#232: Claude Watermarking, AI’s Environmental Impact, OpenAI Talent Drama & Anthropic’s Hidden Advisor

Welcome and Episode Overview

0:00So when you present an AI system's ideas and words as your own with no critical thought, that is AI plagiarism to me. It's like, that's the problem. It's not that you're using AI to help you. It's that you're not putting any critical thinking into it yourself, and the words aren't yours. Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Reitzer. I'm the founder and CEO of SmarterX and Marketing AI Institute, and I'm your host. Each week, I'm joined by my co-host

0:33and 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 232 of the Artificial Intelligence Show. I'm your host, Paul Reitzer, along with my co-host, Mike Kaput. We are recording this at an unusual time, Mike. It is Friday, August 14th at about 2 p.m. Eastern time. We

1:08usually do this on Monday mornings. The week got crazy, and then I'm gone Monday for an event. And then, shout out to Kathy McPhillips, our chief marketing officer who sometimes joins us on AI answers to co-host, who just delivered my laptop that I somehow forgot at the office today. So, it has been like a, literally a crazy week. I was in the office this morning. We were working on some stuff, team meetings, and then I shoot home to my home office to do, because where the

1:39podcast studio is, and I could have pulled my laptop out 10 minutes before Mike and I were about start recording this. I'm like, yeah, we got a problem. So, I messaged Mike. I'm like, it's my laptop by chance sitting at the office still. So, yes, thank you, Kathy, who was heading this direction anyway. It worked out great. She was on her way to a coffee shop, and yeah, she Ubered my computer to me, and so here we are recording. Otherwise, we would have been doing like a Sunday morning thing, Mike. Yeah. Okay. So, actually, semi-related to transition into this

2:11week is brought to us by MECON, the AI Conference for Marketing and Business Leaders that is happening October 13th to the 15th. The meeting I was in this morning with Mike and Kathy and Ashley and some of the others on our team was talking about MECON, and we decided to do something extra special for our podcast listeners during that meeting. So, here we go. So, if you have already registered for MECON using the Pod 100 offer, congratulations, you have already received what I'm about to

2:47explain. So, the way MECON works is Tuesday is a pre-event workshop day. There's five workshops you can elect into when you're registering, and then the main conference is Wednesday and Thursday. And so, what we decided to do is on Thursday, which usually by that point, I'm like hiding in the team room, like just trying to decompress momentarily. We are instead going to do a private lunch with me and Mike exclusively for podcast audience listeners. So, what's going to happen here is if you register

3:20using the Pod 100 code, we have a limited number of seats available for this lunch, but Mike and I are actually just going to hang out in the room and answer whatever questions the attendees have. So, if you've already registered with Pod 100 for MECON, which again is October 13th to 15th, you're in. The team will reach out to you with details about it. We have a limited number of remaining seats left in that room. So, if you go to MECON.ai, that's the event site, use Pod 100.

3:52Not only are you going to get the $100 off the ticket, you're going to get to be a part of that exclusive lunch for podcast listeners. So, never been a better time to get in at MECON or if you're not a marketer, tell the marketers in your organization about it. We'd love to have them join us. Ticket prices go up August 22nd. So, best pricing happens now and a chance to be in the room on that Thursday, October 15th. With me and Mike, it is a brand new thing. We've never tried this before and we figure, what the hell, let's see how it goes. So, VIP lunch, everything, all the lunch is

4:27provided. You just show up and network and we'll just hang out and answer questions and talk about whatever you want to talk about. So, again, MECON.ai. Mike, am I missing anything? Because, like I said, this is about two hours old that we decided we're doing this. No, I think that covers it. It should be, I think we're doing about an hour of lunch. So, you know, I'll have a good amount of time to chat through any of people's questions, topics, things they want to chat through. Yeah. So, it should be fun. It's always cool for us to get to meet the podcast listeners. You don't really know who the podcast listeners are

4:59until you go to these events and someone comes up and introduces themselves. And so, it's always awesome to do it. So, we figure, hey, let's get them all together if we can while we're already there. So, that is happening again, October 13th to 15th, MECON, in Cleveland, Ohio, which is our hometown opening night party on Tuesday at the Rock and Roll Hall of Fame. The event itself is at the Convention Center. It's going to be amazing. We would love to see you there. You can go check out the lineup. I think all but three or four speakers have been announced. So, the vast majority of the lineup, speaker lineup and agenda is there, including Dan Slagan, Mike, who you just did the

5:33AI Transformation Spotlight with, right? On Thursday. Yep. So, August 13th, if you missed it, episode 231 was with Dan Slagan of Zapier. And he told a crazy story about how they're building a second brain at Zapier. And I won't divulge everything, but go listen to that. And then Dan's actually going to be on the main stage at MECON. So, you can come and hear him talk as well. Okay. AI Pulse. We always start off with an informal poll each week. We're actually going to let last week's run. We're trying to increase the number of people that are taking this to try and

6:05get more projectable data. So, we'd love to actually start moving beyond just the informal poll and get like a formal survey that we can actually use the data for. So, last week's, we asked Mike about AI agents. Is that right? Yeah, that's correct. Yep. How people are using AI agents in their work. Great. So, you can go to smarterx.ai forward slash Pulse. It is one question. It'll take you all of about 15 seconds to do this. So, we would love it if you could go to smarterx.ai forward slash Pulse. No contact information gathered. This is purely just like go in there, answer the question,

6:36and get out. So, we'd love to have you do that. And then with that, Mike, there was a last minute topic that we threw in here that I don't even, was not on my list of top 500 things I thought would be talking about on the podcast this week, which is an apparent like media hit job on Dario Amadei's wife. But we are not leading off with that, but we are going to come back around to a wild story that is unfolding Friday morning as we are recording this.

Anthropic Details Content Watermarking

7:04Well, Paul, we are starting out with some Anthropic news because this past week, Anthropic detailed how it's going to start marking content generated by Claude. So, the company is actually going to like watermark the content that Claude produces. The company is using two techniques to do this. There's going to be an invisible watermark embedded directly in generated text that is produced by Claude and digitally signed metadata attached to generated files. So, this text watermark

7:34is imperceptible. Anthropic says you will not see it and it doesn't change the meaning, quality, or readability of Claude's responses, but the mark travels with the text when it's copied and pasted and it may persist through some editing. Now, for files like PNG, JPEG, and SVG images, Claude attaches signed provenance metadata that follows the C2PA open standard. This is an industry framework we've talked about in the past for recording how digital content was created or modified. These markings

8:07apply to users worldwide and cover Anthropic's products, including Claude apps, the API, Claude code, and its cloud platform integrations. This whole thing started or is driven by the European Union's AI Act. Anthropic signed that law's code of practice on transparency for AI generated content and says Claude models launched in the EU on or after August 2nd, 2026 will support machine-readable marking at launch with older models to be updated during a transition period. There is some commentary

8:43online of how this kind of text watermark actually works. I actually just really quick want to read, Paul, the excerpts of a post from our good friend Chris Penn at Trust Insights on the subject since he explains this far better than most people could. And basically, here's what he says of how this actually works. He says, every time AI generates a word or token, it calculates probabilities for what word should come next. The most probable terms are usually what AI picks from, hence why bad prompts lead to slop because slop is high probability. But when a company implements watermarking,

9:16they introduce a secret key. So instead of picking strictly at random from the top options, the key subtly quote unquote loads the dice. It uses the previous few words to pseudo randomly boost the probability of certain candidate words over others in a statistically meaningful sequence. This basically creates a measurable pattern in the text at the paragraph level. And so he says, can humans detect it? Not without assistance. And you need access to the model that created it to be able to detect

9:46it. To detect a watermark, a tool has to look at the exact underlying log probabilities and apply the secret key to see if the statistical quote loaded dice pattern is present. So you also, he notes, need access to the model that created it because each model has its own way of writing its own probabilities. For instance, Claude Haiku knows what the probabilities would have been for any given word. Claude Sonnet, for instance, would have different measurements because it's a different model. And just a few

10:16final words from Chris will link to his full LinkedIn post, which you should definitely read here. Does this mean AI detectors are suddenly good? Nope. Unless the detector software has been given access to the base model to do the analysis and given the secret key to decode it, they're actually likely to perform worse because now the statistical patterns they're trained to detect are slightly less predictable. They're still dangerous and inappropriate to use in any punitive context. Will the watermarks apply to text these systems edit like transcripts? Yes. Can you beat AI watermarking?

