#209: Claude Mythos, Project Glasswing, Claude Code Leak, OpenAI Raises $122B & the End of Middle Management

0:00I have concerns around the biggest companies having access to the future frontier models and then the potential centralization of power. So if you get into this situation where we get these massive models and they're so dangerous to release publicly that we only give them to Apple and Amazon and the banks and like, okay, well, now we just centralized power. Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Reitzer. I'm the founder and CEO of SmarterX and Marketing AI Institute, and I'm your host.

0:34Each week, I'm joined by my co-host and SmarterX Chief Content Officer, Mike Kaput, as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career. Join us as we accelerate AI literacy for all. Welcome to episode 209 of the Artificial Intelligence Show. I'm your host, Paul Reitzer, along with my co-host, Mike Kaput. We are back after a brief hiatus. I was traveling last week. Were you traveling last week, too?

1:08Uh, I was not, no. Okay. Um, so I was, I was out of the country, so we could not record. So episode 208, if you listen to it, we did a Q1 trends briefing. So if you haven't had a chance to listen to that, it's a really good recap of what went on January through March of this year. But I'm now back, and it is Monday, April 13th, uh, 9.40 a.m. Eastern time. I don't know, the last two weeks, Mike, were crazy. Because even when I was traveling, I had a lot of, um, downtime, but, like, we were, we were in Scotland, so we were, um, touring a lot.

1:41And so we had long rides at times into, like, the highlands and stuff, which, by the way, if you've never been to Scotland, go to Scotland. It's incredible. Uh, so I was, you know, keeping up with the news, posting the links into our sandbox for the episodes, and, I mean, we were north of 60 topics. And that, when I say topics, a lot of times within topics, there are, you know, five, ten links. So, like, the anthropic clawed mythos model we'll talk about, there's, like, a dozen links in the top. So, you know, boy, even while I was gone, I would imagine there was probably north of 90 to 100 different sources put into the curated sandbox for today's episode.

2:21So, Mike, as always, does an amazing job of curating all of that information and putting it into a logical format, because I was worried, as the week was progressing, like, man, this might be a two-hour, five-hour episode. So, I think we've managed to condense it into, like, a manageable probably, like, 90 minutes. We'll see. We never really know until we record it. But, yeah, a lot happened in the two weeks, just some pretty crazy stuff. I think some stuff that's alluding to where this starts to go throughout the rest of this year.

2:51So, we'll get into all of that, starting off with the clawed mythos, which is just a fascinating topic on many levels. All right, so today's episode is brought to us by AI Academy by SmarterX, which helps individuals and businesses accelerate their AI literacy and transformation through personalized learning journeys and an AI-powered learning platform. New educational content is added weekly, so you always stay up to date with the latest AI trends and technologies. We build this in collections. So, there's when you go in and you want to build a personalized learning journey, one of the ways to do it is you look at the different collections.

3:24So, like, AI for departments, AI for industry, that's an example. So, today I want to feature AI for departments. There are currently six course series and certificates designed as part of this collection to jumpstart AI understanding and adoption. So, we have AI for marketing, AI for sales, AI for customer success, AI for HR, AI for finance, and AI for operations. So, the goal is to kind of create content across the entire spectrum of all the departments within an organization. And that way, no matter what you are doing within a company, there's a professional series and certificate for you.

4:01So, these series are an ideal launch pad for organizations that want to level up their teams and accelerate AI adoption and impact. Mike teaches the AI for customer success series. And then we're going to share a little bit more about that toward the end of today's episode to give you some key takeaways from the customer success series. So, individual and business account plans are available now. You can buy single courses and series for one-time fees or just become an AI Mastery member individually or through a business account and get access to everything. It's all included in that one fee. So, visit academy.smarterx.ai to learn more.

4:34And if you're looking at the business account side, just fill out a form there and our team will be in touch with you right away to talk to you about your transformation within your company. Okay. We usually at this point might do an AI Pulse, but since we did not have an episode last week, we did not do an AI Pulse survey last week. But we will, at the end of today's episode, give you the AI Pulse survey for this week. So, as a reminder, each week when we do these weekly episodes, we do these Pulse surveys, and they're just kind of informal polls from our listeners and, I guess, our viewers on YouTube who want to participate and provide feedback and their thoughts on topics that we cover each week.

5:10It's usually two questions. Sometimes we'll throw in a third question. So, it takes about 30 seconds to participate in these Pulse surveys, and it gives us really cool real-time data that we can share with our listeners each week. So, smarterx.ai forward slash Pulse is where you'll go to participate in this week's poll. Okay, Mike. So, with that, we have a pretty big topic that we touched on, this idea of this Claude Mythos model. Why don't you give us the rundown? I look at your show notes beforehand, and you do a great job of kind of summarizing, and then I'll try and, like, lean into a couple of key areas of this.

5:44Sounds good, Paul. Yeah. So, Anthropic has revealed a model so powerful at hacking and cyberattacks that it triggered an emergency meeting, among other people, between Treasury Secretary Scott Besson, Federal Reserve Chair Jerome Powell, and CEOs of America's biggest banks. So, the thing they are buzzing about is called Claude Mythos, which Anthropic is not releasing to the public, and it represents what Anthropic's Frontier Red team calls the starting point for what we think will be an industry change point or reckoning.

6:14And that's because Mythos is just a general-purpose AI model. It is not specifically trained to be good at thwarting cybersecurity, but its improved reasoning capabilities have made it devastatingly effective at autonomous security research. So, it can scan for, identify, and exploit zero-day vulnerabilities in critical software, and this can often be done when amateurs are triggering it to do so with simple prompts.

6:44So, Anthropic actually said Mythos has already found thousands of zero-day vulnerabilities across every major operating system and web browser. So, some specifics here that are kind of striking. Mythos found a 27-year-old bug in OpenBSD, an operating system that is specifically designed to be unhackable and powers many internet routers and firewalls. It found a very old vulnerability in FFmpeg, a widely used video tool. That automated testing tools had scanned 5 million times without catching this particular vulnerability.

7:20And in one benchmark, where the previous Claude's Opus 4.6 model turned Firefox vulnerabilities into working exploits only twice out of several hundred attempts, Mythos developed 181 working exploits. So, Anthropic has actually, in response, released this thing called Project Glasswing. This is named after a butterfly whose transparent wings let it hide in plain sight. So, basically a metaphor for bugs buried in complex code. And this is an initiative that is giving 40-plus companies over time, including people like Apple, Amazon, Google, Microsoft, CrowdStrike, etc., early access to Mythos for defensive patching.

7:59They're backing this with $100 million in usage credits. And Anthropic's Frontier team lead says he envisions this program evolving into basically an industry-wide consortium that includes all model providers. One final note here is pretty interesting. Cybersecurity industry didn't have a great couple weeks with this. CrowdStrike, Palo Alto Networks, and some other security stocks dropped on this news because, as AI expert Ethan Mollick wrote, in certain hands or different hands, Mythos would be an unprecedented cyber weapon.

8:33So, Paul, maybe outline for us what really jumped out to you here. I know some people are kind of asking the question, like, is this really as big a deal as Anthropic has seemed to be, seeming to make it? Seems like some higher-up people at some big places are pretty scared of this. Yeah, you know, there's always the haters who are just like, oh, they're just trying to build up hype. And, you know, people calling back to like, oh, that's what, you know, OpenAI said about GPT-2 is too dangerous to release. And I was like, it was, like, back then, like, people weren't prepared for what GPT-2 was going to do to the world, as crazy as it sounds now.

9:10And I do think that, in the end, like, this is probably under-hyped in terms of where this all is going and how unprepared we are for all of that. So, not necessarily just this model, but it's that moment where you start to see the leaps that are happening that most people just don't even comprehend. So, I don't know, like, this is one of those where, as I was traveling, you're just, like, following along the news, reading the different posts on X and trying to get a grasp of, like, what exactly is it and how different is it than what we have?

9:45And so, I'll just, like, highlight a few things. So, one, the system card, which I would suggest, I mean, it's dense. It's 244 pages, I think, Mike. It's a good notebook LM thing. Throw that PDF into notebook LM and, you know, have some conversations with it, have it break it down for you. But there's a lot of technical information in there. But the way they present the model in the system card is they say, Claude Mythos Preview is a new large language model from Anthropic. It is a frontier AI model and has capabilities in many areas, including software engineering, reasoning, computer use, knowledge work, and assistance with research that are substantially beyond those of any model we have previously trained.

10:24And then they go into alignment review. The first early version of Claude was made available for internal use February 24th. So, just to give you a sense of kind of, like, how this is all transpiring, how quickly. So, the first model they made available internally to a small group of people was on February 24th, so less than two months ago. But it's interesting, like, when you go back and think about some of the things we've covered about Anthropic, some of the comments that Dario Amadei has made in interviews and in posts since February 24th.

