EP 447: AI's Technical Leaps: Memory, Models, and Major Changes (2025 AI Predictions - Vol. 5)

This is the Everyday AI show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business and everyday life.

In 2025, we're going to achieve AGI. Large language models may not really exist, at least how they do today. They're going to get infinitely smarter and dumb AI is going to be a little less dumb. All right, you've made it. If you've joined us all week, thank you. Today we are wrapping up our 2025 AI predictions as we are literally laying out the roadmap for you and your company to succeed. Succeed at least when it comes to AI in 2025. Today we're going to be tackling AI's technical leaps, memory models and major changes. That's right. Spicy. It's. It's gonna get extra spicy today, y', all because, because I'm feeling it. I'm feeling in my fields, been talking AI all week, giving you the predictions and I'm excited to cap it all off with the last, the last show here. So. So if you're new here, welcome. My name is Jordan. I'm the host of Everyday AI and we do this thing every day. So it's a daily live stream, podcast and free daily newsletter helping us all learn and leverage generative AI to grow our companies and careers. I want you to be the smartest person in AI at your company and we are your cheat code and your website is where you do it. Our website, your everyday AI dot com. So make sure to sign up for our free daily newsletter. We're going to be recapping today's show there. If you miss anything, don't worry. As well as has, we have more than now like 450 back podcast episodes sorted by categories, no matter what you care about, right? Hr, legal, tech, ethics, advertising, whatever, mid journey, it's all on our website, sorted by category. So you can learn from the world's leading experts for free. You can listen to all the podcast videos. It's all there. All right, so if maybe you just stumbled upon this show and you're listening for the first time ever and you're like, what's going on here? Well, I, I'm super lucky. I get to talk to hundreds of the smartest people in the world on AI. And usually what happens, right? Because this is a live stream. This is unedited, unscripted, right? And a lot of times I go back and I reflect because I write the newsletter with my fingers, right? I don't hand that off. To an AI model. And then later I'm connecting all these dots and throughout the full year I'm writing down big trends that I'm seeing that no one else is spotting. Right. Because guess what? I, yeah, I get to talk to all the people from Microsoft and Google and OpenAI and IBM and all these other companies, but they probably don't talk to each other a whole lot. I'm lucky enough I get to and you know, I'm helping, you know, people hire us to help them, you know, learn, chat, GPT or Copilot or whatever it is. So I'm in a very unique position where I get to talk to a lot of the world's smartest people, soak it all up. But I spot these trends and that's what this week's series is all about. Our 2025 AI predictions and we are on volume five of five. So if you are tuning in for the first time, you missed a lot but I'm going to give you a very quick update on what you did miss. All right, ready? And you can go back and listen to all of these episodes in the theming of things Preparing you for 2025. Trying to keep all these episodes to 25 minutes or less which if you listen a lot, you know that's a miracle. So here's what you missed. Ready? Volume 1 I'm just going to give you the predictions if you want to go listen to this. If something in volume one catches your ear, it's going to be in the show notes, it's going to be on our website. Just go look for it. All right, it's four episodes ago so volume one number prediction. 25 agent orchestrators will be a growing position. 24 public companies will post jobs for AI agents. 23 company reasoning data collection. That's going to be huge. 22. Two high end professional services will go through pricing crisis. 21. UBI becomes a household conversation. And then in volume two start with number 20. Open source surges. Open large language models will temporarily overtake proprietary models. 19. Chinese AI will dominate and cause confusion. 18. Perplexity will pivot, get acquired or get squashed. 17. API prices are going to drop like they're hot. 16. Embodied AI will be an exploding sector. The then in volume three here's what we covered. Number 15. AI video tools will one shot five plus minute HD videos and we'll have advanced personalized media. 14. The future of traditional Internet comes into question. 