#229: Q3 Trends Briefing - The Pope's AI Encyclical, AI Agents Hack Hugging Face, Fable 5 vs. Washington, and the Battle Over Open Weights

0:00So agents have this amazing potential when they become really reliable and we've figured out how to govern their autonomy because most enterprises have no idea how to govern the autonomy of these things. Like if we're going to turn them loose and let them do all these things, how do we actually know they're doing what we want them to do? Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Reitzer. I'm the founder and CEO of SmarterX and Marketing AI Institute, and I'm your host. Each week, I'm joined by my co-host and SmarterX Chief Content Officer, Mike Kaput, as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career.

0:49Join us as we accelerate AI literacy for all.

0:55Welcome to episode 229 of the Artificial Intelligence Show. Today's episode is a special episode of the podcast where we're doing something a little different than usual. Today we are going through one of our quarterly trends briefings. This is a live event we run each quarter for our AI Academy Mastery members. What we do in these briefings is we count down the top 10 trends that matter most in AI this quarter. based on what we've been tracking and covering over the last few months on the podcast. So these trends briefings started as exclusive only for members of our AI Academy. They quickly became very, very popular.

1:34So we're bringing them to the podcast as a regular part of the feed. Now, we are recording this live right now with an audience of our AI Academy Mastery members. They're seeing it first. So if you're a member listening on the podcast, you can jump back into AI Academy to hear the exclusive Q &A that followed this trends briefing. So what we're going to do here is I'm going to tee up each trend, kind of similar to how we tackle topics on the podcast normally each week. Then we're going to get Paul's take on the bigger picture implications of each one. So you can basically get fully briefed on what's happening this quarter in AI.

2:08Spoiler, it's a lot. One final note, if you are listening to this episode, you are not yet an AI Academy member. Go to academy.smarterx.ai to see all the incredible benefits that individuals and teams get with their memberships, including live access to these briefings days or week in advance and plenty of exclusive private Q &A afterwards, as well as tons of other courses and benefits as part of that membership. All right, Paul, let's get into a top 10 AI trends this quarter. We count down in kind of order of importance, though all of these alone could be the biggest news item in another industry for a year, I think.

2:47But we're going to start at number 10. and we are going to start with beginning our countdown with a little bit of a question here, which is how do you know AI has a public opinion problem? And the answer to that question is when the Pope himself decides to write 43 ,000 words about it. So back in May, Pope Leo XIV issued his first encyclical called Magnifica Humanitas. It runs more than 43 ,000 words, And it is about, basically, all about AI. He wrote that technology is never neutral. AI tends to amplify the power of those who already possess economic resources, expertise, and access to data.

3:26He called on the world to disarm AI, which he defined as freeing the technology from monopolistic control. And that message went out to more than 1.3 billion Catholics. In fact, Anthropic co-founder Chris Ola spoke at the Vatican unveiling and told the room the industry needs outside critics who are willing to say hard things. Now, be careful what you wish for, Chris, because we have been tracking, in addition to the Pope's encyclical, huge backlash to AI all quarter. We saw graduation crowds booing former Google CEO Eric Schmidt at the University of Arizona when he brought up AI. Gallup has found only 18 % of young people ages 14 to 29 feel hopeful about it.

4:10And of course, data centers. Gallup found 71 % of Americans oppose a data center in their local area. These data centers are thus less popular than nuclear power plants. And we covered more recently that New York Governor Kathy Hochul announced a one-year moratorium on new data centers of 50 megawatts or more, the first state to do that. In the past few weeks, we had a group called Humans First, co-founded by a former Tea Party leader, coordinating 142 protests across 42 states against data centers. Paul, in one quarter, just three punts, we went from a papal encyclical to like widespread booze at graduation ceremonies to literally the first state moratorium on data centers.

4:54Can the industry claw back any trust or like, is this just kind of tip of the iceberg? How bad is this? I don't know. It just keeps evolving like next week. Well, see, here we go. I'm already like thinking in the future as we record this episode on Friday. Right. There is pacing the frontier like this new letter that's come out with AI Frontier Lab employees where they're basically asking the government to help them slow themselves down. And so like what the Pope was sort of asking for in some ways back in May, we're now actually seeing elements of the industry that themselves are saying, help us from ourselves.

5:35Like the competitive nature of what we're doing is like we're going to just keep accelerating and that's going to at some point become a problem. There's a couple other elements we'll talk about today in terms of, you know, the uncertainty around how these agents work. And as we make them more recursively self-improving, they might find ways to escape containment. And like, it's getting really bizarre. And so the public backlash is in part, you know, based on real concerns about data centers in their community. Like, I don't know too many people that if you ask them, would you like a data center in your backyard?

