Edit images and video by talking to a chat model

Welcome to the AI Academy with Steve and Dustin. How are you doing, Dustin? Good. How are you, Steve? I'm tired. It was Canadian Thanksgiving this week on Monday, so that's our traditional time for closing up our cabin at the lake. And it was more problematic than expected. As you winterize things, you discover new problems, and then you have to solve them, or they're still there in the spring, and they're bigger problems in the spring. So I had a very long day yesterday, ended up not getting back to Vancouver till well after midnight, so I'm a little bleary-eyed.

So I'm hoping you're going to do the heavy lifting today. And indeed, I will. I will do my best. Excellent. So what do we got today? Well, today, really two catalysts for what we're going to talk about today. Number one, we had our Monday meeting with my development team and my support team as well. And something that we've noticed is coming up a lot in customer support tickets is it would seem like a lot of people have a mistaken idea of how AI image and AI video

and AI audio differs from large language models, which are the chatbots. Many people still have yet to learn that there is a fundamental difference between how you interact with a language model versus how you interact with a media model around image, video, or audio.

The two are not the same in that you can have conversations with them. We assume that we can have conversations with AI image models because most of us have only interacted with image models through a chatbot. And so today we're going to talk about the fundamentals here. This is another pragmatic fundamentals lesson in what's the difference between these models, why do you have to interact with them differently, and why doesn't it work? And we're going to start with a fun example of some of the frustration that one of my friends encountered this past week.

Now mind you, this friend of mine, she's brilliant. She's a brilliant marketer. She's known me since almost the beginning of Magai, and even she has struggled with learning these AI models. Despite the fact that I talk with her on a weekly basis. She had these issues as of this morning. So let me show you some of these fun interactions she had struggling to figure out how to use the image models in Magai.

Let's see, where are we at here? And forgive me, I'll have to share my entire screen because it's just easier to make sure I get the right screens. So first one here is we'll see she had this beautiful graphic that she's creating for her new podcast, and she wanted to make this square image in a 16-by-9 dimension. And that would've been fairly easy to do.

Just take this image, and she's telling the AI, make it 16 by 9. Now, the problem here is she was working with an image model, not a conversational model, but somehow the AI image model was able to interpret that correctly and actually did give her the right result. Now, here's the mistake. She then replies to the image model, "Put a bigger space between the photos in the circles." Now, Steve, do you have any idea why this next one absolutely failed?

It's a colossal failure. Because... I don't know. No, I don't know. Because when you interact directly with an image model, they don't have conversations. Image models don't receive the history. They don't retain the history of a conversation because it's not a conversational model. Image models, video models, audio models, they are input, output, end of discussion. No history, no prior context, nothing.

So every time you send a message directly to an image, video, or audio model, all that image, video, or audio model knows is what you just sent them in that singular message. So when she sends the words, "Put a bigger space between the photos in the circles," she assumes, rightfully so, she assumes that that image model is going to be able to see the previous image and the previous message that was sent.

But that's not the case, because image and video models and audio models are input, output, no prior context. So all the image model sees is, "Put a bigger space between the photos and the circles." And so it has to treat that as its own prompt with no prior history, and this is the image that it came up with. And so Christine was naturally a little frustrated, and she figured out, okay, maybe I need to attach that image again.

So she attached it, and she said, "Add a bigger space between the photos. They're too close together." And so there it did. So, probably not ideally what she wanted, but now it had that photo and was able to use that photo. And so Christine kind of figured it out from here. Now, here's another example- So but sorry, just a minute. It wouldn't have continued on with the thread.

So if you had said, okay, if your next command had been make the photo slightly larger with less space, it would again have not known. You would' had to have pasted this photo and then done it again, right? Would've had to attach that image. Because it would've had to retain context. Yes. It needs the image because it has no history, so it doesn't know that it just performed that job for you.

You have to give it the context every time. So it managed to do this one for her, and then so she made the mistake again of not attaching the new image. So she said, "The host name is misspelled, her name should be Christine Deen." And so it created a brand-new photo because it had no prior history, Host Christine Deen. Here's another one. I thought this one was pretty hilarious.

So she started with a really great prompt here, multiple images. And she got her square image, and she made the mistake, "That's not the correct size for YouTube videos." So again, image model, no idea what she's talking about because all it sees is, "That's not the correct size for YouTube videos." So it generated this image, 16 by 9. Landscape is the way.