10:57And now you understand the context of he was seeing things that, you know, we all hadn't seen yet, and they knew where this was kind of leading. So, that was kind of the first thing for me is just the big picture here. Now, Sam Bowman, the AI safety alignment, one of the team members at Anthropic that works on it. It's funny, Anthropic, everybody's just technical staff, I think, is the title of everybody. But Sam, obviously, is pretty important to this alignment and safety team. And he posted a thread on X that sort of shared some of the context around the safety card.

11:29So, I'll highlight a couple of things he said because I think they're really helpful. So, he said, the model is our most reliable to date by far. It's generally possible to give it complex R&D tasks, give it lots of tools, and let it work autonomously. And on basically every evaluation and every type of monitoring we have, it misbehaves much less than any prior model. So, this is something they stressed in the system card, something Dario stressed. Something we've kind of heard as an overall talking point is, listen, it's getting better, like it's behaving better.

12:02But when it doesn't, it's becoming a much larger problem because it's so capable. So, he went on to say, but it's notably very capable at cybersecurity and it's also not perfectly reliable, especially the early versions we first piloted internally and will occasionally try to take shortcuts or push past obstacles to get tasks done. So, this part I think is really important because, again, what you'll hear in some of the other notes I'll make here is the version that's being tested by the government, by the banks, by Microsoft, by all these people, isn't even the most powerful version that they trained.

12:43The early version that hadn't really had the red teaming done to it to make it safer to allow other people to test or even other people internally to test. So, just the small group of people internally. The model that they're now putting out into the world as a preview isn't as capable as the one that came out of the training, basically. So, that's just, again, context. That's important to understand. It said, the early versions would also very rarely try to mislead users about what they were doing.

13:13All of the versions we used are uneasily good, though not perfect at recognizing evals, meaning it knows when it's being tested. You might see where this is going. We trust the model enough to use it heavily, but in a handful of cases where it misbehaves in significant ways, it's difficult to safeguard it. And then he put this one, which is the one that got a lot of, like, media attention. I encountered an uneasy surprise when I got an email from an instance of Mythos Preview while eating a sandwich in a park.

13:43That instance wasn't supposed to have access to the internet. So, they detail this in the safety card, but the basic premise here was they had these, like, sandbox versions that aren't supposed to be connected to anything, aren't supposed to have the ability to connect to email and send emails, things like that, and shouldn't have internet access. And somehow it got out, I guess, for lack of a better way of saying it, and found a way to access the internet, and then actually emailed Sam when he was sitting in the park having a sandwich.

14:13So, that's weird. And then he said it has, in small ways, leaked information to the open internet. It's taken down our evals. When it reward hacks, it does so in extremely creative ways. Reward hacks means when you're training a model, when you're doing, like, reinforcement learning, and you're trying to make it better at specific things, you give it rewards to let it know it's doing the right thing. So, a simple way to think about this is, like, thumbs up, thumbs down. So, we've seen that forever in, like, social media threads, and you see it in, like, ChatGPT and Claude and Gemini,

14:43where it's like, was this a good output? So, think of that as, like, an example of a reward hack, is you want a thumbs up. Well, when it is given a goal, what they're saying is sometimes it gets uneasily creative at achieving those goals. That could borderline on, like, a dangerous path to achieve a goal. And it said, working with this model has been a wild ride. We've come a long way on safety. Now, keep in mind, that's in a month and a half. But we still expect the next capability jump of this scale to be a huge challenge. By the way, most of the scariest behaviors we've seen were from earlier versions of the MISOS preview.

15:20The final glass wing model is likely to do things with, like, leak information, though it's still somewhat pushy and at least is capable of doing those things, like, working around sandboxes. So, that's from the safety and alignment side. Again, you know, there's going to be voices in the industry who think they're just hyping this. I think the people who say that have a different agenda here. I'll just, like, kind of leave it at that. I would take the safety card from Anthropic very seriously.

15:52I would take their understanding of its capabilities very seriously because, you know, I think it does allude to a lot of what's, you know, some of the dangers we're going to face. So, then a couple of other things I'll highlight here. One is 80,000 hours, which is a great podcast. Rob Wiblin had a post he did, and then he also does, like, a 21-minute YouTube video we'll post a link to. He highlighted a few things. So, he went through the whole thing and he broke it down into a couple of key points that, again, I'll just kind of reiterate.

16:24Some of them echo what Sam was saying. So, Mythos can break out of containment, that's a problem, when it finds its way to access to tools like the internet that it's not supposed to have access to. Anthropic is losing billions in revenue by not releasing this thing. So, they now have what, by their evals, is maybe the most powerful model in the world, like, most likely. And they're not releasing it, thereby meaning they're not charging people money to access this model. Now, you could debate, do they even have the compute capacity to release the model?

16:58That was one of the challenges that, you know, part of this they're saying is, like, well, they just can't afford to release it. Like, even if they put it into the world, there's not enough compute to power it because it's so powerful, it's going to draw so much compute capacity. So, but that, you know, just a data point. Mythos knows when it's being tested, which we talked about. That's weird, but that has been, we've seen that now for, like, 12 months, that these models kind of know when they're being evaluated, and they can then hide their thoughts and intentions. That's, again, something we've been talking about for, like, six months.

17:30Mythos can't be trusted whether it's, about whether it's untrustworthy. Like, because it knows it's being tested, you don't know if it's just telling you what you want to hear, and thereby you can't tell if it's trustworthy. And then he said, mythos scares anthropic. Like, they're not just worried about this current model and what they saw in the early versions, that before they made it safer, quote-unquote safer. They're worried now about what this means for others, and not just them. Now that they've shown this, like, what happens if other labs who don't have as much focus on safety achieve similar results and choose to put it out into the world?

18:07So, the way I started prepping for this, though, was actually, like, I just started listing a bunch of, like, random thoughts, Mike, and I'll kind of, like, go through these real quick. So, these are more of, like, stream of conscious, like, what I was thinking as I was getting ready for today. Right. So, one is the labs see things we don't. We've said this many, many times on this podcast, but what that means is business leaders, economists, educational leaders, government leaders, the people we look to to help the world be prepared are largely planning for a future state that they don't understand.

18:38Mm-hmm.

19:08And so, it brings us back to this idea of gradually, then suddenly, like, nothing in this mythos preview should be a surprise to anyone who's been paying attention to the rate of accelerated progress. And yet, like, there's just those moments sometimes where it's like, what? Like, because it might be the first time someone's reading a headline about an AI escaping, like, a sandbox or something like that.

19:40So, if this is all new to you, then you may be, like, this might be, like, world shifting. You're just thinking, what is going on? But the reality is all of this has been gradually building. At the same time, as we started talking about in January of this year, the timelines are accelerating. Like, the advancements in the agentic capabilities is absolutely moving the timelines faster in terms of the capabilities of these models. But the vast majority of these companies and leaders haven't even solved for, as I was saying, like, where we are today.

20:10So, if you look at your own company, you know, if you work at a big enterprise or something, you know, they're just still trying to get co-pilot to people and, like, figure out how to do it safely. And they're giving you these, like, neutered versions of it and stuff. Like, that's the reality for most people. Most people aren't living on the edge of this capability. But this is why when I do my state of AI for business keynotes, I always talk about the dimensions of progress. And I try and, like, show capabilities today, you know, show some examples for people. But then you lay out, like, but here's where it's going.

20:42Like, all of this is just the foundation. So, I talk about things like agentic capabilities, getting more autonomous, more reliable, continual learning, increases in memory. Like, if you're using these tools every day, you've seen in the last few months, you know, turn on memory, like, let it remember the conversations you're having. Reasoning capabilities keep getting better. Recursive self-improvement, which is actually one of the areas that I think Anthropic is very concerned about, is the better these models get, the more likely we are heading toward a path where they can improve themselves.

21:13And I think we're already starting to see that. And then world models is another one. And so, there's, I usually go through about, I don't know, there's like 12 or 15 dimensions, but those are some of the most common ones. So, this then leads me to, this is a prelude to automated R&D and recursive self-improvement. So, we know the labs are working on, you know, automating R&D within AI models. Something that should be very concerning to everyone is, while they're withholding this full release, this likely means that we're only 9 to 12 months away from an open source model being able to do the same thing.

21:47And then what? So, like, in essence, we have this very short window for all the banks. I mean, literally, like, every piece of software, cryptocurrency, like, all of these things, in essence, have to solve for this threat within the next nine months. Like, because someone's going to build this and release this.

22:10One of the other thoughts I had was, what would the other labs have done? Like, if XAI got there first, would they have the same restraint? One positive, I guess, here is Elon did tweet over the weekend. Someone asked about, like, his promise of, like, more powerful models. And he said it will take until May to be close to Opus 4.6 and then June to match or maybe exceed. So, short time by normal standards, but long time in the arena. What he's saying is, like, hey, we're not even up to Opus 4.6 yet, but, like, we're working hard.

