Loading summary
Transcript79 lines
- [00:01]
Michael Stelzner
Looking to transform your marketing with AI? Social Media Marketing World 2025 has all the AI training you'll need. Join me and top AI experts in San Diego this March. Get your tickets now@socialmediamarketingworld.info and save big.
- [00:22]
Matt Wolf
Welcome to the AI Explored podcast, helping you put AI to work. And now, here's your Michael Stelzner.
- [00:34]
Michael Stelzner
Hello, hello, hello. Thank you so much for joining me for the AI Explored podcast brought to you by Social Media Examiner. I'm your host, Michael Stelzner, and this is the podcast for marketers, creators and business owners who want to know how to use AI. I'm really excited about today's show. I'm going to be joined by Matt Wolf and we're going to Explore the top AI tools for creators in 2025. If you create any kind of content, written, audio, video, images, all that fun stuff, you're going to find today's show absolutely fascinating. We're going to dig deep into all the greatest tools and why you might want to consider them. And if you're new to this podcast, be sure to follow this show so you don't miss any of our future content. Let's transition over to this week's interview.
- [01:20]
Matt Wolf
With Matt Wolf, helping you simplify your AI journey.
- [01:25]
Michael Stelzner
Here is this week's expert guide. Today, I'm very excited. Be joined by Matt Wolf. If you don't know who Matt is, you got to know Matt. Matt is an AI tools expert who helps marketers and content creators embrace The Future. His AI focused YouTube channel has more than 700,000 subscribers. His site, Future Tools curates the latest AI tools and news. He also hosts the Next Wave podcast. And Matt will be presenting at Social media Marketing World 2025, coming up in March. Welcome back to the show, Matt. How are you doing today?
- [01:59]
Matt Wolf
I'm doing good. Yeah. Thanks for having me excited about Social Media Marketing World this year. It's going to be a fun event and thanks for having me on the show again.
- [02:05]
Michael Stelzner
Well, thanks so much for coming back. We're here today to talk about the top AI tools for creators and, you know, marketers that are creating any kind of content. You're going to find value in today's interview. I guess my first question, I want to speak to those who are not yet using AI that are creating content. Why should creators be using AI tools to help them?
- [02:27]
Matt Wolf
Yeah, well, I think there's really a few ways to think about that. Right. If you're a creator that just really loves the creative process and you don't want to use AI, then you don't have to use AI. I don't think it's 100% necessary to use AI. However, if you're a creator that loves to sort of like push the boundaries and explore the tools that are out there and really try to find new ways to like, up level your creativity and what you're capable of, I think AI is one of the best ways to do that. Now, if you're a business owner and content is your business and you're looking for the most efficient way to get high quality content out into the world, if that's the scenario you're in, I think using AI is a must because there's really no more efficient way to create content than to leverage some of these tools out there that are really kind of a shortcut to creating really, really high quality content.
- [03:20]
Michael Stelzner
I would imagine in your case, Matt, because you own a business and your business is all around creating content, I would imagine you would have had to had a much bigger team to be able to do the things you're doing. Would you fully agree?
- [03:30]
Matt Wolf
Oh, yeah, yeah. I mean, if I was trying to do this a few years ago, I wouldn't be able to accomplish a lot of what I'm accomplishing because I'm, I'm doing my YouTube thumbnails myself because of AI. I run a whole entire website that curates AI tools that leverages AI to help run that website. I create content on pretty much all of the social media platforms. I'm putting out pretty much daily videos between Instagram and YouTube and all of it. I'm leveraging AI to make sure that I can keep this pace. And to be quite honest, I don't feel too overwhelmed by it. I feel like I'm, you know, living a normal life, working, you know, maybe 40 hours a week. But because I can leverage AI, I feel like I've got a team of 20 people around me when really I'm only working with about two other people on my team.
- [04:14]
Michael Stelzner
Excellent. Okay, we're going to talk about research, we're going to talk about images, we're going to talk about video, we're going to talk about audio, and we're going to talk about some other tools as well. So let's start with your favorite tools for research and why.
- [04:27]
Matt Wolf
Sure, yeah. So the AI tool that I use the absolute most, the one that I find myself going to multiple times per day, is perplexity. I just, I love perplexity. And if you're not familiar with perplexity, imagine if you took Chat GPT and you took Google and you just sort of smashed them together into a single tool. That's sort of what perplexity is. So when you ask it a prompt like you would ask Chat GPT, it's first going to go and search the web for answers to your questions and then it's going to take whatever it finds and then use that as essentially part of the prompt. So it's taking your prompt plus all the information that it found and then it's responding based on all of that. So when you ask it a question, it will actually essentially do some fairly surface level googling and searching for you. It's not actually using Google, but it'll do some surface level searching for you before it gives you the answer.
- [05:21]
Michael Stelzner
There are some users that are power users like you and me that are using search inside of ChatGPT. So kind of compare the two. What's the upside to perplexity over just using ChatGPT's built in search functionality?
