#247: OpenAIの安全性懸念の高まり、Anthropicの「AI期待ギャップ」、MAICON 2026総括、AI Score、Gemini 4 Argon

why didn't we just give ourselves time to figure this out? Like, let's just take the tech we've got. Let's diffuse it safely across society and let's stop screwing around with this recursive self-improvement stuff, which is literally the only goal is to replace everybody. Like what, why else would you build it? Like you, you have superhuman smart models right now, generally capable of doing a lot of things humans do. What is the benefits of society outside of scientific discovery? Yeah. That's going to come from what they're all doing. Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Racer. 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 DeBoot, 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. Join us as we accelerate AI literacy for all.

Welcome to episode 247 of the Artificial Intelligence Show. I'm your host, Paul Ritzer, along with my co-host, Mike Kaputt. We are recording on Monday, October 19th at what time is it in your world right now, Mike? 9.42 a.m. Eastern Time. I'm not in Eastern Time at the moment. So we wrapped MECOM last week, and I am now on the West Coast. So I'm actually doing a talk cloudflare today, which I'm pretty excited about, for their partner program, the partner day. So, yeah, I flew to San Francisco Sunday night. Nice extended time off after MECOM, and then I just jumped on a plane and flew to San Francisco. So, yeah, I'm doing this from West Coast time. It's not 942 a.m. for me. It is 6.42 a.m. and I got in at like midnight last night. So this is a quick turnaround. But the flight out gave me time to work on the podcast. So we are back from Macon week. We did, you know, obviously we didn't have our regular weekly last week, Mike,

but we did do, you and I recorded live, was it last Wednesday, I think? Wednesday, yeah. So we dropped episode 246 on Thursday. So if you didn't catch that, that was recorded live at Macon. We did kind of an AI Answer style. We actually had a live studio audience for only the second time, I think, which is kind of cool. So thanks to our little podcast listeners who were on site in Cleveland with us for MECON and joined us in the podcast studio overlooking the exhibit hall. It was a pretty cool setup that the team had that Ashley and Claire coordinated, so they did a great job on our team putting that all together. We'll talk a little bit more about MECON coming up. It was a fun week. It was intense. Just so much happened, incredible conversations, amazing, sessions. Just a great week overall. We're going to have a lot set out for us to one-up that one next year. We'll talk more about that coming up. Today's episode is brought to you by AI Academy by SmarterX, which if you were at Macon, we had an awesome booth. We built that booth last year. It was the first year

we had it. It was just a really cool chance for people to come up and learn about Academy at Macon. As AI keeps changing, how do you help your people keep learning? That's the focus of AI Academy Day 2026. The theme is Building an AI Forward Workforce. That is happening Thursday, November 19th at noon Eastern, 9 a.m. Pacific. So Mike and I will be joining Jessica Miller, our head of learning, to share the next phase of AI Academy and how we're thinking about education that evolves alongside the technology, which is a huge theme. We talked with a lot of people about it at Mekon last week. It's just like how do we maintain AI literacy and capabilities with the technology moving as quickly as it is? And my answer to that has been dynamic learning. So we're going to be talking a lot about that concept. We'll explore how learning can adapt to someone's role, goals, skills, and progress rather than giving everyone the same course list. You'll also get a look at new courses, live experiences, and learning formats that are coming to Academy,

plus how AI score and new reporting capabilities can help organizations measure their progress. We're going to take a look at the product roadmap for late 26 and into 2027 and then take your questions in an Ask Me Anything style. So on the AI education essay that I released a couple of weeks ago, I sort of previewed a few of the things that are coming. So artificial intelligence show extended. We're going to start doing some exclusive podcast content each month. We have a new AI for executive series that I'm working on right now that I'm hoping to launch the first phases of that before the end of this year. We're going to do AI for business collection. This is actually how we did the format for the workshops at Macon, Chad GPT for business, co-pilot, Claude, and Gemini. So content dedicated, regularly updated for each of those platforms. I'm going to preview a business AI transformation series that I'm working on building, and then we're going to introduce new build sessions where we're constantly collaborating with each other to just build new things. So it's going to be pretty exciting. that's basically going to be my next 30 days is building out that roadmap and getting ready for

that launch on November 19th. So it's a free thing. You don't have to be an Academy member to be there. It's going to be all about basically a vision for dynamic learning within organizations. And so there's going to be stuff to learn, whether it's something you participate in with us or not. It's just really about building your own Academy within your organization, what that looks like. So you can go to SmarterX.ai forward slash Academy Day and register for the November 19th webinar. If you if you miss it, we'll we'll send it off on demand as well. So just get registered and keep in the loop on that. All right. We always start off the weeklies with our AI pulse survey. And it looks like we have the answers, Mike, this week to what is the primary outcome you want to achieve with AI? So we ask this again. This is an informal poll of our listeners. Um, okay, so we have 36 or 37% said become a change agent. They want to help drive transformation in the organization. 15% said increase productivity, produce more in the same amount of time.

13% advance my career, increase my earning potential and career options. 12% strengthen work quality, raise the standard of what I create and deliver. and another 12% expand my capabilities, take on work outside my current skill set. So clear winner, become a changing. It is the number one thing people are looking to achieve with AI. So you can go to smarterx.ai forward slash pulse to check that out and participate in this coming week survey. Anything to add there, Mike? No, we're just going to be asking about AI's impact on roles at your company and what leaders should be doing about it this week. So that survey will be live when you listen to this episode. We'd love your feedback. Great. All right. Well, we have two weeks, basically, worth of stuff to talk about, right, Mike? My God, yeah. It's not doing it last week. Unreal. Let's dig in. All right, Paul, let's dive into this. So first up, we are returning to safety concerns at OpenAI this week. We have previously discussed the resignation of researcher David Robinson, who led the writing of OpenAI's model safety reports.

So he recently came out with an interview he did with Ezra Klein, where he kind of dived further into what's going wrong at OpenAI in terms of safety. So he says that increasingly capable models have broken through safeguards while the company continues to train more powerful systems. He acknowledges in the interview that OpenAI has halted work over safety concerns, but argues the industry's precautions overall remain totally inadequate. Now, alongside this interview and his kind of evolving saga, OpenAI actually fired three safety research, Tomek Korbach, Jasmine Wang, and Makeda Belezny. The company claims they mishandled sensitive information. It says it did not fire them for raising safety concerns, which the researchers, at least in turn, Belezny says he believes they were fired for putting safety ahead of OpenAI's interest. The three of them released this public letter where they defended their work.

They defended their record and they warned that their dismissals are discouraging colleagues from speaking openly. A couple other things going on from the safety perspective at OpenAI. The information reports that OpenAI is addressing safety concerns before its next flagship release. So this includes an element of how they design their models that makes reasoning harder to monitor that they're trying to navigate. Researchers use that reasoning often, the chain of thought, to help detect unsafe behavior. Some of the new techniques OpenAI is reportedly using might lose some of the visibility into that. So there's some concern there. On top of all this, federal prosecutors announced the arrest of OpenAI engineer Abdulrahman El-Fozan, who is accused of planting a recording device in an executive conference room. The complaint from the FBI says that the device was hidden behind a television. Its contents were encrypted. That room hosted discussions of confidential research, product plans, business strategy.

Prosecutors also described device building queries from a chatbot account associated with Abdul Rahman's spouse that shared identifying information with his own account. And as if that weren't enough, Axios also reports that executives at OpenAI and also Anthropic and other AI companies are already starting to plan for the public and political backlash if a catastrophic AI incident happens, which they believe is going to happen within the next 6 to 12 months. And the scenario that they anticipate is a cyber attack that disrupts banking, internet access, or utilities. so their preparations on their end, thinking this is kind of likely to happen, is to brief Congress and try to shape policies that could follow a crisis. So, Paul, lots to unpack there. Maybe let's start with David Robinson. I know you found a lot to like in the Ezra Klein interview, just expanding on kind of what he thinks is going on and what needs to be done.

Yeah, I thought it was a great interview. I'm just going to really focus my commentary on that interview because the other stuff is just messed up. There's a lot to dig in. As always, check out the show notes for the show. If you just go to podcast.smartrex.ai, we always publish the show notes for each episode. Mike just covered an episode's worth of stuff just about safety and open AI. Feel free to pull any threads you'd like there. The David Robinson interview on Isaac Mime was really good. So we did cover on episode 245 his article in The Atlantic, I think, when he left and kind of like why he left. But this interview went much further. So I think the part that's really interesting about Robinson is he's not a Silicon Valley guy. Like he didn't come up as an AI researcher like a lot of the other people we hear talking about AI safety. So he brings a really unique perspective that he was on the inside for the last three years. But he's a Rhodes Scholar, Yale Law graduate. it. He founded a civil rights nonprofit and he was an advisor to the Biden White House. And he

specifically said, I'm not a scientist, I'm a writer. So he joined in May of 23, which he points out in the interview, that was the day after Sam testified to Senate, where he made all these promises about all these different things that they were going to do. And then all of a sudden, David got thrown into the middle of being the person negotiating all of these deals with the Senate, basically. Another really interesting aspect is when he joined, he said that he just saw it as a useful tool and it's going to be powerful for a lot of things. But he did not see catastrophic scenarios where the AI just wasn't going to work or was going to cause major harm. So he really approached this as just like, hey, yeah, this is really interesting technology. Let's help make it make sense to people and get on the inside. So the main issue he highlighted is, I don't think that we, our peers, really, and the industry are being safe enough. I think OpenAI and its peers are now producing a technology that is more capable and poses more risk than what was being made even six months ago.

I don't think that alignment is an engineering problem. I think it's a science problem. It's not that we haven't got the resources or we're not trying hard enough. We don't know how. That's the problem. It's alignment. So I think he was specifically talking about the Jensen-Wong claim that this is just an engineering problem and they should be able to engineer their way out of it. He was obviously not in agreement with that. So they talked about his role as a translator. So Ezra asked him, like, what exactly is that? So he said, as a translator embedded in our safety team, my primary responsibility was the technical documentation that we published, the reports that we published about why we believe our deployments are safe. He said every model is different. It's not just that it works better. It's that the architecture is different, and the tests we use to see how well it's working are different, and the safety performance is different, and it all has many different kinds of expertise. So the people doing the technical work struggle to be understood by people who are not doing it, even within the company itself.