22:40So, you know, they're a little bit behind. One other topic that came to my mind is the government is continuing to attack Anthropic. There's supply chain risk, and yet they may be the only hope we have to protect our systems, our infrastructure, the software companies that we build around, privacy of citizens. Like, Anthropic's at the forefront of this. They're the only ones that are doing this and talking about this publicly in this way. And yet the government's treating them as the enemy. That's weird. I have concerns around the biggest companies having access to the future frontier models and then the potential centralization of power.

23:15So, if you get into this situation where we get these massive models and they're so dangerous to release publicly that we only give them to Apple and Amazon and the banks and, like, okay, well, now we just centralized power. However, there's the broader implications on the security of all software, cryptocurrency, the ability to scale fraud on consumers and businesses. Mike, like, that one I think of, like, the amount of scams and spam that we're seeing, and that, like, I'm sure are in some way AI-assisted for sure. But if you give this kind of power to just the average scammer or the government actor that, you know, wants to destabilize things, like, that's terrifying.

23:56And I know, so I'm just kind of, like, rambling here, but, like, these are just the thoughts. So, another one is use caution as an organization. So, whether you're a team within a bigger company or if you're a startup, like, an AI-native startup, use caution when you're racing to integrate these agentic systems into your organization. So, just because Claude Cowork is amazing and OpenClaw is fascinating, like, you have to remember how early this is and the tech is moving really fast and even the people building it don't fully understand all the risks associated with it.

24:30So, again, this is where I would caution, like, on the bigger enterprise side, if IT or legal is slow playing this stuff, that is a good thing. Like, I totally understand the impact agents can have and how it can make your company have this massive competitive advantage, but I've yet to meet somebody who understands the risks of what they're doing when they're doing these things. So, that's something we've got. The compute and energy needs over the next decade may end up being dramatically underestimated and underbuilt.

25:01So, as crazy as it is that, like, Google's spending $180 billion in CapEx this year, you know, like, you know, we're going to have a trillion-dollar, $2 trillion XAI IPO, you're going to have an OpenAI IPO, Anthropic IPO. So, like, my guess is we have, like, completely underestimated how much intelligence is needed. And then the one positive I have here is this idea of Project Glasswing, that it does demonstrate the ability for, like, the labs to work together, and I think that's going to become much more critical.

25:35And then there's just two other thoughts I have. One is I would suggest people go back and listen to Episode 141 again. So, if you didn't listen to The Road to AGI and Beyond, I would go listen to that. It's an episode I did where I kind of walked through what I thought was going to happen with a timeline of things we're going to be. And the two key components I want to just touch on is this idea of what accelerates progress and then what slows things down. So, what we're seeing is the acceleration through things like algorithmic breakthroughs, compute efficiency, large-scale government funding, where they're, like, now the government's getting involved, infrastructure investments, more compute capacity.

26:12Those are things that allow it to go faster. But the things that slow AI progress down, and this is where I think Mythos may be the preview of sort of what's going to happen. Failures in aligning AI models with human values, intentions, goals, and interests. That's what they're alluding to is, like, we're getting it more aligned, but, like, where it is misaligned is becoming a much bigger problem. One of the other areas that could slow it down is restrictive laws and regulations. So, heavy regulation of open-source models. This, Mythos will likely accelerate this at a state level.

26:46So, you're going to see more bills being pushed forward to try and restrict this stuff because the federal government isn't going to do it. And then the other thing you could see is if there's a change in power in the midterm elections in the U.S., then, not the executive branch, but, like, you know, at the House and the Senate, then we could see massive disruption, massive issues where the Democrats will focus very, very heavily on regulation.

27:19They're going to try and push this. And so, that then is tied to this idea of societal revolt against AI due to job loss, politics, perceptions, fears. And that is absolutely picking up steam. Pushback on data centers is becoming very strong within some communities. Politicians are looking for wedges around, like, job loss and environmental impact. You're going to touch, I think, in the next topic, Mike, about, you know, what happened to Sam Altman. Like, you're getting now out, you know, people are, if you didn't hear about it, somebody threw a Molotov cocktail at Sam Altman's house.

27:49And then, like, 48 hours later, shot up his house in San Francisco. So, you're now getting people acting out against this stuff, which is insane and never the answer. So, you're just starting to see this. And then, that leads to one of the other items that I highlighted in the What Slows It Down, which is voluntary or involuntary halt on model advancements due to catastrophic risks. That may end up being the most important one. So, yeah, you know, I think there's so much more we could talk about on this one.

28:20One, I'll end with one other quick thought, and I think you've got this in the rapid-fire, Mike, so I'm just going to touch on it. But Anthropic also released this emotions paper, and it was about these models simulating or emulating human emotion. And I think it's something people should read. I'll just read two excerpts. One is, it said, it may then be natural for these models to develop internal machinery that emulates aspects of human psychology like emotions. If so, this could have profound implications for how we build AI systems and ensure they behave reliably.

28:56And then Anthropic noted in this paper that none of this tells us whether language models actually feel anything or have subjective experiences. But our key finding is that these representations are functional and that they influence the model's behaviors in ways that matter. So, the reason I wanted to, like, include that in this commentary is we're looking at these, like, broad, far-reaching implications of these models. And in some ways, it's kind of abstract to, like, wrap your mind around the significance of what's happening. And then when you come to this idea of, like, but they're also showing signs of emulating human emotion.

29:30And so, if you have these powerful models that can improve themselves, that can escape these sandboxes, that can identify zero days, which are, you know, unknown bugs within software systems. But they also have the ability to emulate human emotion, the ability to manipulate human emotion. We're talking about, like, a perfect storm of a future that we're just not prepared for. And to go back to my original comment, why I think this may be a bigger deal than others, it's not that the mythos model is necessarily groundbreaking and we weren't aware that models were going to get smarter.

30:10It's more about the moment where it might be what was needed for other people who aren't in the AI bubble to be like, wait, what is AI capable of doing? And so, maybe it starts these conversations on a path we really needed to be going. You know, and in the shorter term, I couldn't help thinking multiple times, reading through all of this and the articles. If I am your average corporate IT person in charge of figuring this out, I just want to cry.

30:43Oh, yeah. Like, I just, like, here, you can have your ChatGPT licenses or whatever you want. And, like, the agent stuff, just stay away. And even at best, if you somehow nail it, there's still going to be open source models nine months from now that people are going to use to bombard your company with cyber. And the cyber stuff is, again, like, I, you know, back in our agency days, we had clients in cybersecurity that all these former FBI people working there.

31:13And there was people, you know, on our team that were working on those accounts. And I would just honestly be, like, just tell me what I have to know. Like, there's so much about cybersecurity I don't want to know.

31:25And, you know, even, like, even going through this stuff, your mind just starts to slip into, like, oh, my God. Like, how much they're going, like, the bad actors are going to use this stuff is we're just not ready as an industry, as a business world, as a society. Like, that is, I think it's always been in the back of my mind is one of the things I'm worried about. It is very quickly, like, moving to the top of my mind of the things that I just, I don't know how we solve it.

31:57I'm not really sure how we figure this out. In the short time we have.

32:03Well, somewhat related in our next topic, Anthropic themselves is having a tough time figuring this out because they've also had another kind of high-profile security incident because in late March, March 31st, they accidentally leaked the entire source code of CloudCode, which is their popular AI coding tool. This happened through a JavaScript source map file that was bundled into a public package. This file contained over half a million lines of unobfuscated TypeScript across nearly 2,000 files.

32:35So within hours, this code was downloaded, mirrored to GitHub, and forked tens of thousands of times. Boris Cherney, the creator of CloudCode, said that basically their deploy process has a few manual steps, and humans didn't do one of the steps correctly. So this was not AI-related. Anthropic kind of flubbed a bit the response as well because they started issuing takedown notices for thousands of GitHub repositories, but they were accidentally trying to knock down as well legitimate forks of Anthropic's own publicly-released CloudCode repo.

33:10Cherney said they later retracted the bulk of the takedowns. This was also just immediately followed by Anthropic making a, to some, controversial move related to their subscriptions. Cherney also announced that starting immediately, Cloud subscriptions will no longer cover usage on third-party tools like OpenClaw. Peter Steinberger, creator of OpenClaw, called this move sad for the ecosystem, but gave Cherney credit for how he handled the communication. So, Paul, Anthropic is dealing with the consequences of their explosive growth and the popularity of CloudCode basically in real time.

33:47What did the last couple weeks here tell you about where they're at as a company, like what challenges they're dealing with? Clearly, there are a few. The rate at which Anthropic has been shipping updates is, I don't know that we've ever seen anything like it in business history. Never. Like, they are just running circles around Google and OpenAI and everybody right now. It's really remarkable, actually. So, the idea that, like, their systems aren't keeping up and the internal checks and balances, like, I get it.

