different way. It's just an interesting direction. But it'll probably be a blip. 

Yeah, it's used. Okay. Yeah, absolutely. Right. 

Once it's in the database, it's there. It's represented. Moving on. Yeah, indeed. And then let's see here, Perplexity. We talked a little bit about this earlier. Going to be selling ads, which I think is interesting. The site actually says that search should be, quote, free of the influence. Advertising driven models. But hey, everybody got to make some money. 

So, you know, that's exactly what Google said. Exactly what Zuckerberg said. They both did one ads when they began and then look what happened. It's just inevitable, man. And Google's also talking about putting at some level of their AI, putting up paywalls. So they're all looking for revenue to figure it out, which is to say that it's not an obvious business model. It's the same mess that media is in. 

Yeah, yeah, yeah, indeed. Not a whole lot there to dive into, but as a Perplexity user, I just think that's interesting. And then I'm like, well, I'm paid, so I guess I won't see them. I hope I won't see them. 

I would like to not see ads considering that I'm paid. And then finally, last week, I took a look at stable audio 2.0 for music generation. I had a lot of fun with that. I'm not going to go into it very deeply this week, but you sent me a link to Suno, which is a service that a lot of people are passing around right now as another example of music generation. And actually, it's creating some really cool stuff. 

I'll play you something in a second. But with the free basic plan, you get 50 credits every day, so which is basically like 10 song generations. And I mean, you can see there's just tons of stuff here. I'm not sharing right now in the right way to play you audio, so I'll get there. But I came across a video and I'll show you this video because it really just kind of blew my mind. Because when we talk about music generation and AI, often where we land is, yeah, but human music has a certain feeling to it, has a certain humanity to it, and you can feel the emotion and everything. Robotic music doesn't have the emotion. So then I came across this video. 

This is Futurepedia on YouTube and he did a generation, and I don't know how many steps it took him to get to this point of a blues song. And check this out. I hope you can hear it. This morning, feeling so low, got the delta blues down in my soul. I mean, what? Yeah. And you know, this is like a minute and a half long song. 

It's just, it throws me away that this is where we're headed. Like, it really does. Because it totally has, you know, it has a hard to it. Like, I was, I listened to this whole thing. I was like, that's actually a really good song. And I don't know music at all well, but it would, after that first, it shifted a little bit. I don't know if it was key or tempo, but it surprised you, right? It did something that I didn't expect to happen, which with AI, you always expect the expected. 

Yeah, totally. You kind of expected to be bland and a little boring and whatever and vanilla. And then right there, I think what you heard was him going false out of or whatever. 

Right after that too, it changed a little tempo. Oh, okay. 

Yeah. Yeah. Yeah. And I mean, the whole spectrum of that music and the rest of the video is good. He goes into other examples and everything. And I haven't really had much time to play with Suno myself, but anyways, I can't wait to see you do it. I'm pretty impressed by that. 

Because you're a musician. You can know when it's, when it's BS and when it's good. Yeah. 

I mean, I think what's interesting here is with Suno, you have the ability to put in kind of structure like around like the lyrics and everything. So you can write your lyrics, you can tag it as verse, verse two, chorus, whatever. And it might not get it perfectly right, but it factors that into how it generates the thing. And my understanding is it takes a while iterating on an idea before you get to the final product, you know, the final like music track that is kind of the wow-y like that one is. 

But anyways, very interesting stuff where we're kind of, you know, we're seeing signs of the possibility that some of these generations don't always necessarily automatically have that robotic quality to them that they can have a little bit of soul, dare I say. Yeah. Without getting pelted with rotten tomatoes. 

And just surprise. Yeah. And an aesthetic pleasure. Yeah. Yeah. It has to go past. And the interesting thing too is, Jason, I think our standards wisely rise before that a fact that it could do anything could make any picture, even if it had 10 figures on each hand. We said, we gave it allowance. We said, okay, that's pretty cool. 

Look, I can do that. And then our standards rose. We expect five figures on a hand and our standards rise. 

We expect a little surprise in the music. Right. And I think that's a very healthy thing to go on. And that's the race I want to see the AI people go on is to surprise us and delight us. However, that's also going to freak people at the same time. 