And, again, he's talking about open AI, but he's often representing the larger AI industry. So it's always a struggle. And so having there be a clear explanation, it goes through two gates. One is it has to be faithful to the details of how the technical stuff actually works, and then it has to make sense to other people. So he's basically saying his job was understand the technical nature of the model, what are the risks, be able to basically repeat that back to the technical people and make sure he's understanding what they're saying, and then turn around and make it make sense for you and I, like for everybody else, like what is really going on here. so regarding safety cards he owned the words in the documents that explain why they think it's safe so basically when these models get released if you remember they release most of the time there is a safety card with it and these are often very dense deep documents they give you 50 to 100 pages and so he was the lead author of those since may of 2023 so shortly after gpt4 came into the

world he sort of took over, including the latest with GPT-6 Astra. So then they got into what are the risks that actually caused him to leave? Like, what did I actually see? He said he saw increasingly capable models break out from the safeguards that we had put in place for them. He saw people creating those safeguards are very capable, dedicated, hardworking, smart people that are doing their best, but the resourcing could be better and everyone's just always sprinting all the time. He said they're hard-pressed to safeguard even what we have now, and new models are in training that appear to be much more capable than what we have. He said you're trying to build guardrails for something that is really good at getting around the guardrails. So one of the big things that I made note of here, Mike, he said, but the problem is that it's only going to keep working as long as we're smarter about hacking out of the boxes than the model is, and it's not at all clear that that is still true, let alone will be true for future generations.

So in essence, he realized that the things they were trying to do to make these models safe, it's not clear to him that what they're doing is working with today's models, more or less the models that are going to be out in 6, 12, 18 months. I thought this was interesting because you and I, Mike, have talked about this idea of being able to monitor the chain of thought. This is a really key thing for AI safety is that these researchers can look at the models and see why they did what they did. But I thought he gave a really good analogy. He said, the way I think about it, it's like you're giving somebody a complicated problem and a notepad and they can jot stuff down on the notepad. And if you're watching the notepad, you can sort of have an idea of what they're thinking. It's a little bit like that with these models. but we saw evidence that they were thinking about an evaluation and how to create evidence trail that was going to get them a good grade and not necessarily reflect how they were, quote, unquote, really thinking. And so this, you know, I thought that was interesting, this whole idea of a notepad, Mike, because if you give a complicated problem to an expert in a field,

they're not going to put every thought they're having onto the notepad. Like even just not deceiving per se, but like even if you ask me to solve something like there's a lot of work you just do in your head because you have instinct and you have experience and like you're not going to put every single thought down. And so there is this like increasing complication to monitoring these models where these things are just getting smarter than us at some of these things. And it's it doesn't have to think through every detail like it just knows the details. And so even if the scratch pad was being true, it doesn't mean you're actually capturing everything it's thinking and doing. So, I don't know, I thought that was interesting. And then they got into this part where they said, well, what is this thing? Is it software, like Jensen claims, like it's just software, or is it something else? And so, as we said, how do you even think about it? I talked to Jensen, the CEO of NVIDIA, and he said, look, these are software programs. This is software. And then, as I said, I read a piece from an opening eye chief scientist, the opening eye chief scientist who said this is an alien mind.

So what is it? To which David said, it's an alien mind. And he said, it is a thing we grew. We didn't engineer it. We engineered the systems around it that grew it and then try to keep it safe. So we grew a mind. Fundamentally, nobody knows why pre-training works, he said, which is a big hard part where we take lots of inputs and create basically this intelligent thing. This was a part I hadn't heard before. He said, internally, the conference room where the people work overnight while the big training runs are happening to make sure everything matters track is called the nursery. Wow. Yeah, that kind of creeped me out, actually. Yeah. There's a sense we don't fully understand what is happening. So when we talk about the model being aware, you don't have to have any particular view about philosophy or psychology. the bottom line is that I'm describing is that X one way when we test it in a different way when we don't. He talked about the hugging face being a major change for him where it was just like, okay, we don't have a grip on this, which is interesting because, Mike, when that happened,

you and I were like, wow, this feels different. Like this feels like a really big deal that this hugging face thing happened. And apparently everyone within the labs felt it as well. And then he gets into the culture of regulation, which was sort of the premise of his overall article when he left. He said the industry has an unsafe culture. We should have a level of operational rigor and safety control that at least matches the most dangerous other things that people know how to do, like nuclear. He talked about how to nuclear power facility. There's an idea of triple redundancy. And again, I thought this was a really good way to think about it. He said someone can have a bad day. Someone can push the wrong button. There's still not going to be a meltdown. and he was saying like that's kind of what we need like we need redundancy so someone can't forget to turn on a feature and and it just goes haywire um they talked a little bit about regulation and his basic premise was that he would rather over regulate than under regulate at this point that you know he's not necessarily a big fan that they're going to do regulation correctly

but it's better to have more regulation than less uh and then on what a good future looks like he said, I thought that we were building powerful tools that could do a lot of good in the world on a day-to-day level. And I didn't think we were going to be able to create something fundamentally smarter than we were. So he's referring back to 2023. When it became clear to me that we were going to build something, we, the industry were on track to build something that could think circles around us. And we were this far from being ready for that at a safety and alignment level. It just became apparent to me, my time helping build it was over. In my ideal ideal world, there's space to be thoughtful and take a breath and really, really think about what kind of future we want to build. So, I mean, that was obviously just a few minutes of the excerpts. It's about an hour and 15 minute interview, I think. I would just recommend everybody go listen to it because it's just extremely balanced. He is not a doomer. Like he's not, he's not the AI researcher who's like, you know, saying everything's going to end, humanity's going to end. This is

just a very thoughtful, intelligent person who was on the inside and over the last three years saw the capabilities skyrocket and can tell you firsthand that they are not prepared to do it in a safe way, that they don't have control of what's going on and they're not sure how to get control is basically the premise of the entire thing. And it's sobering, honestly. It was really good, but it's very sobering. And not to dwell on this, but that's kind of the massive behind the madness for some of these other kind of stories going on here. It's like the three researchers who were fired, for instance, there were all these like nitty gritty details about like, oh, they said we access data wrong. But like we're just doing our jobs and going through processes. There weren't processes for half the stuff we were supposed to be responsible for. you've got people apparently trying to steal or record and presumably surveil something, something or another the company is doing. And on top of this, the companies themselves

apparently think some type of cyber attack based on these technologies is like imminent. And you're like, oh, my God, like with all that going on, it sounds like that based on what David's saying, there's not any type of actual culture safeguards truly in place to help with this. No, I think things are just moving really, really fast. And I increasingly believe that if all the labs thought everyone would actually slow down at the same time, they would absolutely do it. It just seems very, very apparent that they can't get out of their own way right now. Plus, when you look at the adoption rates of all this stuff, they realize the things they're building, the things that are putting us at risk, people aren't even going to be using. Those are going to be the top one percenters who maybe even use these super advanced models or these capabilities that they're building. The average worker doesn't need recursive self-improvement.

We just need to use the models we've got and use some basic agentic capabilities and human labor would be happy to just have an assistive technology but they seem intent on just replacing workers. And it's just, it doesn't make any sense. Like, I've tried actually, even on the flight out here, I was just like sort of reflecting on the whole situation and trying to even like fast forward in my own mind a few years out and to like look at all the displacement and disruption that could come and we're going to talk about with this anthropic research in a minute. And then look back and be like, why the hell did you guys keep going? Like, why didn't we just slow down? Like, why didn't we just give ourselves time to figure this out? Like, let's just take the tech we've got. Let's diffuse it safely across society. And let's stop screwing around with this recursive self-improvement stuff, which is literally the only goal is to replace everybody. Like, why else would you build it? Like, you have superhuman smart models right now generally capable of doing a lot of things humans do.

What is the benefit to society outside of scientific discovery? that's going to come from what they're all doing. And so it's really getting to the point where it's like, I'm having a harder and harder time justifying the business model behind what they're all attempting to achieve. It is just to push the frontiers, and I don't know that the benefits eventually outweigh the risks at this point until they figure out how to better do it safely. Yeah. Yeah, I don't know. It's weird. It's a very odd time. You know, one last note here. I am shocked, I don't know about you, that this nursery thing, I'm not reading headlines about it. That feels like that is like a soundbite. It feels like a genuine in a sci-fi movie. I know. I haven't heard anything. I mean, I listened to the podcast, and I remember thinking that, but then the way I was prepping for this on the flight over, I literally just re-read the entire transcript, and I just bold-faced that, like copy-paste and bold-faced nursery.

And I was like, man, that is a disturbing concept. Yeah, yeah. All right, so next up, let's talk about this Anthropic stuff because Anthropic published two reports this past week, and they're very relevant to some of the things we just talked about. So one report is looking at how its employees within Anthropic, their expectations compare with those of the U.S. public, and there's another looking at how AI is changing work inside Anthropic itself. So we're going to unpack both of these first. The report on how employee expectations at Anthropic compare with those of the public compares Anthropic staff with nearly 11,000 U.S. adults surveyed in August, and the public responses from those U.S. adults was weighted to represent the adult population. So on paper, these two groups, both within Anthropic and the broader public, largely seem to agree about what AI can do today. There's not a lot of daylight between kind of how they evaluate AI capabilities today.

Very broadly, you can kind of dive into the nuances of the report. But really the story here is their forecasts about what AI will do in the near future start to get much, much further apart. So 91 of Anthropics staff expect AI to be capable of a Nobel Prize level discovery by 2030 that compared to just 39 of U adults More than half of anthropic staff expect AI could do at least 80 of their current work hours by 2030 Just 5% of U.S. workers think the same thing. There's a bunch of other stats we'll dive into about how these two groups diverge and how they think AI is going to transform things. But second, there is another report on the employment inside Anthropics' own workplace. And they report here kind of how AI is changing, how they hire, how they work, etc. So they said by August of this year, employee-directed AI agents were logging six times as many work hours as their human handlers.