- [05:32]
Matt Wolf
Yeah, so with the ChatGPT search functionality I feel like it doesn't give you nearly as much depth. So if you were to go and ask a question to ChatGPT and you were to basically turn on the search functionality to have it search for you, it will kind of just like Google for you, it'll give you a few search results and maybe like a single sentence for each of the search results that it gave you. Where perplexity is going to give you like a much deeper explanation of everything. So imagine Chat GPT search functionality as more of just like a, almost like a replacement for Google but with like a little one to two sentence AI summary for each thing it finds. Where Perplexity is almost like more doing research for you. It's going and looking at a handful of websites, summarizing, pulling all that information together and then giving you a summary of everything that it found. So to me perplexity is just more useful because it's going to give me a lot more depth of information when I actually use it. Where when I use the Search GPT, it's kind of a little too surface level for me.
- [06:31]
Michael Stelzner
Well, we also know that Google has their AI tips at the top, right? Do you find this is even better than what Google is providing you when they provide a summary of your search stuff at the top?
- [06:39]
Matt Wolf
I definitely do, yeah. I mean the, the summary that Google will give you is usually like one paragraph and oftentimes it's actually not even that accurate. I mean there was some stories back in the early of when it first came out where it was suggesting people put glue on their pizza. Somebody was like, how do I keep my cheese from falling off the pizza? And it was giving summaries of like, have you tried putting Elmer's glue on your pizza to keep the cheese from falling off? And people were asking questions like, how many rocks should the average person eat per day? And Google was responding with, according to our research, you should be eating at least 10 rocks per day. And they basically was finding information from memes. So Google is actually not very good with their genitive response because it doesn't really, really know the difference between like a meme and actual helpful information. So it will oftentimes pull in responses from memes that it finds on the Internet and act as if it's fact. Where if you're using something like Perplexity, Perplexity is actually going to look at a whole bunch of different sources and only kind of pull in the most relevant information from those sources. Where Google oftentimes is only going to like summarize one source that it found and give you back that information. So perplexity tends to be much more accurate right now. That's not to say Google won't get it more dialed in in the future, but as of right now, Perplexity is definitely a lot more accurate.
- [07:59]
Michael Stelzner
But Google does have another tool that you were about to mention, so go ahead and mention that one.
- [08:03]
Matt Wolf
Yeah, so Google has a tool called Gemini Deep Research and it's unfortunately it's only in their paid program. So I think it's about 20 bucks a month to get into the actual Gemini Advanced plan. But if you have Gemini Advanced, they have Deep Research, which is like Perplexity but to another level. Right. So when I would do a search, like I did a search for example the other day on Quantum Comput, like tell me as much as you can about quantum computing. It actually went and searched 247 different websites, pulled in all of the information from 247 different websites, and then wrote me up a giant essay explaining quantum computing in all sorts of sections like what are the benefits, what are the downfalls? Where does it stand right now? Where is it going? What companies are working on it? It just pulled all of that information in and wrote me up like a, a 12 page document of everything that it came up with on quantum computing. And what's really cool is because it's Google, there's one little button at the bottom of it that says send to my Google Drive. You press that button, it creates a Google Doc Pulls that essay in, cites all of the sources at the bottom of the doc. And so it's very similar to what Perplexity does. Just like a whole nother level of depth. Like literally it will go and find information from hundreds of sites for you. So if you're trying to do any sort of really, really deep research, Perplexity is not going to there. Perplexity is just going to give you a sort of surface level. This Google Gemini deep research is going to go really in depth. It's going to essentially give you an essay back and again, it'll cite all those sources. So I found that one to be really, really helpful for any sort of deep dive research, which if you're a content creator and you're going to make a YouTube video about a specific topic, well, that's going to give you the depth that you need. And then if you want, you can take that whole sort of research paper that it gave you and have it sort of summarize it back into bullet points using Clot or using Gemini or using chatgpt and sort of summarize the main points that it found. So if you want to sort of bring it back down to like an easier to consume amount of information, then you could just pull it into one of the other tools and ask it to sort of summarize the main points for you. But it'll have a lot more depth from the start. Love it.
- [10:16]
Michael Stelzner
Okay, so so far we've talked about Perplexity and I'm assuming you can do most of what you need with their free program, is that correct?
- [10:23]
Matt Wolf
With the free version of Perplexity, you can definitely. Their free program, I believe, uses the Llama model. If you upgrade. I think there's two main things if you upgrade Perplexity. I think Perplexity has started to add ads into Perplexity. So sometimes you'll see some ads within there. If you're upgraded, you won't see the ads. And if you upgrade, you can also choose a different model. So instead of defaulting to like a Llama model, you can actually set the default to use Cloud or to use G or one of the other models that are out there. Those are really the only benefits of upgrading right now.
- [10:53]
Michael Stelzner
Perfect. Okay. Images. There are a bazillion tools out there and you've probably tried them all. So let's talk about the ones that you prefer the most and why you prefer them.