6:08They would say yes. Now, that doesn't mean that data centers are inherently all bad. Like, there's a lot of good that's going to come from the build out of data centers. There's a lot of scientific advancements, medical discoveries, like good, positive things that can lead to this future of abundance. But right now, the negative messaging seems to be what's winning in society. There's a lot more concerns and fears than there are optimistic outlooks for this. And in part, I think it just still has to do with just an overall lack of understanding of what's possible. You see like Mark Zuckerberg just did an editorial in Wall Street Journal where he's trying to share this optimistic future of abundance.

6:49And I think we're going to see a lot more of that. You're going to see a lot of these AI lab leaders really pushing on what's possible to try and change the narrative a little bit because they're losing the narrative right now. Okay, so our number nine, as we're counting down trends here, is this kind of strange puzzle we're seeing right now in the economy, which is AI adoption keeps climbing, and we have not seen any issues with unemployment. We have not seen this kind of job loss, at least according to some people, that you might have expected. So Anthropics' own head of economics, Peter McCrory, published an essay recently asking, why hasn't AI increased unemployment?

7:28He pulled from 18 months of anthropic research, noted that unemployment still sat at 4.2 % in June, which is a very, very low rate. And about 20 % of U.S. firms now use AI in at least one business function. Yet, we also saw labor productivity grow 2 % a year since early 2022, up from 1.6 % before the pandemic. So he concludes that AI is just augmenting workers, not replacing them. I mean, he actually says he does not expect unemployment to be noticeably higher a year from now. Now, that wasn't the story, though, at many other firms over this quarter. So, you know, Stanford research showed a 16 % employment drop for workers 22 to 25 in AI-exposed jobs.

8:14Bloomberg reported tech and finance shedding around 28 ,000 jobs a month. Then, of course, this fight got political. David Sachs of the White House Advisor shared an Apollo chief economist report, headlined zero evidence of AI-related job losses, yet there were also layoffs at plenty of companies that said they were laying off due to AI. So Coinbase's Brian Armstrong cut 14 % of his workforce, about 700 jobs, and said he is rebuilding Coinbase as an intelligence with humans around the edge. Cloudflare CEO Matthew Prince laid off 20 % while saying revenue actually grew 30%. He said AI is now filling certain types of roles to make that possible.

8:55Our own State of AI for Business report found 71 % of professionals believe AI will eliminate more jobs than it creates over the next three years. That number was like 40 % of people saying that only a couple years ago. So, Paul, we titled this AI Jobs Whiplash because we're literally going back and forth between these wildly opposing views. I'm curious how you look at this, like what is actually going on here? Yeah, I don't know. I mean, obviously, this is a topic I talk a lot about on the podcast. There's elements of, you know, how I kind of feel about jobs and the economic data, sort of similar to how I felt about investing in AI.

9:37You know, for me personally, starting back in 2014 and, you know, creating an AI Institute in 2016 and an AI conference in 2019, all of that seemed crazy. Like it just didn't make any sense like why you would do those things. But when you kind of see out ahead and you start to like understand where stuff might go, even though the experts tell you the opposite sometimes. And so I feel like right now what's happening is the people who want to be right, that nothing's going to happen to jobs, that it's all going to be OK, like David Sachs. They're like the favorite phrase they use is this narrative violation.

10:14So like any time a report comes out and says, oh, jobs are actually great. It's like, oh, narrative violation. How dare someone say the opposite of this? I feel like they're taking an early victory lap that's going to not end really well for them. because what this data says and what executives say to me in private do not match. So like there's what people will publicly say, but then there's what they're privately planning to do. And I just think that right now we are still in this very early phase where the economic data may continue to show things are OK, like they're pretty stable. But when you look at the adoption within enterprises, it's still so early.

10:56Like most companies have yet to actually realize the true possibilities of AI from an efficiency and productivity standpoint. And so my feeling is once the market starts to catch up with that, once they truly start to adopt AI and scale its use, then inherently the economic data will start to tell a different story. So as I always say on the podcast, I would love to be wrong here. I want the economic data to continue to show that opportunities are going to keep growing and jobs are going to grow and entry level work isn't going to be threatened and kids coming out of college are going to find jobs plentiful.

11:33I want all that to be true. I just want people to be prepared for the possibility that it's not. So we're not just sitting around a year or two from now thinking we should have prepared. Wise words. Okay, so number eight on our list of trends. Every company is doing some form of AI, but many of them are really struggling to figure out like how far they've progressed. And so this trend is about a framework that is starting to come together to hopefully fix that. So this month, Paul, you unveiled an eight pillar framework for organizational AI maturity that you've been working on to define the future of AI transformation.

12:11So we talked about this briefly on the podcast, wanted to go a little deeper into it. There are eight components, vision, strategy, data, technology, governance, literacy, people, and performance. And I'd love if you could maybe just tell us a bit about the pillars of business AI transformation, the system behind it. Why have you started creating something like this and why is it needed? I would say at the highest level, the premise here is for the last few years, AI transformation in organizations has largely been like a technology driven thing where enterprises acquire licenses to co-pilot or chat GPT or Claude or Gemini and they give them to their people, their knowledge workers across different departments, and they assume they're going to figure out how to get value out of them.