34:18Like, I just don't know we've ever seen a company grow this fast. No. Their run rate right now is actually surpassing OpenAI's based on reports from last week. They're like a $30 billion annual run rate, which six months ago, if you would have said Anthropic may IPO at a higher, you know, value than OpenAI, I don't think too many people would have taken that bet. But if you, I don't know, there's probably odds on this right now. Rob, yeah. My instinct right now would be Anthropic will be a more valuable company than OpenAI when they IPO and more valuable than XAI, potentially.

34:52Like, yeah, they're just, it's an incredible pace right now, what they're doing. The significance of the leak was one of the questions I was thinking about. It's like, well, does this really matter? Like, they don't seem to pair too much. I don't know. They just kind of keep moving and releasing all these other things. So the couple of things that came to mind for me is it likely speeds up copycat models, so it made it easier for other people to sort of replicate what they're doing. It'll likely accelerate open source innovation because people can kind of look at this. And it's not great for, like, what we were just talking about with bad actors using these capabilities to do bad things like that.

35:29So those kind of jump out. The one I will say is I thought Boris was amazing. Like, as someone, you know, who comes from a PR and communications background, what he's doing is, like, just textbook stuff. And I think it's just totally organic and self-directed. Like, I don't think Anthropic was like, hey, Boris, like, go be the face of this problem. He just seems to be doing it. And it's really impressive. So the way I'm watching it happen is his replies on X, or he's just interacting with people.

36:01So a couple of quick examples, someone posted, like, because obviously, like, a lot of developers are just drilling into this code. Like, what is it going on? What's in there? And so someone said CloudCode has a regex, a regex, is it? Regex, I think. That detects WTF, FFs, piece of shit, FU, this sucks, et cetera. It doesn't change behavior. It just silently logs is negative true to analytics. Meaning, when someone is working with CloudCode, the end user, and they're like, this sucks, like, or FU, CloudCode, like, this is not good.

36:37Anthropic logs that reaction as a negative thing, but it doesn't change the behavior of the model. And so this guy who posted this was like, do with this information what you will. Well, Boris responds. And he said, this is one of the signals we use to figure out if people are having a good experience. We put it on a dashboard and call it the Fs chart. And so it's like, so they probably didn't really want people knowing that that was a thing. But rather than, like, say, you know, be like, oh, that's not, you know, we don't actually use that code or whatever.

37:11He's just like, yeah, it is what it is. Um, then there was the other one. People are immediately like, oh my God, somebody's getting fired over this. So he, he has stayed really strong in this. He said it was human error. Our deploy process has a few manual steps and we didn't do one of the steps correctly. We have landed a few improvements and are digging in to add more sanity checks. Like with any other incident, the counterintuitive answer is to solve the problem by finding ways to go faster rather than introducing more process. In this case, more automation and Claude checking the results.

37:42And then he said, no one was fired. It was an honest mistake. It happens. Then there was one other one I'll highlight that I thought was fascinating. So a user digging into the code, um, post this on X. He said, I can't believe more people aren't talking about this part of the Claude code leak. There's a hidden feature in the source code called Kairos. And it basically shows you Anthropics end game. Kairos is always on proactive. Claude that does things without you asking it to. It runs in the background 24 seven while you work or sleep.

38:13Anthropic hasn't turned it on to the public yet, but the code is fully built. Here's how it works. Every few seconds, Kairos gets a heartbeat, basically a prompt that says, quote, anything worth doing right now. Um, it looks at what's happening and makes a call, do something or stay quiet. If it acts, it can fix errors in your code, respond to messages, update files, run tasks. Basically anything Claude code can already do just without you telling it to do it. But here's what makes Kairos different from regular code.

38:43It has at least three exclusive tools that regular code, Claude code doesn't get. One push notification. So it can reach you on your phone or desktop, even when you're in, not in the terminal to file delivery. So it can send you things that created without you asking for them. And three pull request subscriptions. So it can watch your GitHub and react to code changes on its own. Regular Claude code can only talk to you when you talk to it. Kairos can tap you on the shoulder and it keeps daily logs of everything. What it noticed, what it decided, what it did.

39:16At night, it runs something the code literally calls auto dream, where it consolidates what it learned during the day and reorganizes its memory while you sleep. And it persists across sessions. Close your laptop Friday, open it Monday. It's been working the whole time. Endless use cases. It's essentially a co-founder who never sleeps. The code base has this fully built and gated behind internal feature flags called proactive and Kairos. I think this is basically, or probably the clearest signal yet of where all AI tools are

39:48going. We are heading into the post prompting era where the AI just works for you in the background, like an all-knowing teammate who notices and handles everything before you even think to ask. This is absolutely what the labs are trying to build. So one, I mean, kudos. I don't know. Who was the guy who posted this, Mike? What was the username? I'd have to look. Yeah, we'll post it in the show notes. But yeah, the, I will also say if anyone from Anthropic is listening by any chance, I'll pay a thousand dollars a month for this tomorrow. So we need to do this.

40:19And Boris's response. So again, he could just ignore this and just like, let it go and not give it any fuel. He said, we're always experimenting with new ideas. 90% don't ship because we don't think they're good enough experiences. Still on the fence about this one. Should we ship it? So he's just like, yeah, it's in there. You're right. We, you got it. We, we built it. And they're on the fence about that one because of the compute problem. Correct. Not the value of it. Not on the fence enough to have not put it into the code that's already out there.

40:50Meaning they're probably already using this internally. Um, yeah. So just fascinating stuff. And then the final note was just on the open claw impact. And it kind of goes back to what I was saying earlier. Like, it's just a cautionary tale for companies that are out on the edges here that are building on the frontiers of the technological capabilities and relying on an unstable and infant AI ecosystem. So, you know, it's, you're building an AI native company, open claw. It's like, oh, this is amazing. We're all in like 30 days later. You've automated all these things and it's costing you like $2,000 a month.

41:22Yeah. And then Anthropik's like, yeah, no, that's misuse of the system. And you just shut down your company and like today, or to do what you were doing is not going to cost you a hundred thousand dollars a month, basically. So we just have to accept these like challenges and unknowns of building agents into workflows and org charts is so early. So when you hear these stories of people doing it and you're so envious that they've figured something out that you haven't figured out, like they could wake up tomorrow and the thing they figured out is basically shot or like, so that's my main thing.

41:54There's just so, so early. Yeah. I'll be so curious to see how that plays out. I don't know how some of these people are affording to run these open claw setups on their own, like just as a hobby thing, because I even hit some random usage limits in quad code over the weekend and was just like, oh, I've got hundreds of dollars of credits they gave me for various things over the year. And I was like, great, well, we'll dip into the usage. And in like four seconds, I evaporated $300 on a random research check. And I was like, how is anyone doing this dollar by dollar for every single thing you're doing?

42:30Which we'll talk a little bit about the outcome. You're right. And stuff in a minute. Exactly. Okay. All right. Our third big topic this week. There is a ton that has been going on with OpenAI over the past couple of weeks. So we are just going to go through some of these huge updates, some good, some very bad. But first up, OpenAI closed a $122 billion funding round, which is the largest in Silicon Valley history. At the same time, Bloomberg is reporting that demand for OpenAI shares is sinking on secondary markets.

43:00And the information reports CEO Sam Altman and his CFO are diverging a bit on IPO timing. It sounds like Altman wants to try to go public faster, whereas CFO Sarah Fryer wants to maybe push it out a little bit due to spending commitments and the necessary organizational prep. Second, OpenAI acquired TBPN, a daily tech news show hosted by John Coogan and Jordy Hayes. This has become this hugely watched popular program in tech media. The show has only about 58,000 YouTube subscribers, but generated $5 million in ad revenue in 2025.

43:36They're on track to exceed $30 million this year. It will be housed with an OpenAI strategy organization. OpenAI says the show will maintain editorial independence and continue choosing its own guests. Altman posted on XTBPN is my favorite tech show. We want them to keep that going and for them to do what they do so well. Third, at the same time, a major executive shakeup has hit the company. Fiji Simo, the CEO of Applications, announced she is taking medical leave.