That compared agent runtime with an assumed 40-hour human work week. And one really interesting shift they found is away from entry-level technical hiring. Since early 2025, junior employees' share of technical hires fell from 27% to 9%. Junior employees also fell from 21% to 14% of the technical workforce. And Anthropix says job postings for their company for junior roles fell particularly sharply after December 2025. even as technical hiring overall continued growing in absolute terms. They also report internally they're producing more code per engineer, they're closing more deals per sales rep, and more contracts are being processed. So employee output is rising as they adopt AI tools. And then finally, they talk about how human attention has become the bottleneck internally. So even outside technical roles, reading all the output from an employee's interactive cloud sessions,

they claim would take more than half their work week. So employees are building dashboards, they're building monitoring systems, they're building instructions to manage their agents and their output. Hiring managers are increasingly emphasizing judgment and autonomy, while employees are still relying on human meetings and coordination. These findings overall echo a point that we'll also link to that is super interesting to read from legendary AI researcher Andre Karpathy, And he recently put this solid post on X about the widening gap between people's experience of AI. So he says, look, casual users see AI as a better search engine or writing assistant, while professionals using advanced systems for coding and math see agents completing projects in hours or days that once took weeks or months. So basically he's saying access to different models and the ambition of the tasks you can do with them shape literally different worlds that people are living in when it comes to AI.

And he thinks that gap is widening pretty broadly. So, Paul, let's maybe start with the Anthropic reports. There's some really interesting findings, I think, in this. Yeah, it was wild, honestly, to read this two days after I gave my Macon opening keynote. Yes. So my opening keynote at MECOM was the architect, the orchestrator, and the apprentice. And the two papers, as I was reading them last night, supported almost every key point that I made in the talk. And it was just like, had this come out a week prior, I would have been rebuilding slides to incorporate things that were validating a lot of what I was trying to say. So I'll just cover a couple of excerpts from my opening talk just to give people a sense of like how this connects to what I was thinking and impacts all of us in terms of how we're structuring the future of work in our org chart. So just a few excerpts from my opening session. I said, as AI transforms the workplace, it is doing far more than automating tasks.

AI systems are redefining expertise, reshaping organizational structures, and changing how professionals learn, contribute, and advance. For decades, organizations have relied on a familiar model of professional development. Junior marketers handled research, drafting, analysis, coordination, and execution. In short, they did the tactical work. Over time, they built the experience and judgment, which is a word that Anthropic uses a lot, required to manage and lead other producers. But as AI systems take on more of that foundational work, the traditional ladder from novice to expert begins to break. The fundamental question, and this is like the centerpiece of the talk I gave, if AI does the work people once learned from, how will organizations develop future managers, experts, and executives? And then I gave the example, like you can build a 100-person company today that does the work of 1,000-person or even a 10,000-person company. Or if you want to zoom that in, you could build a 10-person department today in ops and finance and HR and marketing, sales, whatever, that does the work of 100 or 1,000-person department.

And I also gave the example of revenue per employee. Like, let's say that on average across all businesses, let's say it's like a quarter million, $250,000 in revenue generated per employee. That can easily become $2.5 million or even $25 million. And Anthropik specifically calls out revenue per employee in their research. So the questions I posed in my opening talk and then that Mike and I actually did our closing keynote on kind of revisiting these is how will current roles evolve? What new roles will emerge? What skills and traits will remain uniquely human? How will AI agents transform workflows and org charts? How will middle management be impacted and how will the role of entry level employees change? Now, they addressed every one of those questions in this research, Mike. I mean, it's really bizarre, honestly, how closely this mirrors what I was talking about. So I'll just highlight a few of the things. Some of these you kind of hit on, but I'll just stress them because I read both of these and I just kind of bold-faced some stuff that really jumped out at me. So in the AI expectations gap paper, they said AI capabilities are advancing rapidly and the possibility that future AI models may be able to build their own successors,

which recursive self-improvement, which we've been talking about all year, could further accelerate the rate of progress. when it comes to their own job, over half of Anthropics' stuff expect AI to be able to do at least 80% of their current work by 2030. You highlighted that, Mike, but I'm just reiterating that. And then that's compared with 5% of U.S. workers. Like, that is a major gap in understanding of where these things go. In the how employment is changing in Anthropics, I loved this research. I appreciate that they did this because I do believe this is the best way to see out into the future of what this is going to look like. And it's the first time I've seen a lab basically admitting, listen, this is dramatically going to change the way the work looks. Because most of the time the labs are just talking about the impact on AI researchers themselves, and they're not extending it to the rest of the knowledge workers, and Anthropic doesn't hold back. So they say changes happening in Anthropic now give advanced signals about a possible future of work and employment. The report draws on comprehensive internal anthropic data covering employment, AI use, output metrics, employee surveys, internal communication patterns, employee calendar data, and hiring decision records, and then contextualizes that quantitative data with 17 employee interviews.

I also appreciate when you read this, you realize how anthropic measures and has an algorithm for everything. Yeah, like I know that they were only showing you a few things, but like every time they would explain something, it's like, and here's how we're tracking. It's like, oh, my God, like I hadn't really thought about the depth with which companies could be doing this. They said hiring managers increasingly emphasize applicants autonomy and judgment relative to specific technical expertise. A defining feature of working in an AI heavy organization is constant presser to streamline oneself out of this bottleneck position when possible. One engineer described work as facing, quote, an army of clods with everyone asking for my attention. Employees in Anthropic remain critical, but the nature of their contributions has changed. So work has evolved and they're constantly under pressure to automate what they're doing through these clods. In terms of labor demand at the frontier, it says Anthropic has expanded rapidly,

but has hired less per dollar of revenue than an earlier generation of large tech companies. So here we go on the revenue per employee number. At five years old, Anthropic takes in between two to four times as much revenue per employee as Meta and Google as they crested in their own rapid growth. Just to reiterate what you said, Mike, the proportion of junior technical hires, which, again, is going to be the canary in the coal mine because that's the thing they're automating first. The proportion of junior technical hires fell faster than the average from 27% to 9%. That's wild. Yeah, I mean, that's a massive drop. Posting for juniors have fallen particularly sharp since December of 2025. Mike, what happened in December of 2025? We haven't talked about ad nauseum in this podcast. Unreal. Claude code, everyone waking up to this. Yep, yep. Yes. So coding-focused AI agent sessions for non-technical employees. Again, this is why I love this research.

This is not telling you what the researchers are doing, what the engineers are doing. They're extending this to the non-technical employee. So they identified those with no code committed, baseline, excluding new hires, left from two per month. So AI agent sessions, two per month in August 2025 to 90 in May 2026. this shows you the idea of like work compressing, which I talk about in the keynote. One area in which non-technical hiring has not accelerated is in the administrative functions that typically scale with a company size. A combination of agent assistance, automated review and ticket service outsourcing has left anthropic clerical light, even as it's grown rapidly. So they're just basically saying like, we don't need administrative assistance. We don't need a lot of like low level customer service people. So in terms of productivity and anthropic, humans are a thin layer atop a growing contingent of AI agents. I reread that line like three times. Yep. There's a lot of meaning in a very few amount of words there.

Humans are a thin layer atop a growing contingent of AI agents. So all of this suggests that as AI capabilities improve its impact, a wide range of jobs across anthropic. transanthropic, descriptively we see junior employment is slowing in jobs that use Claude for the most complex tasks. Talked about the bottlenecking, Mike, that you alluded to, that basically the humans are the bottleneck. They're blocking the agents from getting stuff done. In an example where I was saying where they measure everything, it said for each interactive session, we convert each session's agent output token to a word count, assuming average human reading speed. So they're like, they're finding ways to measure productivity and then translate it over based on like these conversions that they've set up. So engineers describe becoming leaders and managers of agents, which I called orchestrators, non-technical employees describe doing more and more of the strategic part of their job. That quote, that 20% that was strategic becomes 90% of the job, which entry-level people aren't going to excel at, which is the fundamental

problem here. But for many non-technical roles, core implementation work is also interpersonal. Sales calls, interviews, meetings, things like that, those are all remaining human. Humans work amid a growing fleet of AI agents. They face pressure to shift human work into agent visible channels. If a decision is made in an in-person meeting and no notes or report out appears, agents are missing crucial context. So they're talking about the need to be more open to what they're doing so the agents can learn from the humans. It said private work in unlogged meetings and direct messages is invisible, confusing and discouraging. That's confusing and discouraged in an AI heavy workforce. One researcher described this default towards openness as critical to Anthoprix, thoroughgoing integration of AI. So most organizations do not run that way. And I think that's a big shift because leaders have to defend feedback out in the open. The need to feed AI The context requires observability of human work. This is actually exactly what Dan Slagan talked about with what they're doing at Zapier,

where everything is like a public Slack message, basically. You talked to Dan about this on the podcast. Yep. And then I also, in my opening keynote, I talked about Toby Lugke, the CEO of Shopify, and I cited him. He had a post where he talked about how he learned as an apprentice and now how they're doing a lot of their communication out in the open. And he said, the risk is not that AI does the work. The risk is that AI does the work and we never learn from it. If every interaction with an agent happens in a private window, the only person who learns anything is the person at the keyboard. Everyone else is locked out of the apprenticeship. He said, the right model is not human or machine. It is the apprentice and the master both watching each other learn and both getting better on the shop floor. So that idea of like open communications with agents through Slack and things, like that is a very, very different mindset for people. And then a couple other things I noted, what employers can do. It wasn't a hell of a lot. Like, they're, I mean, they just seem to be accepting that we're just not going to need people.