10:50Yes. By doing your own work in whole. So Paul, I'll leave it there. Just wanted to kind of like really give people a sense of what we're talking about here. There's a lot of commentary about this online, not just in AI circles. What are the most important implications of this for people using Claude regularly? Every once in a while, there's a topic, Mike, that I'll see right away and throw it in the sandbox. Like, yeah, maybe we'll get to that. And then you just get surprised by what gets people going. And

11:22this was one of those where I was like, whoa, what is happening? I was seeing people who usually don't even comment on AI stuff, like getting really pissy about this. I was like, this is tech that literally we've known about for like four years. Like we, I mean, Google, I'm pretty sure has been doing this for like two years. Yeah. Chris's post also mentions Google Synth ID, which has been around for years at this point. Which we've talked about like a dozen times on the show. Yeah. So I was kind of taken aback, honestly, by the visceral reaction from some people to this topic.

11:57And I actually thought I was missing something. I was like, what are they doing that we didn't already know was being done, was my question to myself. So I don't know, like I'll try and like give a little context here. And maybe I'm just reiterating stuff that, you know, we previously said, but, um, I was trying to comprehend this. So to summarize Chris's, which I love Chris and he, he does an insanely good job of breaking things down and providing a lot of like technical meaning to like what's going on. Um, I often will also try and take Chris things like, okay,

12:31what is it saying? Like, what is like the, the one sentence way to like explain this. And so the way I thought about it is when you go back to understanding what a GPT is, a generator, pre-trained transformer, what the transformer that got invented in 2017, that made all this generative AI possible. It's all about predicting. It predicts the next word or token, as Chris said, in a sequence based on its learning from all human data that it's consumed. And so in essence, it's just altering that prediction slightly in a way that really only the model knows. That's like

13:06probably what most people would need to take away is that the way these things, right. You know, it seems like magic, it's actually, actually math and it's just making predictions based on probabilities of what the next word will be. Um, and it just does that thousands of times per second and you get your emails and summaries and strategic briefs. And that's in essence, how these things work. So, um, yeah, I mean at a high level, that's, what's going on. Open AI, as we said, this has been known stuff for a while. So in, um, on episode 216, uh, this is May 23rd of

13:39this year, Open AI announced that they would be conforming to the C2PA, which stands for Coalition for Content, Providence, and Authenticity. That's what the, the acronym means, um, that they were adding Google DeepMind Sith ID invisible watermark to images. So it's not, you know, they didn't announce the text part, but then they also announced on July 31st that they would be doing it to audio as well. So watermarking metadata, like it's, it's a thing like open has been doing it and other modalities. Um, but if we go back to Mike episode 110, so this is two years, almost to the day, this is

14:15August 13th, 2024. So two years ago, um, this is, uh, what we talked about then. Open AI has a method to reliably detect when someone uses chat GPT to write an essay or research paper. The company hasn't released it despite widespread concerns about students using AI to cheat. Um, so again, just time timing wise, March 23 was GPT four. So we're, you know, a little over a year or so after that moment in time, we're now the use of chat GPT was coming more widespread within schools

14:49and within, um, you know, businesses. So this article then said the project has been mired in an internal debate at open AI for roughly two years. So they had text watermarking capability in 2022 before they released chat GPT, they had the ability to do this. And someone internally at that time said, it's just a matter of pressing a button. Like literally open AI could have done this in 2022 and they just chose not to. So the anti-cheating tool under discussion at open AI would slightly change how the tokens are selected. Sounds somewhat familiar. Those changes would leave a pattern called a watermark.

15:24So the reason I'm bringing all this back is to actually provide the context as to why we didn't have this already. Like maybe the big deal is that Anthropic put it out into the world where open AI to my knowledge to date still hasn't. Um, and there's different reasons why they didn't do it. So at the time 20, 2024 opening, I said, uh, if too many get access, bad actors might decipher the company's watermarking technique. That's going to happen regardless. Like I give it like a week before someone cracks how they're doing it. Yeah. Open AI employees have discussed providing the detector

15:57directly to educators or to outside companies that help schools identify AI written papers and plagiarized work. Google has developed a watermarking tool that can detect text generated by Gemini AI called Sith ID. It is in beta testing. Again, that was in 2024. It is fully go now. Um, in early 2023, this is interesting. One of opening eyes, co-founders, John Shulman, who, if I'm not mistaken, Mike jumped ship to Anthropic recently. I don't know how recently it was, but I think he did go to Anthropic. Yeah. That sounds right. I think he's at Anthropic now. So, uh, which may be,

16:31there's a connection here. Um, he outlined the pros and cons of the tool in a shared Google doc. Again, early 2023 opening eye executives then decided they would seek input from a range of people before acting. They also said opening. I needed a plan by that fall to sway public opinion about, uh, around AI transparency as potential new laws on the subject, the internal documents show. Um, and so again, back then we were already dealing with this. And then in May of 24, they published open. I published

17:02an understanding of the source of what we see here and see in here online, where they explained all of this, including the ability to do metadata, which may actually be more protective than watermark, and then Google has been very public about their Sith ID. You can go read about that. We'll put some links to that in there. And then Mike, we just talked six episodes ago, uh, Substack launched an AI detection feature. That's going to using Pangram. That's going to tag AI generated content. So again, I, I, I'm not sure what I'm missing here. I don't know why this created such reaction from people.

17:38Um, and even the AI act, like article 50 of the AI act, that's kind of, well, it's, it's voluntary, quote unquote, but like what's seems to be the trigger for, for Anthropic doing this is because of the AI act, the European AI act, even that we've known for like a year that this was the rule. So nothing seems new here to me other than Anthropic released the thing that we knew existed for four years that still doesn't work. Like, so the problem we have now is, and Chris addressed this a little

18:11bit, is that false positives are still a thing. Like you're still going to give these tools to teachers and professors and just general public who wants to criticize people for using AI. And they're going to like present Anthropic saying something was written with Claude as like fact, a hundred percent fact when Anthropic itself says it's not like you highlighted Anthropic cautions that a detected mark doesn't prove Claude authored the content and that the absence of a mark

18:42doesn't prove a human wrote it. Um, so I don't know, like going to this, what does it all mean? So I jotted down a couple of quick notes before we jumped on. So AI is increasingly going to be part of how people write either to help them draft, edit the work or to inspire creativity. Like that's, I use it that way sometimes like just to help me like inspire some things. Um, I don't think about like how the tokens were predicted. Cause at the end of the day, like I still rewrite everything or so like, it just helps me get going. So it's like, okay, like am I like a criminal because I'm using AI

19:15in some way to like help me inspire ideas. So some people and maybe many people will use Claude and others and probably already are as a replacement to having to write and think for themselves. I think that's the biggest fear in schools. So even though we have this tech, or even though this tech is now being shared with the world, the key is still to read to teach responsible use of AI. So not using it at all, isn't the answer, but we need to be able to test for critical thinking and writing skills. So AI should be able to accelerate that when taught properly. So just saying, oh, we're going to check

19:51all the students to see if they used AI, like that's not helping anybody. Like they should be encouraged to use it in the proper ways, unless there's specific instances where you want them to use pen and paper and just prove that they have critical thinking and writing skills, which is a logical thing. There should be checkpoints where it's like, okay, no AI this time. This is purely writing. We're actually going to get together. We're going to do this in class with no computers. And like, we just want to see, have you learned? Are you like at a checkpoint where you've now actually made progress? So what we need to reduce the importance of is AI slop with no critical human

20:25thought. Like that's the issue to me, which then bring me back to the idea of like, well, what is plagiarism? So when you present an AI system's ideas and words as your own with no critical thought, that is AI plagiarism to me. It's like, that's the problem. It's not that you're using AI to help you. It's that you're not putting any critical thinking into it yourself and the words aren't yours. And so that's what we see all the time with people who are like that. I mean, I'll use the AI industry as an example, like AI influencers or like people who are very prolific all

20:58of a sudden on LinkedIn and Twitter. And it's like, I don't mind it if it's actually your words. If you're literally just going to Claude and using it to create something because it's going to get you likes, that's like AI slop, AI plagiarism, in my opinion. It's doing no good to society. You're not actually adding, it's not additive to anything. So just to, you know, build on the plagiarism for a second. So common types of plagiarism, which by the way, I used plagiarism.org and Merriam Webster to source the information on plagiarism. Copying text word for word without

21:31quotes or credit, that is traditional plagiarism. Paraphrasing, where you're changing a few words from a source without naming the original author. Idea theft, where you're literally just using someone else's concept and claiming it at your own. And then submitting unacknowledged text or code produced by generative AI. So like what I think we need to get to, Mike, is just more of like an agreement with society that it's okay to say you collaborated with Claude on the words, but the ideas are your own. Or that something was co-authored with ChatGPT, but like maybe in the

22:02article provide context as to like how your original ideas or questions is what drove it. Like, I don't think we should make people feel bad for using AI in their writing. I think we just need to get to a point where we accept it's part of writing, but we just need to be transparent about it and use it in a responsible way and not lose, you know, not have cognitive decline because we forgot how to think for ourselves. Yeah, I couldn't agree more. I worry about those words, agreement in society, though. That's the tricky part. I know. I know. We can hope though, right? We can try.