- [11:04]
Matt Wolf
Yeah, so the one that I use the absolute most is called Leonardo. And you know, just to disclaim, I am an Advisor to Leonardo. So I do have a little teeny sliver of equity in them, but I use that one the most. That is genuinely the one that I use the most because it's got a whole bunch of different models built into it. So right now, when it comes to AI image generation, if you're going for realism, there's nothing better out there than Flux. Flux is the best sort of underlying AI image generation model to get realism. Well, Leonardo has Flux built into it. So if you want to get that realism, you just set it to the Flux model and it'll do that. Leonardo also has its own model called Phoenix, which is a sort of rethinking of stable diffusion. So stable diffusion is another underlining AI image generation model. Phoenix was sort of iterated and sort of built from that originally, but it's actually a really, really good model for really creative, really colorful, high contrast images. I feel like it Mid Journey used to be the king at that. Now I feel like the Leonardo Phoenix model is even better than Mid Journey. Again, like really colorful, high contrast, creative looking images. So that's the model, that's the tool that I use the most right now for AI image generation. If you ever look at my YouTube channel, almost all of my thumbnails were created in Leonardo. It even got to the point now where Leonardo does really good text inside of your images. So I can say generate an image that has a colorful background, has a person in the forefront and it says AI News off to the left. And a lot of times I'm pulling that image straight out of Leonardo and using that as my thumbnail. So I don't even have to do a whole lot of extra work after the fact now because I can just have Leonardo generate the image with the text already on it. That's my number one.
- [12:51]
Michael Stelzner
I've got a question about Leonardo. I know they've got a bunch of different models, right? So is Leonardo built on top of like open source models? Is that kind of the idea? Like it's an interface that uses Flux because Flux is another one of those models where there's a lot of different tools that have access to Flux. And I'm with you. I Think Flux Pro 1.1 and 1.1 Ultra are some of the most photorealistic ones that are out there. Just so I understand Leonardo, they got a whole bunch of models, right? So is the idea there that it's just kind of like a easy to use, centralized interface that works on top of open source models or what's the story there? I'm Just curious.
- [13:28]
Matt Wolf
Yes and no. So it does have a whole bunch of models in it, but Leonardo also has their own proprietary foundation models. Right. So that Leonardo Phoenix, the Leonardo Phoenix is one that they developed themselves very, very similar to what stable diffusion does, but they developed it themselves, sort of modeling what stable diffusion did, and then flux is built into it. The older stable diffusion models are available in it. So if you really like Stable Diffusion XL or stable diffusion 2.1 or one of those other models, all of those models are also available. So if you get into Leonardo and you go to use it, it's going to default to the Leonardo Phoenix model, the one that they created. But there is a drop down to select any other model. So it does become that user interface to be able to use all those other as well.
- [14:12]
Michael Stelzner
Got it. All these tools are paid, presumably. Right. There's really nothing that's free out there, I would imagine. Is that correct?
- [14:17]
Matt Wolf
Leonardo has a certain amount of free credits per day. It's something like 100 free credits per day. So you can generate. And every image, depending on how high of quality you want to go, uses different credits. And I'm not a big fan of the credit model because it's really hard to determine how many credits you're going to use because the. The larger the resolution you make, the more credits it's going to use, the more images you generate in one go, the more credits it's going to use. But on average you're probably looking at like 4 to 8 credits per image. And I believe they give everybody 100 free credits a day. So theoretically you can create 20ish images a day or so for free before needing to upgrade. So depending on how much you need to use it, you actually can get away with using their free model quite a bit, actually.
- [15:03]
Michael Stelzner
And they have a really good AI upscaler, do they not?
- [15:06]
Matt Wolf
They do. They do. It's one of those AI upscalers that if you want it to, you can have it sort of hallucinate extra information. You hear the term hallucination a lot in the AI world. In some areas, a hallucination is a feature, not a bug. In other areas, when you want something factual, you know, hallucinations are a bug and not a feature. When it comes to AI art, hallucinations are what you want. Hallucinations are what create that creativity. And they have an upscaler that will actually sort of add extra stuff into the image for you and hallucinate extra information.
- [15:40]
Michael Stelzner
Like textures, for example. Right. It'll add textures where there are not any textures.
- [15:43]
Matt Wolf
Right, Exactly. Exactly. Yeah. Or I'm not sure if you've seen any of the images where somebody took like the old pixelated version of Laura Croft, for instance, from Tomb Raider, and then they ran it through one of these upscalers and then the upscaler actually made her look realistic. It didn't just make a much higher resolution version of the pixelated Lara Croft, it made a higher resolution but realistic version of the Lara Croft. Right. So if. If an image is like more pixelated, let's say you upload a screenshot of Minecraft or something where it's all made out of cubes, you can actually take it and tell it, hey, make this look more realistic. And it will upscale the resolution while also at the same time enhancing all of that stuff to make it actually look. Look more realistic. And it's sort of guessing what the more realistic version looks like, which is technically a hallucination, but yeah. So Leonardo has that built into it. So if you want to upscale something and actually tell it to look more realistic, it will do that. Or you can tone that version down and just tell it to upscale it and make it a much higher resolution. It will do that as well. So it's got a few of those options in there.
- [16:50]
Michael Stelzner
Yeah. And folks, upscaling is kind of a big deal. Like, you could take an AI image from another system other than Leonardo, and you could just use Leonardo to upscale the image. Right. Because, yeah, for example, I use Flux somewhere else and sometimes I need to upscale those images. And I like how it adds the little details to faces and stuff like that. Okay, so outside of Leonardo, there's another tool that you mentioned when we were prepping for this Ideogram. Talk to us a little bit about that.
