Ep 676: 2025 AI Roadmap Rewind Human vs Machine, AI Models Shrink, and AGI No One Noticed

2025 AI Roadmap Rewind: Pinpoint Shifts Every Business Leader Needs to Track

The 2025 landscape for artificial intelligence evolved faster and more tangibly than many business leaders anticipated. This year-end summary focuses directly on which developments had real business impact, drawn from a systematic review of 2025’s boldest AI predictions and their outcomes. Each insight below corresponds directly to tech and regulatory events tracked by AI-focused thought leaders and business strategists.


Copyright Settlements: Towards Pay-to-Train Models in AI

One of the year’s pivotal moments was the settlement of a landmark $1.5 billion copyright lawsuit against Anthropic for allegedly leveraging published books in its AI training data. This event signals a transition toward “pay to train” agreements, providing clear pathways for companies using third-party intellectual property within machine learning models. Major players such as Disney and OpenAI exemplified this shift, securing direct licensing deals involving equity and cash—critical for enterprises that rely on copyrighted data or partner with IP-rich industries.


AI Influencers and UGC Content: Displacement of Human Creators

Brands and marketing firms are already mainstreaming AI-generated influencers and avatars for user-generated content (UGC), with platforms like TikTok and Meta enhancing their suite of generative ad tools. AI avatars with multi-million dollar followings and major brand endorsements are no longer hypothetical—large consumer firms are integrating these into UGC campaigns at scale. Recent analyses suggest up to 80% of influencer-like content may soon be AI-generated, compressing costs and accelerating content velocity for marketing teams, but raising profound questions for agencies reliant on human UGC.


AI Coding and Software Creation: Non-Technical Staff Build Apps On-Demand

The market saw the normalization of “vibe coding”: non-technical users leveraging low-code and AI tooling to build full-function applications on-the-fly. Real workflow cases show solutions like Google’s Disco and AI Studio enabling staff to generate new web apps via simple commands or even through their browser activity—no prompt engineering or coding background required. A recent study found 70% of new enterprise apps in 2025 were built via these accessible tools, democratizing the software creation process and accelerating internal innovation cycles.


Virtual Machines for AI Agents: Cloud PC Adoption in Enterprise Automation

AI agent deployment increasingly relies on virtual machines or desktops, evidenced by innovations like Microsoft’s Windows 365 for agents and Google’s browser-based Project Mariner. These systems support agents that automate operational tasks and decision-making workflows without physical installations. Businesses strategically deploying cloud-based PCs tailored for agent use free up IT resources and facilitate rapid scaling of automation initiatives.


Political and Regulatory Shifts: From US Deregulation to Global Stringency

AI entered the policy mainstream, especially in the US, where federal actions actively preempted state-level AI regulations. This top-down deregulation approach stands in contrast to the European Union and countries like Singapore and Italy, which enforced rigorous, innovation-constraining AI governance. For multinational companies, this has meant segmented strategic rollouts—US markets accelerating new feature deployment, while EU and APAC plans face delays and stricter compliance demands, often impacting the availability of advanced AI capabilities.


Emergence of Narrow Agents and AGI Benchmarks

General-purpose AI agents failed to gain mainstream traction due to performance gaps, but vertical- or role-specific “narrow agents” dominated departmental automation investments. Salesforce, Oracle, and GitHub led with tailored agents optimized for CRM, ERP, and coding, signaling a $7.3 billion investment trend toward departmental AI rather than broad AGI.

Notably, while most organizations didn’t witness an AGI arrival “event,” LLMs (Large Language Models) achieved near-genius averages on official reasoning and IQ benchmarks. On OpenAI’s GDP Val, AI models were judged by industry experts as equaling or outperforming human professionals in three out of four critical economic tasks. For business process outsourcing, enterprise enablement, and knowledge work, this data represents a pivotal inflection point in human-versus-machine productivity debates.


LLM Memory, Model Shrinking, and Mixture-of-Models in Production AI

Persistent context and “memory” became front-and-center for all major LLMs—OpenAI, Google Gemini, and Claude—enabling enterprise-use chatbots to recall previous queries and personalize outputs over time. In parallel, the rise of “small” and domain-specialized models (ex: OpenAI’s 20B-param GPT-OSS surpassing last year’s giant GPT-4 family) meant more efficient, cost-effective deployment at the department or use-case level.

Perhaps most quietly, “mixture of models” or “ensemble” architectures moved into practical use: platforms like Zoom orchestrated multiple off-the-shelf models in parallel to outperform any single LLM on complex task bundles, while Google's Interactions API enabled live interplay between several model families for optimized outputs.


Strategic Considerations for Business Leaders

The granular movement of 2025’s AI space highlights a few hard-nosed realities:

  • IP-related settlements are now standard cost centers for AI projects with third-party data dependencies.
  • Internal and external content strategies require rethinking personnel mix as AI UGC reaches near-parity with human creators.
  • AI/automation scaling, especially in regulated geographies, requires modular design for maximum agility amid changing compliance rules.
  • Organizations prioritizing “mixture of models” approaches and departmental (not universal) agent deployments gain resilience and best-in-class performance.

For boardrooms and functional leaders, staying current means not just reading headlines about “AGI” or influencer trends, but tracking concrete progressions in how work gets done—a process growing less visible, more automated, and increasingly shaped by the regulatory choices of national governments and transnational blocs.