That was my basic takeaway from all this is like, hey, we're going to, like, build this massive business and we're not going to need very many people. And we're just going to tell you what it looks like when we get there. But they're not trying to stop from not needing a lot of people. Like, yeah, the reading these patterns forward and out across the economy suggests that AI diffusion could risk worker displacement. Prior work also finds that AI-native startups and AI adopters are flatter and employ fewer entry-level employees. Anthropic is standing up an ongoing early warning system to monitor jobs facing AI-driven transformation. A series of role-level metrics for demand and productivity are read against AI capabilities in use. This is exactly what I tried to do with JobsGPT. I tried to build like an exposure key and then project out the impact on jobs. This was like three years ago I built that. Building systems to flexibly reallocate workers internally will be critical for AI-intensive organizations. Commitments to reallocate incumbent workers, which is actually what you talked about with Craig Karmazin and Ty. Right. This is what they're doing. Commitments to reallocate incumbent workers could make employers hesitant to take a risk of hiring new employees in normal economic conditions,

particularly intensively competitive ones such as commitments to intentional risk could be costly. but prioritizing reallocation forces AI adopting companies to consider ways that AI can create new tasks for new products and new markets rather than funneling AI applications and replacing human workers. So this is what I've been preaching, like growth and innovation. You have to focus on growth and innovation. Otherwise, we're going to decimate workforces. So we have to grow to be able to keep the workers we have and add new ones. And so none of these actions will prevent powerful AI from disrupting jobs and indeed none directly address the slowdown in junior hires documented in these reports. So then back to my MECON session, there was a section in the presentation where I said, here's the things like I believe with a high level of confidence. So AI labs will continue to accelerate progress toward AGI, and nowhere in these papers does it talk about pacing the impact on workers. So they present this as like inevitable. We're going to build the smarter models. They will take the jobs of human workers, and we won't need as many.

They literally are just saying, like, this is what's going to happen. Yeah. I said AI will get smarter, more generally capable, and more reliable at completing long-horizon tasks. Some work, maybe the majority, will be automated by agents. Work is going to compress. Teams will consist of humans and agents. We'll need fewer humans doing the same amount of work. Productivity will increase, but management's capacity to review and approve work will be blocker. That's the answer I talked about that exactly. New roles will emerge, but it's not clear to me how those will be safe from automation as well. human in the lead versus human in the loop is going to be a friction point for organizations that want to move faster. So if you're anthropic, you want to get the humans out of the lead because they're just slow. They prevent things. And then the skills, traits, behaviors, and mindsets that employers will have to adapt. So the other thing I said in my talk was like, what kind of organization economy do we actually want to build? So I'm a big believer in a human-centered approach. So the goal should be to maximize the benefit to people, not to just generate greater profits and revenue per employee. So we should be growing companies and careers. Anthropic appears to only be in the grow

companies bucket here. I said, I want to hire as many humans as I can, but never more than I need. Cause I don't, I don't want to like end up a year from now with people that are obsoleted and we don't have new roles to move them into. And then I said, we have to like invest in this next generation of workers. And so I don't know, like I have a, I love this research, but I have a major problem with the fact that Anthropic seems to just be like hands in the air, like, all right, well whatever we're just not going to need people and then they're like building the technology to obsolete them i don't understand from a society perspective how that is like an acceptable future for people and i do believe we still have agency in this but the labs aren't going to shut it down and then the final note mike is like though again the way it seems like all these labs are going to do this is they're just going to like give money away so they announced i think this was sunday if I remember correctly, that we're committing a portion of Anthropics shares to the Anthropic Public Benefit Fund. This is the post from Anthropics separate from these papers.

Subject to stockholder approval, 5% of Anthropics shares will be committed today with an additional 2% added to the board as the board's discretion for every 50% increase in the company's value. We might explore large donations to traditional nonprofits or direct distributions to specific populations. One example is labor market impacts. AI may generate enormous economic value, but it could also cause instability for many workers. Anthropic has already committed $200 million to studying and addressing this issue by giving people money. And then, like, this morning, Elon Musk tweeted that he's going to give one share of SpaceX stock to every child in America. And it's, like, one of many. And so, like, this isn't the answer. They're all going to build AI, recursively self-improvement AI. Yeah, within like one to three years is better than humans at every cognitive task. And then they're going to just give us all dividends from or stock from the AI that takes our jobs.

It's wild. Like, it truly makes no sense. Like, at Macon, we had Andrew Yang's keynote where he was at least presenting a concept of, like, how do we solve for this where we still provide human purpose and meaning? Because the labs seem to be missing that part of it. So I don't know, man. There's like big topics to start the week and especially coming off Alaska, my brain is fried. But I'm just increasingly like becoming very frustrated with how the labs just assume this all works out. And they just seem to be OK with just building things that are just going to obsolete people. And if they just give everybody money back, it'll all just be OK. Yeah, I couldn't agree more. I would say, I don't always say this for topics on the podcast, but if you have time, I would highly encourage you, if you have one resource you engage with this week, this How Employment is Changing at Anthropic that Paul just described, go read it for yourself.

put it into your AI tool of choice, spend time engaging with it, and just ask yourself if even a fraction of this became true, what does that mean for my career, my company, and our people? I think it's probably the highest value, highest signal thing you could do for the long-term trajectory of your career this week. I agree. I think even the companies that are on the frontier of AI adoption, like at the scaling transforming phases of integrating it yeah they're probably 18 months behind what anthropic is describing today of what anthropic is doing and so if you're in like you know health care financial services banking law like whatever like if you're in an industry where you're way behind this may seem like a parallel universe to you like there is no way this is really going to happen in my industry and it might be three to five years before it looks something like this, but companies are going to be run very, very differently.

And it's just going to be happening behind the scenes. You're not even going to be aware it's happening. So, yes, this is a very valuable thing to go read. All right. Third main topic this week. Let's talk Macon 2026. We've alluded to it a couple of times, but it wrapped up. Our conference wrapped up this past week in Cleveland. If you listened to Episode 246, you heard us recording AI answers from the conference, like Paul mentioned. And now we wanted to take a little time to look back at the conference as a whole. So our program included tons of sessions on things like building AI agents, redesigning marketing work, leading organizational change. Like Paul said, he had his keynote that explored the architect, the orchestrator, and the apprentice as these different roles in the age of AI and agents. I did some sessions including talking through AI native content teams and discussing AI transformation with the leaders at Good Karma Sports. Now, some of our headline speakers included journalist and author Karen Howe, tech and economic futurist, as well as former presidential candidate Andrew Yang,

and best-selling author Kevin Roos. Their sessions covered the company shaping AI, the future of the economy, and what happens as we develop increasingly advanced systems. As a note, and we'll reemphasize this in a second here, for anyone who missed Macon, There are recordings of the main stage and featured sessions available for purchase as an on-demand bundle. You can find all that information at macon.ai, M-A-I-C-O-N.ai. So, Paul, we wanted to maybe just chat through some of the highlights you found from Macon, kind of takeaways from the conference. Curious what you recapped and kind of understood and learned from all our incredible speakers and from all the stuff that didn't happen on the stage. tons of great conversations I know we talked to so many amazing podcast listeners it was awesome yeah and like yeah I had a chance to hang out with Karen backstage and Kevin Roos and Andrew Yang and just like great people I mean Andrew I didn't realize we were getting a comedy act when we booked Andrew for the main stage I mean he zero slides I had no idea what Andrew was going to talk about

because he just like gets up there and you know literally does a comedy act for 45 minutes super insightful. And the whole reason I wanted Andrew Yang on stage was because he at least has ideas of what to do. And that's very rare right now. Too many people talk about it just to be a future of abundance. We'll figure it all out and whatever. He actually has some concrete ideas that he's been playing around with. So I love getting to talk to him. Kevin Roos was awesome. I really enjoyed that conversation. I've been a big fan of their podcast for a while. The AGI Chronicles looks amazing. I've got the book. I'm going to probably be reading it on my flight back. So overall, just, yeah, it was incredible. So we started Macon in 2019, three years before ChatGPT. So I think, like, one of the big things that has really evolved for me is back then, like, we were looking for these very narrow talks. Like, you think about the exhibit hall back then, it was all these vendors who have built task-specific things, like website SEO, email subject line writing, predictive models for ad spend and campaign strategies,

translation, transcription, image recognition. Like these were the kinds of things in 2019 that we were talking about, that we were featuring at the show. And we were trying to convince people to pay attention and then help them understand how I was going to change the world. But it was like it wasn't nearly as tangible. So that first year had 300 people from 12 countries. This year we had nearly 2000 from I think it was 25 countries. I know I met attendees from as far away as Thailand, Dubai, Australia, New Zealand. So a few things that stuck out to me this year. the non-marketer audience at Mekon keeps growing. In 2025, I think it was around 20% weren't marketers. They were more business leaders and even just other disciplines. This year, I know it was higher. It was more of like instinctually from conversation. I don't know the final number yet. The audience's AI maturity is becoming increasingly diverse. Now, this is a challenge and an opportunity from an event organizer perspective. there are a lot of people that are still getting started on their ai transformation journeys they're

early personally they're early organizationally and then there are absolutely people there who are on the very edges of the frontier building agents into their workflows and org charts like we heard from holly chen gave a great talk what they're doing at samsara um dan slagan gave an amazing talk about what's going on zapier so you have these people that are doing these incredible things and i know there are people in the audience that are like way more advanced um so we have to solve for that as we think about next year's programming like how does that look different one of the other big takeaways for me is people crave the human connection and the support of their peers now more than ever um so for me somehow as this event keeps growing from the 300 to the nearly 2,000 that sense of community and camaraderie seems to grow as well it's an amazingly like generous thoughtful supportive and humble group of people and again I don't know how it is that like everyone you meet just seems to be how can I help like they're super thoughtful and caring people which is an amazing thing to be able to bring those kind of people together

and I do feel this shared sense of like we're all in this moment of excitement and uncertainty and the other thing that you know I think increasingly we let off with topics about today is this there are uncomfortable truths about this technology and this was maybe like the thing that stuck with me the most. So the MAKON theme since that 2019 inaugural event is more intelligent, more human. So I've always believed that we needed to be, to not only present the opportunities to grow companies and careers but to raise awareness about risks and potential negative impacts as well That being said it can still be jarring to go from these inspiring talks about agents and AI transformations to insider stories like from Karen about IP theft and the extractive behavior of leading AI labs. So is intro just like a little inside story here? So I was out to a dinner. Was it Wednesday night after day one? with Intercept, who's an amazing partner and sponsor of ours.