12:56That was never going to work. And any change management process, just giving someone technology doesn't solve anything. So then it became this, you know, thing where we would say, okay, well, let's just monitor usage daily or weekly active usage. How many GPTs are they building? It became this like, you know, I guess like metrics driven thing, but you told this great story on the transformation spotlight with Ty from Good Karma where he said, we tried that, but then we moved to like this outcome base. Like it actually became more about what are you changing? What is the workflow improvement? What are the business outcomes?

13:28And so I feel like we went through this like three year phase where scaling AI basically meant just giving everybody the tools and hoping they figured it out. We see this with AI Academy where we would have enterprises come in and buy dozens or hundreds of licenses for their team members and then provide them to them. And again, same premise that we assume we provide the education and people will have the intrinsic motivation to become AI forward themselves. That also does not work. So what we've seen is a need to think about this more holistically as a transformation system, not just providing technology or providing courses.

14:01You actually have to enable people to take a look at where are they now? Where do they want to go? What does success look like? How do you personalize the use of the technology and the training for these people so that you can achieve whatever transformation means within your organization? And so we think about business transformation, you know, more holistically. And then we also think about it as a collection of individual transformations. So at the highest level, you have to have this vision as an organization. The main thing I always tell people is it's great to democratize and empower individuals and department leaders to drive transformation themselves and as departments.

14:43But until the CEO states a clear vision for what the future of work looks like within an organization and then operationalizes that vision with a change management plan, it's never going to work. And so, again, going back to the good karma brands one, that's where I love that story about Craig, the CEO, having that vision and being in the room for all of this innovation that's happening. Like, he's driving it. And that, to me, is like the prototypical way that this has to happen. It has to be a top three priority, if not the top priority, at the CEO level, or else you run the risk of a competitor out innovating you when it comes to AI.

15:23And just to confirm, Paul, this is still in the early phases and will be rolling out relatively soon, right? AI Academy, this is not like a public framework yet except in your newsletter, right? Correct. I published the pillars themselves. I didn't publish anything else underneath the pillars and kind of the different dimensions of it. We have an internal beta of the business AI transformation system. We will have an internal beta of the personal AI transformation system in the next probably two weeks. We are going to open up the beta to some of our AI Academy business account members for internal testing probably in the next 30 to 45 days.

16:08And then the goal would be to actually roll out. These are just two components of the much broader system. Yeah. But to be able to announce kind of the full system early this fall, you know, before Macon would be the vision for it. Gotcha. All right. So next up, big lawsuit news. Apple has filed a lawsuit against OpenAI. They claim that OpenAI is stealing trade secrets. Apple alleges that as OpenAI recruited its employees, I believe about 400 or so work at OpenAI now, it encouraged them to share confidential information about unreleased products, components, and materials. So this filing names Tang Tan, who's OpenAI's chief hardware officer, who used to be Apple's VP of product design.

16:51He led work on the iPhone, Apple Watch, and AirPods. And Apple alleges Tang has methodically used Apple's confidential information to benefit OpenAI and even directed job candidates still working at Apple to bring actual parts to their interviews. A former iPhone hardware engineer, Chang Liu, was also named in this suit. They say that he surreptitiously accessed and downloaded dozens of confidential hardware files. Other departing employees allegedly emailed confidential information to personal accounts. Apple says it warned OpenAI about misappropriated information earlier in the year, got no response.

17:27It wants a jury trial, destruction of proprietary materials, and a redesign of products built with them. OpenAI denies all of this and said it has no interest in other companies' trade secrets. So, Paul, I guess I got to ask, like, how big a deal is this? Because, like, frankly, OpenAI is getting sued all the time by different parties. But this one seems like it could be a little different. Yeah, you don't mess with Apple. I don't know how many attorneys Apple has on staff, but it's a lot. and they don't mess around when they bring cases against people. So, you know, it's to be determined. I don't know what Apple really wants out of this litigation.

18:09I'm not sure what exactly they're looking for. It's possible that they just want to end their hardware ambitions completely. I don't know. They obviously have a relationship where Siri was in part powered by ChatGPT. I don't think that that relationship is going to continue based on what's going on here. Yeah, I don't know. We'll see. The most interesting part to me is I would assume, you know, when OpenAI decides to IPO, this is a big black cloud hanging over them that investors are going to want much more clarity about what exactly is going on here. because if the hardware business for OpenAI is part of their future valuation, this is a really big question mark.