13. Social media makes deep fake AI problems way worse. Number 12 first big copyright case is decided Number 11, AI influencers are going to start killing off human UGC content. Here we go with the top ten. This was volume four, first from yesterday. Go back and listen if you want to. Number 10, non techies will build on the fly software. Nine, reasoner rappers will hit the scene. Eight, virtual machines become all the rage. Number seven, AI becomes overly political. And number six, global regulation around AI titans, but not in the US. And here we go. Without further ado, here's volume five. Our last set of predictions. AI's technical leaps, memory models and major changes. Here we go. I'm out of breath. Number five, narrow AI agents will be achieved. Number four, LLM memory becomes a major focus. Number three, LLMs become small language models and SLMs dominate. Number two, mixture of models becomes a thing. And number one, AGI is achieved, but no one notices. All right, same as always. Whether it's live stream audience, you let me know. Podcast audience. I leave my email in my LinkedIn. Just tell me you're from the podcast. Just connect with me, but let me know which one of these is the most likely to happen and which one is the least likely to happen. Maybe I'll put it in the newsletter, but let me know. You can just put like if you think number three is the most likely, just put three most or five least or one most. Whatever you think. All right, Because I want to hear from you. This thing's. It's not just me yelling at the camera in my home office. I'd love to hear from you. We'd love to feature, you know, some of your great comments in our newsletter as well. So let's get into it. Number five, narrow AI agents will be achieved. Let me tell you what that means. First, I have to quickly explain a little bit of the difference between AGI and ani. So artificial general intelligence versus artificial narrow intelligence. So artificial narrow intelligence has been achieved for quite some time, right? Depending on, you know, your perspective, your background, it could have been for, you know, more than 10 years, maybe even longer. Right? That's when a certain AI system beats the world's leading person in a certain task or a certain field, right? You can go back to, you know, go on Jeopardy, or, you know, AI beating chess, grand masters, whatever. Ani, we've been past that, but, you know, that's not really what we're about. You know, everyone's always talking about now agents and AGI, right? Two things that I think we are on the cusp of. But I don't think we're going to have agentic AGI anytime soon. And I think it's actually, I'm not going to say it's a problem. It's more of people are talking about it like that's the thing that's going to happen first, right? Where we are going to have agentic AGI and what that means. Or like, let's just start at agentic AI, right? We'll break one piece of Alphabet soup off the, the spoon at a time. All right. Agentic AI is when an AI system can go and make and execute decisions on your behalf. It has access to tools it can access, generally can access the Internet and your data. Right. So think of it like, oh, you know, like a very new intern, right? You give it some directions and then it can go play with multiple programs and it can make decisions on your behalf, access your company's data, you can give it guidelines, etc. Right. AGI, that's, that's more of a. This, this ever changing threshold. More on that later. Right. But I think when we think of agents, everyone's thinking of agentic AGI. No, no, it's going to be narrow AI agents. So everyone thinks that they're going to have one agent that just becomes their personal does everything. No, I think the, the rise of agentic AI is in 2025. That's a lot of rhyming, right? It's, it's on narrow applications, right. I don't think you're going to have a super agent like a Jarvis that does literally every single thing. I think you're going to have 10 different narrow Jarvises, right, that are just good at one very specific thing. And there's going to be an agentic router of sorts. More on that here in a minute as well. A lot of my predictions today actually kind of bleed into each other and I did that on purpose because I'm the most excited about this one. It's going to be very hard for me to wrap this show in 15 minutes. All right? But I don't think in, in, in. To keep it short, I don't think we're going to get these general AI agents, which is what most people think we're gonna get. We're gonna have narrow AI agents because plenty of people, the big companies as well, they're going to try to release general AI agents. They're not going to do too well. All right. The same way that early large language models, you really had to tune them and work with them a lot to get them to be good at one thing. Right. That's why in our you know, prime prompt, polish course, which I know y' all been hitting me up. It's gonna be back on soon, I promise. Right, but that's why, you know, we teach you, you really have to work with a chat to get it good at one thing. Right? Same, same thing applies. True. At least right now until we kind of quote unquote get to AGI or agentic AI that uses reasoner models. But until then, I think you got to focus on one agent and you're going to have many agents. One of our first predictions was actually agentic orchestrators, and that's why, because