Topics Covered in This Episode:

  1. Anthropic $1.5B Copyright Lawsuit Settlement
  2. AI Influencers Replace Human UGC Content
  3. Non-Technical Vibe Coding Software Trend
  4. Enterprise Reasoner Wrappers in AI Adoption
  5. Rise of Virtual Machines for AI Agents
  6. Political Impact of AI on US Policy
  7. Global AI Regulations and EU AI Act
  8. Narrow AI Agents Drive Business Value
  9. LLM Memory and Context Advances in 2025
  10. Shift: Large Language Models to Small Models
  11. Mixture of Models Approach for Enterprises
  12. AGI Benchmarks Surpassed, Yet Goes Unnoticed




Episode Transcript 

Midroll [00:00:00]:
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.

Jordan Wilson [00:00:16]:
Did AI end up being a political force this year? Did AI start to kill off influencers? And maybe more importantly, did we achieve AGI in 2025? Well, more than eleven months ago, we laid out 25 kind of bold predictions for 2025 that covered those questions and a lot more. So now as the year comes to a close, we're fact checking ourselves. Why? Because when we give predictions, they're much more than that. We think of them more as a roadmap for busy professionals who are trying to make sense of AI and grow their careers. And we're lucky enough on this podcast to be heard by millions of people. So usually around this time each year, it's when business leaders reflect on what's worked that year and then plan for the year ahead. So on today's show, that's exactly what we're gonna be doing. We're gonna reflect on the advice that we gave you and fact check ourselves in part two of our AI road map rewind.

Jordan Wilson [00:01:19]:
Alright. I'm hope I hope you excited. I am too. Let's get into it. Welcome to Everyday AI. If you're new here, what's going on? My name is Jordan Wilson, and Everyday AI, it's for you. It's a daily livestream podcast and free daily newsletter, helping everyday business leaders like you and me make sense of these AI updates and how we can extract the important insights to grow our companies and our careers. So this is a little bit of a special episode.

Jordan Wilson [00:01:43]:
So like I said, in January 2025, we laid out our AI prediction and road map series, and now this is part two of fact checking ourselves. So if you missed part one, you can go click, back, you know, two episodes. Go listen to episode six seventy four, and, you can go ahead and, check on part one. So on part one, we were going reverse order. We went predictions 25 through 13. And then on today's show, we're gonna go predictions 12 through one. So, and we're gonna be recapping it all in today's newsletter. So if you miss, certain fact or stat, maybe you're out, walking your dog or running the treadmill or running your dog on a treadmill and you miss something, it's all gonna be recapped in our newsletter as well as all the other AI news that you need to know for today.

Jordan Wilson [00:02:33]:
Let me set the stage here. So maybe if you didn't listen to episode one or maybe this is your first episode ever, well, welcome. So I've been lucky enough over the past three years to talk to some of the smartest people in the world, when it comes to AI. So every everyone from, you know, people who work at the big companies, you know, like, Google, you you know, Microsoft, etcetera, to small business owners, you know, who are making use of this technology, to Fortune 100 CEOs. So you hear a lot of these conversations on the podcast, but there's a lot more that happen off air that you don't hear about. So I have been lucky enough over the past three years to talk to hundreds of the smartest people in the world. So when I put together these prediction shows in January, like I said, I take a lot of care in them. And at the time, they may seem kind of bold and crazy, but as maybe you'll find out in, part one and in part two, they actually ended up being not too crazy.

Jordan Wilson [00:03:31]:
But, you should go back and listen to them anyways. I still think that they're really valuable, and I want you to fact check myself because, I'm gonna be coming out with our 2026, AI prediction and road map series here in a couple weeks. So if you wanna go listen to the full episodes, go listen to episode four forty three through four forty seven. Alright. Let's get straight into it. Like I said, we already went through 25 through 13 in part one. So part two, we're picking up with number 12. So number 12 was the first big copyright case would be settled.

Jordan Wilson [00:04:02]:
So here's what it was. Well, yes, that happened. It was a $1,500,000,000 settlement, where Anthropic had a class action lawsuit re regarding, could, allegedly just stealing information from books. Right? That's what was alleged against them, and there was a $1,500,000,000 settlement. So, this is, I think, one of the first big, pieces that eventually we're gonna move to a pay to train model. Right? And another good example of that was from, well, this week, when OpenAI and Disney, you know, went into a, agreement where essentially Disney gets some equity in OpenAI, and they pay a billion dollars, and OpenAI gets to use, their IP. So this was the first big copyright case, that was settled. Obviously, the biggest domino that I've talked about has still been going on for, more than two years, the New York Times versus OpenAI case.

Jordan Wilson [00:05:03]:
The Anthropic case wasn't the only one against the authors, you know, of these books. It was technically called Bart's versus Anthropic, the class action lawsuit, but we got the first big copyrights, case to be settled. And there's been a lot of other lawsuits filed, since very similar to that one. And, you know, for whatever reason, I think at the time, Anthropic was maybe a sitting duck. Right? Technically, you could probably make this a a a similar case against some of the other, big tech giants, but, you know, they're kind of a smaller, technically, a smaller company even though they're enormous. Technically, one of the biggest private companies in the world, but they were still small enough. Right? Smaller than Anthropic or or sorry. Anthropic is smaller than OpenAI, smaller than Microsoft, smaller than Google.