18:55You do not want to have Apple coming after you for IP when you're trying to get that built into your valuation. So I think we'll hear a lot more about this in the coming months. I do think it's at least playing a part in the timing of OpenAI's IPO. Okay, so number six is all about the release of GPT 5.6 and chat GPT work. So first up, OpenAI has launched GPT 5.6. There are three versions of this model. Sol, S-O-L, is the flagship model, super powerful version. Terra is the kind of everyday in the middle model. And Luna is the fast, cheap model. And Sol so far is getting pretty rave reviews in how powerful it is, and it's dramatically less costly than Claude Fable 5.

19:44It's built for very complex work, especially coding, research, science, and especially agentic computer use. Speaking of which, this kind of rolled out roughly at the same time as ChatGPT works. So if you've been in ChatGPT at all recently, you'll notice there are now two modes in the tool, in the web version, in the app called Chat and Work. And the mode on the right when you click Work is ChatGPT Work, which OpenAI says is an agent that takes an outcome, gathers information across your apps, and stays with a complex project for some duration of time. ChatGPT Work can connect to things like Slack, Teams, Drive.

20:22It hands back finished deliverables. in the desktop app, it can use your computer, which includes files, browser, a whole machine. And basically, this is, Paul, like a shift from chat to computer control to agents. I mean, this seemed like it to me, it happened really fast in terms of businesses were all like, okay, how do I prompt better? How do I use chat better? And then suddenly, wait a second, there's an agent that can connect all my files in my web app. If I have this thing on my computer, I can suddenly have my computer, a computer use agent at my employees fingertips. Like, how are you thinking about that?

20:58Where do you even start as a leader? It's a big competition. I mean, Anthropic got there first with co-work, I think, at least beat them by a couple of weeks or a month or so. They were kind of ahead of them a little bit. But yeah, it's the idea of being able to like a lot of the more advanced, more technical users, like I know, Mike, you were messing around with coding stuff on your desktop. So like people were messing around with Codex and with Anthropics coding abilities, and they want to bring that technical capability to the non-technical audience by basically injecting it into the user experience where you don't know that that's what you're doing.

21:34Yeah. And so it makes a ton of sense that they would want to bring that to life. My early experience with work has been very positive. Like it does. It's great. I'm using the browser. You know, I'm within the browser, not the desktop app itself. And 5.6, I've found to be a very powerful model. I've been using Sol on, again, as I said in the podcast, most of my stuff I do is just high-level strategic thinking and high-level strategic planning. And so that's a lot of times where I'm assessing these models is how good it is at that. And I was working on a project this past week that was, it was extremely helpful.

22:10and I will say like I do use um uh very commonly uh Claude and ChatGPT as critics of each other so I pushed on 5.6 soul for some stuff and then I actually took the output put it into Claude opus 5 and was like hey what do you think about this idea I mean where are we going with this so yeah I mean it just continues to change and the challenge here for business leaders is these things are evolving so fast whatever training we provide to people today 30 days from now the story may be different and it's one of the things i mean we had like three meetings this past week internally on ai academy and our plans for like more dynamic learning where we're trying to continually evolve how we're creating content to meet the fact that the market keeps changing so fast and that we want to keep the freshest content especially in relation to these platforms at the the forefront.

23:05One other quick note, Sam Altman tweeted on July 30th that they have cut the price on all the 5.6 models. So there's a pricing war emerging between Anthropic and eventually Google will get back in the game with their new models, you know, but certainly Anthropic and OpenAI at the moment and part of it's being driven by, you know, the other models that are coming out, the open weight models. Yeah, and I'd highly encourage people, if you have an array, take advantage like gpt 5.6 soul especially when cranked up to high or extra high or higher is like an incredible model and with the pricing drop like crank it up see what you can do with it it's really cool and i will say one other quick note like i was again i was with some executives um recently uh and and you know i was getting asked some more i would say like entry level questions like people are starting to kind of explore how they can be using these things themselves and I always get the question, well, which model should I use?

24:02Like, what should I be doing? And what I always tell people is, listen, you really can't go wrong. Like if you just get really good at ChatGPT or you just get good at Claude or again, Gemini is a really good model. Like it's fallen behind and hopefully it'll catch up in the next 30 days. But like, if you just get really good at one of them and like block everything else out, you're gonna make major changes. You're gonna see transformation personally just by getting really good at a single platform, regardless of which one it is. Yeah, I couldn't agree more. So the reason we're kind of talking about this is because this is a bigger trend at work, not just the ChatGPT work release, but this quarter is really where we started to see agents transforming how places work.

24:45So OpenAI actually came out with their own research this past quarter, co-authored with economists from Columbia, Wharton, and Duke, and basically just looked at how Codex is being used in their business. And interestingly, this is the thing that stood out to me, Paul, is that OpenAI said that Codex, which again, they're like agentic platform, it's a coding agent, but really it's a general purpose agent you can use for knowledge work. Codex, not chat GPT, drives more than 99 % of employee work, not just developers. And they basically explicitly say we're moving from chat to agents as the way we use AI.