- [17:16]
Matt Wolf
So Ideogram was really the first tool to come around that actually figured out how to get text inside of images. Up until Ideogram, you know, whenever you tried to get an image that had text inside of it, it would just kind of look like gibberish. It didn't look like real letters or it would misspell things. Ideogram was really that first one to come out and go, look, we can actually generate text inside of images. So I started playing with that one a couple years ago because of that ability to add text to the images. And over time, that one's just gotten better and better and better. So if I can't get the image that I'm really looking for out of Leonardo Ideogram, the next one that I go to, and it's gotten really good, it's its own foundational model as well. So they built it from the ground up. It's not built on top of stable diffusion or flux or anything like that. It is their own internal model that they built. But it's just gotten really, really good over the years. And they also have a system where I believe you get a certain amount of free credits per day. I don't know exactly how many, but you can generate a certain amount of free images per day before being required to go and actually pay to. To, you know, generate a lot more images. But yeah, it's just a really, really good image generator that will do both realistic, probably not as good as flux, and it does good, creative, colorful images as well. So, you know, I'm kind of switching between the two all the time. But yeah, that one's another really good.
- [18:35]
Michael Stelzner
One I have in my notes. Flow state mode. I don't remember what that was we were talking about. What the heck is that?
- [18:40]
Matt Wolf
Yeah, so flow state mode is actually another mode that was inside of Leonardo. And what that one is is so basically, you give it a prompt. Let's say you don't really know in your head what you want the image to look like, because as you're probably aware when you're prompting images, the more detail you can give into the image, typically, the better the image is going to come out. But let's say you don't really know what you want the image to look like yet. You know, you want an image of a cat with a colorful background, for instance. You can go and prompt a cat with a colorful background. And what it will do is it will generate an infinite scroll of images of a cat with a colorful background. And you can just keep on scrolling and scrolling and scrolling until you find one that you like. And then there's a little button that says more like this. You click on the more like this, and it opens up another infinite scroll of images that are similar to that one that you just clicked on. And you can scroll and scroll and scroll. And then maybe you really like the color palette and the composition of just one of those images. You click on more like this, and it creates another infinite scroll. And you can kind of dive deeper and deeper and deeper. And so you can start with a very, very rough idea for an AI image and then let it figure out, you know, what it could look like. Let it give you a whole bunch of ideas, and then just drill down. When you find one that you like, maybe you go a little too far. There is a button to go back to the sort of previous level of infinite scroll and continue to scroll through that. It's just a completely different way of prompting where you don't need to have much detail or much of an idea of what you're looking for. You can give it a rough idea and then let it sort of flesh out a whole bunch of ideas and then dive deeper and deeper and deeper. When you find an idea that you like, it's a really, really cool feature. In fact, that's how I make most of my thumbnails now, is using that flow state model.
- [20:25]
Michael Stelzner
Do you have any tips on image generation prompts? Because so many people mess up their prompts, you know, and they just like get random garbage. I mean, have you found, like, because you've done so many of these, have you found any certain kind of little quick tips on how to increase the likelihood you get what you're looking for?
- [20:40]
Matt Wolf
The more detail you give, the better, right? So if you're using something like the flow state that I just mentioned, you really don't need to worry about it. You can just sort of dive deeper and try to find an image that you really like. But if you're trying to generate something very specific, you want to give it as much detail as possible. Saying that most of these AI tools now have like a feature that's called something like Magic Prompt or Prompt Improver or, you know, they all kind of name them different depending on the tool you're using. But a lot of these tools now, you can give it a very basic prompt, click the button to turn on the prompt improver, and it will actually flesh out the prompt with a lot more detail for you. Both Ideogram and Apps, you know, have that.
- [21:17]
Michael Stelzner
You could probably go to Chat GPT and ask it to do the same thing, couldn't you?
- [21:20]
Matt Wolf
Yeah, yeah. In fact, before, before a lot of these tools rolled out the prompt improver, that's the way I used to do it, was go to Chat gbt, Hey, I need a colorful picture of a cat. Can you give me some. Give me five really interesting prompts to do that. You read the prompts and then pull that in. Now, a lot of these tools that have the prompt improver, they're literally just kind of doing that step for you. They're probably tied into like, you know, the, the Chat GPT API or the Cloud API or something like that, and they're just taking that step off your plate, essentially.
- [21:48]
Michael Stelzner
Here's the thing about AI, it's not just hype. It's transforming everything we do as marketers if you're ready to master AI for your business, I've got something special for you. At Social media Marketing World 2025, our dedicated AI track will give you the exact playbook you need to thrive as an AI enhanced marketer. Learn from experts like Matt Wolf, Chris Penn and Nicole Leffer and many more. Don't get left behind. Join us in San Diego this March. Visit social mediamarketingworld.info and get your tickets now while you still can. Save Big okay, one more question. If you're trying to retain some certain consistency between a lot of your images, like for example, I don't know, I think you have a fairly consistent look and feel for your thumbnails. I've heard people talk about seeds and stuff like that, which I think are some sort of unique identifier. I don't even know exactly what that is or how that works, but is that something that you could explain?