Jordan Wilson [00:05:52]:
So they may have been an easier, target per se for this first class action lawsuit. So let's go on to number 11. So anyways yeah. That that I don't know. To me okay. I I also need to grade the, the prediction if it came true or not. So this one, I'll say yes. It came true.

Jordan Wilson [00:06:11]:
I didn't think that the, New York Times versus OpenAI would be that big case. So, yeah, it ended up being still a pretty big one. $1,500,000,000 settlement, to these, hundreds of authors, that their books were allegedly used for anthropics training data. Alright. Let's go to number 11. AI influencers are gonna start killing off human UGC content. So my actual prediction was that AI influencers or avatars are gonna start replacing human user generated content at scale. So I'm talking specifically UGC influencer type videos.

Jordan Wilson [00:06:46]:
So although there isn't a huge AI, like, there's not a huge study on this, I think it's happening. And the crazy thing is is most people don't know. Here's why. I'm not a big person on social media. Right? This everyday AI gets live streamed to to LinkedIn, YouTube, but I'm I'm not, like, on Instagram or SnapHub, whatever those things are. Right? However, I do know we've, right, my other company, you know, I have another company. We do, you know, marketing for small businesses, and I know the tools that are out there. And I know people that are using these, and big companies are now using just AI influencers for UGC campaigns, user generated, content.

Jordan Wilson [00:07:34]:
Right? It's it's those videos. If you're scrolling your feed, it's just a random, you know, random person pushing a a a beauty product or, you know, a new workout, whatever it is. Right? You don't know who it is, but someone's saying, oh my gosh. This thing, like, you gotta check it out. Right? So many of those are fake, and AI influencers are becoming a huge thing. So, as one example, hopefully, I get this right. There's an AI influencer, Lil Mikaela. It's who has, nearly 3,000,000 followers across different social media, networks, earns up to $10,000,000 a year and has partnerships with real brands like Samsung, Calvin Klein, Prada, and a lot of others.

Jordan Wilson [00:08:19]:
So these companies obviously know that they're working with an AI influencer. Right? A lot of the fans or the followers know as well, but some don't. And I'm telling you, y'all, this is gonna become the norm. I actually had a, a a friend text me, earlier this week, you know, asking me about this because she saw an article, you know, about something like this, in the media. Right? Like, is AI gonna kill off a certain breed of influencers? And I said, hey. Probably by the end of next year, I'd say maybe 80%, you you know, of of what you might see, online is going to be AI. Right? And the the platforms are also pushing this and enabling it. Right? Whether you think that's a good thing or a bad thing.

Jordan Wilson [00:09:04]:
Right. As an example, TikTok expanded their Symphony, Gen AI ad tools. So, essentially, TikTok is providing a similar service. Meta, similarly, has introduced and expanded their AI ad tools use. Right? So these big social media platforms are giving brands the tools to use AI in their ads. Right? Not full blown avatars just yet, but it's getting there. And those, full blown AI avatar companies, HeyGen, Synthesia, Hour One, there's a lot of them. You know, they're improving vastly.

Jordan Wilson [00:09:39]:
Right? And then you have tools like Veo, like Sora, right, that, are primed now for making UGC content. I think, Cameos and Sora is a great use, for making UGC content. Right? Maybe of yourself, you know, because I think that's all you have the rights to do. But, you know, know, that's a whole another, a whole another issue. But I would say this has definitely been a fact. AI influencers are starting to kill off human, UGC content. There's actually a Yahoo Finance article, here. It said Gen z job warning as new AI trends set to destroy 80% of influencer industry.

Jordan Wilson [00:10:19]:
Alright. Our next and sorry. I'm a little hoarse. Yeah. Been getting randomly sick here at the end of the year. It's the Chicago weather. It's like I think it's it's, today is gonna be, like, 50 degrees, but, like, four days ago, it was negative five. So yeah.

Jordan Wilson [00:10:39]:
Sorry. If I sound a little hoarse, it's because I am. Alright. So, number 10. This is I said non techies are gonna be able to are gonna build on the fly software. Y'all this was, this is how fast time flies in AI world. This was before vibe coding, right? Vibe coding was literally not a thing. And I just kind of in January, I just kinda predicted there's gonna be this thing called vibe coding.

Jordan Wilson [00:11:06]:
Right? Except I said it's gonna be nontechnical people building on the fly software. I think vibe coding is a much better term. Right? So the term vibe coding was actually coined, by, Andre Capathi in February, right, a month after this AI prediction. Right? But at the time, I'm like, these tools are coming out. People are just gonna start building their own software, and it's not gonna be hard. Obviously, this one is extremely true. You know, and and it's it's so true. It's now gonna be baked into browsers.

Jordan Wilson [00:11:37]:
Right? A recent announcement, Google's disco, experiment. You can generate many web apps just from the tabs that you have open in your browser. Right? And you don't even have to prompt anything. You just click a button. Right? It's it's it's a fee, an experimental feature that's coming out. But, I mean, talk about things like Gemini Canvas, Chad GPT Canvas, Claude artifacts, let alone all the, you know, higher level, you know, IDE, Vibe IDEs, you know, Cursor, Windsurf, you know, the new anti gravity, from, from Google codex. Right? There's great kind of vibe coding platform, but then there's things that are, I'd say, more like disposable apps, right, that you can just create in a front end large language model, and that's what so many people, are doing. I actually had a great, conversation, couple months ago, with Paige Bailey, one of the heads there at Google.