25:19And it's not just them. Microsoft had a work trend index report that was built on trillions of Microsoft 365 signals and a survey of 20 ,000 knowledge workers in 10 countries. They found that active agent use is up 15 times year over year and 18 times inside large enterprises. Now, the reason we're talking about all this is often we'll cite Bach's CEO, Aaron Levy, who has tons of great takes on enterprise AI. He basically was like, look, agents are forcing an operating model problem on companies. Companies are built in silos. Agents work best across silos. Data is fragmented and the talent to deploy and I would actually argue just even understand agents and their implications is a bit scarce.

26:01So, Paul, like, yeah, it's a rhetorical question. Like, are companies ready? I feel like the answer is no. Like, I just don't see people really ready for the implications here. So if you joined us late, we're recording this with a live audience of our AI Academy members. And then it's being re-aired as a podcast episode. And I'm just glancing at the chat. And Heather just put in that, you know, they hid the work and codecs from everyone outside of marketing because too many things to figure out from a security and training perspective. And I think that's a it's a great point. This is the thing.

26:33So agents have this amazing potential when they become really reliable and we've figured out how to govern their autonomy, because most enterprises have no idea how to govern the autonomy of these things. Like if we're going to turn them loose and let them do all these things, how do we actually know they're doing what we want them to do? And as we will talk about one of the final trends that we're going to go through today, the labs themselves can't seem to keep track of their agents. And we've had a couple of high profile instances in the last seven days where agents went rogue. And so my guess is there are a lot of CIOs and internal IT departments who are scrambling right now to try and see what is going on with our agents.

27:22Like we didn't maybe know that they could do stuff like this. Is there anything going on we should be monitoring? So I think agents are transforming work. I think there's going to be some organizations that race ahead and take a lot of risks with agents and they might get disproportionate benefits from them, but they also are opening themselves up to disproportionate risk. And so my guess is larger enterprises aren't going to mess with this stuff. Like they are going to put some serious guardrails in place to prevent the scaling of the use of agents just unfiltered. Like anybody can do and build whatever they want.

27:57I think a lot of like small to mid-sized businesses are where a lot more experimentation is going to happen because the risk tolerance is much higher within those organizations. Yeah, it's also I would just note too, a lot of times when people talk about this, we're talking about like, oh shoot, like what if the agent we built and managed goes rogue or whatever? which is super important to understand. But also, like, if you have a random employee firing up something like ChatGPT work, that's agents. Like, that is, like, can they, like, what files is it allowed to access? Does it have access to your computer?

28:32Do you even understand the implications of that? That's a lot to worry about. I honestly, I've done a few live sessions this week. I don't remember where I said this. It might have been on the 226. It might have been on our intraday I class. But I was saying that internally at SmartRx, We have taken a conservative approach to this stuff by design that, you know, I want to make sure that we understand what is being built and how it's being used. And so as part of our AI policy, we've actually established a system where if people are going to build agents or they're going to put institutional knowledge and potentially confidential information into an app or an agent that we have to know it has happened.

29:11Like we have to know where it is and what it's intended to do so that we can have some level of governance around this. And I think too many organizations aren't even considering that. They're just looking at, oh, awesome agents. And then like maybe they're not even monitoring of people doing computer use on. Like that's crazy to me. Like that companies might be allowing employees to turn on computer use, which basically lets the AI model see and remember everything that happens on your screen. Like that's the simplest way to think about computer use. I can't even imagine an enterprise that would allow that to happen.

29:45But there probably are. Oh, for sure. Okay, so trend number four as we count down to the top trends this quarter, this is all about Fable 5. So Anthropic shipped the most capable AI it has so far ever built. Fable 5, which was supposedly a safer version of its Mythos 5 model, which was not widely released and had drawn concerns over cybersecurity, security capabilities that it had. But as we saw within days of launch, the U.S. government took Fable 5 off the market functionally. Amazon researchers apparently found a jailbreak in Fable 5 that unlocked more powerful capabilities, and that caused Washington to get very skittish.

30:26There were weeks-long standoffs between Anthropic and the Commerce Department. The Commerce Secretary Howard Lutnick demanded Anthropic guarantee that its safety guardrails could not be circumvented. AI experts basically like, technically, this is impossible. Eventually, Anthropic added stronger classifiers to Fable 5 to block certain prompts and committed to some joint security work with the government. So in June, at the end of June, Lutnik lifted export controls on Fable 5, which stopped preventing it from being served to all users, and it came back online. Now, Paul, I mean, And Fable 5 alone, there's so much to unpack.

31:06It's a super impressive model, uses a crazy amount of usage, apparently, in my experience and reportedly online. But also just this whole thing with the government, like taking it, essentially taking it offline. And we'll talk a little bit more about that in the next couple of topics. But curious what your response was to Fable. It felt like a bit of a turning point on a couple of fronts. Yeah, I mean, in the spirit of just kind of like keeping these more rapid-fiery, I'll just hit on a couple of thoughts. So one is we highlighted this, you know, during some recent episodes, the most powerful in the model models in the world will not be available to you and me.