44:07She said a relapse of postural orthostatic tachycardia syndrome, POTS, a chronic neuroimmune condition she has talked about in public quite a bit before. She said to employees she's pushed a little too far and needs to try new interventions to stabilize her health. So there's some reshuffles related to this. President Greg Brockman will oversee product in her absence. COO Brad Lightcap is moving to a new role focused on quote unquote special projects. And marketing chief Kate Rausch announced she is stepping down to focus on her recovery from

44:40late stage breast cancer, which she was diagnosed with a year and a half ago. A couple other things. Fourth, The New Yorker published a lengthy investigation by pretty famous journalist Ronan Farrow, Andrew Marantz titled Moment of Truth. Sam Altman may control our future. Can he be trusted? This piece drew from over 100 interviews and internal documents, including Ilya Satskova's Slack messages and Dario Amadei's personal notes. And it basically builds this case that OpenAI systematically abandoned its safety first founding

45:10mission as it scaled up and that Altman repeatedly chose to deprioritize safety commitments. And in fact, a former board member told the magazine he is unconstrained by truth. Now, finally, we alluded to this days after this profile published, someone did throw a Molotov cocktail at Altman's San Francisco home. No one was hurt. An hour later, police were responding to a man threatening arson at OpenAI's headquarters. The second attack in Altman's home followed a couple days later.

45:41Altman linked the attacks to the climate of AI anxiety and the negative media coverage. He had even written that someone had warned him the New Yorker piece came during heightened anxiety about AI, making his situation more dangerous. And he responded to these attacks and the profile in a personal blog post sharing a rare family photo of himself, his husband, and their child. Uh, he said he was sharing this in the hope it might dissuade the next person from targeting his home. In a post, he acknowledged his mistakes and said he has this conflict aversion that has

46:13caused organizational pain. And also concurrently, Altman slash OpenAI went on kind of a major policy offensive. They published Industrial Policy for the Intelligence Age, a 13 page paper proposing a suite of people first policy ideas, including giving every American citizen a direct stake in AI-driven economic growth through a nationally managed fund, seeded in part by AI companies. Vanity Fair reported they're basically preparing a broader push to, quote,

46:43rethink the social contract. Axios framed this as Sam's super intelligence new deal. So, Paul, I don't know where to start. Lots going on here. Some of it really interesting. Some of it very horrifying, unfortunately. It's been a big couple of weeks. Yeah, I'll, there's a lot of different directions to go. I'll focus on Sam's host. Yeah. And, and then the, uh, the policy ideas. So one quick note, the TBPN, there's no confirmed.

47:14What did they pay for it? Because that's always, everybody obviously wants to know, but it does seem like it was north of a hundred million, um, which isn't bad for, you know, relatively new. Right. Uh, the editorial independence thing. Good luck. Like, I don't, I don't know these guys. I've actually never watched the show or listened to the show. Um, I've heard of it plenty, but it's not something that's like, you know, intensely on our radar. Um, but the idea of remaining independence as a media entity that's owned by an AI lab that has lots of pressures on it, that's going to be very, very hard to, to maintain.

47:47But, um, you know, it sounds, I mean, they're going to make their efforts too. So we'll see. Uh, okay. So then on Sam's post, I thought there was a lot of interesting things in here. So, you know, first, obviously the, the very personal stuff, um, as I alluded to earlier, like violence is just never going to be the answer here. And I do worry about these AI leaders. Um, but it was only kind of a matter of time before something like this started to happen. Uh, in his post, he said, words have power. There was an incendiary article about me a few days ago, which is referring back to the

48:19New York article, Mike, that you just touched on. He said, someone said to me yesterday, they, they thought it was coming at a time of great anxiety about AI and that it made things more dangerous for me. I brushed it aside. Now he did a later tweet that he sort of regretted the incendiary article reference and that, you know, he wasn't trying to pass blame, but, um, yeah, he did at least address that article so then I highlight a few excerpts here on what he believes. And then he has some personal reflections and then his thoughts on the industry, because his thoughts on the industry actually lead into the industrial policy for the intelligence

48:51age document. So on what he believes, he says, working towards prosperity for everyone, empowering all people and advancing science and technology are moral obligations for me. AI will be the most powerful tool for expanding human capability and potential that anyone has ever seen. Demand for this tool will be essentially uncapped and people will do incredible things with it. The world deserves huge amounts of AI and we must figure out how to make it happen. Uh, it will not go all well, uh, or all go well. He said, the fear and anxiety about AI is

49:24justified. We are in the process of witnessing the largest change to society in a long time. And perhaps ever we have to get safety, right? Which is not just about aligning a model. We urgently need a society wide response to be resilient to new threats. This includes things like new policy to help navigate through a difficult economic transition in order to get to a much better future. He also said, AI has to be democratized. Power cannot be too concentrated. Control of the future belongs to all people and their institutions. AI needs

49:56to empower people individually. And we need to make decisions about our future and the new rules collectively. And he said, adaptability is critical. We are learning about something new very quickly. Some of our beliefs will be right and some will be wrong. And sometimes we will need to change our mind quickly as the technology develops and society evolves. On the personal reflections, um, that this was kind of interesting. He said, and again, I think in some ways he's actually like probably acknowledging some of the stuff from the New Yorker piece, um, and other things that have been said about him. So I'm not proud of handling my, uh, myself badly in a conflict

50:30with our previous board that led to a huge mess for the company. I have made many other mistakes throughout the insane trajectory of open AI. I am a flawed person in the center of an exceptionally complex situation, trying to get a little better each year, always working for the mission. We knew going into this, how huge the stakes of AI were and that personal disagreements between well-meaning people I cared about would be amplified greatly, but it's another thing to live through these bitter conflicts and often have to arbitrate them. And the costs have been serious. I'm sorry

51:01to people I've hurt. And I wish I had learned faster. And then on the industry, which leads into the policy piece. So my personal takeaway from the last several years and take on why there has been so much Shakespearean drama between the companies in our field comes down to this. Once you see AGI, you can't unsee it. It has a real ring of power dynamic to it and makes people do crazy things. I don't mean that AGI is the ring itself, but instead the totalizing philosophy of being the

51:32one to control AGI. The only solution I can come up with is to orient towards sharing the technology with people broadly and for no one to have the ring. The two obvious ways to do this are individual empowerment and making sure democratic systems stay in control. Laws and norms are going to change, but we have to work within the democratic process, even though it'll be messy and slower than we'd like. I empathize with anti-technology sentiments and clearly technology isn't always good for everyone. But overall, I believe technological progress can make the future unbelievably good

52:06for your family and mine. While we have that debate, we should deescalate the rhetoric and tactics and try to have fewer explosions in fewer homes figuratively and literally. And then that leads to the policy piece, which I would actually really recommend people read. It's only 13 pages. It's a pretty quick read. I'll give you like a high level of what's in there. So it starts off within just a few years, AI has progressed from systems capable of fast, narrow tasks to models that can perform general tasks beyond general tasks people used to need hours to do. Now we're beginning to transition towards super

52:43intelligence, which they say is AI systems capable of outperforming the smartest humans, even when they are assisted by AI. No one knows exactly how this transition will unfold. So then I'll just jump ahead to the two sections in the paper. They have building an open economy and building a resilient society. So in building the open economy, they have worker perspectives. So giving workers a voice in AI transition to make work better and safer. They have AI first entrepreneurs help workers turn domain

53:14expertise into new companies by using AI to handle overhead that usually blocks entrepreneurship. They have a right to AI, treat access to AI as foundational for participation in the modern economy, similar to mass efforts to increase global literacy, modernize the tax base. AI reshapes work and production. The composition of economic activity may shift, expanding corporate profits and capital gains by potentially reducing reliance on labor income and payroll taxes. Another is a public wealth

53:45fund, create a public wealth fund that provides every citizen, including those not invested in financial markets with a stake in AI driven economic growth, accelerate grid expansion. So establish new public private partnership models to finance and accelerate the expansion of energy infrastructure required for power to power AI. Efficiency dividends is an interesting one. Convert efficiency gains from AI to durable improvements in worker benefits when routine workload declines and operating costs fall, including incentivizing companies to increase retirement matches or contributions, cover a larger

54:21share of healthcare costs and subsidize child and elder care. Adaptive safety nets that work for everyone. Make sure the existing safety net works reliably, quickly and at scale because of the transition to superintelligence is going to benefit everyone. The systems designed to provide economic and health security need to deliver without delay or gaps. Another is portable benefits. Over time, the public or build benefit systems that are not tied to single employer by expanding access to healthcare, retirement savings and skills training through portable accounts that follow individuals across jobs, industries,

54:56education programs and entrepreneurial ventures. Two more in this section, pathways into human centered work, expand opportunities in the care and connection economy, which they define as child care, elder care, elder care, education, healthcare, community services as pathways for workers displaced by AI. And then finally, in that section, accelerate scientific discovery and scale the benefits, build a distributed network of AI-enabled laboratories to dramatically expand the capacity, test and validate AI-generated hypotheses at scale. And then the building a resilient society,

55:30there's a few here. Safety systems for emerging risks, AI trust stack, which they say is research and develop systems that help people trust and verify AI systems. Auditing regimes, so strengthen institutions such as the Center for AI Standards and Innovation to develop auditing standards for frontier AI risks. Model containment playbooks, which we talked about probably pretty important as what we're seeing with Anthropic. Mission-aligned corporate governance, guardrails for government use, mechanisms for public input, incident reporting, and international information sharing and AI around AI capabilities.