So frustrating. And mind you, Christine has a square image. And then so she replied, "No, make the image you created into a 16 on 9." She's assuming the image it created, it knows the image it created because she's assuming it has history, but it does not. She's just talking with Ideogram here. It's just the image model, and so it created this random gradient image. And then she said, "Recreate at 16 by 9." Again, doesn't have the history, so it just created a brand-new image, 16 by 9.

So, this can be very frustrating. Here's another example. She tried with the same thing. Again, didn't give it the attachment, so it doesn't look like the dimensions for a YouTube video and the title of the episode should be at the top, i.e., From Soldier. And there we have From Soldier 2. No idea what the image model was thinking.

So our lesson today is understand that LLMs, those are the ChatGPTs, the Geminis, the DeepSeeks, the DLMs. We have them labeled here in Magai, kind of sectioned off as chat models. And then we have image models, and we have video models. And coming this week, hopefully, as long as all the bugs are worked out, we will have audio models as well from ElevenLabs.

And then after that, we'll be adding more audio models that you can generate audio with. But the reality is, make sure that you understand how to use these models properly, because otherwise you're going to have some frustration headed your way. So let me switch over to our official team here. Before we do, Mike actually asked a question that's kind of right on point about this. "I've been having this exact issue with image models.

When you want to make a small edit to a 95% good image, what do you suggest? I've re-uploaded the 95%- Yeah ... good image, but the response will have to recreate the entire image from scratch, which could change a lot of items." And we've all experienced that frustration. We're so close, and then we just seem to get farther and farther away. Yeah. We're going to go through exactly that. I'll show you how to understand how to work with these.

It's actually more simple than you think. And then we'll talk about a change that we made in Magai several months ago that makes it so that you actually don't need to select an image model anymore. You can just talk to the LLM, and the LLM will use the image and video, and eventually audio models for you. But just a quick sort of visual breakdown of this. There are language models, and you give these language models questions, ideas, and context, words, and reasoning. And it sees the conversation. It reads the full thread every time you send it a new

message. It has that thread in its history. Image models. What you give them are visual descriptions. You give them the exact thing that you want it to give you. It gives you back one image. Sometimes, some apps will give you the ability to have multiple images. It does not see the conversation. It does not see the request.

Every single request you send it is brand new, as if it's never heard of you before. Video models, same thing. You give it a scene, you give it an action, and a camera move. It gives you back a short clip, and again, it does not have the history of a conversation. So if you have it produce one shot, you can't follow up with it and say, "Now give me the next logical shot after that," or make edits to this, unless you hand it that clip again.

Audio model, same thing. You give it a script, a sound, a description. It gives you back a voice or music file. And know, same thing, it does not have history. So if you want it to know the thread, you have to continually give it the thread. And so what does that look like inside of Magai? If I were to go in here and pick an image model, and I will pick Nano Banana 2.1, the newest version of Nano Banana. It is fantastic, by the way.

Let me just go in here to our AI Academy. Make sure we're in here. I'm just going to quickly type this out. By the way, short code or slash commands, when you're trying to find the right model, a big time saver. We got 2.1 selected here, so I'm going to say, "A cat playing cards." Now remember, our image models are just, you describe what you want, exactly what you want, and the image model gives you what you want.

Nano Banana's fairly fast as well, so this should happen pretty quickly. So it's going to output the image. It's not going to output text. It's not going to reply to us. It's not going to talk to us, because all this AI is capable of is images. That's the one thing it can do. Now, if I wanted to make edits, again, we've seen that if I just tell it, "Make edits to that image," it doesn't have that continuity.

If you said, "Make it a Siamese cat," it would just have a Siamese cat sitting on a fence somewhere. It would create a brand new image of a Siamese cat. Yeah. Correct. Because I'm not giving it this image as context. So in Magai, we built several tools to help you continue that image edit. So you have to get it into your input somehow. First way you can do that is there's an attach button.

If you hover over the image, you click attach, and it attaches itself to the prompt box, just like if you were to drop a document in there, right? The other way is through the edit mode button. The edit mode button will automatically attach the last generated photo. Mm. As long as this is activated, you can say, "Make it a Siamese cat."

And now- I see you're back to typing. Well, yeah. I could've dictated, I suppose, but it was a short one. But because I attached the photo, now it knows this photo, this prompt. It has that. And so it can make it a Siamese cat. Same photo, right? Brilliant. Yeah.