Jordan Wilson [00:12:31]:
And she said, you know, the people that are winning hackathons are nontechnical people, and they're just going into you know, in her case that she was talking about, they're going into AI Studio. AI Studio has a great build feature. You just you don't need to know anything about coding. You say, make me an app that does a, b, and c, and it just happens, and you can just use it, and it just works. Right? So the reality is, well, a a recent study said 70% of new enterprise apps in 2025 were built using low code AI tools. Alright. So, yeah, crush that one. Number nine, reasoner rappers will hit the scene.

Jordan Wilson [00:13:09]:
So my prediction back in January was that tools are gonna emerge that wrap reasoning models in enterprise data to drive decisions. So, essentially, how the transformer old school models work on structured data, I said, well, there's gonna be a a kind of movement, for reasoning data or a company's kind of how they make decisions. Right? So I'll say this one didn't hit right, a little bit maybe, but it didn't hit as hard as the other predictions. In this case, though, I'll say that I'm definitely not wrong. It is starting, maybe just a little early on this prediction, but there is probably this new agency layer, right, kind of what what I and other people call it. Right? But this is a a business agency. This is a business's decision making, their expertise, you you know, their subject matter experts. It's what happens in their head.

Jordan Wilson [00:14:05]:
Right? So my thought was back in January that there's gonna be a big push, for collecting and curating essentially companies IP, their decision making process. Right? And there is a little bit. You know, there's tools like Letta, that have become pretty essential, to parse and structure the hidden thought process, that pairs models, these reasoning models with human or company reasoning. Right? I still think it's gonna be a multibillion dollar industry, you know, whether it's next year or the year after. But there is this kind of enterprise layer growth, and I think the strongest public evidence is kind of the explosion of agent frameworks and observability or tracing, which is exactly what these wrappers need to function reliably in business settings. So maybe didn't hit on this one kind of close, but I don't think it's going anywhere. I think as reasoning models become the default, I think that companies are gonna eventually, smart companies are going to eventually realize, wait, We need to feed these reasoning models with our company reasoning. Number eight, virtual machines will become all the rage again.

Jordan Wilson [00:15:16]:
Alright. So as a reminder, this is before there was even a a chat GPT operator. Poor chat g p t operator. It was announced, and it's already it's already dead. Right? But my prediction was that because agents need a computer, that virtual machines or virtual desktops become trendy or at least they become a thing again, and that's definitely happened. Right? So even like Chad GPT's agent, it uses a virtual machine. So maybe companies aren't out there, you you know, putting out virtual machines, but it's actually gonna start happening and it has. There's a recent announcement.

Jordan Wilson [00:15:52]:
Right? Like, some of these predictions, there's, like, big news that came in November and December, and I was like, yeah. Right? Like, as I've been planning the follow-up show, I'm like, yeah. I knew I knew something was coming. Right? So, obviously, Google's Project Mariner, that's positioned as a browser based agent. But the big one actually came in November when Microsoft announced Windows three sixty five for agents at Ignite. And, essentially, they have now cloud PCs that are specifically designed for AI agents. Right? That's it. So, yeah, it happened.

Jordan Wilson [00:16:26]:
It took a while, but, yeah, this was true. Kinda seemed like a, like like a niche, prediction because at the time, there was no general use agents from the big companies. Right? You had Copilot, you had Copilot Studio, and that was it. That was it. And Copilot Studio was kind of brand new, but no one was talking about virtual machines, and now they're definitely a thing. Number seven. So AI becomes overly, political. I said, AI is gonna become deeply entangled in politics and policy conflict.

Jordan Wilson [00:17:00]:
Yeah. That's happened. So, you know, and I'm looking at this. I'm based in The US. Our biggest, listenership is here in The US. So looking at this mainly through The US's point of view. But in December, so, this week, president Trump signed an executive order blocking states from regulating AI, framing deregulation of AI as a critical critical weapon, in the AI race against China. So there is literally no bigger way to, for AI to become political than at the federal level saying, hey, states.

Jordan Wilson [00:17:35]:
You can't make laws on this, which is very, not in the same vein as mostly everything else that, president Trump and the Republican party. Right? They like to hand as much power over to the states. You you know, that's kind of been one of their big talking points, and this and AI is completely the opposite. Right? So if nothing else, you know, this went against the grain of of what president Trump has traditionally done, in his second term in office so far. So yeah. Definitely, AI has become, I'd say, way too political because even outside of that, even outside of the geopolitical, even just, in the, domestically, it's become political. Right? So in, July 2025, the Trump administration signed America's AI action plan. And in there, there was an executive order, preventing the federal government from procuring AI models that include, quote, unquote, ideological biases or social agendas or woke AI.