22:36Yeah, but I couldn't agree more with the overall perspective.

AI Environmental Footprint and Data Centers

22:39Well, speaking of maybe disagreement in society, our next topic is about the AI industry's environmental footprint, which came under some new scrutiny this past week. There were some developments about corporate commitments, disclosure requirements, and some growing public backlash. So first up, OpenAI sent a letter this past week to Texas Governor Greg Abbott, committing to build AI infrastructure responsibly in the state, including pledges to pay its own way, protect residential and small business customers from added costs of data centers

23:13as they build these, conserve water, and provide accurate information about its electricity and water usage. Governor Abbott announced that OpenAI will comply with the data center standards he established for Texas earlier this summer. Bloomberg also reported that this past week, top AI companies, including OpenAI and Anthropic, have not disclosed their greenhouse gas emissions, made net zero pledges, or published sustainability reports, even as both companies prepare for IPOs. That may soon change because a California law known as SB 253 begins to take effect in November,

23:48requiring companies with more than a billion dollars in revenue that do business in the state to report emissions from their direct operations and energy use. According to Bloomberg, Anthropic is already working with the carbon accounting platform watershed to measure its footprint and comply with that law. We also saw a some public backlash on display in or being the subject rather of a recent episode of the Ezra Klein show in which Ezra Klein of the New York Times interviewed writer Jasmine Sun about the

24:18growing movement against data centers. And the episode notes that an overwhelming majority of Americans oppose having data centers built near their homes and that New York governor Kathy Hodgell has imposed a moratorium of up to one year on new hyperscale data centers, which we've talked about with more than a hundred similar proposals across the country. Not everyone agrees this backlash is justified. In commentary published this past week, the vice president of general economics and trade policy at a think tank called the Cato Institute argued that data centers are not the problem, bad policy is,

24:53and they wrote that much of the opposition rests on exaggerated claims and they cited figures showing data centers use just 0.3% of the U.S. public water supply in 2023. So, Paul, we've talked a ton about the backlash against data centers, the environmental footprint. I'm kind of curious, not only do you see the public's opinion on this changing, but also I've heard more and more conversations or gotten more and more questions like, do leaders need to be talking to employees about the impact AI is having on the environment or

25:26the perceived impact at least? I definitely think it's a growing group of people that are very, very curious on these topics. And some people have moved to the point where they're very passionate about their beliefs on these topics. Just from my own experience, having now been on the public stage talking about AI dozens of times a year since 2015-ish, up until last year, I could count on one hand how many questions I probably got about data centers in the environment. Every talk I do, regardless of who it's

26:01for, I get at least one or two now, every single time. So, just that tells me it's definitely moved past. People ask questions about business of AI and applying it to marketing, whatever. But oftentimes, in my state of AI talks, it just immediately starts going to impact on education, impact on the environment. What about data centers? So, it's definitely kind of been risen up in terms of awareness. It's logical. I mean, data centers are a hot button issue. They've obviously, like Bernie

26:32Sanders and others, have made them a very political issue. A lot of local communities don't love them, which I can't blame them. Like, the way I think about it, and I'll get into this a little bit with some of Jasmine's comments from the Ezra Klein Show. Like, if you polled anybody, do you want a data center in your backyard? My guess is you're going to get a lot of no's. I think there actually is some data on it, and it's pretty high. But I think if you said, do you want an Amazon warehouse in your backyard? No. Do you want a solar farm in your backyard? No. Do you want an industrial parkway

27:04in your backyard? No. Like, I don't want any of those things in my backyard. So, I don't know that asking that question about, do you want data centers in your community is really, like, telling us that much. It's like, they just don't want industrial things in their backyard. Okay. So, I'll get into that in a second. So, I'm just going to zoom in on the Ezra Klein Show, because I think there's some incredible excerpts from this worth shining a spotlight on. And then I would recommend to people, if you are interested in this topic, go listen to this episode. It's really good. She does an incredible job, and she actually went and

27:37spent time in the communities talking to people and talking to leaders. I think she actually just got back from a trip to China, and she even provided a perspective about how are Chinese communities feeling about this. Like, do they have the same reaction to data centers and AI? Yeah. So, the lead up in the podcast, in the summary, Ezra writes, it says, what is big and ugly and has united Republicans and Democrats at a time when it felt like nothing could? AI data centers. Talks about the polling data, about DeSantis in Florida, proposing legislation related to the

28:11AI Bill of Rights. Then you got Bernie Sanders on the other side calling for a data center motorium. So, it's like, everybody just seems to hate these things, except for the AI labs and the electric companies, basically. So, okay. So, then I mentioned the data centers versus other industrial buildings. That's a topic they do talk about, like fulfillment centers and things like that. And it's like, yeah, people just don't want those things there. And so, it becomes this like, maybe it's just the AI industry overall, though, that's the problem. So, one of the things I hadn't really thought about that she talked about that I thought was intriguing is when the labs and these

28:46hyperscalers go into these communities to build these data centers, they have everyone from the trade leaders to the city council members sign NDAs. So, no one can talk about this thing. But that ends up, what ends up happening is, so, say you get the leader of a local labor union signs an NDA. Well, they have to go then talk to contractors and subcontractors and laborers. And, like, eventually, word gets out, especially in smaller communities, that someone's bringing a data center to town. And then they come to council meetings and they complain about data

29:18centers and the council members can't say a thing because they're under NDAs. Then you lose trust. And it's like, well, now they're hiding stuff from us. It's like, well, yeah, they sort of are. But the thing that's interesting is a lot of this backlash started a few years back when the labs and hyperscalers didn't realize the public was going to hate AI and data centers so much. And so they did their usual NDAs, like they put everybody in our NDAs for everything. And so that was just standard business practice, not realizing that they were going to lose the trust of all these people. And now the NDAs were going to come back to bite them because everybody's

29:52going to know they were doing it anyway. The water issue is an interesting one because that is one where I feel like there's quite a bit of misinformation online about water. I think in the earlier days of data centers, it was a much larger issue, but there's been a lot of effort by the data centers and the companies behind them to solve for this. So specifically, Jasmine said they do require some of it, primarily for cooling the data centers because these chips and servers run really hot and they need AC. The thing that's gone a bit wrong in the water debate is that today's new data

30:27center construction is almost all closed loop systems in the same way that air conditioning is closed loop, which means that they recycle the water within the system and they use a fraction of the water that say a golf course would use as an example. And yet, you don't hear too many communities like complaining about golf courses. Right, right. She also related inference or like the use of Chad Chippity and other tools to YouTube videos and said YouTube videos use way more, like watching Netflix, watching any shows on Amazon Prime. That all uses more energy and water than Chad Chippity queries, but it's just a public perception

31:05thing. Electricity, concerns are real. They use a ton of electricity. We don't have enough electricity in the grid to provide where this is going. So that is a real issue. And so one of the things is they come in and they try and say, hey, electrical bills aren't going to go up. And the problem is like nobody believes them. Nobody believes any of these people, any of the tech people, any of the politicians. So when they say this, they don't believe it's true or they don't believe it'll be true perpetually. And so it's just like, it's a hard thing to message against.

31:35The positives, tremendous tax revenue. There was one she cited in a small community where Microsoft was going to provide, I think it was almost 20 million in tax revenue, which is massive for that local community, what it can do to its schools and public systems and things like that. Jobs. A lot of people think, oh, they're data centers. They don't have a lot of people working there. It's just like the labor for the year or two to build them. And then it goes away, which also isn't really true. Like they do create high paying jobs. They will sustain for probably at least seven to 10 years or beyond that. And so you do have good jobs going into these. But

32:11I think the thing, Mike, that just kind of hit home to me is that overall, the AI industry seems to have far more of a reputation and trust problem. And the one thing I thought, I really liked that they explained her and Ezra both is when you deal with industrial facilities or solar farms or things like that, there's this obvious benefit. Like, okay, like you're going to put a car factory in, I use cars, cars are helpful to the society, whatever. But if you say I'm putting a data

32:41center in somewhere, the average American and really average anybody around the globe is like, well, what does that mean to me? Like, what do I get out of a data center? Like, well, you get to use ChatGPT. Okay. Like, I don't really use ChatGPT that much, or I didn't really find it that great when I used it that one time. So there's this lack of understanding of the good. And so what you have is all these AI Silicon Valley billionaires, as they said in the article and in the podcast, who benefit from all of this, but what the individuals get out of it isn't very obvious.