Now it's Siamese. So if you wanted to give it a dealer's visor, you would just tell him, "Give the cat a hat." Right. And notice, because I had edit mode on, it automatically put that new image into the prompt box, so it's already ready to prompt. Add a visor hat on the cat. And I can continue prompting, and as soon as this one's done generating, you'll notice it pops right into the editor because I have edit mode on.

So this solves the 95% problem. Well, but you have to know to turn edit mode on, right? Yes. And that's kind of our problem. Our big challenge is not everybody wants to keep editing, right? Maybe that first photo is right, and they don't want that photo to be automatically in there. So we default it to off. We assume people are going to get an image and then want to move on to the next image that they want to generate.

And the same thing applies to video models, right? You have to give the context. Now, my favorite way to attach it, like I said, there's multiple ways to get that image into the box. You can just click and drag it. It's that easy. Just click it, drag it right into the box. You can go back up here. Maybe I want to start with this, and just click and drag it and drop it into the prompt box.

That would work really well, especially if you asked it to give me three options for this. The way we do say for a subject heading. Mm-hmm. Give me three options. Although that's a lot of processing power for it. But here's what would happen there is, again, these image models, they can't process natural language requests. You give them input, they give you one output. So a language model, on the other hand, would say, "Oh, this user wants three images."

Oh. "So I need to make three different image generations." Whereas an image model doesn't have reasoning. It sees text prompt, and it turns that text into an image, one image. So it wouldn't be able to process multiple. Now, again, like I said, some tools out there will, by default, make three different image calls or three different requests from one.

But we have to know that that's not the way that the image model actually works. Now, the last way is, some of you may not even know that this files icon exists when you generate images or files. In a chat inside of Magai, you'll notice the files icon appears in this right sidebar. You can click it, and you can see all the files that exist in this conversation. And again, if you need to go back to one, you can click it, drag, drop it into the box to get it into the box.

Oh, that's very useful. If you get that 95% and then it's a swing and a miss, you can go back to the original one. Yes. Because we often, I have historically, just layered on top of my miss, gone down the wrong path, basically. Right. Yeah. And so one of the changes we're going to be making is going to fix this. But it is still good to have that education to understand how these models are fundamentally different. All AI is not the same.

It doesn't behave the same way. You don't interact with it the same way. Even though most of the applications we're using, like Magai or ChatGPT or Claude, they are making it seem like we treat them all the same. But really what's happening is this, and I'll show you the change that we made in Magai just several months ago, is we gave all the LLMs access to all the image and video models. So now I can open up, let's say,

try Claude Haiku. And I can say, here you go, "Create an image of a cat playing cards using Nano Banana 2.1." And now Claude, because Claude, through Magai, has access to these image and video models, Claude is going to use the image model for me.

And I can just talk to Claude the way that I want to naturally talk to the image model, and have Claude handle all those little intricacies of handing off the previous image and basically doing all the settings that I would need to do manually if I'm working directly with the image model. So I can say, "That's great, but make the cat Siamese." No attachment, no nothing.

Claude knows what it has to do to fulfill that request with the image model. So if you are doing this with all of the context of having, knowing your corporate colors, your corporate logo, all of that kind of stuff, it would all naturally flow in if you did this through the LLM instead of doing it through the image model.

Right. So, the LLM has the ability to have all that context, all that history, the back and forth, and your workspace context, and whatever context that you add to the agent. This is another reason why I created and I've demonstrated several times the art director agent, who is exceptionally good at image and video prompting because it's trained specifically on every different model and how to prompt each different model.

This is why it really helps to have that agent and the LLM just do all the image generation for you. So you should no longer be manually picking from the model selector which image and video model you want. This may be good to just know which models are available. But you can even ask the LLM, "What other image models do you have available?"

And oftentimes, even if you prompt it to generate an image, sometimes the AI will just ask you, "Well, which model do you want me to use?" And it'll list them all out for you. So it knows through Magai which models it has access to, and it will list them out for you here. And you can even ask it, "What video models do you have access to?" And it will tell you all the video models it has access to.

And you can even say something like, "Use Gemini OmniFlash 1.1 to turn that Siamese cat image into a video, just a five-second clip, and make it 16x9 aspect ratio." Again, I don't have to toggle edit mode. I don't have to pull the image in.