Jordan Wilson [00:18:32]:
Right? So, essentially, the federal government said, yeah. We're not gonna do business with anything that we don't like. Right? Anything that we say is woke. So y'all, I know, like, I, I already know just from what I said there, I'm gonna get, you know, people who are Democrats, who are angry at what I said, people who are Republicans angry at what I said. I know I'm gonna get hate mail just from both sides on this. Right. I'm just speaking facts. Right? This is literally what happened.

Jordan Wilson [00:18:58]:
Alright. So just FYI. Also, it doesn't get more political in AI than the, the cozying up of big tech leaders, with the federal government or, president Trump. So as an example, for the president's inauguration campaign, everyone donated, like, a million dollars. Right? So Meta, Mark Zuckerberg, personally, this is according to reports. NVIDIA, Google, Amazon, Apple, Tim Cook, Microsoft, OpenAI, with, with Sam Altman, right, Broadcom, Adobe, Perplexity. Right? Essentially, everyone, just about any big tech company donated, to, president Trump's inauguration. Not not really normal.

Jordan Wilson [00:19:44]:
So, yeah, another sign that, yeah, AI became extremely political in 2025. Alright. Prediction number six was global AI regulations tightened, just not in The US. Right? And, yeah, that happened. And it it honestly impacted how users worldwide were able to interact or not interact with their favorite large language model of choice. Right? Obviously, in the EU, very restrictive with the EU AI act, and a lot of features like OpenAI's long term memory upgrade, got really delayed, in The EU and The UK, because of this strict AI regulations. And there's rules for general purpose AI models that became applicable in The EU starting in August with penalties that can reach up to £35,000,000 or 7% of global turnover. Also, Singapore launched its global AI assurance pilot in February, and Italy also had their national AI law that took effect in October.

Jordan Wilson [00:20:51]:
So you have a lot of big nations or groups of nations that have different laws. Right? I'm not gonna spend time talking because it would take hours, but a lot of just very strict regulations on how AI can be used. And in many cases, certain features or certain AI tools just can't be used in certain countries, right, which obviously, you know, stifles their innovation. And so, yeah, the EU act went into full enforcement in 2025. Meanwhile, The US did absolutely nothing for the most part to block, innovation. There's no real, regulation in The US. Alright. Number five, narrow AI agents or narrow AGI also achieved.

Jordan Wilson [00:21:36]:
But, anyways, my prediction when it came to agents that there wasn't gonna be a runaway general purpose agent. The general purpose agents were not gonna be good, but narrow agents would absolutely dominate. Right? And that prediction came true because you can there's no one great agent out there. Look at the big players. Right? Microsoft Copilot, I'd say is probably the closest to having a general agent, but even those are built on a company's certain use cases. Right? So they're narrow. They're not general for the most part. Right? OpenAI's, agent mode, not really good.

Jordan Wilson [00:22:19]:
Right? I'll be honest. Sorry. Friends at OpenAI who are listening, it's not very good. It's slow. It's clunky. For certain tasks, it's okay. It's manageable, but general agents aren't there yet. They will be.

Jordan Wilson [00:22:32]:
I think especially with some of the recent model updates that just came out, like, looks at watch in the last ten days or the last couple of weeks, specifically, Gemini three, Gemini three Flash, and GPT five two, and then, Claude, Opus four 5. Right? On the agent side, that really changes what's possible, but we haven't really realized that yet. But the agents that have been making the biggest splashes, I'd say, are just narrow agents. Agents built around a certain vertical, a certain task. So as an example, Salesforce's agent force two o, you know, handles end to end CRM workflows. Oracle deployed 50 role based AI agents in fusion cloud apps, GitHub Copilot, coding agents, obviously, Anthropics, Claude Code, right? A lot of coding agents, right, with narrow use cases. And Menlo Ventures report show that there's a $7,300,000,000 enterprise investment in departmental AI, which is essentially narrow AI use cases. Alright.

Jordan Wilson [00:23:35]:
Number four, prediction. LLM memory becomes a major focus. This prediction obviously came true. So at the time, right, and I I didn't wanna harp on this on every single one. I thought it was funny to talk about the the vibe coding one because it feels like vibe coding has been around for, like, five years, but it wasn't even talked about until February, which is after this. But think back to January. Aside from ChatTBT had a very early version of memory for some users, but aside from that, LLM memory wasn't really a thing. It wasn't.

Jordan Wilson [00:24:11]:
People talked about it. Some people talked about it. Right? I wasn't the only one, but it obviously came true. So OpenAI, had a handful of big updates to their memory and chat history. Google Gemini rolled out a memory, upgrade or, you know, memory, to its users. Claude just went right there. The last of kind of the big three, that finally rolled out chat memory. So, yeah, now the three big front end LLMs, in chat GBT, Gemini, and Claude finally have memory.

Jordan Wilson [00:24:42]:
Right? Like I said, it was gonna be a major focus, and it obviously happened. Also, context caching on the developer side has become a standard API feature, whereas before, wasn't really a thing. Right? Which is a kind of a, you know, context caching is a form of memory. It's a it's a form of, you know, not having to pay over and over for things that have already been processed. It's a it's a more of a technical, memory mechanism. And then also, in with Gemini three, they introduced an, architecture that allows for agents to recall user preferences across months of sessions. Right? So, memory is taking shape and form in many different ways. But, I mean, if you just look at the, the progress, on the front end, it's been one of the biggest things is has been memory.