33:15And so that becomes the larger issue is that for years, the Silicon Valley leaders communicated to the public about AGI and solving math problems and this future of abundance when they should have been making the benefits real and tangible to the average consumer and worker. And changing their tone, like they all did like six weeks ago, simultaneously on jobs to where, oh no, it's going to be great. We're just going to create all these jobs. That's not going to cut it. Like that was, I'm sure part of a comm strategy that someone told them all to do. But that's the real

33:45problem to me is they have a communications and PR problem. But to Ezra's point, they have a product problem. Like the value proposition of a data center is unclear to everyone, but the electric companies and the trades and the companies themselves that are building them. So that, I don't know, it was fascinating. Like it just, it opened my mind to a lot of angles that we haven't talked a lot about on the show. And I thought she did it. Both of them did it in a very approachable way. Yeah. There's a lot of really helpful nuance here. And I just wonder, I keep coming back to,

34:17and I don't have a great answer for this, but it's like, if your employees or people that are customers of yours or clients or anyone within your organization are coming with these perspectives of like, these things are terrible, what's the use of it? How on earth are you supposed to achieve any type of AI transformation with those folks? Like, is there any way to kind of message that, not even messages, just to educate around it? You don't have to have a strong perspective, like, oh, let's be pro data center, but more, how do you talk about it?

34:48Yeah. And I think it just goes back to even within the companies, just communications and transparency. Because again, like if I go to a private event for a company, I mean, this, I did a private event for, let's just say one of the companies that's building the data centers recently. So it was like 400 of their executives. I got questions about the environmental impact of their own technology from the people within the companies building it. Yeah. Like, worried questions, like what's going to happen in our local community? What does this mean? Like, what's going to happen with jobs? So there's a lot of people who don't understand

35:21what's going on, even within the companies you would think would understand all of this. Because it is, it's all moving so fast. And a lot of the people building it aren't thinking about what does this actually mean to the different stakeholders in the community, in our own company who worry about these topics. And maybe it actually affects their willingness to use the AI themselves because they worry about the impact it's having on the environment. So, I don't know, just because you work at a company that wants to be AI forward doesn't mean all your employees are on board with it. And there could be a number of different reasons, including some of them just really have

35:54concerns about this stuff.

OpenAI Hiring Controversy and Personnel Shakeups

35:56All right. So our third big topic this week is kind of an interesting, dramatic story where OpenAI is taking kind of a disproportionate amount of heat from the White House over its hiring of Dean Ball, who is someone we've talked about at length on the podcast. He's an AI policy writer and former Trump administration official who joined OpenAI earlier this summer as its head of strategic futures. He writes a widely read AI policy newsletter. And last year, he spent four months as a senior policy

36:28advisor for AI at the White House Office of Science and Technology Policy, where he says he was the primary staff drafter of the administration's AI action plan. Now, this past week, the New York Post reported that the White House officials are warning OpenAI that the hire could damage the company's relationship with the administration. Three officials told the Post that Ball exaggerated his role in the AI action plan, and they described him as a junior to mid-level policy analyst whose ideas were regularly ignored. One official said he was,

37:03quote, at best a nuisance and at worst irrelevant. And, you know, this is building on, you know, Ball suggesting on X that the White House should create regulatory risk to discourage American companies from using Chinese AI models, to which White House AI czar David Sachs at the time asked whether he was confessing to a regulatory capture strategy. Defense Undersecretary Emile Michael called him the AI world's supreme village idiot. An OpenAI spokesperson defended the hire, saying Ball's

37:34role focuses on research, not lobbying or political outreach. Ball himself has kind of brushed off this report, mostly with some jokes and, like, tongue-in-cheek commentary. On top of this, this past week also brought some other OpenAI personnel news. OpenAI's longtime Chief Operating Officer Brad Lightcap, who moved into a special projects role earlier this year, announced he is leaving after eight years to start something new. Former OpenAI Chief Product Officer Kevin Weil is raising $150 million for a new AI science startup as well. So a couple personnel shakeups here, Paul. But really,

38:08this, like, Dean Ball thing is kind of interesting. For some reason, it seems like he's really gotten under the White House's skin, despite the fact they keep saying, like, he wasn't that important. Why is there this disproportionate amount of attention being paid to him? You know, I was, like, half joking to Mike as I was leaving the office today without my computer, apparently, that, you know, we basically host an AI soap opera show. That was when we decided to put the Dario Amadei wife article into today's episode. But this certainly fits into that category. It is not

38:41intentional. So if you ever feel like this is a soap opera, it is. We just do our best to commentate and make it explain why this matters to talk about this stuff. Dean Ball is very influential. He was a very high profile hire for OpenAI. He definitely made some enemies in the Trump administration prior to joining. We covered on episode 222 his June 26, what should be done post, which I think probably did not help things. I would imagine this inflamed some already high tensions within people in the Trump

39:15administration. When he was joining OpenAI, when he announced it, he claimed he was going to be able to continue to share his thoughts openly. What I said at the time was, like, I hope that's true. But that would mean that OpenAI remains comfortable with what he has to say and that the government doesn't exert pressure on OpenAI if they don't like what Dean Ball has to say. And I think we have now run into a case where an administration that doesn't mind throwing its weight around, especially if there's

39:47people that they feel aren't toeing the company line, I guess you could say. They don't have a problem with trying to make your life miserable and trying to get you fired from places. So I would guess that there is quite a bit of pressure already at OpenAI to move on from this experiment. I will be fascinated to see if OpenAI stands their ground on this one. If the Trump administration got Anthropic to silence Dario, the CEO of Anthropic, one of America's most important companies, they

40:20basically sidelined him from talking to the Trump administration because they didn't like him. I don't think it's a far fetch to think they could get other people silenced if they wanted to. So the New York Post said that the tensions between Ball and the White House first came to a head in February, I think as you were referring to with this high profile spat with Anthropic. At the time, Ball called the Pentagon's decision to label Anthropic as a supply chain risk a psychotic power grab and almost certainly illegal. So that could definitely trigger some issues. He also did

40:50an interview with Ezra Klein, Ezra again, that we did cover at the time where he talked a little about this. But I'm going to zoom in, Mike, for a second on that June 26 essay. He had 35 things that should basically happen. Number one, when President Trump signed earlier this month the executive order on cyber and AI, which claimed to establish a voluntary testing program for frontier models, it was really establishing a de facto involuntary licensing pre-approval regime for frontier models. This analysis has proven correct. First, the administration revoked public access to Fable.

41:27Now it appears that OpenAI's GPT 5.6 is being limited to only a small set of US companies. So that probably wasn't looked upon kindly. He's right. That is what it is. It is a de facto involuntary pre-approval regime, even if it's not what they're calling it. And then number five on that list, this is probably the one that really pissed some people off. Nobody I know in the Trump administration has any frontier AI experience. Just a few months ago, someone with experience at both OpenAI and Anthropic was hired to run the Center for AI Standards and Innovation, but he was fired by

42:01senior administration officials within days. The lack of technically expert staff is one of the main reasons to doubt the near-term ability of this administration to produce a high-quality safety standard anytime soon. The New York Post, when they asked for comment this week about this, OpenAI spokesperson pointed to a June 18 tweet by the company's chief strategy officer, Jason Kwan, quote,

42:31really glad Dean is joining OpenAI. He spent a lot of time thinking seriously about the biggest issues frontier labs need to get right, risk governance, frontier policy issues, and what comes next. We won't always agree on everything, which is a good thing. This is a really important moment for these debates, and we'll be better for having him pressure test and shape our thinking. So high level shows how political all of this is becoming. Everything within the labs is political, which I think may or may not have something to do with the hit piece we're going to talk about. And then how sensitive the administration

43:02is to criticism is just like, it's hard to watch. So yeah, this is going to get messy. The administration won't give up. If they don't like someone, they don't just decide next week, it's fine, leave them there, it's cool. So they're going to make life pretty miserable for OpenAI, and I could see this not being a long-standing employment arrangement one way or the other. Yeah, call me cynical, but given that Dean Ball, as much as I respect his work, is not a member of