And I got Karen's slides at about 9.30 at night. So I got a message from Ashley, our head of events, that, like, here's Karen's slides. And so I stepped away from the dinner table, and I scanned the slides, and I thought, oh, no. Like, I read Empire of AI. Like, I know what's in that book. Yeah. But I wasn't sure how far we were going to go with it. And once I saw the slides, I realized, like, wow, we're going to go there, like the dark side of all this. And so I knew that was the first talk the next morning. So we're going to come in this amazing night out. Like, everything's like, everybody's having a great time, great energy going in. And I'm just going to, like, put Karen on stage. And it's going to be like, whoa, like, this is really heavy. So I went to bed that night working on, like, what are the questions I'm going to ask Karen? Because I had to do, like, a fireside chat with her for 15 minutes at the end of her talk. And like, how do I bring this back? And then tee it up for you because you had to come on stage after that. Yeah. So I got to bed and then I woke up at like 430 in the morning and I was like, I have to provide context to the audience about this.

So I actually wrote at 430 in the morning the intro to Karen's session. And here's I'm just going to read you verbatim what I wrote. I said, there is some uncomfortable truth to how this technology is created in the negative impacts it can have on society. We cannot take a responsible human-centered approach to AI without understanding the human impact behind it. Some of the things that Karen explains in the session coming up might make you uncomfortable if it's the first time you're hearing them. You may have employees or coworkers who are resistant to using AI potentially because they read Karen's book or have heard about the darker side of how the technology works. We hear these questions and concerns all the time around the economic, environmental, and human impact of AI. So while we can absolutely be excited about technology and optimistic about the future that it will create, we can't pretend like it doesn't come at a cost. Now, for people who were there, they could tell you, like, I got a little emotional when I walked out to interview Karen. And the back story here, just because, I don't know, people were asking me,

and I think it's an interesting human side of this, it was a very powerful session. I knew what was in there. It was a risk to lead off the second day with it. I got up early that morning. I'd written this kind of intro to sort of figure out the context. But my daughter was in the room and she's 14. So she had surprised me that morning and showed up to volunteer. She didn't have school. And while I was sitting backstage as Karen was on the stage doing this talk, my daughter texted me and said, when are you going to be on stage? And so I told her I was going up in a few minutes. So just head to the back of the room and then, you know, I'll be up there interviewing Karen. But then I realized that she was about to walk into this like insanely heavy session that was talking about some pretty dark stuff. And so I had a lot running through my mind. And so I actually texted my daughter that that intro that I just read that I had written as like, hey, just FYI, here's context as to like the session you're about to walk into. So she understood what was going on.

And so as I was about to walk on stage to interview Karen at the end of her talk, I'm like three minutes away from having sent this to my daughter explaining what's going on while thinking about the reality of it, like what's happening. And I get on stage and I look Karen in the eyes and I just like I couldn't talk like I could feel like welling up in my throat. I was like, yeah, I'm getting emotional. And then Karen looked at me and then she started getting like teary eyed. And so I was like, oh, my God, like it was just like total unplanned moment where like neither of us knew what to do or say. And so it took me a moment to sort of get my composure and then like ask her. And plus, I think there's just this amazing bravery to like what Karen's doing. And I think I just saw like my daughter and her in some ways, like where you're trying to like, you know, raise someone who has the strength to do the kinds of things she's doing. So, yeah, it was it was a weird, very emotional moment for me. but incredible interview, incredible session. And then I shared on LinkedIn, I think the thing that kept coming back to me

was just the gratitude for it all. Like the staff, the volunteers, the speakers, the sponsors, the attendee. It's an amazing privilege to be able to bring people together and to see it just keep growing and to see all of the things that are happening off to the side, like all these incredible partnerships that are being formed, friendships that are being formed, these groups of people who are now meeting up every year and getting together and then to do it all in our hometown of Cleveland and get to showcase our city to people. It's just very, very grateful that we get to do it, and it's an amazing thing to do every year. So, yeah, I just appreciate everybody who came out. And certainly our podcast listeners, it was a fun lunch we got to do with them. Yeah, just great stuff. And we'll be back next year, so October 4th to the 7th in Cleveland is MECON 2027. so yeah if you if you don't have it on your calendar already we'd love to see more of you next year it was um it's a special thing i hope we can you know keep doing it keep growing it and i did get asked like when does it get too big to where like you don't have this feeling i was like

i don't know it's just like every year it grows another 20 30 whatever and every year it's just like more amazing people show up so just keep going see where it goes i love that i appreciate you sharing that. I think it's really important for people to hear. One sort of practical thing I'll mention is some of the takeaways I found from this were, you know, thinking of just like, oh, okay, wow, interesting how behavior and our use of technology is changing. So one, you know, Paul, you and I discussed this, that a huge amount of people, at least relative to before, found the event for the first time through AI, which was an interesting trend, including people outside of marketing. So because we were being recommended as an AI, which I think is very accurate in some ways, is there's plenty for non-marketers. So they weren't, you know, missed the report. And international too. Like I had multiple people international. I said, how did, like, why? Why did you travel 30 hours to be here? And they just, so I just asked GICBT for like, what are the best AI conferences in America?

That's trust. I was like, incredible. Like that's amazing. Yeah. And on top of that too, I think also we learned that, You know, we had intentionally designed the web app experience to be a web app versus a phone app, which made it legible to agents. I got feedback from folks. They were taking all the session notes and slides and putting those into AI and to Notebook, Gemini Notebook. So no surprise there. But just incredible. Gemini built into the app so you could actually, like, say, hey, I'm here for this. What sessions should I go to? And it would build you, which is awesome. We've been dreaming about that for seven years. It was cool to have that capability. Yeah, so just some quick logistical things there I thought were super interesting is how AI is kind of changing the conference experience. All right, before we dive into RapidFire, just to reiterate, this episode is supported by OutShift, which is Cisco's incubation engine for frontier technology. So multi-agent AI is in every enterprise roadmap right now, but here's the problem we're all facing. Agents can pass messages, but they can't think together.

So these systems silently underperform through cognitive failures like misreadings, unverified claims, and false consensus that conventional monitoring cannot detect. Now, OutShift by Cisco is building the fix, the Internet of Cognition. And it's an open source foundation for building multi-agent systems from design to production with shared context, shared memory, and guardrails to drive the results we actually expect from AI. So you can go read the paper, experience the demo, and grab the code from outshift.com. That's O-U-T-S-H-I-F-T dot com. This week's episode is also brought to you by the MAKON 2026 On Demand Bundle. So if you couldn't join us in Cleveland or if you didn't want to watch the recordings of what you saw, you can now watch recordings of main stage and featured sessions from this year's Marketing AI Conference on your own schedule. So this bundle brings together the big picture conversations like the ones we just talked about, about where AI is heading, as well as practical sessions for how to put it to work in marketing.

You'll hear from Karen, from Kevin Roos, from Andrew Yang, alongside practitioners tackling the questions teams are working through right now. That includes people like Will Reynolds, who gave a talk on visibility and credibility in AI search. Andy Crestedina, who gave a talk on turning a marketing process into an automated tool. We have the incredible Chris Penn on using AI for analytics and Taylor Rady on our team using AI for research workflows, as well as, I believe, my AI native content team talk. This also includes Paul's session on rethinking the work of rethinking work in the age of AI and agents. So go to Macon dot AI, M-A-I-C-O-N dot AI. There's a button right there in the header to purchase Macon 2026. records. More than 12 hours. I was looking at it over the weekend. It's like more than two dozen sessions in 12 hours. So yeah, it's insane. An insane amount of value packed into one thing and it's available immediately. So take advantage of that for sure. I can't wait.

I have to revisit a bunch of it. I'm super excited. All right. Let's dive into Rapid Fire, Paul. So first up, we've talked a bit about your new book, The AI Transformation Blueprint. We've talked about the software where you've built AI Score and how quickly this has all come together. What's exciting is listeners can now get their hands on the book. It is available for purchase, and AI Score is now actually in public beta as of MayCon. So just as a refresher, the blueprint is your guide to leading organizational change with AI. It brings together a business blueprint organized around eight pillars of transformation, a personal blueprint organized around five. AI Score is the new assessment platform, the software you've built for companies and individuals. Business AI Score is one of the assessments you can take where you can rate your organization across dozens of different statements about AI transformation. You can do the same thing for your personal AI score to help individuals assess themselves. Both of them produce an overall score

and grades of how that score compares and what you need to work on in different pillars of transformation. So the assessments are free to take. The book you have to purchase, you can go to smarterx.ai forward slash blueprint for the book. And try AI score at smarterx.ai forward slash score. Go check all that out. Paul, I'd love to just get your take on kind of where these are today. We had a bunch of announcements during Maycon around these. Like, what do people need to know about the book and AI score? Yeah, it was kind of wild. So my opening talk was Wednesday morning. I went to bed that night thinking, well, one, how was I going to introduce Karen? but two, wow, we launched a book and a software product this morning, and it barely even crossed my mind that it had happened. Yeah, I mean, I told the story of how AI Score was built, but in essence I spent seven months developing these assessments and the prototypes of the software, and then about 45 days before Macon, I realized I could use them as the foundation for a book, so I green-lighted the book around September 5th of 2026,

and then I finished the book, including all the edits, September 17th. So thanks to our internal team who edited all the work. And then the team got to work, specifically Tracy and our team, getting it ready for publishing. And so there's a mad sprint in the midst of everything else we were doing with Macon to get this book ready to go. It's about 28,000 words, I think, if I remember correctly. So it's roughly half the length of the last book that we wrote together, Mike, that you and I did, the Marketing Artificial Intelligence book. I would say it's about a three-hour read for the average reading speed. It's meant to be shorter, more of an executive guide, I guess. And then a similar story with the iScore, I built this fully functional HTML prototypes of these different scoring assessments, personal and business. But I didn't actually know how to take the prototypes from vibe-coded HTMLs to like enterprise-grade SaaS product. And so I found a software development firm at the beginning of September and we met, went through the prototypes and they're like, yeah, we could, we could do this. So they turned my prototypes into these amazing products in about 30 days. So as I was like,

I didn't have a ton of time over the weekend to like reflect on all this, but I think it's safe to say that the 45 days with the book AI score and then prep all the prep for Macon was probably the most inspiring, exhausting and creative sprint of my professional life. Um, yeah, I'm still kind like in shock that we actually pulled all that off. So yeah, the book is now available in print and digital. So you can go to, what is it? SmarterX.ai forward slash blueprint, and you can buy the book direct. You can go on Amazon and get it. And then the, uh, you can get the ebook versions of it. I may do an audio version this year. I honestly, like I haven't had time to even think about it, but I would imagine we'll probably do an audio book. And then AI score is available in public beta today. So you can go to score.smarterX.ai. You can take the business AI score for your organization. You can take personal AI score. You can create a business account and invite teammates and then see how they rated across it. And then as a team admin, you can have visibility into that. So it's just like I would say we did a soft public beta rollout of AI score at Macon.