you're going to have a ton of them and they're going to be narrow, not general. All right, Number four, LLM memory. Large language model memory becomes a major focus. So I'm not talking about context windows. These are kind of two different things. I'm talking about memory. So that's an AI's ability to be able to recall events almost infinitely across the entire organization or across an entire account, and not just in a singular context window. So I know that there's kind of some, some nuances there. And you know, unless you're a dork, it might be a little difficult. But think of it like this. I think context windows and you know, that's kind of the amount that you can throw at a large language model in a single chat. And it can kind of remember it all right, until it starts forgetting things. Context windows are going to continue to grow, but I think memory is actually going to become just as big of a factor. Right. OpenAI's Early Memory feature was actually very novel. It just is not executed well because you can't kind of toggle when you want it to remember something or not. So it's a little finicky. I think some simple UI UX interfaces are really going to take this memory kind of concept and run with it. Also, we've seen reports that OpenAI and Microsoft are working together on AI advancements scheduled for this year, where they're focusing on developing AI models with near infinite memory and also expanded context capabilities. Right. That's one of the biggest problems, you know, aside from hallucinations and people not even knowing how a freaking large language model works. Right. They don't know what a transformer is. They don't know, you know, tokenization, they don't know weights, they know nothing. Right. They're just like, ah, look at, look at this prompt that I put into AI and it gave me this garbage. AI is not taking my job. No. AI is definitely taking your job because you don't know how it works. So yeah, you're first on the chopping block, Bill. All right, I don't know. Sorry. If your name's Bill in listening, I just picked a name. I'm sorry. But no, it's, it's. Most people don't understand the basics of generative AI of large language models, but as context window grows, as memory grows, outputs are going to be better. Then you become more trustworthy. Organizations have more trust in these models as hallucinations hopefully start to become less and less frequent. Yes, hallucinations are a feature, not a bug. I get it. We're talking about non deterministic, you know, technology here. So the whole point is it's supposed to be able to give you something different every time, but not at the, at the expense of being wrong or incorrect. All right, Are you still running in circles trying to figure out how to actually grow your business with AI? Maybe your company has been tinkering with large language models for a year or more, but can't really get traction to find ROI on Genai. Hey, this is Jordan Wilson, host of this very podcast. Companies like Adobe, Microsoft and Nvidia have partnered with us because they trust our expertise in educating the masses around generative AI to get ahead. And some of the most innovative companies in the country hire us to help with their AI strategy and to train hundreds of their employees on how to to use gen AI. So whether you're looking for chat, GPT training for thousands, or just need help building your front end AI strategy, you can partner with us too. Just like some of the biggest companies in the world do. Go to your everyday AI.com partner to get in contact with our team. Or you can just click on the partner section of our website. We'll help you stop running in those AI circles and help get your team ahead and build a straight path to ROI on the gen AI. Number three. Here's a sneaky one. Large language models as we know them today, they're gone. Large language models are becoming small language models, believe it or not. We're going to do some math here, y'. All. All right, let's do some mathing. Let's talk about GPT4. That I think is the epitome of large language model. Right when we're looking at frontier models, you know, you can look at the early or maybe even go, go further back. Early days of the transformers, late 2010s seeping into the early 2020s. You know, these, these models kept getting bigger and bigger and at first, you know, small language models have been around for a long time as well, you know, and generally they're there for more specialized purposes and not these general models. So these general models, we call them more large language models. Right. Then you had your smaller models, and it was judged by the size of in measured in parameters. So not all companies release and say, hey, here's how big these parameters are, or how many parameters are in the model. Right? But it was reported that GPT4, which at the time when it came out, smoked everyone. It's disgustingly good, right, that it was 1.7 or 1.8 trillion parameters. All right, so nearly 2 trillion parameters. Think