43:34technical staff working on the models, I think OpenAI values more its government contracts and access than any one person. Yeah, and I don't know him personally, Mike, but we've certainly followed a lot of his work and writings and interviews in the last 12 months. Yeah, doesn't come across to me as the kind of guy who's just gonna shut up and do what he's told. Like I, right, right. I just feel like if someone at OpenAI comes to him and says, you got to tone it down, he'll be like, all right, man, this didn't work. Thanks for,

44:06thanks for the shot. Like, yeah, I'm gonna go make my millions on the speaking circuit and having opinions like it was fun while it lasted. Yeah. All right, before we dive into rapid fire, this episode is also brought to you by AI Academy by SmarterX. AI Academy by SmarterX helps individuals and businesses accelerate their AI literacy and transformation through personalized learning journeys and an AI-powered learning platform. We add new educational content literally weekly, so you always stay up to date with the latest AI trends and technologies. This episode is brought to you by the AI for Industries collection, which

44:42features eight course series and certificates designed to jumpstart AI understanding and adoption. We have AI for professional services, healthcare, software and technology, insurance, financial services, retail and CPG, manufacturing and education. These are certification series that are an ideal launchpad for organizations that want to level up their teams and accelerate AI adoption and impact. We have individual and business account plans available now through AI Academy, or you can buy single

45:13courses and series for one-time fees. So visit academy.smarterx.ai to learn more. And you can also use the code POD100 for a hundred dollars off any individual membership.

45:25All right, Paul, let's dive into something. Enough of the tees. Let's get into it. Yeah, that broke just before we started recording, kind of weird scenario, very dramatic.

Profile of Anthropic CEO's Wife

45:37Let's get into it. The Wall Street Journal published a profile of a woman named Kami Clark, who is a name you have not really heard in AI, but happens to be the wife of Anthropic CEO Dario Amadei. They called her one of the most influential voices shaping his decisions as Anthropic heads towards an IPO that could top $2 trillion as soon as this fall. Clark has no role at Anthropic, but people close to the company say she acts as a sounding board and strategic advisor.

46:09She brought in a key early investor in 2021, former Google CEO Eric Schmidt, whom she had also previously dated. The Journal reports she also pitched Schmidt on a venture fund called the Mother of AGI Fund, which was designed in part to formalize her involvement in Anthropic. Other co-founders, including Amadei's sister, Daniela, didn't support the plan. It never moved forward. Now, the real story here is like so few details about Clark exist online. It's like literally the first most people

46:40are hearing of her. Most people didn't even know he was married, I don't think. I did not. Yeah, I don't think that was really knowledge. The Journal actually reports that efforts have been made to remove references to her. Amadei's Wikipedia page didn't note he was married until this summer, and Claude itself answers queries by saying Dario Amadei's marital status doesn't seem to be clearly confirmed. Here's where the weirder parts happen. The profile also digs into Clark's

47:10entrepreneurial past. She, at one point, was... This might get us banned. We may not get any reach on YouTube this week when you get into this. Well, we're about to find out what the limits are, but she, at one point, was pitching a woman-focused porn company. Revolutionary porn company. Revolutionary porn company, so you can go do research on that on your own. She unsuccessfully pitched Jeffrey Epstein to invest in. These are according to emails released by the Justice Department. She was also at one

47:46time working on a woman's dieting app that morphed into a woman's healthcare AI company. One critical, probably, piece of context here is this is happening right as Anthropic is hurtling towards its IPO. The Financial Times reported this past week that investors expect the company to go public as soon as October at a valuation of $2 trillion or more, which would be the largest IPO in history. They're citing company revenue projections of $100 to $120 billion by the end of 2026. So,

48:19Paul, I don't know where you want to start, but nobody knew he was married. Nobody knew about Cammie Clark really at all. Why are we suddenly hearing about this now? She dated Eric Schmidt. This is a really fascinating detail. The other thing, Mike, that I just, I think I mentioned to you why, you know, part of me wanted to not talk about this. The other part of me is like, I think we have to now. As I mentioned up front, this is all the makings of a political hit piece. Because the one you referenced, Mike, the Wall Street Journal, I don't know if they

48:53cited the information, but the information had this first, I believe. So, the information has the story, it's timestamp August 13th at 2.09 PM. I don't know what time the Wall Street Journal one was at, but someone obviously did this. Like, someone gathered August 13th at 8.42 PM was the Wall Street Journal. So, they followed on. And it seems like they were directly sourced the information. They're not just re-reporting what the information had. Which tells me, somebody put this package together

49:29and then reached out to very high profile outlets and said, any interest in a story on Dario's wife?

49:39I'm just going to stop there, Mike, because I don't want to get in trouble. It's just very, very intriguing timing. Like I said, it just is the kind of thing you see in political campaigns. And I'm probably just going to leave it at that for now. That's fair. I think we'll probably learn more in the coming weeks.

Bernie Sanders Demands AI Development Pause

50:04All right. So, next up, Senator Bernie Sanders of Vermont sent a letter this past week to OpenAI CEO Sam Altman, Anthropics CEO Dario Amade, and Meta CEO Mark Zuckerberg, demanding that their companies pause AI development. Sanders pointed to reports that AI has been used for the first time to create new viruses, which we covered on the podcast, and to recent incidents in which the company's models escaped their control, including an OpenAI model that hacked into another company. We also covered that on the podcast. He argues that companies are betraying their own commitments. He cites pledges that

50:38each of them made between 2023 and 2025 to pause or stop development if their AI grew too risky. He writes that the moment has arrived and AI capabilities have reached a critical threshold. In the interest of humanity, he said, stand by your words. Pause AI development. It is not too late to avoid disaster. Stop building machines that humans cannot control. He basically ended with a direct warning saying, if you do not take appropriate action now, my colleagues and I in the US Senate

51:09will. At the same time in Washington this week, the White House is reportedly preparing to expand its AI policy and oversight of AI models. A group of House Democrats called for the CEOs of OpenAI and Anthropic to testify under oath about the recent AI-enabled hacks. And Senator Jim Banks of Indiana recommended federal oversight of unreleased AI models. So, Paul, no surprise here we're getting more political battle lines being drawn. Still, pretty strong words from a sitting US senator,

51:41like, how seriously should we take Sanders threat? Like, why now? Why is he doing this? Yeah. Again, I just, I feel like I just like should hit a button that repeats this disclaimer every time we do this political stuff. But like, if anybody's a new listener to the show know Mike and I do our very best always to just remain completely political neutral in these conversations. Anytime we're talking about AI, I'm just straight up looking at it as someone who studies the space and observes it and what I think, you know, is best for society kind of stuff. I could care less who's on what side of the aisle saying whatever they're saying. And honestly, they have

52:16no clue what they're saying anyway. Everybody's like, they're actually agreeing on some AI things, which is pretty amazing to watch. So I'll just comment on Bernie Sanders stuff is absurd. Like, it's not pausing. We're not going to stop building data centers. Like, I don't know. I've never followed his career well enough to know what his shtick is. Like, I don't know what the end game is of saying all of this. Like, maybe it's just to raise awareness and like, move the conversation, which is fine. Like, I have no problem with that if that's how it's done. But an outcome from that,

52:49it's not happening. Like, if we're not pausing AI, and that would be like the worst thing we could do in America is like, just completely pause AI because the other countries aren't doing it. Like, it's just not going to happen. It is not a reasonable, logical thing to even be proposing. That being said,

53:07having open AI and Anthropic testify all for it, man. Like, that was some wild shtick. Like, what just happened with those AI agents, we shouldn't just gloss over as, oh, well, yeah, they broke containment and communicated with each other and build agent swarms. And like, that was weird, huh? Like, no, that was like an inflection point in the advancement of the technology and its integration into society. Like, we should probably stop and have some conversations about that. So, yeah, all for it. And then federal oversight of unreleased models, that is exactly

53:41what I called out last week, that it made no sense that these few select labs and their handpicked, already wealthy partners get to use the most advanced models, which could do horrible things as long as they don't just release them. So, yeah, hell yeah. Like, if you're going to regulate or have oversight on released frontier models, you should do the same thing for open weight models and you should do the same thing for unreleased models. Like, let's just do it. Makes sense.