Jordan Wilson [00:25:36]:
It's been using, more of the context window, in making more of it, if that makes sense. Right? In that context window right? Because outside of the context window, it doesn't really matter much. But in that context window and then being able to, layer in memory, layer in. Right? Also, Gemini has a level of personalization, that can work with your, you you know, your search history. Right? And even just as we get connectors, right, and and being able to work with that data, it's it's huge. Right? So I think that the context plus connectors plus memory, was definitely the biggest leap forward when it came to front end, large language models in 2025. Alright. Our top three last three predictions, here we go.

Jordan Wilson [00:26:24]:
Number three, large language models become small language models. Well, what I meant was my prediction there was that smaller efficient models would become more prominent and capable, than January big models. Right? That's what I was alluding to. Essentially, I said, hey. In a year, there's gonna be small models that are way better than big models, and people are gonna be using small models a lot more than they're using them now. Because, you know, back in January, people weren't really talking or using small models. And in some cases, there weren't a lot of them available. Right? OpenAI had their mini series, but now, I mean, Google so let's get to the reality.

Jordan Wilson [00:27:03]:
The reality is, yes, it obviously happened. There's been a new category, of models. Right? Like, as an example, Gemini Gemini now has their flash, but then they have their flash light, which is an even smaller version of their, kind of, behemoth model, you know, the Gemini three pro. But it's not just that. It's not just the, you know, the the the GPT five mini, the, you know, Gemini three, Flash. It's not even those. Some of the models like GPT OSS. So OpenAI's open source model, a 20,000,000,000 parameter model, more powerful than GPT four.

Jordan Wilson [00:27:46]:
Oh, right. So think about a year ago, give or take GPT four. Oh, was the world's best model right now. You have an open source 20,000,000,000 parameter model that when you look at third party analytics, right, like artificial analysis, it's better. Okay? I'm gonna do the math. I'm gonna do the math for for for those of us that aren't great at math. Parameters. Right? Models are judged in their size by parameters.

Jordan Wilson [00:28:20]:
For the most part, with closed proprietary models, you don't know how many parameters they are. With open source models, you do. That's why with GPT OSS 20 b, we know it's a 20,000,000,000 parameter model. You can think of it like kinda like a hard drive. Right? According to most reports, the earlier versions of GPT four, were 2,000,000,000,000 parameters. So GPT OSS 20 b, OpenAI's open source model, was about 1% of the size of the model that was the most powerful model in the world about one year prior, And it was better. So yeah, the models got smaller. And then, I mean, there's a whole, I mean, the phi series from Microsoft really good.

Jordan Wilson [00:29:05]:
The Gemma series, Gemma three, from Google, which we might see more of here in the next couple of days. So keep keep your eye out. Alright. Small language models, I've always been bullish on them. I'm even more bullish today than I have ever been, because that's the direction. We've seen even domain specific models, from OpenAI with their science models, anthropic, with their financial models. Right? You we don't talk about them a lot unless you're one of the few companies that use them. But I've said this all along.

Jordan Wilson [00:29:35]:
The future of large language models is using many small domain specific models. Alright? Number two, speaking of models, mixture of models becomes a thing. Alright? A year ago, no one's talking about mixture of models. We talk about mixture of experts, which is which is a little different, but my prediction was, that there would be systems that would orchestrate multiple models in parallel as a bundle. So let me kind of explain the difference between mixture of models and mixture of experts. And, also, I even think at the time last year when I was building, my prediction show, I don't even know if mixture of models was talked about. I might have made up the, the mechanism, but, you know, now people are are talking about it. There's there's other other words for it.

Jordan Wilson [00:30:20]:
Right? But mixture of experts or MOE, which a lot of people know, and now there are people talking about mixture of models. You know, there's different names for it. But mixture of experts is essentially a single sparse model that uses a gating mechanism to activate only a specific subset of parameters or experts for each unit. Right? So these big parameters, when you have a mixture of experts model, it only activates, certain parameters. Right? And that's how it can, you you know, it can use a mixture of them faster and better. Right? Whereas a mixture of models is more of an ensemble of independent fully active models that all process the same input, aggregating their distinct outputs to improve overall accurate, accuracy. So this this is another one that, man, I was I was playing in the show, you know, last week, and I'm like, there's not really there hasn't really been one big story, you know, that someone went with this mixture of models, and then Zoom did. For all companies, Zoom.

Jordan Wilson [00:31:24]:
Right? So, yeah, the the video calling company. Right. So Zoom, put together a, essentially, mixture of models, approach and got the highest score on humanity's last exam. But critics were not really happy, and a lot of people are like, I don't know if this counts because all they did. Right? So humanity's last exam, I'm I'm I'm not gonna go into it, because it'll take a long time to explain, but it's essentially a very hard, AI test. Right? We'll just say that, with, questions that make models really think and reason, the answers aren't in their training data. Right? But usually, you give one model. So what Zoom did is they created essentially a mixture of models.

Jordan Wilson [00:32:15]:
So they didn't build a model. They didn't fine tune an open source model. They didn't, you know, they just used off the shelf models, kind of duct taped them together, and then they beat the best score. So, yeah, they essentially use a mixture of models, but there's a lot of, there's actually a lot of movement in this space. You you know, similar to model routing, but a little different. But Google's interactions API from this month, actually, it now dynamically routes queries, and shares the information between Flash Pro thinking in real time. So they're also using it's not just model routing. Right? Because that's usually, you know, a a a router sends your prompt or your query to one.