- [22:52]
Matt Wolf
Yeah. So basically, when you're generating an image, I'm not sure how well I can explain it, but I'll sort of give a very simplistic explanation. When you generate an image, it's basically creating sort of a blank canvas with a whole bunch of noise on it. That noise is based on the random seed that you give it it and then the noise is sort of diffused down to become the image that you were looking for. So you can't really get a consistent image by using a consistent seed. All the seed is doing is it's defining the very starting point of the image, the amount of noise that it's putting in the image before it diffuses down to the image that you are looking for. But if you change the prompt at all and use a different seed, it's going to be a completely different image with a potentially completely different style. The seeds aren't really designed to get like a consistency across images. If you do want to get consistency across images, you want to use what's called Loras L O R A. It's short for low rank adapters. But what a Lora is is basically you can train into the AI model a sort of like another level that sits on top of the AI model. So you can train a Lora on a style. So let's say that you, you really like a certain cartoon style. Let's say you want to generate every image to look like it came from south park or something. You can go and train aura by giving it about 10p screenshots from south park, for example. And then you'll get like a little trigger word and that trigger word, if you add it to your prompt, will make sure that it uses that style that you created. You could do it with a style, but you can also do it with like faces. And you can do this in Flux and you can do it in Leonardo. I don't believe Ideogram has this feature yet. Mid Journey just rolled out a feature where you can do Loras as well. But basically you can go and upload like, like 10 images of your face or 10 images of a, a character that you created. And then every time you use a specific trigger word, it will tell the AI model to use this Lora on this image and it will make sure that it creates this consistent character within the image. So whenever you see any of these like AI influencers or things like that, where every single image looks like the same character but they're in like a different scenario, they're most likely using a Lora that they created on this character. That might have gotten a little too into the weeds.
- [25:12]
Michael Stelzner
But no, I mean, that was super, super helpful. I've heard so many people talk about Flux Laura and I kind of conceptually understood it, but I think that was really helpful. Okay, Video side of things. There was a lot of stuff introduced just before the holidays from a lot of platforms. So why don't you talk about the three big ones that you're most excited about on the video frontier side?
- [25:33]
Matt Wolf
Yeah, so SORA was the one that sort of set the standard of what was going to be capable with AI video. I mean, Sora, they announced it almost a year ago and then took, you know, eight or nine months before they actually released it to people.
- [25:47]
Michael Stelzner
And they, they is who?
- [25:48]
Matt Wolf
That's OpenAI. So OpenAI is who created Sora. OpenAI showed it off about a year ago and then waited about nine months to, to roll it out. And then in December, OpenAI did their whole like 12 days of announcements. And one of the announcements during that 12 days was, hey, we're giving you access to a model called Sora Turbo. Now what Sora Turbo is, is it's sort of like a dumbed down model of the Sora that they showed off nine months ago. So if anybody saw any of the demos of Sora nine months ago, they looked really, really good. There was like an image of a ship on an alien planet with some astronauts walking around and it looked really, really realistic. And everybody thought, wow, this is amazing. When we finally got Sora, they actually gave us a model that uses a lot less compute, it uses a lot less computer power. To generate that image or that video. And so the videos we're getting out of this model aren't nearly as good as all the demos they were showing us nine months ago, because if they gave us that model, it would just be way too expensive to use. The cost to process those videos is just too high still. So they didn't make that one available. But Sora is still pretty good. I think the best on the market right now, in my opinion, is the one that Google recently showed off called VO2. And VO2 is still in early beta testing. I did manage to get access to it, and you can get on the wait list, and I don't know the cadence that they're rolling it out.
- [27:12]
Michael Stelzner
But actually, didn't you end up prompting one of the prompts that I recommended? Because didn't the dog surfing one remember you asked for?
- [27:18]
Matt Wolf
Yes, I did. Yeah.
- [27:19]
Michael Stelzner
So kind of explain what it did, because I think it was pretty impressive, right?
- [27:23]
Matt Wolf
Yeah. So I generated one of a dog surfing in La Jolla, and it sort of looked like La Jolla, right? It looked pretty close to La Jolla. It looked very realistic. It looked like a real dog surfing. But the thing is, when you generate with VO2, just like with something like Mid Journey, it actually generates four videos. And out of the four, usually, like, two of them look kind of wonky and two of them look really, really good. So they kind of give you an option of like, here's four generations that we just created. Hopefully one of these is good enough. When I did the dog surfing in La Jolla, that one, I think, you know, two out of the four looked pretty good, and a couple of them looked a little.
- [27:59]
Michael Stelzner
Were the physics side of it. Because obviously there's some physics involved here, right? I mean, with a lot of these in motion videos.
- [28:05]
Matt Wolf
Yeah. So basically it's trained on just tons and tons and tons of videos. So it's got a whole bunch of videos. It. It understands what's in the videos that it was trained on. And then when you ask it to prompt, it sort of knows what those types of videos look like based on all of its training data. There are a lot of people out there that are talking about how some of these video models are basically like world models. They understand how the world works and how physics work. But if you look at some of these gener, there's a lot where the physics is like, really weird. And even with VO2, you still get a little of it where you see people walking and like, if they're walking sort of by the Camera parallel to the camera, you might see their legs switch places with each other. And there's still some really, really weird physics stuff happening, which means it doesn't totally understand physics and how the world works yet. But that's kind of where they're going, is they're trying to train it on, like, how the world actually works and how physics actually work, but they're doing it by just feeding it hundreds of thousands of videos, labeling the videos of what it actually is, and then letting the training model behind the scenes sort of understand what these videos are doing.