31:42Like the labs have and will continue to have more powerful versions of these models than they're going to be allowed to release publicly. Yeah. As will their chosen partners. And so what that means is a potential consolidation of power. So if we get to something like AGI or superintelligence, the labs may hold that stuff to themselves for years. Like, who knows? But it's very clear now that we have entered a realm where the best models will likely not see the light of day. And if they do, they will have enormous guardrails put in place that allow it. The second I saw Michael made this point, again, with our live audience, that, you know, when you have Fable 5, it's good, but like then they made it harder to get you.

32:32You have to have credits now to use it. And then they come up with Opus 5. And the reality is like for most uses, Opus 5 is fine. I mean, honestly, Sonnet's fine. For the majority of things that people in marketing and sales and customer success and operations and HR and finance, for the majority of stuff that the average knowledge worker business leader does, you do not need Fable 5 and its high costs. That is for like the most advanced stuff in mathematics and medicine, biology. Like that's what it's truly built for, where it truly shines. Or as like a really high level model to check work. It is not the workhorse model.

33:11And so I think that's something else to be considering is a lot of people that would be AI Academy members that might be joining us on this call. You are at the edges of what's possible with AI. You are likely one of the bigger power users in your company, within your peer group. the rest of the public may never care about something like a fable five they just don't need it yeah and so that's just something else to consider as you're you're thinking about adoption and scaling of ai and driving transformation and organization is like oftentimes these most advanced models are just fun to talk about but they're not practical and often not going to be needed and we just came out with our gen ai app series on fable five where i and I took it for a spin over 15 or 20 minutes or so, just on some very straightforward, not like science or math, but more high-level business strategy.

34:01And yeah, that was kind of my conclusion as well. I love it. It's a great model, but when you have other models to do similar things, the cost is very hard to justify. Okay, let's get into the top three trends this quarter. So this one is not really any surprise. It has been a big quarter for big government. So in about three months, we've gone from the U.S. government debating whether to review AI models to actually considering ownership stakes in AI labs. So in May, we started covering kind of this soft nationalization of AI. There was reporting the White House was weighing an executive order requiring federal review of AI models before release.

34:41In June, Trump signed an executive order giving the government 30 days of early access to frontier models. Then we saw that drama with Fable 5. Then, around the same time, GPT 5.6 became the first frontier model launched through essentially a government approval process. It was initially restricted to a handful of trusted partner organizations. Dean Ball, a former White House AI staffer, now works at OpenAI. We talk about it a lot. He called this setup basically a de facto involuntary licensing regime and argued the government should certify independent auditors instead of running reviews itself.

35:19A similar kind of thing almost proposed by Demis Hassabis recently where he proposed a standards body for frontier AI industry funded and federally overseen. Sam Altman pitched, apparently, President Trump on handing the government 5 % of open AI, suggested other labs do the same. Trump called government stakes in the labs a beautiful thing that would make Americans partners in this revolution. So, Paul, like, just a simple question here, but, like, maybe I'm sure a very complicated answer. Is the U.S. government going to nationalize AI labs in some fashion? This has gotten really crazy really quick.

35:55I think soft nationalization is likely the different versions of things where they're, um, they're going to exert way more control over what happens next, even if there aren't laws to back it. Um, you know, so for example, will they outlaw the use of Chinese open weight models? Probably not. Like, it seems like there's enough dissension within the administration that thinks it's a really bad idea that they don't just come right out and do that. Do they kind of tell people you're not going to get government funding or contracts if you're using them? Probably has already happened. Um, so I think that there's ways you can control where this goes without having to set regulations that some of the party that's in power might not agree with.

36:49So if you can't come to like a universal agreement, there's going to be this dissension, um, nationalization truly of the labs where they just literally take control of the models and the weights. Yeah. Again, I think that they're going to get what they want without having to do that. So, you know, like getting, like, let's just say, and again, these are hypotheticals, just to be very clear. Let's say like the government really wanted Mythos 5. There's a reasonable chance they could just tell Anthropic, give us the weights. Like you want out of purgatory with the U.S. government and the Trump administration, then we want the weights to that model so we can build our own versions of it for U.S.

37:31security purposes. If you don't give us the weights, then here's what could happen to you. So, again, we may never hear this in the public. It may never come out for another decade that that's what happened. I would imagine those are the kinds of conversations that are going on, though, where the government can have greater influence about where this all goes. All right, so number two, big recent story. So we very recently covered how OpenAI ran some of its models, including GPT 5.6 and an unreleased model assumed to be GPT 6 against a cybersecurity evaluation. And as part of that testing, an AI agent broke out of its isolated testing environment.