Jordan Wilson [00:32:56]:
This is a little different. It shares it between all three. So, you know, even Google's interactions API for developers is using kind of this mixture of models. Also some other, you know, signals. The IDC reported in November that 70% of top AI driven enterprises will use multi, multi model architectures by 2028. And, another good example, ARIA, or sorry, Aria. They're a recent sponsor on this show. They're a key player.

Jordan Wilson [00:33:27]:
Essentially, they have a mixture of models, architecture. They I mean, they recently raised a $100,000,000. They're a newer company. You you know, they had a $100,000,000, round in September, I believe. So it is a space that is growing, you know, essentially not having to use one model. I think this is, in the future, it is going to become, much more commonplace, than you would think. Right? Because I think the smartest companies are having a modular approach when it comes to their AI strategy. Right? Because sometimes and I think we saw this with GPT five.

Jordan Wilson [00:34:03]:
Right? Going from GPT four o to GPT five. For companies that maybe didn't have a modular approach, they had a singular approach. If they were just using, you know, GPT four o, whether on the front end or the back end, and maybe all of their, you know, processes with a big switch like that. Maybe I mean, maybe some things work better. You hope things work better, but a lot of things are gonna work worse. So that's why I think having a mixture of models approach is really smart. That's what smartest companies are doing. Alright.

Jordan Wilson [00:34:33]:
And then last but not least, the big prediction. I said in January, AGI is gonna be achieved, but no one notices. Alright? So my prediction was that AGI, artificial general intelligence, would arrive in 2025, but daily life doesn't feel any differently, and there's no bold proclamation of, you know, hey. AGI, has been achieved. You know, we're gonna throw up the AGI flag on the federal buildings, and, you you know, now the human's taken that. I think most people have said that AGI is just gonna happen and no one's really going to notice. So here's the reality of what happened, and I'm not gonna fact check myself on this one too hard. I'm gonna let you.

Jordan Wilson [00:35:21]:
Alright? So livestream audience, let me know. Say, AGI, yes. AGI, no. AGI, maybe. Right. Spotify people in the comments, let me know. I'm gonna read them. But here's some of the reality.

Jordan Wilson [00:35:33]:
So in 2025, large language models started surpassing top the smartest humans in the world on almost all elite thinking tests. Right? So any test that humans take and a single AI model takes, AI models on almost all beat the smartest humans in the world. Okay? But, I mean, AGI still doesn't have one agreed meaning, and even the optimistic voices note that today's AI systems can look brilliant on some tasks and be like, wow. This is definitely AGI level. But on other tasks, they can seem very dumb. Right? And let me be let me be very honest. A lot of that's a skill issue, FYI. But one of the biggest things, little dorkier unless you're a math geek.

Jordan Wilson [00:36:22]:
Right? It was the IMO gold medal performance that was, I think, maybe the standout AGI moment example that a lot of people pointed to in 2025. So what is, the IMO? That is the International Mathematical Olympiad. Essentially, it is one of the most difficult, math competitions in the world. And OpenAI and Google DeepMind for the first time ever won gold. They solved all the problems. And then you had, this week, you had Boris Power, the head of applied research at OpenAI. So a lot of people would conclude that Boris is probably one of the smartest people in the world, when it comes to AI and research. He said that winning the IMO gold was seen as an AGI level difficulty problem.

Jordan Wilson [00:37:10]:
Right? It happened. It passed. And no one was like, alright. Life changes. Right? I don't think that there ever will be a standard or agreed upon term of AGI. Right? But there's a lot of sides, and I wanna talk about two of them. One would be the, the AI IQ test. Alright? So this is from trackingai.org, a good, good resource.

Jordan Wilson [00:37:40]:
So, essentially, I have a, on my screen here, and if you ever wanna see the video version, it's always on our website at youreverydayai.com, but I'll describe it. Right? Essentially, tracking AI, they give offline IQ tests to large language models. Right? So the answers aren't in their training data, and then they do they do an offline and they do a MENSA Norway test. Right? So now it's at the point. Right? A year ago, the models had an average, you know, an average IQ about the, you know, the average person, has an IQ of a 100. You you know, I think, bright, you you know, there's different classifications. I guess a bright person is around one fifteen. Gifted is in the one twenties, and I think you get to a genius level IQ, I think between, like, one thirty five and one forty.

Jordan Wilson [00:38:30]:
So now you have, I mean, Gemini three Gemini three pro and not too far behind, GPT five pro, thinking, GPT five two thinking and GPT five two pro. They're at the about a 130 mark for these IQ tests. So they are near genius level. They're in the top 1.5% of humans. Right? Okay. So I I again, I mean, AGI is not an IQ test, but if you know what you're doing, an AI model is smarter than almost any human in the world. And that's just, you know, normally, those humans are smart and they're one field. Alright? And then I will throw out.

Jordan Wilson [00:39:14]:
Right? And I did a whole show on this. I don't know. Maybe a year ago. I looked at older definitions. I looked at definitions of of artificial general intelligence from, you know, 2000, 2005, 2010, 2015. And when you look at all those old definitions and where we're at today, if you were just to look at those old definitions, we've achieved every single old definition. Right? But as AI gets smarter, the goalposts keep moving. So I don't know if AGI has been achieved.