I don't have to do anything except just use my words. Just explain to the AI what I want it to do, and it goes and uses those other AI models for me naturally. And it should be fairly quick. Gemini OmniFlash is a pretty fast video model. But you'll see it handles it quite well. So the real lesson today here is just use the LLMs to generate your images for you inside of Magai or your videos as well, and soon your audio, and let it worry about all the intricacies of making

sure the continuity is there and making sure that the image model has what it needs. I would say he has a pretty good hand because he was purring. I don't know if you could hear that. So, I think... Hang on one second. So we've got, difference between the models.

This is kind of important. So same conversation, different, but here's a normal three-message exchange. So a red vintage convertible on a coastal highway at sunset. That's what the image model sees, and it kind of breaks it down into these subjects. But again, it doesn't see the conversation. Sent to a language model, it sees all of these.

But you send it to an image model, it only sees one at a time. It's pretty obvious. So treating the LLMs are sort of like the middleman between you and the image and video models. And we kind of went through all this already. But that's fairly straightforward. If anyone has any questions, one of the changes, that's where I was headed. One of the changes that we're going to make to Magai is when, in the future, to prevent this confusion from people who haven't gone through this

lesson yet, who just want to naturally speak to their image models, is we will be actually putting a large language model in between your requests. So in the future, you go and you pick Flux and you send a prompt. We're actually going to go ahead and automatically put an LLM between you and that image model rather than going directly to the image model so that this confusion doesn't ever happen again.

But in the meantime, like I said, it's just as easy to go ahead and pick an LLM that you trust and that you like, that you know is going to do a good job, and just have it use all the image and video models on your behalf because it has those tools. It has that ability in Magai to just go ahead and use those tools. Before we get back to talking about the image stuff, there's been many questions people have asked because somebody posted that their replays are going to be awesome.

This is from Henry, "I'm a Magai member. Could you please ask Dustin how to watch the replays?" We have made a change to the replay situation, and Dustin and I were actually just working on it before we began. So the way that the AI Academy is working moving forward is every live session will be continued to be free. Anybody can join us and view. If you want to view the replays, which there's now a great library of inside of Circle, if you are a member of the Grey Wave or if you want to just purchase access to the replays for $10 a month, you can access all of the replays within Circle.

So it's primarily here for our members of our Grey Wave community. For Magai subscribers, we wanted to attach it more closely to your account. So Dustin has now pulled it behind his firewall, and I'll let him explain. Yeah, so we're in the final process. I'm just currently migrating over the videos. We're going to hopefully finish that today.

But from now on, if you're a Magai customer, you will be able to log into our site, to Magai.co. Go to Magai.co/academy. And once all videos are there, you'll have this page where you can just go visit, you can see what the latest video was, as soon as it's been uploaded, and then you'll have the list of all previous videos. You'll be able to click on any of those videos, open it up, and you'll get the video up top. You can watch the video.

You'll have chapters on the right-hand side to kind of skip through the different chapters. You'll have a brilliant AI summary of everything that's covered, and you'll be able to have a text version of that tutorial for every single video that we do here on the AI Academy, along with a full transcript as well. So these will all live on our website behind the login, so it's only accessible by Magai users, and you won't have to go anywhere else.

So you'll also have a link in your... Let me open up that page inside your Magai account. Right now, the AI Academy link here goes to the Circle community. And once we finish and finalize everything on our site, it'll just go directly to the AI Academy on our site. So you'll have that quick link.

You have all access to all the videos. We won't be adding all the previous videos to the new AI Academy home because some of them are outdated and no longer relevant because of where technology's come. But Steve and I are working together to get all those videos that are still relevant, still valuable, and get them into the new library for everyone. Yeah, we purged the ones that are no longer relevant as well. So that's perfect.

So it'll be good. Yeah, it's going to be sweet for that. So yeah, it'll be attached to both accounts as far as if you're a Grey Wave member, you'll always have access to the replays right here within Circle and Magai users within your Magai account. So we think that should streamline things. And as I look through them, there's a lot of value. As those of you who are regulars here certainly agree, there's lots of value in them, and this makes it that much better. All right. So let me check out...

Judith says, "How wonderful is it?" I agree with Judith. Judith is always on point. Can Grey Wave members access the videos? No. Well, Grey Wave members can access the videos in Magai if you have a Magai account. So that would be the way, Agatha Yeah, you will need a Magai account to log into the site and view all the library.