54:12Like, I don't, so again, I look at things as it's just trying to like shake some things up and get people pissed and like get talk and that's fine, but it's not logical versus, well, these are actually like pretty reasonable things to be discussing that could help quickly if we could do them. So,

54:32yeah, this is a lot for Friday, to be honest with you. Like, my brain was not ready for this. Yeah. I told someone on a call the other day, for whatever reason, just with all the stuff going on, it felt like a week of Mondays. Oh my God, yeah. Today feels like another one. You're right. Yeah. And the funny thing is, is like, we were originally going to do this at 9am today. I realize I'm totally sidetracking now, but I guess this is what happens on Friday afternoons. Um, and, uh, I said something last night to my daughter and I was like, I don't even know how

55:02I'm going to do this tomorrow. I'm gonna have to get up at like 5am and prepare for this. She goes, why don't you just do it at a different time or day? And I was like, well, we're already doing on a different day, but maybe we should do it a different time. You're right. And so I messaged Mike and I'm like, Hey, what about like 1.30 instead? Because then we can see what else happens on Friday. Well, we would have missed the whole like Dario Amadei saga if we would have done this at 9am. Anyway, it was a good move. Yeah. So back, back to the podcast, I guess.

XAI Releases Grok 4.6

55:30Well, next up XAI has released Grok 4.6 this past week. This is a new frontier model built for long running AI agents, multi-step coding and interactive and visual work. The company says the model verifies its own work more often before moving on. It can handle up to 500,000 tokens. On Artificial Analysis Intelligence Index, a composite score across nine major benchmarks, Grok 4.6 scores 61. That's up five points from Grok 4.5 and matches OpenAI's GPT 5.6 sole. The biggest jumps came on

56:04agent, agentic work and coding tests. Its score on DeepSWE, a benchmark for real world software engineering tasks, rose. Its score on Apex agents, a benchmark for multi-step agent workflows, that rose. The model is now available through the XAI API, Grok build and cursor, along with third-party platforms, including OpenRouter. This release comes after SpaceX had acquired XAI earlier this year. And SpaceX CEO Elon Musk is already pointing to what comes next. He posted on X that Grok 4.7 is

56:38significantly better than 4.6 and should be ready in three to four weeks. So Paul, I think what kind of caught attention here is at least anecdotally, um, people had kind of, some people at least had started to count XAI and Grok out, but this release is getting a lot of positive attention. The model in some ways may be on par with GPT 5.6 sole, which surprised a lot of people given where XAI was in this race so far. Like is XAI back in the race? They seem to be. And speaking of soap operas,

57:09Elon's been pretty chill lately. Like we haven't had any like crazy Elon stories in a while.

57:15That's probably good. But like, although I did see this morning that, uh, it got leaked that they may, you know, so the Roadster, which was Tesla's first car back in whenever, I don't remember what year they, they debuted the Roadster, but they've been talking for like 10 years about coming out with the new version of the ultra sports car. And apparently now they've been testing a flying version of it. So we might actually get the Jetsons. Like we might get our flying Roadster. Um, there's some rumors that they might actually preview it before the end of August. So we shall see. Yeah. I, I don't know. I wouldn't say I was someone who had written off Grok,

57:48but I would say that when they started, um, leasing out compute in Colossus and Colossus 2 to Anthropic and others that it seemed like they were maybe moving in the direction of just a competing model, but not trying to necessarily be at the frontier because they were giving up some of that compute, but I don't know. I mean, they're moving fast, coming on strong. And I, it seems like Grok's even jumped Gemini at this point. I haven't looked at the data recently, but like, yeah, it's wild how fast this stuff moves, but I would never, never underestimate Elon's ability

58:22to do really big things. Um, when you don't expect it. So yeah, we'll see.

Google DeepMind Leadership Changes

58:30All right. So next step on the last episode, we covered last week, the episode we covered Google's AI leadership reshuffle when the company announced Google DeepMind CEO Demis Hassabis would become DeepMind's chair and chief scientist of Google parent alphabet, uh, with his deputy, Corey Kavakuglu taking over the lab in the days since a little more reporting has come out, come out on what led to the shakeup or some of the details behind it. So this past week Reuters published an inside account based on seven people knowledgeable about Gemini's development. It reported that Google

59:04co-founder, Sergey Brin, who holds no executive title has been informally influencing how the company trains its models. And he urged DeepMind staff at a town hall earlier this year to move faster as rivals pulled ahead. Reuters reports that a new version of Gemini was delayed roughly two months after internal testing showed it lagging rivals in areas like coding. And the staff later learned non-technical teams would move out of DeepMind and into corporate Google. As Kavakuglu takes over DeepMind, he will also

59:36have apparently the final say on major decisions at the lab. Reuters describes this as a further erosion of the lab's autonomy since Google acquired it in 2014. Separately, the Wall Street Journal reported how Hassabis had pitched that new AI oversight body, which we talked about in past episodes, in the months before this shakeup. This was an idea he first made public in mid-July, a US-led standards group modeled on a FINRA, a financial services regulatory body, that would safety test

1:00:06frontier AI models for dangerous capabilities before release. Apparently, uh, Demis discussed the proposal privately with Trump administration officials, executives at rival AI labs, and European policy makers before making it public. So, Paul, a few new details here about all these big moves at Google. Especially interesting, Sergey Brin is getting back in the mix, it seems, a bit. Yeah, we talked about that on, what was that, episode 230, about Brin's, like, increasing role and went back and looked at his comments at Stanford, uh, where he was getting interviewed, um, sort of a prelude to, to all of

1:00:42this. Um, again, I feel like I'm just, like, conspiracy guy today, but this is, this is totally getting leaked. Like, so, they're trying to control the narrative and alter perceptions about Demis' role a little bit, which is really weird to me, considering he's still the chairman of DeepMind and the chief scientist at Alphabet. But, like, in that Reuters article, it said in past years, Asab has worked against some efforts that could have met new revenue for Alphabet or helped it gain better footing in the AI race. Um, I don't know. There's just some things where they're trying to kind of

1:01:15say, like, maybe the leadership we had wasn't moving fast enough and we needed to focus more on product, less on long-term research, and, um, and this is actually a good thing. And I get it. Like, I understand why you would do it. They have to kind of do this. But, um, yeah, the article said, Bryn has used the implicit power he holds as Google's co-founder to push resource allocation towards specific areas like recursive self-improvement. That's interesting, you know, to be calling that out. And I kept coming back, like, over the last couple of days,

1:01:46I was thinking more and more about how far Google has fallen in eight months. Like, it's wild to see, go from Gemini three or whatever, being like the top model to, you know, lucky if they're top 10 at the moment. Um, and I wonder, and this article said, like, they had to delay the release of the next model. And now it sounds like 3.5 just might get buried. Like they're not even come out with the pro. They might just go right to Gemini four, but internal testing hasn't been great so far. I feel like they're gonna, they're gonna need world models to be a key on lock because that's where

1:02:20they're, I think no question still in the lead, um, is on world models, uh, image, video, understanding, you know, physics kind of stuff. And if that becomes a key on lock to AGI and beyond, they could very quickly, like this omni model that they would talk about with all modalities in one. Yeah. They could reemerge pretty quickly and be like, oh, they're back. Like, I expect that to happen. I kind of think it will, but tough stretch to, to watch. Like they're

Revisiting the Responsible AI Manifesto

1:02:50yeah. It's rough. Yeah. All right. So next up this past week, Paul, you posted on LinkedIn revisiting something called the Responsible AI Manifesto for Marketing and Business. This is a document you originally wrote in January, 2023, just a couple months after the launch of ChatGPT. And it lays out 12 principles that guide SmarterX's human centered approach to AI. This includes commitments like the responsible design development, deployment, and operation of AI technologies,

1:03:22a human centered approach that empowers and augments professionals and keeping humans accountable for all decisions and actions, even when assisted by AI. So you also at the time released this under a creative commons license. So other companies can adapt it as a starting point for their own responsible AI policies. But in your post, you said that you revisit these 12 principles periodically to see if they need updates. But so far you haven't felt compelled to publish a version too. But you said you're curious how others think about this, especially as AI agents become more

1:03:53reliable and autonomous, which introduces different types and new levels of risks. So Paul, walk us through this manifesto and why you might be talking about it or thinking about it or revisiting it now. Yeah, when I do my state of AI talks, I'll often weave in, you know, a few of these principles or talk about the importance of having AI principles as an organization. So I like come back to them periodically just to, you know, look at that. And I think I was, I was maybe preparing a deck for a talk this week. And so I happened to be in there and I was like, yeah, I haven't really thought