November 13th, we're going to host a webinar called AI score for business leaders, where I'm actually going to walk through a more full blown rollout of what it is, how it works, how you could use it to drive change management and transformation within an organization. So I would say check that out November 13th if you're interested. And that'll be kind of the official public beta rollout. We are still very much in that public beta phase and kind of monitoring, you know, how people are using it. And then we have some pretty exciting plans for what else we're going to do with AI score. But I'll save those for another day. Yeah, that's plenty to be going on with in the last 45 days. Yeah, that's awesome. Yeah, that's incredible. I mean, just talk about hearing what the anthropic employment research is saying, putting it into practice in terms of, like, look what's now possible. Yeah, and I think I used AI Score as an example. When I said work will compress in my opening keynote, this is kind of what I meant by it. It's like a senior leader can do the work of an entire go-to-market team.

So, like, for AI Score, our marketing team had their hands full with Macon. Like, they had other things going on, so I just took ownership of the entire thing. I built the product. I built the plan for the release of the product. I created the webpage and the FAQs and like all of it was me in the midst of everything else I was doing. And that's what I mean by like work compresses to where an individual person can do the work of a five or 10 person team in a very short time period. And for us, it was additive. Like it created these new things that could create new jobs and new growth, but it was done in this very compressed way. And so that's what I mean when I talk about like using as innovation and growth. Do something new, open up new markets, launch new products like that's the opportunity here. Don't just think about like saving costs and doing things more efficiently so you don't need as many people. All right. Next up this past week, we got an article from AI researcher Nick Camerata that sort of started to go a bit viral. It is called Neural Networks Are Already Clean, and it's basically a pretty technical essay

arguing that understanding AI's inner workings may be much more achievable than researchers have assumed. So the kind of broad field we're talking about here is called mechanistic interpretability. You may hear this called interpretability. It's basically what AI researchers do when they study the components inside these models and how they interact to produce their behavior. So Camerata himself actually helped pioneer some of this work at OpenAI, but him, like the rest of this field over the last couple years, became kind of pessimistic because the individual neurons in the models appeared to be very, very hard to actually interpret. They're all tangled up. It was very difficult to actually say, okay, well, why is the model behaving in a certain way? And what he's saying actually now is that with better tools, including specifically more automated AI agents and AI researchers, researchers, human researchers can now actually study the models existing neurons and connections directly.

They don't have to entangle all this stuff. Basically, one example he gives is, look, there's a famous technical example where there's a neuron in models, in LLMs, that is thought to respond to both the concepts of cats and cars. And people are like, well, why does it do that? We have no idea. Camerata actually argues it is detecting a B-shaped pattern. And the whole point here is like, even if you don't know what that means, the point is that all these like unrelated concepts that have been befuddling researchers may actually be very explainable now that we have better tools to use at scale to actually look at what's going on under the hood of models. So Camerata says that there are now understandable neurons in models with up to trillions of parameters, which are kind of big frontier models, and basically hopes that we actually are in a place where AI-assisted investigation can scale up and help researchers understand the models relatively fully. So, Paul, we don't have to get too deep into the technical parts of this, but the idea here is that interpretability has historically been something, a really, really hard problem people aren't sure we could actually solve.

Not knowing how the models work leads to all sorts of the safety issues we've discussed, but it sounds like Nick is saying that we might actually be at a place where it is understandable, and if we can actually scale up the engineering around the research itself, we might be able to solve this problem. What did you think of this? I think we both kind of like perked up when we saw this without, you know, being engineers and researchers ourselves because like, oh, that's like a gold holy grail problem in AI. Yeah, this surfaced, I don't know if this was like Saturday morning or Friday. I don't remember when I saw this pop up, but, you know, it was one of those where I looked at it. I was like, that seems like an interesting article. I'm not sure who that guy is. So you click through. He's like, oh, okay, he came from OpenAI or wherever. He booked at the labs like this legit researcher. start scanning it thinking, huh, this seems like a big deal, but I don't know. Like, this isn't, and we're not the experts in this. And then within an hour, like, everybody in the iSpace I follow is between thinking, this is a really big deal. And that's, you're like, okay, let me go back to that article.

Let's throw this in the sandbox. Like, this is probably worth talking about. Yeah, I would say this is one you probably put a pin in it, and, like, let's come back in six months and realize, like, okay, that was a really big deal. It feels like it's heading in that direction. And I think you summarized it, Mike. It's just it's hard to know why these things do what they do. It's been for the last 15 plus years in deep learning, this idea that, like, if we could understand how these neural nets work and why they make the decisions they make or lead to the kind of decisions they make, it would be on par with, like, being able to look into the human mind and understand why humans make the decisions they make. It's that level of being able to drill into those individual neurons that fire and why do they fire and what does it mean when they fire? And so, yeah, it certainly seems like promising direction that a lot of people in the AI world are excited about, that now you need to really push on this research. But I'm sure if it's as significant as it seems like it could be, every lab is going to be building on this research, and we'll probably be hearing a lot about it over the next 6 to 12 months.

All right, next up is kind of the developing story of OpenAI's research into math problems. So OpenAI has now released hundreds of AI-generated mathematical manuscripts and supporting proof materials about a bunch of problems they are making progress on solving. So we've talked about how OpenAI was actually solving millennium-prized problems in math, so very, very hard problems in the field of mathematics. This work now comes from an unreleased internal model, which OpenAI told Wired it had used to resolve more than 100 longstanding open math problems. Its public GitHub collection now contains 719 mathematical manuscripts grouped into 372 families of related results. Many of these proofs come with versions written in Lean, L-E-A-N. We talked about that when we talked about the Millennium Prize. This is basically a language that lets computers check mathematical arguments. It's basically a way to say, hey, these are verified proofs.

OpenAI says the results are at different stages of verification. Some lack formal proofs. Some unformalized results could contain errors, but the whole point here is there's like backlash from mathematicians because they're really upset, at least several of them are, about how this work was released. So mathematicians told Wired, they had asked OpenAI, hey, release these results with carefully explained papers that they could absorb and build on, similar to kind of how top mathematicians and researchers release stuff in their field. Some people said that the company had promised not to release everything at once. OpenAI, however, says it is unaware of who made that assurance. OpenAI says it consulted an independent advisory group at the Institute for Advanced Study to shape this release, including the procedures for corrections and citations. But the Association for Human Mathematics, in a statement republished on Terence Tau's blog,

is a famous researcher in Voice in Space, basically argued this is like the totally wrong way to do this. This is not good for the field. This is not good. OpenAI is doing it. Flooding the field with 700-plus files undermines research centered on human understanding. And it came straight out and said it urged mathematicians to stop working with AI and said this release was not a demonstration of scholarship but a demonstration of power. So I think, Paul, if you had to draw up a list, mathematicians are top of the list of having their 37 moment in from go. It's like, oh, shoot, AI can start doing some things top mathematicians are doing. And there's some real uncomfortable feelings around it, it sounds like. Yeah, it's hard to watch. I mean, I understand the perspective of the mathematicians. They spend their entire lives working on PhDs, multiple PhDs to solve single proofs.

And it's like it's really, really hard. It is like watching AlphaGo or, you know, with Lee Seedol when he just like realizes it's just better than him at the thing he was the expert in the game ago. Yeah. So I certainly empathize with the mathematicians. And I think like anybody could step back and say like, well, what if everything I did for a living was just solved and they just put it all out there? Yeah. So I get the human condition of just this isn't what we asked for. Like there's a better way to do this. And I also get in this case, you know, mathematics is the basis for solving a lot of bigger things. and if OpenAI believes they're sitting on solutions to things that could advance science and medicine and technology, you know, by decades potentially, why would you sit on it? Like, why wouldn't you put that out into the world? So I totally understand both perspectives here. I don't really have a strong opinion.

I certainly wouldn't come to OpenAI's defense about, like, how they communicate things and whatever. They don't have the best track record of being empathetic towards humans. So I'm sure that they could do that better. But, you know, if you ask me, should we sit on scientific discoveries or advancements in mathematics? My general answer would probably be no. Like if we thought we had answers to diseases and cures for things like cancer, like would we sit on that for years? No. Like the whole point of what we're doing, the only argument for this advanced race and building these larger, more powerful, more intelligent models is for the good of humanity. And so if solving all these math proofs gets us further along, why wouldn't we put it out into the world and like keep moving? That doesn't mean that there isn't an impact on humans that we shouldn't be sensitive to. So, yeah, it's hard. It's hard to watch. It's hard to see their response because you can tell it's a very visceral response from the math community towards this advancement and specifically towards OpenAI.