Okay. So now we're getting back to some kind of back to basics. And oh, this is the one I was looking for, the full bleed. There we go. So let's get into some specifics on things, because people run into these once that last 5%. Sorry, kind of Michael's asking about it again. I've been trying to create a full bleed 8.5 by 11 ebook cover, and I've gotten close, but not 100%. Do you have a suggested image model that might be best for this type of image and text?

Yeah. Oh, let's talk about image and text. That's a great question, and it ties right into this. How do we use image models? Every single image model that you interact with has size limitations. It can't just create any size image. That's not how they work. They have a preset list of these are the sizes of the images they can generate. And if you select an image in Magai, you'll notice in the settings bar, if you open it up, has an aspect ratio.

These are not a limitation that Magai has put on the model. This is the limitation of the model itself. So, usually the dimensions are based on a 1024 pixel aspect or a max width. Some of the models, like, I believe it's Gemini. Oops. Image models, Nano Banana, they have higher resolutions that they can do, but the aspect ratios are still very much fixed.

Think of it like in Canva or any of these graphic design tools, you have preset sizes. The image model is trained on specific sizes, and those are the only sizes you can create images in. Same with video models. If I were to switch to a video model, you'll see that there are preset aspect ratios. For example, Gemini Omniflash can only do a 16 by 9, or it can do a 9 by 16. Can't even do square, can't even do three four or four three.

Video and image models have set hard-coded sizes that they can do. So if you're trying to get it to do a very specific size to, say, print work, it's not designed for that. It's not built for that. What you can do is partner it with an LLM, and I would maybe say to one of the smarter models, we'll go Opus here because it's kind of a challenging one, and maybe try something like, "I want to turn the Siamese cat image into an ebook

cover. So I want you to create a canvas for me that is in standard letter size format, so 8.5 by 11, and put the cat image on that cover, covering it entirely so that I can then print this as a PDF." And again, I don't even know if this is going to work because sometimes LLMs struggle.

That's the beauty of dictation is we ramble and explain and slowly paint the picture for the LLM of what it is we want. Yeah. So let's see. It's going to use its ability to write a canvas. It's going to build this out of code and it... Well, let's give it a title here. Set the title as "Card Tricks from a Siamese Cat: A Memoir."

My life is a gambler. You got to know when to fold them. By the way, I have to say, we were up at the lake this last week, and with my brother, and we have family up there, and so we often play cards at night. And my brother bought this, I don't know if he got it on Amazon, but he got a card dealer.

You put the cards in, it shuffles the cards, you tell him how many people there are, and this thing spins around like a banshee, spitting out cards, dealing perfectly every single time. It is fantastic. I was enjoying between hands more than the hands that we were playing of cards. Especially when you're playing games where you have 13 cards or 15 cards per person, and you've always got the terrible shufflers at the table that are just brutal at it.

Right. This thing was amazing. I just want one just to watch, and it randomizes as it goes. So it does the "card, card, card, card" and go one, two, three to one person, then one to another person, then two to another person, and it still ends up with the right number. Phenomenal. Hmm. Technology. Fascinating. I completely digress, and the only reason I did is because of the card image.

Sorry. It's your guest fault. No worries. So for some reason, the download as PDF isn't working, so I can just say, "The download as PDF button or save as PDF button and the print button is not working in this canvas. Fix it." And we'll see if it manages to fix it. Now, sometimes there's some issues with, without getting too into the weeds here, when you see the canvas in your chat here, this is an iframe, and so there are some features and functionalities that don't work. This is for security purposes.

But if you click the share button and you get the publish canvas link here, and then you go over to that link, sometimes some of the functions work a little bit better because it's not an iframe and there aren't as many security protocols so that there can't be malicious things happening in the chat with whatever code was built here. So, that's why I'm going to see if it can fix it here.

AI sometimes fails at the whole PDF thing. We'll see what fixes Cloud Opus can build here. But to answer Michael's question, this is how I would approach it. I would approach it with an LLM, not strictly the image model, because again, those image models have hard-coded sizes that they can generate. And so this is a job for a language model and its ability to execute code and put elements together via code functions.

Superb. For images, this will be a game changer. Oh, man, it's really going to town here, isn't it? There we go. Okay. So let me refresh this canvas now. Click on download PDF. And it doesn't seem like these are quite working either.

Well, such is AI. So you can, I think, maybe even... Well, what if I did print? So that might work but yeah. So AI still kind of struggles with this.

There are ways to kind of troubleshoot it yourself. Hopefully, that's a good help. Okay. We have a couple of back to basics questions. Where do I click to find the AI agents? I can't find them now. Yep, they are located...