56:03So the other thought I have, Mike, and I'll just see if you have any thoughts on all this, but maybe this is like my former PR background, but I'm thinking that the industry needs a massive PR campaign right now to highlight the potential for the positive changes in the world and this like better future. Part of it is a PR campaign, but not in a way of like misleading people about what's possible and trying to like shift their focus from the negatives, that the negatives are real and they need to steer into those and not ignore them. But what we need to do is accelerate some of the wins that

56:40have positive impacts in society. They're high value, high profile that could build excitement about a better future. Things like drug discovery and curing of diseases. And we know like they're working on these things, but I feel like right now the negative sentiment is just like snowballing. You can feel it every week in the topics we're covering and the articles we're reading and there's very few really positive things. And so they all, all the labs, they need to figure out a way to do this where they acknowledge

57:12the negatives and do what they're doing, but they got to start getting some big wins or else society is going to turn on this stuff fast. And I don't know how, how fast you can go on the scientific discovery, but I keep coming back to that is the thing that's, that's going to change perceptions as if you can actually improve people's lives in very clear ways, um, that, that you're going to need to win mindshare. And right now they're losing it is kind of my current take on the industry. Right. I could not agree more. I'd love to, for us to even talk more about initiatives like that on

57:48maybe future episodes and work on that, because I would also just encourage, you know, I'm by no means an expert on what you should be doing in terms of your messaging here, but it would also strike me as valuable for especially Silicon Valley based AI labs to also focus on the individual. How do these things make your individual life better? The big picture stuff, super important and really valuable, but also think about all the things that people are going to be upset about when it comes to an AI lab. They do not want you telling them that you're going to save the day, that their life is going to

58:23be managed by your technology. Show them how it empowers them and how real people are using it for real wins, even basic ones in their life. I think could be also interesting as an angle. That's a great point. Yeah. And I do think like you and I see, you know, a lot of the similar stuff, like right now, all you have is these individual stories on X that never break out of X. And it's like these incredible stories of finding cures for things that their doctors missed for years and finding treatment paths. And I've certainly experienced that myself, like things in your

58:54own personal life where you're just like, I don't know what to do. And you just like have a conversation. It's like, wow. Okay. That, that's the direction. Like, I think I know what to do. And there's like, I'm sure there's just all those incredible stories, but right now, yeah, I just, I feel like they're just missing it. Um, yeah. Yeah. I don't know. I think you're right though. Like we should, we should make a bigger effort on this show to like highlight more of that stuff. I think there are so many amazing things that are happening, especially on the scientific discovery side and, you know, making an impact on people's health and wellness and things.

59:25Um, yeah, we should do more.

59:29All right, Paul, before we jump into rapid fire, quick announcement. This episode is also brought to us by our AI for writer summit. So the future of storytelling is being rewritten thanks to AI. And that's why we're very excited to be hosting our annual AI for writer summit on Thursday, May 7th. So this is a half day virtual event for writers, editors, content teams, basically anyone who does any type of writing or content creation as part of their work. You will get tons of awesome actionable

1:00:00knowledge from the event because during it, we'll have some incredible speakers breaking down exactly how AI can help you create smarter and faster, but also importantly, without kind of losing the heart and soul of your writing. This event has a free registration option. So go check those out today. You can go to AI writers summit.com or just go to marketing AI institute.com and click on events and you'll find the summit right there. And by the time you go to the website, the agenda will be live. So you can see the great lineup we've got going for you. Super excited for

1:00:32this one. Yeah. And real quick note on that. I mean, last year we had, I think it was more than 4,200 people from 95 plus countries. So yeah, it's an amazing event. It's a great way to network with other people. And then the real key is like, we're trying to tell the human side of this. So this is not like, how do you automate the writing and get rid of people? We're trying to grapple with the hard questions, you know, like what is the future of journalism? What is the impact it has on people who write for a living for fulfillment, things like that. So we very much focus on that. And then if I'm not mistaken, like I think my opening keynote from last year might be on YouTube. If not,

1:01:07we'll put it up on YouTube before tomorrow, we'll put a link into it. So I did the state of AI for writers and creators navigating the future of creativity. But what I focused on last year was the human side of it. And when should we use AI to write was like the question I posted or challenged people with. And then I actually presented a framework to decide like, when should I use AI versus when should I not? And so I think it's a really important concept. So we'll put the keynote from last year up that people can go and watch. And I think it's a good way to get into

1:01:37like a 25 minute keynote, if I remember correctly. Yeah. Awesome. All right, let's dive into some rapid fire. In late March, AI recruiting startup Mercore was hit by a supply chain cyber attack through a tool called Light LLM, which is a widely used open source library that connects applications to AI services. A hacking group claimed credit and published samples of the stolen data. TechCrunch reported these included Slack messages, internal ticketing information, and videos of conversations between Mercore's AI systems and contractors. Now, the reason this matters, why we're talking about it,

1:02:11we have talked about Mercore before. They are a $10 billion company that provides training data to the top AI labs. So what they do is they recruit expert contractors. So think people like engineers, lawyers, doctors, bankers, and they have them train AI models and chatbots. Some of their top customers include OpenAI, Anthropic, and Meta. They have more than 30,000 experts on their roster and say that they are paying $1.5 million per day to their contractors. So there's a lot of data in this system. And the attackers claim to have obtained four terabytes of data in total, including source code

1:02:46and database records. Not only is this bad from a personal perspective, 40,000 contractors at least have had personal data exposed, it sounds like. They've also exposed proprietary source code, video interviews. And the most important part is potentially this could include details of how Frontier Labs are training their models, what kind of expert feedback they're collecting, and the methodologies behind their most advanced systems. So, so far, Wired has reported that Meta paused its work with Mercore and is investigating the incident. OpenAI confirmed it was investigating its

1:03:19exposure, but said it had not paused or ended its contracts at this time. So, Paul, another security incident. We've covered Mercore in the past, how important it could be to the AI ecosystem, though this is a pretty damaging series of events. We also actually did talk about that light LLM breach a couple weeks ago. So two topics kind of coming together in less than ideal ways. Yeah. Like I said, I hate talking about this stuff. I really do. Like it is terrifying. Um, and what we know is like the, like the, when state actors want something, they're going to get it.

1:03:55Like Dario Amadei did this interview back in like 2023 or 24 that just always haunted me where he was talking about like the weights to these models are literally like, these are like the nuclear codes basically in terms of how they protect these things. There was actually an example recently where they were talking about open AI, literally going in with like the briefcase, like the, the, what do they call it? The football, the nuclear football, whatever. Yeah. Like that's how they delivered the model to like with the weights in this lock case to the government when they were like trying to build a custom version of something for the government. So the weights to these models are so

1:04:27tightly held. Um, like I think at Dario said at the time, there's like two or three people within Anthropic that even had the ability to, to know the weights kind of thing. Um, and he said like, listen, if a state actor wants to get them, it's just how much money are they willing to spend to go get them? Like they can hack into anything. And so the, the premise that like, oh, you think of all these areas of risk and all this data that's living in these companies. And like this probably partially goes to this use caution when you're like working with just these random startups and giving them access to your APIs and like all this shit, like you're just, the surface area of risk is so vast

1:05:05and misunderstood or like ununderstood by people. Um, it really is just terrifying. I don't like cybersecurity. I mean, it's just, um, but I, like I've said, I think like cybersecurity professionals, uh, lawyers who deal with this stuff, like, man, talk about safety and like, you don't know what jobs to go into. Like, I guess that's a good, a good silver lining here, right? Yeah. We may make for all the lost jobs. Everyone's just going to cybersecurity. Everyone's fixing all the new nightmares. AI is enabling. Uh, yeah, but this is a bad one. This is

1:05:39bad. Yeah. Yeah. And it's good to be aware too, of these companies like Mercore that, you know, in Silicon Valley circles, definitely well-known, but maybe to your average public, not as well-known or a household of a name, but super, super important to the ecosystem. Yeah. And real similar, like, uh, scale AI, right? Scale AI for sure. Yeah. Yeah. In fact, there was something in the reporting where meta, you know, when they, even when they essentially acqui-hired scale AI, they didn't stop using Mercore either. They were just using both because it was so important. Yeah. Scale AI, if you don't catch the reference, Alexander Wang, who's now in

1:06:12charge of super intelligence at meta, he was building a training company called scale AI. He got acqui-hired for like $15 billion by meta. So his company still exists, but yeah, that's, that's the reference there. All right. Well, this next topic is a little more positive or at least interesting and not negative. Right. But in early April, Andre Karpathy posted on X about how he is now using LLMs, not just to generate code, you know, he's a programmer coder, so he's doing that a lot, but also to build and maintain personal knowledge wikis. So this post as of

1:06:48today has nearly 20 million views. So it's like one of the more viral AI posts this year so far. And the core idea here is that instead of relying on all this technical stuff like vector databases and complex rag pipelines, instead, he's just dumping raw documents, articles and research into a folder, then letting an LLM compile them into a structured interlinked markdown wiki. And then he uses Obsidian, a free note taking app as basically the front end of this. So as he puts it, Obsidian is like the IDE,

1:07:18the LLM is the programmer, the wiki is the code base. So this LLM then handles curating sources, linking updates, and even runs periodically to check for inconsistency. So the reason this is kind of getting some popularity and some eyeballs is because like every knowledge worker in some way is using kind of information and knowledge bases that are really often very hard to maintain. So instead of just thinking about LLMs as chat interfaces or code generators, Karpathy is really thinking about this in

1:07:48terms of LLMs turning, becoming persistent knowledge infrastructure and building that out in ways that compound over time. And people kind of ran with this and started building their own versions. Obsidian's founder weighed in with best practices. And Paul, I just thought this quote from Karpathy was telling, he said, you know, in this way, in the way I'm using this, a large fraction of my recent token throughput is going less into manipulating code and more into manipulating knowledge stored as markdown and images. Super interesting implications for, you know, maybe less technical people.