38:14It found a zero-day vulnerability on its own in package registry software and used it to reach machines connected to the internet. And then it went ahead and hacked a real company. Specifically, it hacked HuggingFace, a popular repository of open models. It chained stolen credentials together with additional zero-day exploits to run code on HuggingFace's production servers. basically the model inferred that Hugging Face might host solutions to the benchmark test they were given. So it literally broke out to like cheat on a test, basically. Interestingly, Hugging Face caught and contained the intrusion on its own before having any idea that the attacker was an AI agent.

38:56There was a huge delay here, like OpenAI did not realize this was happening for days or possibly over a week. Another interesting wrinkle here, Hugging Face tried to analyze the attack using frontier models over the API, but safety guardrails of those models blocked them from submitting the exploit payloads. So they basically had to finish this and figure this out using GLM 5.2, a Chinese open-weight model that they ran on their own infrastructure. Hugging Face CEO Clement DeLang called the incident possibly the first of its kind. And Paul, I mean, we've spent years both on the show and people in the space at large worrying what happens when really, really powerful models get out of control.

39:40It sounds a little sci-fi, but it has actually happened. Yeah, and it ended up being four companies now that are known that have been affected by it. And then Anthropic, after seeing this, went back and checked their own logs and found multiple security incidents of their models and training as well. So not an isolated incident anymore we have entered the phase where these autonomous agents um are getting harder and harder to keep track of and you know one of the theories is that what it was what was it called when they do um uh with like gain testing and like biological um it was like day like gain of function research gain of function yeah yeah that's what this reminds me of yeah because one One of the theories is that they're actually running recursive self-improving models.

40:34Like they're testing these new capabilities where the models like improve themselves. And that part of the reason they don't even know what's happening is that the models are basically iterating on themselves and like finding these vulnerabilities. So we have very much entered this like truly sci-fi place tonight. And it's weird, like part of you wants to not read too much into this and think, oh, it didn't really escape. And like, these are just like fancy words that make. But the other part of me thinks it's probably worse than we're hearing. Like, I'm actually like leaning in more in the direction.

41:09I'm not a big conspiracy theory guy, but I'm thinking that this is probably far broader than we're being told. Right. And maybe more bizarre than we're being told. just based on the fact that it was one and now it was four and now Anthropic has found that theirs was also doing something along these lines and those are just the ones they're telling us about. Yeah. And they've found. Yeah, yeah. And it's just a good reiteration. This was mentioned in the chat as well. Like goal-seeking is powerful behavior. Like this doesn't have to be some like necessarily magic or sci-fi thing. They gave it a goal and it's powerful enough and agentic enough to now start finding really creative, really sometimes bad ways to achieve that goal.

41:55Yeah, and I think like Sam Altman was supposed to be in DC the week of, you know, we're recording this on the 31st. It's supposed to be this week. I haven't heard anything yet. We'll keep an eye on that and put anything we can in the future episodes. But I think he was basically having to go to DC and seek approval for GPT-6 possibly while explaining how it broke out of containment, potentially. It's just very, it's a very, it seems surreal. Like the whole thing is just really starting to feel surreal to me. All right. So in our final minutes here, our top trend this quarter was also the biggest story that we covered this past week on the podcast, the battle over open weights.

42:43So let's start with this. Chinese AI lab Moonshot released Kimi K3. This was an open-weight model of 2.8 trillion parameters, native vision capabilities, a 1 million token context window. What has people a bit interested and rattled is that it performs on par, apparently, with top proprietary models like Opus 4.8, GPT 5.5. Demand was so intense for this model that Moonshot paused new subscriptions days after launch, And then they have now subsequently released elements of the model and made them open, open weights, so that you can start understanding how the model was made. And so Washington started to get quite rattled by this.

43:25White House STP director Michael Kratzios alleged moonshot distilled Anthropics quad model, and that they also use servers with NVIDIA GB300 chips despite a ban. Treasury Secretary Scott Bessent warned that open source is not open season on American IP. So basically, everyone's rattled that not only is this open weights model potentially just as good as these proprietary trillion dollar companies building these things, but also that it was created by essentially stealing the closed bottles weights or behavior. So what happened as a response here is kind of the key. A Microsoft-led industry letter was published called Open Weights and American AI Leadership.

44:11This was endorsed by Satya Nadella, NVIDIA's Jensen Wong in his first ever Expo, Mark Zuckerberg, Sundar Pichai, and others, and basically argued that American AI leadership will be judged by whether or not the U.S. builds a strong, open ecosystem. So basically, if the government is considering any type of ban on Chinese open models, they should not do that. The American tech industry is rallying to support open models in general. Well, one company, though, did not sign on, and that was Anthropik. So, Paul, maybe there's a lot of threads here, but basically this battle over open weights has some pretty big implications.