Let me close this. So whenever you switch your model, you should see the settings row here. Settings row contains any settings for that given model, but it also contains the agent, if you're using an LLM. Now, if you're using an image or video model, image or video models can't use agents, so there's no agents there. But if you're using a language model, the agent will be directly underneath in that settings row, and you can choose from any of the agents that you see available.

And we're working on getting a button in here that will send you to the agent library. If I copy this real quick. I think I only shared one screen. Hang on. Got to learn. Share entire screen every time.

It's just like three extra clicks Here we go. So you'll have this tab here where you can explore all the public agents that have been created, currently 377 of them. And you can explore them, like our slide deck creator, and then you can add them to your workspaces.

Well, of course I have all these agents, so let me see. Is there one that I don't have? Yeah, there we go. So you can click the Add Agent, and then it's going to ask you which workspace you want to add it to, one or several or all, and you can add that agent. But we'll be bringing that button into the interface this week.

Okay. We have one other question here. I think Michael's on a roll with his questions today. But this is hearkening back to previous trainings that we have done. He's asking, just as you're able to create an agent, form a persona to help you create agents, which we've done, so we've used the AI to help us create agents, can you create an agent to help you configure your workspaces?

Not yet. It's something that's on our roadmap. We started building that out and then we realized that there are some other things that we need to consider, some security and just checks and balances, I guess, that need to be in place to make sure that the AI doesn't change it without your knowledge. So some hardening, I guess, of the workspace context and protecting it.

But eventually that'll come once we've worked out all the angles to make sure that it's safe and secure and doesn't have unintended consequences. Okay. Now next one we've got is Carl is asking, and this is a good question. I like this. Is there a way to take out the ones you don't use? So what he's referring to is the mess that our canvas ends up being with a whole bunch of, if you had three or four images, saying, "Oh, this one's so far out, just don't even remember it anymore. Just delete it.

Clean it up." Yeah. So I'd love for Carl to confirm this, because my interpretation of that is take out the image and video models that you don't use. Oh, okay. I was thinking of in the thread. But either way. So let's answer both. Yeah, we can answer both. So no to both.

That's the easy answer. So in the example of in the case where if you wanted to take out the previous images or the previous videos or even the previous canvases it created, you could use your words to tell it what you want it to use rather than what you don't want it to use. Because it'll only use what it thinks is relevant.

So if you say, "I don't like image one and two, let's create a new one that is separate from those," or, "Let's start over, I don't like any of those." You're just telling it, "I want to use those," or, "I don't like those. Let's do something different." It will obey. It will absolutely do that. Now, if the nature of your question was AI models, can we get rid of the AI models that we don't use? No, there's not currently a way to do that.

However, for some of our enterprise customers and some of our larger team customers, we do provide a organizational dashboard, and if you run a large organization, oftentimes they do want to turn off specific models and not allow their staff or team members to access certain models, whether that's for regulatory issues or just policy issues.

Our organization customers, our org team customers, can turn off certain models. They can control which models are accessible for their team members. So that is an ability, but you have to have an organizational account. If the question is agent related, can I get rid of all the agents in there that I don't want? Currently, no. The preset agents that we provide every Magai user with, they're in there and they stay in there. There's currently no way to turn them off.

We are considering that ability to turn them off. But only once we get that full agent ecosystem sort of integrated well. Until we have that in there, you won't have that ability. Once we have that in there, there will be a much more easy way to kind of control which agents are available in there by default, and you can kind of remove them as you see fit. Hopefully, that covers all the different ways that question may have been intended.

Before we go on, as we completely sidebar, we have a few more questions. I found this. This is the card shuffler. That's Canadian dollars, $179. That thing is amazing. And you're going to want to put googly eyes on it because it looks like it's spitting it out of its mouth when it's doing it. Having said, there was another feature that my brother told me about, and I see here, there's buttons. So if you're playing poker and you want a card sent to you- Oh my God ... you can hit the button and it'll send you a card.

That's incredible. Isn't that something? That is incredible. Yeah. I should have an Amazon affiliate link for this, I tell you. It is just- We don't play cards at our house, but I almost want to start playing cards just so I can buy this and use it. Yeah. We play cards all summer up at the lake. It's the thing for us to do at the lake. So it's a nice family thing. But at any rate, there you go, folks. No charge.