1:04:25deeply about this in a little while. I wonder if I would change anything. And that's when I threw it on LinkedIn just to get feedback from people. I was like, anybody else say anything? Because there's a part of you, it's like, well, I'm probably too close to this. But I mean, this was three and a half years ago, I wrote this, like, this was two months after ChatGPT. It's kind of hard to believe that it would stand the test of time. Like you would think I would have probably missed on something significantly. But I don't know, like I read through and I'm like, I don't think I would change anything yet. Like there's the one that jumped out right away is obviously related to AI

1:04:57agents. I guess there's two of them. So the number two principles, we believe in a human-centered approach to AI that empowers and augments professionals. AI technology should be assistive, not autonomous. I do believe that. I think that there's some instances where autonomy in low-risk environments, you know, where you could in theory get fully autonomous with some workflows. But I would have to like look at a list and like go through and be like, yeah, that's the one I would do. Like I don't off the top of my head know where that would change yet. But right now I still feel like humans have to be in the loop. And then the number three was we

1:05:29believe that humans remain accountable for all decisions and actions. Even when assisted by AI, the human must remain in the loop in all AI applications. That actually goes to maybe a little bit of what we just talked about with Open Ananthropic and they should testify. Like they're responsible. That was their agents that went rogue. Like you did that. You put them in the environment. You gave them access to the services that access the internet. Like they hacked other companies. Like that's you. So I feel like we just sort of as a society move past the humans and organizations are still accountable for their actions. I don't know. That's kind of weird to

1:06:04me. So, and then the other one that has held up and I wrote this one, I remember at the time, very specifically, and I wrote these by the way, in like 30 minutes. It was like back in 2023, I was at the gym. I started like having these thoughts about like, wow, this is going to go really wrong, really fast. And I stopped in between reps and I like started writing these out. And then I got home and I just finished writing. And I was like, we're just publishing it. And I probably went to Mike. I was like, Hey man, could you just put these online for me? It was it. Like there was no long, like drawn out thing and vet these. I didn't use Chad Chibi T to help me write them. Like this was just ideas. So the one I said was, we believe in personalization without invasion of

1:06:39privacy, including strict adherence to data privacy laws, mitigation of privacy risks for consumers. And the key following our moral compass, when legal precedent lags behind AI innovation that has continued to remain true. And it will be true for the foreseeable future. Like legal precedent is always going to be behind the, you know, AI and society. So yeah, like Mike said, these are totally free for anybody to use. The comments I got on LinkedIn, the couple that jumped out at me as a few people did bring up the autonomous agent thing and ask some like following questions about that. And then someone actually mentioned like, Hey,

1:07:11it seems like I might be missing the environmental impact part. And I was like, that's actually a good one. Like, so I haven't added a 13th one, but if I do, it'll probably tie something to, um, you know, the impact they have on the environment and being conscious of that and doing what we can, that kind of thing. So just to reiterate companies can use this for themselves. Um, would they just like, what's the first step? Like you just copy it and start remixing it in whatever way you would want to use. Yeah. The link we'll put in has, you can go look at the whole thing. You can download a PDF, but creative common share alike license is literally means you are free to mix, adapt and build on the

1:07:47work, even for commercial purposes, as long as you credit the source and you license your creation under the same terms. So it's in essence, like open sourcing an idea where I'm not going to like come at somebody for plagiarism because they took our 12 and published them. You're welcome to, right? You can add to them. You can edit them. You can do whatever you want. You just have to do it under the save creative comments license. So other people can build on your work.

AI Agent Hacks Gym Website

1:08:09All right. So here's an example in this next topic of maybe something that is related to agents and some of the security issues around them. So an AI agent asked to book a gym class, instead hacked the gym's website in what ABC news, Australia reported this past week. It's possibly the first known case of an autonomous AI cyber attack in the country. A Melbourne man set up the open source agent software, open call powered by Anthropics Claude in this case,

1:08:40and asked it to book him into a popular morning class at his gym. The agent found a flaw that let it reserve classes further in advance than the gym allowed. So he was sitting fourth on wait list for another class and asked the agent whether it could move him up on its own. It found that the booking system never checked whether a request to cancel someone else's reservation was authorized. So it canceled the booking of the person in first place, bumping the owner of the agent from fourth place to third. The agent told him the site's backend had

1:09:14zero authorization checks on canceling other people's reservations and said it had tested this on the person in wait list position one and reported that the cancellation actually went through. So when the owner of this agent asked it, Hey, like I didn't want you to do that. Undo this change. The agent said it could not. It told him the person it removed was gone from the wait, wait list with no way to restore them. It then drafted a disclosure email to the gym's booking software vendor, explaining the flaw and suggesting fixes, which the owner of the agent signed off on in the agent sense. Anthropic had not

1:09:49responded to any press requests around this as of this past week. So like, this is kind of a small, weird thing, Paul, like, but it, I think it's kind of interesting to talk about. First, we got agents hacking websites from the lab. Now we've got individuals accidentally, it seems almost using agents to hack websites, just trying to do basic stuff using agents. Like, is this a test of what's to come now that everyone has access to things like open call? This is going to be happening every day in companies that don't put the governance in place like this kind of stuff, not like maybe to this level,

1:10:22but agents doing stuff when file folders, they shouldn't have done it in. And like, and I, I said something like a pushback, like I'm an anti agent or something. I was like, no, I'm just a realist. Like we have no idea how this stuff works. Like I get that a lot of people are super excited about it and they're off building on the frontiers and like, it's awesome. Go do it. But like you're, you're, there's tremendous risks and unknowns. Um, I mean, this is kind of funny, like I, I guess, um, I don't know, like it, this is, it's just where we are. We're very, very early in these, these agents and understanding how they work when they just

1:10:57figure out their own plans and they're really good at finding loopholes and cheats. And they don't necessarily know that things are bad. They just have a goal. And it's like, oh, I found a way to do the goal that the human gave me. I'm going to go do the goal. And, um, I don't know, man, like I said, it's kind of funny, but I don't think it's going to be very funny when this starts happening in companies all the time. And then it's, uh, it comes hard to manage. Yeah. And I don't know the details of this gym chain or anything, but I just think of like local businesses in my community where we live, Paul. And like, I'm like, uh, is your local like

1:11:29restaurant or business even equipped to deal with this? Like, again, this wasn't even that malicious. Like if I went and tried to use, uh, open call or whatever to go book a reservation at my local restaurant, I have no idea what system they're using and like how compromised or not compromised it is. Like, yeah, they may never know if they're just going to go higher wire. They think it's like a bug in the system. It's like, no, she's just getting hacked like by an agent or a swarm of agents. And, and, and not maliciously. It's just like someone had their agent be like, Hey, can I get a table tonight for four or something, you know? Right. So who's liable

1:11:59in that case? Like that's what I come back to the thing I mentioned. Like, I mean, he did it, but is it anthropic that's liable? Is it him that's liable? Is it who? I don't know.

AI Use Case Spotlight

1:12:07Good luck. Yeah. All right. Next up, we have our AI use case spotlight. Every week, we kind of give you a quick look under the hood at some real use cases we're exploring, um, in our work or in our personal lives as the case may be. Um, so I'm going to share one real quick ball. And then I know you've got some stuff to go through. So my use case this week is actually more personal experimentation. I have been doing with, uh, an open source AI agent called Hermes, which is basically like open claw, but not open claw. Um, so Hermes is not an AI model. It's an open

1:12:39source agent framework that anyone can download. It wraps around a model. You select like, what do you want to use with it? Like GBT 5.6, quad, whatever. And basically it gives it tools, persistent memory, reusable skills, and the ability to take actions. Now I had experimented with some of these kinds of tools before, but, uh, like several months ago, but I just like couldn't find a real use case for it. But, um, I kind of came back around to it and started to find some interesting ways I could use this because like, if you're listening here, you might say like, well, isn't

1:13:11this just quad code or codex or something like that. And there's a huge amount of overlap here. And that's why I couldn't really find use cases. I was like, this is just worse than like what I use on my computer. However, what's interesting is Hermes operates through in part a messaging service called telegram, like an app that's very popular for messaging. Um, so once you have it running on a local machine, which I have one on my personal computer running at home in a virtual machine and a sandbox, so it's not like running wild, you can actually just like message it via telegram,

1:13:45like, Hey, go do this, go do that, go do this thing, go check that. Like all these little things where I could go do whatever your standard AI model can do. It's literally using the models I use every day through chat GPT. Um, but what's cool is it's on 24 seven. So I have it connected to some personal systems like my personal calendar, personal Asana. And that's really like the use case I found that's super helpful. It's almost like a chief of staff. Like I can go into chat GPT or quad or something and access those systems. But it's so nice to just have in one place on my phone,