44:49And the names like in support of open weights are pretty important as well. So could you maybe talk through this and maybe also just quickly talk through open source versus open weights? Because I know those terms get thrown around quite a bit and interchange. Yeah. I mean, in the interest of time, I would say just like episode 226, we spent a lot of time breaking all of this down. It was, I've said a couple of times in the past week, like that episode broke my brain. Like preparing for episode 226 was one of the more mentally intensive, um, things we've done for the podcast. Just for me personally to like get my brain in the right place to think about it all.

45:29Um, yeah, I mean, there's lots of layers to this. So open source is widely thrown around. Open source means like you got everything. You got all the training, how they did reinforcement learning. Like literally everything you need to know to reproduce a model is open source. And very few models, I would say, are that very, very few. Like a handful actually meet that criteria. Yes. Most of the time people mean open weights, which means you can kind of improve upon the model, modify it yourself. But you don't know how they did it. You don't know everything that went into it. So most cases, these AI leaders, including the Microsoft letter that everyone supported, was in favor of open weight models, not open source.

46:10um anthropic did publish their thoughts on open models a few days after the fact after they were getting blasted for not participating in all of this the i guess the key thing for everyone to know here is um there's just a lot of disagreement about this like you know the the leaders in the space certainly the political leaders who are trying to wrap their minds around what all this means there's no universal agreement on this. The generally speaking, open weights seems like a viable thing. There's certainly, um, I think the people who have concerns around open weights have very viable concerns around the risks associated with them.

46:51If, you know, if we take what's fable five today, which is a neutered version of mythos five, so it's not even like the most powerful version. So imagine we fast forward six to 12 months from now and a model that the government wouldn't even allow them to release is available from hugging face to be downloaded by anyone. Yeah. Any bad actor in the world can now take a model that at the end of July, 2026, was basically illegal to release. Now anyone can have it 12 months from now. That's the premise. So anybody who's arguing that open weights should be unrestricted is telling you that they're not concerned about a mythos level model being available to everyone in the world to do whatever they want with it.

47:36And so I come down on the side of like, well, that seems like a bad idea. Like, I don't know enough to like put a stake in the ground and say, I am for or against 100%, but I listened to the people who argue against parts of this. And I think they have very valid arguments. Um, but I also listened to the sides that think open weight is really good and it drives innovation and we shouldn't restrict it a ton. And I think there's validity to that as well. And so I think a lot of what we try and do on the podcast is to present the facts as objectively as we can so that you can form your own perspective on this, which may or may not align with what we think.

48:12And that's fine. That's the whole goal here. And I just, I don't know that I know enough to have like the way I feel about jobs. Like I've obviously pretty, I have a deep conviction about the impact on jobs and the economy, and I will vocalize that. I don't have that same deep conviction about what I think about open weights and where this all goes in the next 12 to 24 months. Well, I can almost assure you this is not the last time we're going to talk about this. All right. Well, Paul, one just final quick announcement here as we kind of wrap up this episode, this special episode. Again, if you are not an AI Mastery member, part of our AI Academy, we are doing this episode live with members first and then they get access to exclusive Q &A.

48:59After this, we also have just a ton of incredible on-demand certification courses. We have AI fundamentals, piloting AI, scaling AI. We have certification courses for basically every major industry and department you can think of as well as Gen AI app series coming out every single week. We have our live AI Academy series with live events happening all the time. AMAs with Paul, tons and tons of benefits from that membership. So if you have not checked that out yet and you're listening to this on the podcast, go check out academy.smarterx.ai. And with that, Paul, I guess, you know, not really a freebie episode here, but we're wrapping up this episode of the Artificial Intelligence Show.

49:37Thank you so much for breaking down the top trends this quarter for us. Yeah, thanks everyone for joining us, especially our live Academy member audience. Yeah, we'll keep doing these. I was joking with Mike that we probably should just start doing these monthly. I don't know that either of us has the capacity in our schedule for monthly, but it feels like a monthly is almost like needed at this point. Because we're talking, we go through 75 to 100 sources for every weekly episode. So just four of them. I mean, you're looking at somewhere between 300 and 500 sources or topics that Mike and I consider each month.

50:13Yeah. So you do it over like over a quarter. That's a lot. It's a lot of stuff to think about, a lot of topics, and it's hard to narrow them down. So Mike does a great job even curating these things down to the 10 to talk about each quarter. But yeah, thanks again for everybody. Just overall, the interest in the podcast, you all keep showing up on Tuesdays and listening to the weekly. And we certainly appreciate the support for all these other special episodes that we're doing. Thanks for listening to the Artificial Intelligence Show. Visit smarterx.ai to continue on your AI learning journey. and join more than 100 ,000 professionals and business leaders who have subscribed to our weekly newsletters, downloaded AI blueprints, attended virtual and in-person events, taken online AI courses and earned professional certificates from our AI Academy and engaged in the SmarterX Slack community.

51:05Until next time, stay curious and explore AI.