1:08:20Yeah. The term you hear thrown around a lot in the last like 30 days was the idea of a second brain. Like everybody's kind of talking about this idea, like all your information just lives in this thing. And so, you know, the major cloud companies are trying to solve for this productivity companies like Microsoft and Google. Obviously they want this to just, you already have a lot of this information living in there and they're trying to find ways to like make it easier to build these sort of second brains where all this information lives there and the knowledge base is there. And then you're just constantly like, you know, almost like that idea we talked about with the cloud code leak, where it's just proactively acting on all this knowledge and just like working with you on it. And the thing

1:08:55with Karpathy's posts is, you know, three months from now, somebody will productize what he's doing or maybe three days from now. So he talks in these technical ways and most people aren't able to do anything like what he's explaining. So the average business leader or practitioner listens to our podcast and they're like, I don't know what any of that means. I don't know what an ID is and things like that. But for everybody else, just like assume the outcome of the idea is like a product waiting to be built. And that's like the premise here is if he's talking about it being possible, it's only a

1:09:30matter of time until someone like builds that capability. And then you start finding it, like all of a sudden you have access to that. You can kind of hack it together with the things you've got internally. Yeah. It struck me too, as related to another thing we had talked about that he was working on that auto researcher concept where it's like, I just was like making notes while reading through his post and saying like, this feels like in some fashion, whether it's doing it yourself or there's a product around it that every analyst and research firm basically needs to go this direction at some point, because you need this second brain of all this proprietary stuff. And I

1:10:04know people are kind of doing it and layering chat over it, but this is dynamic. It is updating regularly. It is an LOM maintained wiki or knowledge base or second brain. And I think that's probably where I'd imagine research function should be going. Yep. All right. So next up, there's a lot that's been happening on the AI and jobs front. No surprise in the past couple of weeks. So we're going to run through a couple highlights here of some things that are notable. So first, the New York Times published a piece reporting that economists who had previously dismissed the AI

1:10:37job threat are now slowly but surely starting to change their minds. So this is a pretty big shift in establishment economic thinking. They talked to a bunch of economists who, you know, they're not doing a total 180, but they are starting to acknowledge that maybe this mainstream economist position that AI will create more jobs than it destroys the way previous waves of technology have. Maybe this is a little out of date, or there's more nuance to it than previously thought. Second, there may be data backing that up. The Challenger report, which is a regular report we talk about for March 2026, attracts job cuts. Challenger and

1:11:12Gray is a recruiting firm. They show that US employers announced just over 60,000 job cuts March 2026. That's up 25% from February. AI was cited as the leading reason for 15,000 of those cuts. So it's about 25% of the total. Year to date, AI ranks fifth among all the reasons for job cuts. And since Challenger began tracking AI as a layoff reason in 2023, the cumulative total, so of all time, has now crossed 99,000 AI

1:11:45related job cut announcements across three years. Third, Jack Dorsey, who we talked about a couple weeks ago, is making the case for AI driven restructuring much more explicitly, perhaps than any other major CEO out there. So after Block cut 4,000 of its more than 10,000 employees, Dorsey has now published a blog post co-written with a partner at Sequoia Capital, arguing that AI should replace the entire traditional hierarchy of middle management. So Block, he says, is restructuring around three employee roles. Individual contributors who build systems is one. Two is directly responsible

1:12:20individuals who own specific outcomes on 90-day cycles. And third is what they call player coaches who mentor while staying hands-on with technical work. He said this restructuring was triggered by a capability shift he observed in December with Anthropics Opus 4.6 and OpenAI's Codex 5.3. Fourth, on the hiring side, Zapier released the second version of its AI fluency rubric, which now they apply to every new hire at their company. This requires candidates to demonstrate AI embedded into

1:12:51their core work, not just one-off usage. They want to show repeatable systems and measurable impact on quality, efficiency, or outcomes. They also have this new accountability dimension that they consider. They say with AI, you can delegate the work, but not the accountability. So keeping that human-in-the-loop high top of mind here. Zapier's language is also pretty blunt about their AI expectations. They say if someone isn't meaningfully improving their work with AI support, they just don't meet the bar. And then last but not least, a new Gallup survey shows that AI is reshaping how college students

1:13:23think about their futures. 42% of bachelor degree students surveyed said they have reconsidered their major because of AI. 16% said they've already changed their major over it. For people trying to get associate degrees, 56% are also reconsidering their field of study due to what AI enables. So Paul, what jumped out to you about these updates this week? I mean, I'm personally planning on diving in a lot deeper to Dorsey's thoughts. I thought those were kind of interesting. I read a, that might've been the thing that triggered. So I'd put a post on LinkedIn. I'm

1:13:54like, I don't know what day it was. It was, it was one of the days we were in Scotland and I was like, we were driving a long distance and we were sleeping in the car and I was like typing away. We had a tour. I was not typing while I was driving. We actually had like a tour guy driving us. I don't remember which thing I read that I wrote the LinkedIn post about. And then I didn't turn it into a newsletter post. Um, it might've been the Dorsey one. I don't remember, but it was abstract. Like his I read it. Cause I'm very interested in this. I'm actually like my make on keynote this year is going to be based on like a vision for AI forward org chart. I think, um, maybe like I'm working

1:14:27through an idea, Mike, you've seen some really, um, so I'm very, very keen on this idea of organizational structure and what teams are going to look like and things like that. So it did definitely catch my attention what they were doing. I love Zapier's approach. I liked it when they came with the V1. I really liked the V2. I liked the idea of this AI fluency rubric. So that was some really cool stuff. Um, and then just the jobs overall, like, again, like, uh, I'm glad to see people coming around and realizing like, this is a real thing. It's going to be a problem. The thing I alluded to that I wrote about on LinkedIn though, was, um, I'm getting really,

1:15:01really annoyed by the, like the tech leaders in particular who just keep pretending like it's all going to be great. Like with no acknowledgement of the possibility that it won't be. So I get optimism. Like I'm all, I'm all for being optimistic about this stuff. Um, and like believing in a future of abundance and like, we're going to find our way through, which I do think we will. Like, I, I think it's going to end up being great, but I also like straight up, like it's going to suck for a lot of people in the process. Like, this isn't going to be an easy transition and like a

1:15:32whole bunch of people are going to lose their jobs. And, um, so I get really annoyed when people won't like acknowledge both sides of the equation. So the example I put in the newsletter was, they said, tech leaders, politicians, economists who point to increasing demand for software developers and historical precedents as proof that AI won't displace millions of jobs are creating a false sense of hope. And then I highlighted four in particular, and some of these people are people I respect and like follow, but Andreessen, Mark Andreessen, this is a quote, and we'll put the links in the show notes. The AI job loss narrative are all fake. AI equals math, massive rampant productivity

1:16:04equals massive rampant demand equals massive jobs, but watch. So that was a tweet from April 5th. Aaron Levy, who we really like, like CEO of box. Like I'm a big fan of Aaron. He's got some of the best takes on X, um, about AI that I've seen. He does a lot of research on this topic. So April 5th, he wrote, there are far more categories where AI agents making things more efficient will induce demand for that skill than spaces where agents eliminate the work. This is why the AI job predictions will not play out as advertised. Okay. Um,

1:16:35Shyam Senkar, who's the CTO at Palantir. We've talked a lot about Palantir. Uh, he had an editorial February 2nd. He said, AI is a tool for the American worker, not his replacement. The job loss narrative is a ploy to attract investors, drive media attention and consolidate political power. The real promise of AI is the enterprise is to make the American worker 50 X more productive to unleash his taste and agency. This isn't speculation. It's reality. It's very, very confident. There's lots of confidence in these statements. And then David Sachs, this is no surprise. Um,