1:14:17like, Hey, go add this thing to Asana while I'm like running to a meeting or something like that. So it's kind of function this, like these tiny little gaps in my day, especially around like project management and planning or like, Hey, what's coming up on the calendar in the next few hours? Like, do you have any recommendations for how I might be able to structure my day better? Things like that, which are again, things I can ask if I'm in front of a computer or something. I found it really interesting to be doing this 24 seven with this persistent agent that also Hermes is interesting because over time, apparently it's self-improved. So it learns on its own from you

1:14:52to like, do things better, create skills that might be helpful. I haven't used it enough to see all that at play, but it's been a kind of fun little, uh, experiment, I would say. Sounds like what Siri should be like. Siri should obsolete. I think that's what, what they're trying to get to. And I've never unfortunately been a Siri person. So now that they've updated it, it might just do this for me. I just need to get in there. I don't think they'll do it yet, but like maybe by the fall. Yeah. Um, all right. I'll just do a quick spotlight on a few things I prompted last week. Cause I always talk about how I use it primarily

1:15:26as like a thought partner and strategy guide things. So I had a big meeting with, um, my legal team, my accounting team on a major business thing I'm working on. So I had received six very dense legal documents, um, on things that I am not an expert in. So this is my actual prompt. I have a meeting today with the attorneys and accountants regarding the project. I'm going to paste the email received, and then I'm going to upload the related documents. I need you to review everything, summarize the key points for me, highlight the primary decisions that I need to make, proposed questions I should ask to help me make the decisions and call out any additional

1:15:57information that would be relevant for me to consider in order to quickly move forward. I then shared the output from that. So it was amazing analysis that was done by 5.6 SOL, GPT 5.6 SOL. I took the output, I sent it to the advisors, and then we use that as the basis for the discussion on the call. So I had like zero time. I got the documents. I think the night before I had a meeting at three o'clock and I had no time in my calendar. So I was either going to go in completely unprepared or I was able to do this analysis and send it to them. And it was great. It was like, I didn't nail everything, but like the actual experts were like, this is actually really

1:16:29good. This is a good talk. And that's how we did it. Uh, another that I thought was amazing. I had two presentations I had to build this week and Mike can attest, anyone who's ever done public speaking can attest. Um, the amount of time that we used to spend looking for images in clip art or like, uh, stock photos to create a nice looking image for your cover slide. I literally gave it the deck and I said, create a cover slide image for this presentation. It nailed it. And so then I did

1:17:00it in Gemini too. And it looked like clip art vomit, but like, so somehow Gemini got worse at images. I don't understand what happened to nano banana, but like GPT 5.6 nailed it. Um, top level design, Gemini did not. So then I went into GPT 6 5.6. I said, okay, now I have a presentation for this organization. Here's that deck even better, like crushed it. Mike, can I tell you like, it was awesome. It was a really cool. Um, so then I, uh, let's see. Oh, I had to visualize this crazy user flow for this product I'm designing that I'll tell people about in like a month. Um, but

1:17:35it's this insane thing where you have to visualize, they come to the website and they can go down these different paths based on who they are. And I was like, I can't just have this two page outline. So I took my two page outline of how I envisioned the workflow going. And I said, help me visualize this. I gave it to fable five and GPT 5.6. Like here's the actual prompt. Help me think through the user flow and experience. I've put a rough draft together. Can you evaluate and then help me visualize the final concept in a flow chart? So 5.6 gave me this insane mermaid chart, which I didn't even know that that's what they're called, but it was awesome. I took that. That would have probably

1:18:08taken me 10 hours to try and create on my own in Apple keynote or PowerPoint or whatever, sent that to the developer. And I was like, here you go. Like, this is what I'm basically envisioning. So, I mean, collectively, just those like three examples I just gave, I probably saved 20 hours this week, like just doing that. And that's why I always say, like, I'm all for the agent stuff. Like, I want to learn it all. I want to do what Mike's doing in my like work life and my personal life. Like, but it's like, I don't have time to do it, but what I do have time to do is just use it at a very high level really well for strategy and advising and things like that. And like nine

1:18:40times out of 10, it's good for me. Like, that's what I just needed for right now. And I'll figure out the other agent stuff and like more advanced things later on. I love that Mike's doing it and other people in our company are doing it because right now I don't have time to be the one experimenting there. Yeah. But what you're also using it for is literally the highest leverage possible thing to amplify. It's like, you don't need to automate a bunch of tasks. Yeah. I'm trying to do like really big things that create massive disproportionate value. And so for me, that's just being able to talk to something at all times that can help me think. It's awesome. All right. So we'll wrap up here with

AI Product and Funding Updates

1:19:15some AI product and funding updates. I'll run through these real quick as we wrap up this week's episode. So first up, OpenAI expanded its cybersecurity program called Daybreak. It gave vetted partners like CrowdStrike, Cisco, IBM, Accenture, and Palo Alto networks access to its models through two tiers, including a GPT 5.6 cyber model that is new and trained for advanced authorized work like finding zero-day vulnerabilities and validating exploits. OpenAI also launched a feature called Computer History, an opt-in feature in the ChatGPT desktop app for Mac

1:19:48that turns a user's activities across apps and websites into memories and a timeline ChatGPT and Codex can reference, recording clicks, typing, and app switches, but no screenshots or audio that is rolling out to pro, business, and enterprise users. SpaceX officially closed its $60 billion all-stock acquisition of AI coding startup Cursor, with Cursor announcing it will join the SpaceX AI team to help make Grok the world's most useful AI, and improving products including Grok Build, the Grok API, and Cursor itself.

1:20:20Anthropic is meeting with potential investors as we have discussed to shore up confidence ahead of what could be the largest IPO in history. The Wall Street Journal reports that could be as soon as September or early October. Investors are pressing the company about cheaper Chinese AI models, their tensions with the Trump administration, and the growing public backlash against data center construction. Anthropic is also reportedly in talks by something called Decart, an AI startup that makes software to cut the cost of training and running AI by helping chips work more efficiently. They are in

1:20:52talks for about $6 billion, which Bloomberg reports would be the company's largest known acquisition. Igor Babushkin, co-founder of XAI, this past week raised $1.1 billion for his new startup River AI. They are building tools and hardware that let people and businesses train and run open source AI models on their own data and their own devices. NVIDIA is reportedly developing a new family of open models called Nemetron 4, with the largest version expected to have at least one trillion parameters as

1:21:24part of a push to build the world's best open source AI. Jeff Dean, the long-time Google AI leader who just left the company, which we talked about last week, is reportedly in talks to raise a billion dollars at a roughly $10 billion valuation for his new startup Discovery Loop, which aims to use AI to automate parts of the scientific process. And finally, Manus, the AI agent startup that Meta acquired late last year announced it will soon return to operating as an independent company to comply with regulatory requirements and said that data some users created after the acquisition will be deleted as part of the

1:21:59separation. All right, so that is it for this week's AI news. Paul, one quick reminder here, go take that AI pulse survey, smarterx.ai forward slash pulse. We're continuing to run last week's survey. So if you have not taken it yet, please go take 10 seconds to do that. Paul, thanks again for breaking everything down. I think we're going to get a, it'll be an interesting week or two moving forward here. And we'll try and get back to Monday recordings because I feel like, I feel like I'm toast by Friday at three o'clock. Yeah, crazy week. I'd like to see if the soap opera

1:22:34continues next week. Maybe some new models. Yeah, should be fun. But everybody have, well, you're listening to this during the week. So everybody have a great week. Do we have another, episode next week, Mike? Is there a second episode? I don't know. I don't think we've got an AI Answers episode planned for next week. But the week after we will have another AI Transformations episode. Okay. Yeah. So again, if you haven't checked out the AI Transformation series that Mike's been doing, a couple of amazing ones to kick off that series. So go check out those bonus episodes that are in the same feed as this weekly one is. So thanks again.

1:23:08Thanks, Mike, for doing this on a Friday. And thanks again to Kathy for dropping my computer off so we can make this happen. All right. Bye, everybody. Thanks for listening to the Artificial Intelligence Show. Visit smarterx.ai to continue on your AI learning journey and join more than 100,000 professionals and business leaders who have subscribed to our weekly newsletters, downloaded AI blueprints, attended virtual and in-person events, taken online AI courses and earned professional certificates from our AI academy and engaged in the SmarterX Slack community. Until next time, stay curious and explore AI.