of that like, like a hard drive, right? Like two terabyte hard drive. Right. For a comparison, two trillion parameters. Huge. Okay, so Microsoft just came out with some research recently that apparently cracked the code on the size of some of the newer proprietary models. Guess what, its successor, GPT4O, 200 billion parameters. So a tenth of the size and much more powerful. Still a fairly large model. But when you look at GPT4O mini, this is some of the biggest news, I think, of the last couple of months, that literally no one's talking about GPT4O Mini, which is a very capable model. Right. And people generally who are using the API, right, who are using OpenAI's API, many of them are using GPT4O, not just, sorry, GPT4. Oh, mini. Not just for chunking, but for being the workhorse because it's an extremely capable model. Right. It is near the capabilities of the original GPT4 model. That was 2 trillion parameters. I'm not great at math, but let me just go ahead and compute. But that's, that's not even 1% of the size. That's a half percent of the size. Okay, what are you getting at, weird guy? Well, the definition of large language model used to be, you know, it was always, you know, a moving definition, but it was like. Oh, it's when it's hundreds of, you know, hundreds of billions of parameters and then it's like. Oh, well, no, it's actually when it's trillions, trillions of parameters. Okay, what happens when even frontier models, the big ones. Right. What happens when there's a GPT5O that's like couple billion parameters. Right. And can fit. I'm just going to grab something random on my desk here. Right. And can fit inside of these headphones as an example. Right. I need obviously, proprietary models. You can't download them, but I'm putting out a point, right? Think of Meta's Llama, right? They haven't released the large version yet of 3.2, but the 3.1 version was 405 billion parameters. So I assume the, if they do come out with the large variation of llama metas, llama 3.3.2, which you can download and run locally and fork it and all that good stuff, right? Probably fit it on, you know, Nvidia's new digits computer, right? Large language models are becoming small language models and I think we are going to see. Speaking of Llama, right, Here's a little fun fact for you. Nvidia has fine tuned Meta's Llama and this was their 3.1 in it, out benchmarked, met his actual model, right? So I think what we're going to see is more of those types of builds, right? I'm very excited for either Llama 3.3 or you know, when they do, if they do come out with their 3.2 large to see what happens, I think some of these other big tech companies like Nvidia did are going to start to tune models. And I see, I don't see this yet happening in 2025, maybe toward the end, but early 2026 I see proprietary models even going domain specific, right? And that's going to set me up into prediction two here. Real nice. But bear with me, think if these models Llama, right? And I know there's going to be some other open, even truly open source models that are going to be benchmarking as like a top 10 frontier model in the world. So when that happens and then when you, eventually when they're small enough that you don't need a research institution in a team of essentially data scientists to fork them to build new versions of them, right? Pretty soon anyone with $3,000 could, you know, go and download the world's most powerful open source model and play with it on your computer, right? That's nutty because it's like, you know, when you think of like three, four, five years ago, you had to have a freaking data center, right? You had to spend probably millions of dollars to be able to actually run and use and, and run inference on these models. Like it's crazy. So I think this whole concept of large language models, you know, for, for a point there I was calling them jumbo, jumbo models. I think they're just going to be small models now, right? I don't think, you know, in, in three, four years from now, I don't think we're going to have any models that are hundreds of billions of parameters, right. I think even the biggest models in theory are going to be able to fit on a device, right? Which is big. That cuts down energy consumption, right. So we don't have to have, you know, 500 nuclear plants, you know, here in the U.S. producing nuclear energy because we're out of energy. Right? Because when you send your, all your queries or all your company's queries to the cloud requires a lot of energy. Right. Also data and privacy concerns. You know, everyone wants edge models, everyone wants on device AI and I think we're going to get there, right? I mean, think of the fact that GPT4O mini is reportedly 8 billion parameters and it's out benchmarking, you know, so many models that were more than a hundred times that size. It's wild. And then I think, like I said, in the future, I think we're going to start to see hundreds of small language models that are very capable. Nvidia's