The better it gets, the more it's going to freak people. Right. Because something like that that we just played for some people is going to be like, whoa, that's amazing that technology can do that. 

And for others is going to be like, nope, that's too far. Now you're now you're intruding on my human right to create human sounding music. And, you know, that that 

you're well, and when he when he goes to blues, is it cultural appropriation now? Not by a white person, but by a white person using a machine. 

That's an interesting. Yeah, that's an interesting angle. 

I don't have the soul that rhythm, but I can get the machine to do it for me. Yeah, that's interesting. 

Oh, boy. Interesting stuff. Well, we've reached the end of this episode of AI Inside. Hopefully, hopefully you've enjoyed at least a little bit of the music that we played there at the very end. Jeff, tell us a little bit about all the books that you're working on. I love the cover, by the way. 

Yeah, thank you. Thank you. So this is this is out in October, the web we weave. While we must claim the internet for moguls, misanthropes and moral panic. But today I want to plug something a little different. I did a long morning, you folks, long paper. commissioned by the California Chamber of Commerce on the California Journalism Preservation Act, which is, and if you go to medium, JeffDrivers.medium .com, I put it all in one post, but you don't read it there. It's a 79 minute read. But this is a link to the PDF. Neiman Lab put up two excerpts today. And it's my analysis of this legislation, which is an effort by the lobbyists for the newspaper industry to get money from Google and Meta. 

I think it's terrible legislation. I go through the history of it. And I go through the history of consolidation in news and news and copyright and so on. But at the end, I propose a whole bunch of different solutions that are being proposed as better alternatives than this. It's not just affecting California because there's also a version of this law federally. There's another one that was just proposed in Illinois. I think they're all bad. 

I think when we learn from what happened in Canada, it's miserable. And so I learned a lot of researching and writing this. There's some really interesting stuff in here. There's some geeky stuff in here, too. There's lots of footnotes, lots of links. 

But if you're all interested in these issues about copyright and the news and both AI and the platforms, then it's not for everybody, but some people might find this interesting. 

Excellent. JeffJarvis.medium.com. 

They can find the link there. Or if you go to Neiman Lab, last I looked up was the top story with two excerpts there. Excellent. But if you read just the excerpts, you miss out on some of the fun history stuff, which I really enjoyed doing. Yeah. 

Right on. Excellent. Everybody should check that out. Thank you, Jeff. Thank you. Yeah. And usually you can watch this show as we record it live at the Yellow Gold Studios YouTube channel at Yellow Gold Studios on YouTube. For whatever reason, this week StreamYard was not playing nice with YouTube, so you couldn't see it live there. 

But it will be published there. So just go to YellowgoldStudios.com and you will find everything that we're doing here with AI Inside, at least the video version. If you scroll down, I've got the little playlist for all of our podcast episodes. But of course, aiinside.show is the page that you can go to subscribe to the audio podcast, which is what most of you actually do. 

And we appreciate all of you. And then, of course, patreon.com/aiinsideshow. That is direct support line to keep us doing this show week after week, learning about AI along with you. Jeff is going to be out next week. So I will miss you on the show. 

Same back. But I've got a little bit of a kind of a special episode. I figured instead of getting another, just picking a random guest co-host, I'm going to do two interviews. And they're both. See, it takes two people to replace me. 

I just want them to know. That's right. Exactly. 

That's what you should take away from this. Going to talk with Jeremy Toeman, who's the CEO of a service called Augie, which is all about, it's for video creation. If you've got a script and you're recording your A-roll and you need B-roll to match it, it can analyze it and give you B-roll as you go for your script. A very interesting technology. And then, Surafel Defar, who's the founder of Revoldiv, which is a free transcription, AI transcription service online that I actually rely on for transcribing the shows that I do. And so this is a cool opportunity to talk to people about, you know, what they've created, how and why, and all of the inside stuff, as far as that's concerned. 

I find that stuff super fascinating. So that's next week's episode. Looking forward to that. 

And yeah, we do this every Wednesday. Thank you for watching and listening. Please subscribe. 

Please tell people about AI inside. And we'll see you next time on the show. Take care, everybody. Bye. Bye. you you