- [29:09]
Michael Stelzner
So what I heard you say is that Google's VO2 was the one you were most impressed with, but it's not as of this recording out to the public yet. And it did look incrementally better than the stuff from Sora. And I would imagine, because Google owns YouTube and people forget that they've got access to the world's largest library of videos. So of course they're probably going to take the lead on this, as my guess.
- [29:33]
Matt Wolf
And then also just to comment that too, YouTube just rolled out a feature where you can actually check a box saying, yes, Google can train on my videos, or no, they can't. So I think early on, they just probably pulled all the videos from YouTube and trained on them. Now they're actually giving you the option. So if you don't want future YouTube videos that you created to be trained into the model, you can actually uncheck that box and they won't train on your model.
- [29:55]
Michael Stelzner
How long does it take? Because I know you've experienced both of these from the moment you put a prompt in to get a video out. Is this like a really long process because it's so Compute intensive?
- [30:05]
Matt Wolf
With VO2, it's about four minutes on average from what I've found. But it'll let me open up like five instances of V2. So I can go and prompt five different tabs and have five different videos going, but each one will take roughly three to four minutes to prompt. Sora, I think, might be a little bit faster. Sora is probably a good three minutes. I mean, I think all of these models are really based on how many people are using them all at the same time. I don't know if you remember the day during, like the 12 days of releases that they actually released Sora. That first day that they released it.
- [30:38]
Michael Stelzner
Nobody could get in, right?
- [30:39]
Matt Wolf
Well, people couldn't get in. They actually turned it off. I actually managed to get in, but it was taking 45 minutes to an hour. And a half to generate one video just because so many people were all trying to use it at once. So they've since increased the bandwidth with. But a lot of these are. Are hitting the same server. So there's a lot of people all trying to use it at once. It's going to be a little bit slower than if during times where not a lot of people are trying to use it. And that's going to be pretty much across the board for like any video model, any image model, if, you know they're bandwidth constrained, they're going to be a little bit slower.
- [31:10]
Michael Stelzner
When we were chatting, you also mentioned a new one to me, at least called Luma Ray 2. Talk to me about that.
- [31:17]
Matt Wolf
Luma was actually a company that did sort of 3D object scanning. So they have some tools where you can take your iPhone phone and go out into the world and sort of scan around an object and it will create a 3D representation of that object and then you can pull it into, you know, Blender or Unreal Engine or whatever you use.
- [31:33]
Michael Stelzner
I think they call that ray tracing or something like that.
- [31:36]
Matt Wolf
I'm not sure if that's the same concept or not. It's basically 3D modeling. Basically what they're doing is they're creating what's called a Gaussian splat. I'm not sure if you've come across that term, but it's basically making like a point cloud of whatever you're looking at. And then the points are sort of expanded like, like a putty to like mold the image. I'm oversimplifying this because we could probably get like super into the weeds if you wanted to, but that was like really what Luma's like bread and butter was. That's what they were known for. However, they started releasing this video model called Dream Machine. And Luma's Dream Machine probably came out like nine months ago. And it was okay. It was decent. It wasn't the best video model anybody's seen so far, but it was okay. But just within the last few weeks, they actually came out with this new one called Ray 2. And Ray 2 is really, really good. I mean, it's probably not quite as good as VO2, but it's like, better than Sora, not quite as good as V2. It's sort of that middle ground. But it is accessible to everybody right now. It's open, it's available, Anybody can get in and use it. I think it's a paid model. You might get a certain amount of generations for free before you have to upgrade, but that one is open and accessible, where VO2 is still kind of in beta at the time of this recording. It is really good and it's, you know, about the same 3, 4 minutes to generate a video. But they do look really, really good. And they're probably a little bit closer to like this world model concept, because a lot of the training data they have are these 3D scans of the world that people have been using Luma for for the last, you know, five, six years. So Luma has all of those 3D scans that all of these people have used. And when you use Luma, you're basically giving them permission to, you know, train on whatever you scan into it. So a lot of their video is trained on both a combination of real world video, but also all of these scans that LUMA has collected over the years.
- [33:28]
Michael Stelzner
This is such an interesting topic right now because anyone who's messed around with even photorealistic images, like the stuff flux can do, knows that there's always problems, right? There's always things that are off. And now imagine like 24 frames per second, right? Because that's really what we're talking about here, right? And things in motion and the laws of physics and the way light refracts and reflects and stuff. I mean, this is an extremely complex undertaking, is it not, to create video, just level set people, like, especially the marketers in the room here who are listening. Is it going to be a while before we actually have something that's quite legitimately impressive to people?