Llama Nematron esque type models. There's going to be hundreds of them, right? State of the art models. Hundreds of them, not just, you know, a dozen or so that are at today's capabilities. All right, number two, I'm gonna get a little technical, but I think I just set you up there with number three. So number two, we have mixture of models become a thing. You might be thinking, oh, Jordan, you're a dweeb. We already have that. No, we don't. I made it up. No, we don't. See, we don't have it. We have something called mixture of experts, right? So let me just break it down super simply that mixture of experts is essentially, it's activates one expert at a time. So there's a, essentially there's a gatekeeper, right? So in this, in this system there's a gatekeeper. So you give a, you give a prompt, there's a gatekeeper model. And then, you know, let's just say there's, you know, a hundred different smaller models and the gatekeeper model is like, oh, okay, let's send it to this model. All right? And then for the most part it activates one expert at a time. It might give it to multiple experts, right? There's different moe setups. I don't see that happening anymore. I see something different even within the same system. I see something called a mixture of models. That's when it just, you can run a prompt and the same system will be able to run multiple specialized models in parallel, not a gatekeeper handing it off to the best model or handing it off sequentially and it runs one at a time, but running it through multiple specialized models in parallel, each working on different parts of the task simultaneously. Right. So it's kind of similar to how mixture of Experts works, but also completely different. So this is what I see. This is what I see. I see as an example, on the front end, you send a very advanced Query to chat GPT, right? I have the. The Pro plan. It's pricey, 200amonth, but I think it's definitely worth it. All right. And I can choose right now, I can choose Zero1 Pro, right? I love to handle. I like to see how long I can make 01 Pro think, right? Sometimes 10, 12 minutes. I see when I give those very hard tasks off, it's going to give some of it to 01 mini right away. It's going to give some of it to. To GPT4O maybe because, you know, the 01 models yet don't have the web. Right? So in parallel, I see this as how it's going to be happening in the future, right? Or think of like Google, right? A centralized place where hey, here's, here's my company, here's the issues that we're working at. You have all my data. And then Google, in this mixture of model scenario, it might send some of your things to its. To its Imagine AI image generator. It might send something from there in parallel at the same time to its vo. It might send something to a specialized large language model. It might send something to, you know, Flash thinking. It might then send something to Deep Research, right? But keep it all in the same interface, right? And essentially similarly to how a mixture of expert scenario works up. But this is a mixture of models. It runs them all simultaneously and it just chunks off the different pieces, does it all at the same time, and then gives you their output. So similar, but a little different because studies show that right now organizations are deploying on average three or more foundational models in their AI stacks. And oftentimes it's more than a dozen. All right, here we are at the last one. I would love to spend 30 minutes alone on this, but I'm trying to keep these at like 25 minutes. And I just hit the 25 minute mark. So it's gonna be quick prediction. Number one, AGI is achieved. No one notices. No one notices, right? So there's already been an argument with OpenAI's O3 model, which is the successor to O1, right. Some people are like, oh, well, you Know, hey, it passed the ARC AGI challenge. So AGI is achieved, right? And there's a lot of now internal talk, especially at at OpenAI that, you know, oh, now they're focusing on super intelligence, you know, because they've said, hey, they've quote unquote, kind of figured out AGI. So if you don't know what AGI is, it's artificial general intelligence. So that is when one AI system can essentially do any task, any knowledge based task that a human could do, but it can be better than all humans at all tasks, right? So, you know, I gave the example earlier, like, oh, an AI can beat someone at chess, right? So think of any knowledge based task that you could do well right away. One model is better than every single other human on the planet at just about anything. The definition of AGI constantly changes. I'm a dork, y'. All. I did a show on this about three or four months ago. I went back and I used archive.org and I looked all the way back to like 2005, right? Because you can go look at the web from like 2005. And I was reading the definitions of artificial general intelligence from 20, I think I looked every five years up through 2015, and then I looked every single year from 