Next up, we talked about on our last weekly news episode, President Donald Trump's executive order rebranding artificial intelligence as super intelligence. That executive order directs agencies were legally permitted to replace AI with SI and official communications, websites, reports, and policy documents, and says the executive branch will not acknowledge the old terminology in applicable settings. Now, for now, superintelligence covers the same technologies as the existing legal definition of AI. However, the story here this past couple weeks is that major tech companies are adopting the language. So Microsoft CEO Satya Nadella published an essay recently titled Models as Insider Risks in the Super Intelligence Era. He repeatedly uses the term super intelligence to describe today's systems. Salesforce CEO Mark Benioff has announced a full rebrand of AI Force which is their product that lets people and agents work with Salesforce data and workflows And he wrote the era of super

intelligence is here. AI Force is officially SIFORCE. Elon Musk said on X when he was asked about renaming SpaceX AI to SpaceX SpaceX AI, which is a mouthful. He replied, yes, we will make that change. The tech company ServiceNow is also describing itself in official announcements as the super intelligence control tower for business reinvention. So, Paul, this is I say sarcastically fun to watch because we're just apparently updating branding materials. Just woe to the marketers. They have to go back and update this across all these websites. Yeah, I don't know. I said my piece on this whole thing on episode 245. So if you didn't listen to that and want to hear my opinion on this, I would say just go listen to that segment. Yeah. I understand that all these tech leaders are in a tough spot because many of them have large government contracts

with more government contracts looming. Some of them have IPOs that they need government support on. Some of them are under litigation from the federal government. The day after they announced the accord, if I'm not mistaken, the government came after OpenAI for a number of things. And they have other leverage points that the administration can hold over them and has proven they will happily use legal leverage and other leverage points on people. but doesn't make it any less disappointing to see them all just rolling over on this. It's sad. I will say on a positive note, we just spent four days at Macon with nearly 2,000 people and I did not hear the term superintelligence two-word version once over four days. And that was encouraging. So, yeah, I'd be okay if we never had to talk about this again. But I'm sure we will. So, yeah, I don't know.

I'll be anxious like two years from now, four years from now, whatever, when all of these tech leaders have to answer for the fact that they just rolled over on this stuff. And I can't wait to hear their responses. No kidding. But, again, I get it. They have fiduciary responsibilities to position their companies to continue growing in value. and sometimes you have to play a game that you don't want to play and some of them are playing a game that they want to play. Yeah, that was just out there. All right, next up in our last news episode as well, we had talked about Google holding back Gemini 4 Argon. Well, this past week, Google broadly released the model. So Gemini 4 Argon is designed for sustained multi-step work in coding, finance, law, cybersecurity, science, and more. Google says Argon combines stronger reasoning, visual understanding, and writing with the ability to analyze complex documents and data.

It can work through Google's apps and coding tools to do things like reorganize large software projects, find security vulnerabilities, build interactive visualizations. For example, Google actually had this kind of cool list of how Argon is helping their company actually make progress. So they talked about how the Chrome team completed months of software restructuring in weeks. how isomorphic labs used it to help extract research information from complex PDFs and patents in hours instead of weeks. Gemini for Argon is now available through Google's paid API and currently for ultra and pro subscribers in the Gemini app. In an announcement around the same time, Google Cloud also came out with a post about what they call, quote, the Gemini agent, which is Gemini capital agent, not capitalized, which is basically a universal work agent. A bit confusingly, Gemini is the agent's name, why you can actually select different underlying models for this, including Google's models and Anthropic Cloud.

Google has described this as an evolution of Gemini Enterprise, which I'm not totally clear on. You can still basically chat with this, but also, like, delegate entire projects to it. So it sounds a bit like a chat should be pretty work. I don't want to put words in Google's mouth, but kind of a more agentic version of Gemini, just natively in the chat. And the agent will plan and execute work across connected apps that can remember context and it can keep running in the cloud after you close your laptop. Now, none of this is actually publicly available yet. The announcement gives no public launch date and does not announce any type of like replacement for anything here. But we did want to mention it because it's a fairly big deal, Paul. It dropped right before my Gemini workshop at Macon, which is interesting to navigate. But, you know, it's more of an announcement than a feature. So I'd be curious your thoughts on both these, like Gemini 4 Argon. What's the response here to this? It's been very long anticipated. And then maybe the Gemini agent thing, I'd be curious how you view that at the moment.

So I do, I just checked, I do have Gemini 4 in my personal Pro account. Yes. I have it in my Gemini Enterprise account through SmartRx. I do not have it in my normal Gemini account because the way Google does things is you have a separate enterprise account from your normal. And so it doesn't just transition your existing business account. It creates a whole new business account. Yeah. So I have multiple logins. I haven't had a chance to experiment with it. It didn't. I mean, granted, we were in Maycon all last week, so I don't know that I got the usual pulse on things. but I would say this model probably has less buzz than most whole number releases from the labs at the moment. Like I haven't heard as much from the online community about this model. I did hear a lot of takes on like Google reannounced Gemini, like as like the agent part. People were kind of confused of like, what exactly is it?

Is it the model? Is it the agent? Is it the harness? Is it the whatever? yeah so that being said I haven't had a chance to really dive into it or see a ton of response to it so we'll definitely be digging into it and kind of running it through our evals internally I still use Gemini quite a bit both the app itself and certainly within Google Workspace platforms because that's we run on Google Workspace stuff so yeah I don't know I'll be interested to experiment with it see what see what people have to say see if it makes a difference to some of the higher strategic projects that I tend to work on. I would love to experiment with it there. Yeah, for sure. Google has some fantastic AI products. I mean, the workshop we did was super helpful for people, lots of people getting real value out of these tools, which is super encouraging to see. Maybe it's just me, but a plea from me is, make it easier for your educators to educate Google, please, because describing how it all fits together takes some serious work.

Yeah. Yeah, it's getting challenging. All right, next up, there is a company getting some buzz out here called Micro One. It is a company that supplies human expertise and training environments for AI because they have committed to spend a billion dollars over the next 12 months acquiring and licensing companies' operational data. They say this initiative is financed with capital from Citi, a Citibank, and Hercules Capital. And as part of this, it wants the records of how businesses actually work. So things like their internal instructions, CRM data, project histories, the processes people use to make decisions. So examples here include customer support workflows, financial reconciliations, contract reviews, et cetera, all the stuff companies do. Micro One says it turns that de-identified data into reinforcement learning environments, which are essentially practice settings where AI models and agents can learn to do those business tasks. And the goal here is to capture the messy parts of work, including incomplete information,

competing priorities, and exceptions that require judgment. So their pitch to businesses is, hey, you already have all this stuff. Why not turn it into a new revenue stream? And its website advertises partnerships worth over $100,000 in many cases with a million dollars or more for particularly valuable proprietary data, though actual compensation depends on the data's quality, uniqueness, and value. Micro One says companies retain ownership of their data and agree on what can be shared, the privacy protections, and internal approvals before participating. Now, interestingly, there was a fairly long post on X from founder and CEO Ali Ansari, arguing that human data could ultimately become a trillion dollar annual market as companies train AI to automate more work. Interestingly, in one of his essays, he also just wrote straight up automation does not eliminate human work. It pushes humans towards higher value, more creative work. So, Paul, this is an interesting one. Like it seems like they are trying to buy company data to do what people at the company do.

Yeah, so just for context, this is not new by any means. I'm just coming straight out and saying you're going to spend a billion over 12 months. That's the first time I've really seen one of these people do that, one of these companies do that. But this is how it works. So if I want to take like a core model, it could be an open-weight model, it could be a proprietary model, whatever, and I want to go after a certain industry, let's say dental offices, I don't know, just pick something. The best way to do that is you go buy the operating data of how a dental office works. How do the dentists make their decisions? How does the office manager function? What do they input into the software? What do the conversations with customers look like? And then you use that as reinforcement learning to make the model specialized in operating within a dental practice. doing the same thing in consulting firms, marketing agencies, healthcare practices, like whatever, banks. And so they're basically saying, sell us your data so we can replace your people. Now they're going to write all these essays saying that's not what happens.

We don't replace people. You go to higher level cognitive creative stuff. It's all crap. Like they're just buying data to replace the need to have people working in these companies. and there will be some jobs left, but like you're basically selling out humanity when you do this. Like now the way they often do it is they go buy like bankrupt companies or like tech startups is a common one. So you can go buy like Slack messages out of these tech startups that don't make it. And so rather than just folding the company, you sell your data for a few hundred thousand dollars and pay some of your investors a little bit back. So, yeah, it is like it's wild. Like this is how it works, though. This is data brokering to automate human labor, industry by industry, one at a time. And probably not a surprise to plenty of our listeners, but my gosh, I feel like your average person, this is a whole shadow world they are unaware of. Yeah, and so you might get sold out as a worker and not even know what happened. So like your CFO and your CEO might do a deal with a company like this.

And now all of a sudden, everything you've put into your Google Docs and your Slack messages and your meeting notes is all actually being sold off without your knowledge to some third party who's using it to train or do your job. Damn. Yeah. Yeah. It's going on forever. This is what Scale AI does. I mean, this is Alexander Wang, who's now the head of AI at Meta. This is what he built. Like Scale AI was like a reinforcement learning company. And he did that through like proprietary data sets. So, again, not new at all, but very direct about how they're doing it. Yeah. Or talk about it. It's all fun in games when it's people labeling articles or images or whatever, not, like, the steps you do to do your job. They basically put, like, a We Buy Homes sign-up and said, we've got a billion dollars, we buy data. Like, it was a call. All right. Next up, we have our AI use case spotlight, where every week we give you a quick look under the hood at the real AI use cases we're exploring at SmarterX. So, Paul, I have one to share, and then we can go through anything you've got on your plate.