Jordan Wilson [00:39:42]:
Right? But at what point will it be achieved? I don't know. I guess when we have a definitive definition, for me, I would say, yeah. Probably. Right? And one reason why, I'm gonna show you here in a second. But, Sam Altman in 2018, in the OpenAI charter defined AGI as AGI is by which we mean the highly autonomous systems that outperform humans at most economically valuable work. Let me repeat that last part again. Outperform humans at most economically valuable work. So benchmarks have never really measured this.

Jordan Wilson [00:40:22]:
Right? And I think that AI companies may have gotten kind of smart at overfitting their models to score really well on the benchmarks. Right? That doesn't always help them when it comes to human preferences like LM arena. Right? But now there's a new, a new benchmark that's judged by humans. And I'll probably, do a dedicated show on this once because I think it's really interesting. So there's a, it this is an open AI. I did talk about this, couple days ago. So this is an open AI benchmark called GDP Vow. Right? So right now, their model, g b t five two, pro has the best score on this.

Jordan Wilson [00:41:02]:
But when they released it in September, it didn't. Right? So a lot of people are like, oh, yeah. Of course, OpenAI is gonna have, right, the best score on their own benchmark. Well, when they released GDP Vowel, they didn't. It was clawed four five. But GDP Vowel test models on real jobs with economic impact. Right? Exactly what, Sam Altman's definition. It's when autonomous systems can outperform humans at most economically valuable work.

Jordan Wilson [00:41:31]:
So GDP Bao test models on real job tasks from 44 jobs in nine big US economic sectors. So tasks use real work files like spreadsheets, slides, PDF images, and then they produce those actual products. Right? So previous versions, even of g p t five, weren't any good at producing, spreadsheets or, slides. Well, now g p five two, thinking in pro, they really are. So how this is? Essentially, a human who is an expert in the field judges this, and they compare the model's work to that of a human's without knowing who did what. Right? So, yeah, it's like, oh, hand in a spreadsheet. Smart human, expert human, hands in a spreadsheet. Model, hands in a spreadsheet.

Jordan Wilson [00:42:23]:
And then an expert who I think they said on average had about fourteen years of experience. So a domain expert judged the two ones, and they didn't know who did what. Right? And then the scores are averaged to get the model's overall win rate. And then grading, the human grader bases it on usefulness, format, and quality. And so here's an example. Right? This is an actual example from GDP valve. So in a retail task, so this is in retail, they're given a past sales and promo calendar in Excel, and they have to produce a forecast and a one slide summary. So then the graders compare the models, spreadsheet and PowerPoint slide to the humans for math, readability, assumptions, and presentations.

Jordan Wilson [00:43:07]:
So what happened? And, well, let me first read, OpenAI's kind of definition here. So they said GPT five two thinking is the best model yet for real world professional use on GDP val. And if in eval measuring well specified knowledge work tasks across 44 occupations, g b t five two thinking sets a new state of the art score and is our first model that performs at or above human expert level. Specifically, GPT five two thinking beats or ties industry professionals on 70.9% of comparisons. Alright? And then if you look at g p t five two pro, the wind tie rate is actually 74.1%. Okay? Let me repeat that. On economically valuable work, judged by experts, one model versus smart human. They have to produce spreadsheets, PowerPoints, writing, etcetera.

Jordan Wilson [00:44:17]:
The AI did better three fourths of the time on economically valuable work. Wait. What was that definition That Sam Altman, the CEO of that same company said AGI, by which we mean highly autonomous systems, yes, that outperform humans at most economically valuable work. 44 real jobs across nine major US sectors in the AI model wins or ties three fourths of the time. I don't know. To me, my prediction that AI AGI is achieved, but no one really notices, kinda seems like it came true. But guess what? As the models get better, we're just gonna keep kicking the goalpost. But don't worry.

Jordan Wilson [00:45:04]:
As the models get better, we're still gonna be here every single day keeping you up to date. Alright. That is a wrap on our 2025 AI road map rewind. So volume two, like I said, if you missed volume one, make sure to go check it out. That's episode six seventy four. I hope this was helpful. And, yes, we are gonna be, announcing dates soon, for our, prediction series for 2026. And we got something else special kicking off here real soon.

Jordan Wilson [00:45:37]:
So if one of your personal goals or if one of your company's goals is to double down and go all in on generative AI in large language models, trust me, we have something special coming. So if you haven't already, please make sure. Number one, you're gonna wanna go subscribe to the podcast if you haven't already. So if you're listening on Apple or Spotify, thank you for that. Number two, you're gonna wanna go to our website. Trust me. Youreverydayai.com. Go sign up for the free daily newsletter.

Jordan Wilson [00:46:04]:
We're gonna be re recapping today's show and a whole lot more. Thank you for tuning in to our recap, our rewind series 2025. Even though it's not over, it's been a great one. Thank you for your support. Thank you for tuning in. Hope to see you back tomorrow and every day for more everyday AI. Thanks, all.


Gain Extra Insights With Our Newsletter

Sign up for our newsletter to get more in-depth content on AI