2015 to 2024, right? Guess what? By definitions of 10 to 15 years ago, we've already achieved AGI. I think AGI is constantly, it's the, the definition of what it means is changing just as fast as all the models right now. Microsoft came out and said, oh, well, AGI is achieved once, you know, one AI system is capable to generate $100 billion in profits, right? That's part partly because of their current kind of relationship with OpenAI as a reported 49 equity holder in the company. And there's kind of this AGI clause. So they're trying to define it one way. But here's the thing, y', all, we can keep defining it every single day. Let me ask you this, let me ask you this. And, and once, as an example, once o one gets tools, right? When we see agentic AI in the combination of a reasoning model that has tools, when those things happen, all the individual pieces are there. When that happens, that's AGI, right? It doesn't have to be this big moment, right? Like OpenAI CEO Sam Altman kind of said the same thing. He's like, it's going to happen and then everyone's just going to go on about their lives. I feel the same way. I think all the individual pieces are there. It's agentic AI plus reasoning model, but also GPT model. So it's the combination of those two models working at the same time, plus tool use. Right. So being able to access, access the Internet, access, you know, data analysis modes, all these other things. AGI is definitely going to be achieved in 2025, whether people admit it or not. I don't know because I think the definition is going to keep changing. But do me a favor, go, go look@archive.org look at the definition of AGI from 2015. I did a full hour episode. We've already achieved it. But I think the common consensus in 2025 will finally be, all right, we've achieved AGI, and then we're all going to find out nothing happens. Right. I don't think there's this. This huge life shift, business shift. I think we're going to see a lot more layoffs, and I think things are going to get weird when it comes to traditional jobs. I don't think nine to five jobs are going to be a thing in like, three to five years. Right. But I don't think it's going to be this. This mind in the sand and the whole world shakes. It's not going to be like that, but it's going to happen in 2025. All right. I hope this was helpful, y'. All. We made it. This is volume five, our last set. So this is more than just a show of predictions. Right. I probably should have named it the Trends Ahead, because that's what this is. That's what this series has been. This has been the culmination, y'. All. Certain days, I spend 10, 12, 15 hours reading about AI, learning about AI, talking to people about AI. I don't make these predictions lightly. Right. These are your blueprints. All right. You need to be learning from what I said because. And I'm not trying to say that in a way of like, oh, listen to what I say. No, because I just steal all the information from all the smartest people in the world. But I'm giving you all that platform to say, this is what's happening. This. These are the trends ahead. All right. Some of y' all know, like, my background, actually, I have a couple different backgrounds, but I spent 10 years working at a nonprofit. I love helping people. That's why I do this every day. Right. I could probably be making a lot more money doing something else than teaching you all about AI every day. Right. But I feel the need to. Right. I was a journalist also before that. Right. So it's, it's kind of this culmination in what I've been doing for the last two and a half years, doing this every day. I honestly want to help. I want to help people. I want to help companies not just survive this AI, but I want them to thrive. I want to help companies thrive. It's not going to be easy. Things are only going to get weirder. They're going to get harder. Job loss. I'm telling you, I don't want to end on this note. It's going to pile up. Right? But if you're here, if you're listening, if you are actively taking a part in using AI every day, in learning every single day, putting it into practice every single day, you're going to be fine. All right? So please go listen to all of these episodes. We put a lot of work in. I hope they're helpful. If so, let me know. Please reach out. I always, like I said, put my email in the show notes. If you're listening on the podcast or my LinkedIn, just let me know you're from the podcast. Otherwise I don't know who you are. All right, if you're listening on the live stream, let me know. I hope this is helpful. We got a lot planned for the rest of this year, 2025. It's going to be exciting and I can't wait to be able to show you guys a lot of what we have coming for you. So thank you for tuning in. I hope to see you back next week and every day for more Everyday AI. Thanks, y'. All.

And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going for a little more AI magic. Visit your everydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.