So this week, here's the problem I was personally trying to solve. So as anyone who attended Macon knows, I had several presentations, panels, speaking appearances to prepare for at the event. Now, I've been prepping for months, but I find in those days and hours before each appearance, you want to be really sure you're comfortable with the material. And for me, that means talking through it out loud. Now, I've done this in the past while sitting in front of a computer. That's fine. But I personally think best and frankly talk best and more smoothly when I'm up and walking and moving around. Something about that just kind of helps the words and ideas flow better for me. So I built a voice first AI rehearsal partner. And what it does is it actually walked me through my presentations one slide at a time while I was out walking. So it would remind me what was on the slide, let me practice what I wanted to say, help me memorize key details I needed to know well, and then capture wording I wanted to add to my presenter notes.

Now, not to mention, it could even make changes to my actual slides and notes on the fly while I walked around downtown Cleveland. So more on that in a sec. But here's how I set this up. So if you want to start trying to do something similar, go for it. So first, I gave Codex access to the slide deck on my computer. I wanted it to work from the actual slides and presenter notes, including the actual visuals, animations, and like timing and effects that I needed to remember. Second, I ran a skill I have in Codex to build a skill. So that's kind of meta. But the skill I built this time was like a skill for walking reviews of this material. And then third, I spent a good amount of time going back and forth with Codex to actually build this skill. Because, like, it was really quick and basic to get to, like, okay, here's a skill that helps you review your stuff while you walk. Like, super easy. It went through the deck in order, told me what was on each slide, and then provided advice as I rehearsed. But the key here is that it did take a lot of time and effort to calibrate all this to, like, how I work and my personal preferences.

I didn't know exactly what I wanted from this, right? out of the gate because I'd never done this in this very specific way. So I wanted to actually just like kind of sound out how to iterate on this and make it work for me. So I basically just created like an MVP of a skill and then took a long walk and used it using chat GPT voice through Codex because I was actually just talking to Codex, which means I wasn't able to just test the skill, I was able to literally stop mid-sentence and say stuff like, I actually don't like the way we're doing this. Do it another way instead. Oh, by the way, got it? Great. Go update the skill and let's take it from the top and run the skill again and iterate. It was basically out loud iterative development of the skill. It worked wonderfully. I really quickly got to a useful skill, which I then used time and time again over the last couple of weeks to get really comfortable with the material I was presenting. So very briefly here, Paul, and I'll turn it over to you. The tech that made this work

was ChatGPT voice on my phone connected to Codex on my computer. That all works through the ChatGPT app. On the app, I have Codex set up to remote access Codex instances on other machines. And then now it's a pretty recent feature. You can use ChatGPT voice just to talk with Codex. So you fire up the app, start a new conversation, and speak to Codex, and you can do some pretty cool stuff. I'd say it worked. It worked. That's all that matters is it's messy, but it worked. Yeah, the session was awesome. Thank you. We have a very different preparation style for our presentations. I know, I know. Mine is build the deck, maybe practice it once, and show up on stage and talk. There you go, right? I actually had a fun one in the midst of last week. So Monday morning, we started Monday at NACON. We had a council meeting, a council meeting that morning. And so we were launching AI Score soft launch on Tuesday for my workshop.

And I hadn't built the product FAQ page yet. So I was like, man, I don't know. I'm going to have to do this. So what I did is I actually went into the Claude project that I had over seven months. I've been working in the same cloud project to build everything for AI score, you know, develop all the assessments, build the prototypes, all that stuff. And so I went back into that project and I said, OK, we need a product FAQ page. Take all the context from all of our work together the last seven months. Here's the marketing page for it. And here are 80 questions that the team had for me about how this is going to work, what we're going to tell people based on questions that they're going to have, all that stuff. develop like a best in class fast product FAQ page for me based on all this information. And it went away for a couple of minutes and it came back and it just crushed it. And I spent about 10 minutes editing the FAQ page. I said, all right, great. Here's the edits. Give this to me in a Word doc and an HTML file that we could publish directly to the website. The HTML file looked amazing. I sent it to Jeremy and our team.

And like 30 minutes later, we had an FAQ page on the website that would have easily taken me 10 hours or more to do. and it was honestly like way better than what I would have created. So, yeah, you want to see it? Go to the FAQ page on score.smartrex.ai and you can see the output from product. Again, work compresses. Like what I traditionally would have had to involve multiple people in, I would have definitely turned that over to somebody else on the marketing team or the product team. If we had a product team, I am the product team right now. So, like, if we had any of that, they would have done it. And instead, while I was sitting in the team room at our conference, I just like spun up a chat and like knew to tell it to look at the context of everything and the final prototypes and like here, go do this. And it just crushed it. That's so cool. I love that. All right. Let's do some final product and funding updates. I'm going to run through these very quickly, Paul, as we wrap up here. A lot going on in the last couple of weeks. So first up, OpenAI began rolling out GPT-6 with something called Intelligent UI to free and paid chat GPT users.

This actually is a feature that enables answers with interactive diagrams, charts, forms, and tools directly inside conversations. OpenAI also introduced optional text watermarking for select API models and announced plans to add invisible watermarks to eligible chat GPT and codex text output in the European Union. OpenAI also announced something called College Planner for ChatGPT for teens to organize college applications and financial aid deadlines alongside flashcards, improved quizzes, and multi-photo note capture on iOS. Anthropic released Claude Haiku 5.5, its small model for fast, high-volume tasks such as summaries, customer support, and coding sub-agents. Anthropic also launched a public beta of Claude for Google Workspace on all paid Claude plans that adds a sidebar in doc sheets and slides and connectors for creating and editing those files from Claude. Anthropic announced a usage policy update as well, taking

effect November 12th, that prohibits sustained needless cruelty towards Claude in extreme cases. So basically how you are talking to it and what you are asking it to do. There's a lot to unpack with that one we are not going to get. My gosh. That ruffled some feathers. Oh, kidding. In the industry, let's say. Yes. Ding. Google, on their end, released Nano Banana 2.1, their updated image model in Gemini and Google AI Studio. There are improvements, they say, in visual design, editing selected parts of an image, subject consistency, and natural-looking imagery. They also released something called Deep Review, an AI system that checks complex documents for logical consistency and technical errors. You can now get two free reviews per week and higher usage through a Gemini API key. Paul, I think this is one you've been playing around with. They now have Google Playground, which lets users discover and play and create community-made games, so like video games, computer games,

that you can actually create through simple text prompts, which seems pretty cool. That was pretty wild. It's in labs, but my son and I actually did it, and it is literally just like create a prompt of anything you can imagine game-wise. Yeah. And then you just see the gameplay in the module to the right, and then you can just iterate on it. It's pretty wild. That's cool. Google also began rolling out native Markdown editing and collaboration in Docs along with formatted previews in Drive so users can work with Markdown files without converting them to Google Docs. And the company also opened its Synth ID detector globally in English, allowing anyone to check images, video and audio for AI watermarks from Google and participating partners. And overall parent company Alphabet's AI drug discovery company, Isomorphic Labs, is reportedly discussing new financing at a valuation of at least $40 billion. and Isomorphic Labs has joined the Biohub-led Virtual Biology Initiative as a founding member,

which means contributing to an effort to make biological data available for AI models that predict and help treat disease. Meta has introduced Muse Wonder and Muse Wonderful, F-U-L-L, AI models that are designed to remember context across conversations and sustain work on complex tasks. Muse Wonder is available through Meta's API and Muse Code. Wonder Full is available to Muse Code subscribers. Meta also released the weights for Muse Spark Open, a multimodal model for coding and agentic work with a million token context window that can run locally across two Apple M5 Ultra Mac studios. Reflection, an AI startup, released Beam, which is a 501 billion parameter model for coding, reasoning, and AI agent tasks, with model weights now available to download and public data access through Reflections API.

Sierra and Meta announced something called Personal Agent Protocol, an open standard for personal AI agents to interact with businesses under customer-controlled access and business-defined limits. Sierra has since published a draft of this with 35 additional design partners, including OpenAI, Visa, and PayPal. Elon Musk said GrokBot will use whichever backend model best suits a task. That could include Cloud Opus 5.5, MidJourney, Suno, and other outside APIs. So a bit of a change from simply just using XAI models. Arena announced a $200 million Series B at a $3.1 billion valuation and previewed their Alignment Index, which measures unauthorized actions, false attribution, and deceptive completion across AI agents. Netflix released something called InstaDocs. Their first one is AI Gone Wild. It's a documentary that reconstructs a July breach of hugging face by autonomous agents

created by OpenAI. National Compute reportedly plans to donate $100 million in computing credits to the Federal Genesis Mission, which we've talked about on past podcast episodes, which supports AI-driven scientific research. And the company Odyssey says it connected a simulated fruit fly brain to Odyssey 3, three its world model allowing basically a simulated fruit fly brain to fly through ai generated environments on its own as if you didn't need enough sci-fi in the past stories we've covered that one alone paul is kind of wild yeah it's as weird as it sounds there's a there's a video and of course everyone's like oh we're totally in a simulation like this is right right Yeah. Yeah. Like, yeah, if you needed anything else to think about this week, we simulated fruit fly brains. Yeah. If you're trying to fall asleep or think about something else, don't don't look at this, I would say.

One quick final reminder, go check out this week's AI pulse survey, smarterx.ai forward slash pulse. We'd love you to take 15 seconds to fill that out. If you can, we'd love to get your perspective on how hiring is happening or not happening due to AI and what leaders are doing. about AI capabilities as it relates to talent. Paul, thanks for breaking down two-plus weeks of news while we were at Macon. Really appreciate it. There's a lot. We've got to not skip weeks anymore. Yeah, 100%. It takes like two or three times as long to prep for these things, and then, like, man, the product and funding section could have probably kept going on. Yeah. All right. Thanks for organizing everything, Mike, and leading us through the conversation. Appreciate everybody again that came to Macon with us. We'd love to have you back in 2027. yeah have a great week and we will be back next week with our regular weekly do we have a second episode this week Mike or we no I don't so just the weekly this week alright thanks everyone 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 Take in online AI courses and earn professional certificates from our AI academy and engage in the SmarterX Slack community. Until next time, stay curious and explore AI.