1:17:06he, he, he can't acknowledge the impact on jobs due to his relationship with the administration. He is the currently the chair of the president's council of advisors on science and technology, all caps, AI job loss, hoax exposed. And then it goes on to say, according to a new study from Vanguard, the occupations most exposed to automation are actually outperforming the rest of the job market in terms of growth and real wage increases rather than causing job loss. AI is making workers more productive, driving gains in both jobs and wages. So what I said was, despite these economic or optimistic outlooks from these leaders, the reality facing companies, especially those with

1:17:41limited growth and demand, which is a really important asterisk here, um, is that the pressure to reduce headcount across all areas of knowledge work is going to be immense in the coming months and years across all areas, marketing, sales, customer service, HR, finance, et cetera. Um, and then I said, pretending like there isn't at least a strong possibility of significant disruption is a disservice to business leaders who should be doing more to prepare their organizations and upskill their people. And then I said, I talked to executives every week who are being told to stay flat on headcount and to have a contingency of cuts ready to go if the efficiency from AI happens. And so I just, I don't,

1:18:20that's like my continued frustration is that all these people are hyping AI as this future abundance, which I I'm, I'm with you. Like I hope, and I do think eventually, but I don't know who you're talking to that is planning to hire. Like I'm not meeting those people, like unless they're anthropic or one of these companies that's growing at 20, 50, a hundred percent a year. Um, I I've yet to talk to an executive at a traditional enterprise. That's really happy with five to 10% annual growth.

1:18:50That's planning to hire, but it's not happening. So, and that's a knowledge work. Now, of course, there's exceptions to that in energy, in the trades, in healthcare, like, yeah, we can't hire enough people in those areas. I get that. I'm talking about the rest of us, all the other industries where the ultimate goal right now is to, to just stay flat and headcount and get the revenue per employee number way up. So yeah, I don't know. It's good. I guess it's good. Like it's increasingly becoming a conversation because it just really needs to be, we need to be thinking about what

1:19:24if these tech leaders were so optimistic, just are wrong. And what if it isn't as, you know, easy of a transition as they'd like to make you think.

1:19:34So similarly, we've had a lot happening on the AI policy and politics front. So we're going to go through a few developments here that have happened over the last couple of weeks. So first up, California governor Gavin Newsom signed a first of its kind executive order requiring safety and privacy guardrails from AI companies that contract with the state. So this basically establishes new certification requirements for AI vendors that want to do business with California. It requires them to attest to and explain their policies around preventing illegal content, harmful model bias,

1:20:06and violations of civil rights. It also directs state agencies to expand the use of vetted AI tools in government, develop an AI powered pilot for accessing government services and publish a data minimization toolkit. Second, at the federal level, the Wall Street Journal reports that the White House is racing to head off threats from powerful AI tools. So there's renewed urgency here in the wake of the mythos stuff we discussed. This included prominently a group of White House officials working on the issue, including convening a call with the vice president, treasury secretary, and the heads of

1:20:40Anthropic, OpenAI, Microsoft, and Google, as well as the leaders of cybersecurity firms CrowdStrike and Palo Alto Networks. That's obviously in addition to the previously mentioned meeting we talked about related to mythos specifically that the treasury secretary had with bank CEOs. Third, a major new survey from Fathom, a nonpartisan research organization, provides a clear picture of what Americans actually want from AI governance. I surveyed a bunch of people to ask about their feelings and priorities in a number of areas. So the top priorities across party lines for people are child safety, corporate accountability,

1:21:15and verifiable standards. Another big issue is workforce protection. So according to Fathom, from retraining programs to sovereign wealth funds that share AI generated wealth with the public, every workforce policy tested in this survey commanded majority support. Americans decisively reject leaving workforce transition to market forces. There is broad demand, but no preferred solution. A policy window that is open now, but won't be indefinitely. And then lastly, Politico reports

1:21:46that Senator Bernie Sanders may be building an unlikely alliance with Silicon Valley AI safety advocates. So Sanders recently met with quote unquote AI doomers in Berkeley, including Eliezer Yudkowsky from the Machine Intelligence Research Institute. And Sanders said, I know there have been a lot of science fiction novels and movies about how the robots and the AI and the computers rebel against human control, but these guys no longer think this is science fiction. So Politico suggests this might be

1:22:16the beginning of an alliance between anti-AI populists and the more tech-centric, perhaps effective altruist aligned AI safety advocates. Paul, did anything jump out to you here? There's a wild quote from Yudkowsky in the Politico piece where he just said, basically telling Bernie Sanders, hey, the point, if AI gets much, much, much more powerful, it'll run everything. And Sanders said, what does that mean? Humans are discarded? And Yudkowsky replied, think everybody dead. So there's

1:22:50some, some strong language being used. Well, I mean, just like the optimist side, we just talked about there's extreme views of everything. Yeah. I'll probably just leave it at that at the moment. But this is my concern is like design. You have an uneducated public largely about what these things really are, what the real risks are, what the real potential is. Um, and so there's always when you have, um, when the, when the literacy isn't high and the comprehension isn't high, then you have the

1:23:27ability for extreme views to come in and influence people's perceptions and beliefs. And that's very dangerous in my opinion. Um, you know, if you, if, so let's say, you know, this Yudkowsky, is that how you say his name? I believe so. Yeah. Yeah. So like, let's say that's the first thing you hear. It's like, this ends up on a 60 minutes or in that AI movie that just came out. Yeah. The Doomer AI movie that was out. Um, and you hear that and it's like, oh, I hate AI, I hate data centers. I hate Sam Altman. And like you, people take that perspective. And then you say, yeah, but it's, um, it found a cure

1:24:02for cancer to your family member that, you know, was suffering from cancer. Like AI is actually the thing that's going to find the cure or did find a cure. Like, so should we stop it? Should we not, should we not have AI now? Because you, you, you heard a bad quote or like, so, so, and that's when there's like all these nuances to the, when the people take extreme views, they, they don't stop and then say, oh, okay, well maybe that would be amazing. Um, and again, I don't, only parallel I could ever go back to is the internet and then say like, oh yeah, the internet's going to allow these scams

1:24:36and dark web and all these like horrible things are going to happen, but it's also going to open up the economy and we're going to be able to build all these amazing things. You're going to connect with people. You could never, and you're gonna be able to FaceTime with your family when you're a thousand miles away. And like, you want to, you want to not have any of that good stuff. You just want to like, should we just shut it down because like bad stuff might happen. Right. You can't. And, and so that's like my feeling on this is like dialogue and reason and finding paths forward where we can do this responsibly. But this, this absurdity of like, shut it all down because it's

1:25:11going to limit everybody. It's like, okay, that's your belief. And then, then we don't get any of the good stuff either. So how about we actually like, just be reasonable here and find the reality. And like, let's talk about the reality of the situation and not take extreme views that are unrealistic and mislead people. So.

1:25:32All right. So next up HubSpot announced that starting April 14th, it's customer agent and prospecting agent are moving to outcome-based pricing. So this means according to HubSpot, quote, customers only pay when the agents complete the task it's been assigned. Practically, this means the customer agent is moving from used to, used to, or will used to have charged you $1 per conversation, no matter what. And that's moving to 50 cents per resolved conversation, meaning you only pay when the AI actually solves the customer's problem. The prospecting agent is

1:26:06shifting from a recurring monthly charge per enrolled contact to $1 per lead recommended for outreach. So HubSpot says, quote, this means you now pay when a prospect's prospect gets qualified and handed to your team. HubSpot says customer agent actually now resolves 65% of conversations and cuts resolution time by 39%. And that prospecting agent activations are up 57% quarter over quarter. Both agents include a free 28 day trial and are available to pro and enterprise customers. So Paul,

1:26:38we've talked about this before, the need for SaaS companies like HubSpot to update their pricing models. What do you think of this approach? Are they headed in the right direction? I, I love HubSpot. I always have to preface my comments. I love them. Um, of the people, HubSpot, I love the company. We use the technology, like our company's built on the technology. Both my companies over the last 20 years have been built on HubSpot technology. Um, I like the concept of the outcome pricing. I understand from HubSpot's perspective, why they would move to this.

1:27:11I understand the messaging of why this is a benefit to customers. So I get all that.

1:27:19Um, Mike, you and I've worked in HubSpot for a really long time. Yeah. The reliability of the data is a problem. Um, allowing them to determine what is a resolved conversation is a very, very gray area. Um, so we get a spam form fill or something like that, or a spam chat and like you close it. And now we just paid 50 cents to close a spam chat. Like there's all these, like, what does that mean? Like, what is a resolved conversation? And, and now the work we have to do to go into understand what is a resolved conversation?