Come to the AI Academy and- ... buy a bunch of new gadgety things. New toys. Actually, and Carl confirmed that he did mean agents when he- Yeah ... for the question that you just asked. I'm not too sure what this one is. D**k is asking, "If I'm a member through The Great Discovery, do I have access to the library?" Yes.

What is The Great Discovery? So The Great Discovery is one of our partners. Oh, okay. They are offering Magai to their customers. And yes, so as long as you have a Magai account that you can log into, which is what The Great Discovery provides, you will be able to get The AI Academy library. It's all it requires, is as long as you have an active Magai login account, then you'll be able to access it.

Oh, this was one. I missed this one. Sorry, Renette. I did miss this earlier. When you use your LLM to prompt a photo or video, does it not use double credits? So the LLM charges whatever amount of credits are used by the LLM, and the image model's going to charge whatever credits are used by the LLM. It's not necessarily double, but it is two different AI models consuming two token rates, or, well, images are not tokens. They charge completely differently.

But the heart of that question, I think, is something I really want to emphasize. We really need to stop worrying about cost. We absolutely do. Using the gas analogy, if I need to get to the store to buy something for groceries, I'm not counting the cost of the gas, right? If my car needs more fuel, I just put my gas in it.

We're not counting the cost of, well, I need to go have this business meeting. I need to drive there, and if I drive there... We're not counting that cost. We should not be ruled by the token costs of our AI machines. If the task has value, we just do the value, or we just do it and get the value.

And really, if it's working for you and you're getting your productivity out of it and you bump your head on the ceiling, then it's a decision point. Is it worth this much money for me to do these tasks? Yeah. Then it becomes evident. Rather than worrying about how much each one, I don't want to do this process because I want to retain my tokens for something else.

Right. You should be measuring it based on the output, I'd think. Yeah. I agree. And the question I always ask myself, because I'm someone who out of my own pocket I pay well over $1,000 a month for the AI work that I have done for my business, for my personal projects, for my hobbies, for my personal life. I never ask that question. It's a matter of if I need to get this done, if this is something that needs to be done, if it's a task that requires being done, am I going to spend

30 seconds with an AI model, or am I going to spend an hour of my own time? At the end of the day, I'm going to choose the AI every single time because I value my time more than what any AI model is going to charge me in tokens over the course of 30 seconds. So we really need to put that into perspective. Now, obviously, if you have a hard budget of I can only spend $40 a month on AI, and you do three or four tasks and all of that $40 is eaten up, then maybe you have to consider, is the work I'm doing valuable? Is the work I'm doing worth $40?

Is it bringing $40 back to me, or is it saving me an hour of my time? Yeah. All day, every day, you should value your time- And it's also- ... more than the AI subscription. It's also learning to use the tool. Brad says here he disagrees. Last month, I went over my account twice and saw that I was using an expensive LLM. That's experience, but that's not to do with saying, "I don't want to do this because it's going to use credits." Use your credits wisely. Make sure you're using the right tool.

Yeah. If you've got a Toyota Corolla, you're not going to tow your trailer to the lake. Use the right vehicle for the right task. Right. You're not going to pull out the Lamborghini to go get your mail, right? Yeah, exactly. You go to the mailbox that's around the corner. I don't know if they do that in Canada, but they do it here. For some reason, the Bakersfield, our mail has, it's around the corner. We've got to go to this big mailbox.

Oh, yeah. The group mailbox. There are some. Yes, they have them. Yeah. So I'm not going to pull out the Lambo to go around the corner, right? It's learning how to use the tools well, use them responsibly. But again, if it's a business task, and if I spent $40 on one subscription and it saved me five hours, that subscription was worth to me, because I won't spend an hour with anyone for less than $300. That's my going rate as a consultant.

If I spent $40 and my time is worth $300 an hour and it saved me five hours, three times five, that's $1,500. That $40 AI subscription just paid for itself- Multiple times over ... well, 15 times over. So it's a matter of understanding what value that that subscription is bringing you.

Yep. Okay. I think we'll call it a day at that point there. We're at 50 minutes. A good kick of the can and some good, useful information today, as well as some good entertainment-based information. So we covered the entire gamut. Folks, we appreciate you. Dustin, we appreciate you. Those of you that are waiting for the replay to appear in the Magic, be a little bit patient. We're going through a couple of little technical challenges to make sure that it's all going to work properly and that we have a sustainable system for the future, but it'll be done shortly. So with that, I thank you so much.

Till next time, have fun storming the castle.