- [34:08]
Matt Wolf
I'm always impressed by the new stuff that comes out, but I think the biggest issue right now is a lot of these video generators are only generating about five to 10 seconds. I mean, the best one out there right now is only going to generate 10 seconds of video footage for you. Now it really depends on what you're using it for, though. Like, if you just need B roll for your videos, typically B roll only sits on the screen for, you know, three to five seconds at the very, very most. You don't typically have B roll in a video for a whole hec longer than that. You know, watch anything on tv, watch movies. You'll notice they never stay on a single shot for more than three to five seconds. And so these models that can generate, you know, five to 10 seconds, if just three to four of those seconds are usable, you've got some solid B roll that you can use in your videos. And so I've actually done that with a handful of videos where B roll will pop up for Just three or four seconds, but it'll be AI generated B roll and people don't even realize because I'm just taking like the best three to five seconds of that actual generation. Now. If you're asking like, are we going to get 30 minute long videos that are all generated with AI that are actually entertaining to watch and nobody can even tell their AI. Personally I still think we're a lot further away from that. A lot of people debate me on this. I've commented before that I think we're pretty far from that still. And a lot of people have said no, no, we're going to get really close to that. I think the biggest bottleneck is just the compute power. I think it'll be right now it's just way too expensive and require way too much GPUs to be able to do that right now. I don't think it's far off to think that we'll have that within the next couple of years, but I don't think we're going to see like a AI generated 30 minute video where it was just like a single prompt and the whole video was generated with AI and it actually looks good and impressive and nobody can tell it's AI. I just don't see that coming within the next year or so.
- [35:57]
Michael Stelzner
What's your thoughts about what Adobe's doing? I don't know if you've had a chance to kind of see some of the stuff they're doing with existing footage and expanded, extending it and stuff like that. I mean, because I would imagine this is going to be an obvious thing for someone like Adobe to be working on only because so many video people use Adobe products, right?
- [36:13]
Matt Wolf
Yeah. It's funny you bring that up. I was actually on an Adobe briefing call this morning before our call, so I was actually getting some demos of some of the stuff they're rolling out. Unfortunately it's still under embargo, but I think a lot of the like generative, like extend stuff is really, really cool. I'm excited about that. I mean, I don't know how many times I've been editing a video and I need my B roll to go 10 seconds, but I only have a clip that's 8 seconds and I, you know, I have to like slow down the speed of the footage to get it to go that extra two seconds. Well now with what Adobe's doing, you're not going to have to do that anymore. You can just like stretch the clip, the extra two seconds. It's going to use AI and try to figure out, like, what the continuation of that video looks like. And that's, to me, really, really cool. I also really, really love Adobe's Generative Fill in Photoshop. I use that all the time.
- [37:03]
Michael Stelzner
I think I love it. And Generative I use all the time too, right?
- [37:06]
Matt Wolf
Yes, yes. It's even great for video, right? So I can take a screenshot of my room right now, like a freeze frame and extend my office out and add like extra elements around it and things like that. So, yeah, the Adobe Photoshop Generative Fill and Generative Extend are really, really, really cool. We're going to see a lot of that stuff likely inside of, like, just the Firefly web app pretty soon. So you don't even need to use Photoshop to do that kind of stuff in the near future here. But yeah, they're doing some really, really cool stuff. They showed off a demo at Adobe Max here where you can sort of arrange a whole bunch of 3D objects, sort of like what you'd create in like a blender or something like that. Arrange a bunch of 3D objects and then arrange the camera wherever you want it. And then based on where the camera's pointing, click a button to generate. And then it'll generate a realistic image from that. So you're going to be able to get like a shot of like a environment and then move the camera around the environment and get a consistent looking.
- [38:03]
Michael Stelzner
That's cool.
- [38:04]
Matt Wolf
Shot every single image. Image. Because you can just move the camera to different spots in that environment and just generate based on the 3D objects that filled out the space.
- [38:12]
Michael Stelzner
Okay, audio. Next frontier on the audio side of things. There's a lot going on there. What are the tools that you love and why?
- [38:19]
Matt Wolf
Yeah, so my personal favorite that I've been using a lot lately is Notebook lm. This is another Google tool, totally free to use. Anybody can use it right now. But this is a tool where you can give it a PDF document, you can give it a URL to a website, you can type your own text into it. You can basically give it information. And. And there's a few things you can do with that information. Once it's in there, you can actually chat with the information. So you can go in there and say, you know, based on these PDFs that I uploaded, what's the answer to this question? And it will chat with you about that. It will automatically create an FAQ around it, things like that. However, the. The feature that I think most people are excited about is they've got this generative Podcast feature in there where it'll take all of the information that you dumped in there and then create a podcast where there's two hosts speaking back and forth about all of the information that you uploaded. And it sounds surprisingly natural. It actually sounds really like two real people having a conversation. And they also just added a new feature now where you can interrupt them in their conversation. You can press a button, interrupt them and ask your own question. And it's like somebody calling into a radio show, right? So you, you interrupt them, they'll say, hey, Matt, what's up? Because, you know, they know your name or whatever, but they'll say, hey, what's going on? And then you can jump in, ask a question question, and then the two hosts will start to have a discussion about the question that you just asked when you called in. And all of this is downloadable, so you can download the podcast when it's done. I know a handful of people that have actually started podcast feeds over on itunes and Spotify, where every day they're having Notebook LM create a podcast on like a different topic within a niche. So like there's an AI podcast that every day it's talking about a different concept like machine learning and neural networks and diffusion models and whatever. And every day it's like a different topic in the AI world. And all they're doing is going and rounding up a whole bunch of information about that topic, throwing it into NotebookLM, having NotebookLM generate a podcast, they download the podcast, upload it into their RSS feed, and they have a daily podcast about a very specific topic. So that's been really, really cool, really fun to play with. It's a great way to get information and sort of take it in in a fun, entertaining way. Right, like, like you're listening to a podcast as opposed to reading an article.
- [40:33]
Michael Stelzner
We had our notes SUNO and 11Labs, if you want to talk about them as well.
- [40:36]
Matt Wolf
Yeah. So 11 Labs is a really great tool for text to speech. It's probably the most realistic text to speech out there. So if you're creating a video and you need to sort of, maybe you misspoke and you want to fix it without going and re recording it. You can actually train your own voice into a layer. Eleven Labs type stuff out and then it will give you like an audio file of you speaking that out and then you can go back in and sort of like replace it in your video. I actually did that once with like a sponsorship ad read where I misspoke on the sponsorship ad read and they Gave me some feedback to go and change it. I just typed up a new sponsorship ad read with 11 labs and overlaid it into my video. So 11 labs is really, really cool because you can actually have your own voice trained into it. They also have a lot of really, really good voices is already built in. They have a voice library where you can actually go and pay to use other people's voices. And the people that train their own voices in there actually earn revenue when you use their voice. So a lot of really, really cool features like that. You know, there's been a big trend lately of people making, like, faceless YouTube channels where it's just a voice speaking and then they overlay a bunch of B roll information. So you can use 11 labs for that kind of thing as well. And then they also have a sound effects generator. So if you need a sound effect of like a cat meowing or a door slamming or an explosion off in the distance, you can give it a prompt. It will actually make that sound effect for you, which is really great. If you need sound effects for a video and you don't want to go, use like a stock sound effects website. And there's a whole lot of other features inside of 11 labs as well. They actually have their own feature called Gen FM that's very similar to what we just talked about with Notebook lm, where you can upload articles and it will actually create a podcast for you about it. I still think Notebook LM does it a little bit better, but 11 Labs also has that built in. It's just a really cool sort of all in one audio platform where there's a lot of different audio features. They're also going to be rolling out a music feature where you can generate music directly inside of 11 labs. I interviewed somebody that works at 11 labs on our next Wave podcast, and he was sort of teasing out that there was going to be a music feature. But right now, the best music tool in my opinion is suno, which is the other one that you just mentioned. SUNO to me, is just. It makes really fun sort of poppy songs. You can either put your own lyrics in there and say, generate a song with these lyrics and it will go in. Do all the background music and the drums and the guitars and whatever, and it will actually use the lyrics that you put in. Or you could just let AI generate a song for you. You can say, make a song about my dog Bowie and how he goes crazy running around in circles in my backyard. And you just give it a prompt like that, and it'll make a whole fun song about that topic. And if you need like background audio for your YouTube videos or something like that, you can flip a switch and say no lyrics and it will actually make good like sound beds for your your videos. I use that feature all the time when I just want to have a little bit of background music in my and I don't want to go hunting for through a stock video site. I just kind of tell Suno, here's the sort of tone I want from this video, don't include lyrics and it will create this sound bed background music for my videos. So lots of really cool features in Suno as well.
- [43:39]
Michael Stelzner
Matt, we have just scratched the surface of the world of AI and all the crazy tools that are out there. If people want to follow you on the socials or YouTube, where do you want to send them? And if they want to connect with you and your business also, where do you want to send them?
- [43:53]
Matt Wolf
Yeah, so the best places would be YouTube. You just search out Matt Wolf on YouTube and you'll find me. And then I tend to spend most of my social media time on X these days. That's where a lot of the AI community hangs out. So I'm Mr. E Flow on X M R E F L O W on X there quite a bit. Those are probably the two best places.
- [44:11]
Michael Stelzner
Awesome. And Matt will be in San Diego so you get a chance to talk to Matt directly. Matt Wolf, thank you so much. Oh, San Diego, ocialmedia Marketing World. I forgot that important little distinction. Matt, thank you so much for coming on the show and sharing your insights with us today.
- [44:24]
Matt Wolf
Yeah, thanks Michael. It's been so much fun.
- [44:27]
Michael Stelzner
Hey, we took all the notes for you over@socialmediaexaminer.com a41 and be sure to follow this show, by the way, on whatever app you're listening to. We've got some great content coming your way. If you've been a listener for a while, we'd love a review on your favorite platform. Do check out our other shows, the Social Media Marketing Podcast and the Social Media Marketing Talk Show. This brings us to the end of the AI Explored podcast. I'm your host, Michael Stelzner. I'll be back with you next week. I hope you make the best out of your day and may AI help you become more successful.
- [44:59]
Matt Wolf
The AI Explored Podcast is a production of Social Media Examiner.
- [45:06]
Michael Stelzner
If you're serious about learning more about AI and marketing, I'll see you at Social Media Marketing World 2025. Go to Social Media Marketing World.info and secure your spot today.