#210: Stanford 2026 AI Index, OpenAI Internal Shakeups, What Agents Mean for Business, Claude Design & Dwarkesh vs. Jensen

0:00The brightest minds in AI don't necessarily agree on some of the very fundamental things about the industry, about like definitions of AGI, about whether or not large language models are the path forward and are going to get us where we want to go. Or if we need world models, like Jan LeCun would said, whether or not jobs are going to go away or more jobs are going to be created, whether we should sell chips to China. Nobody knows. But the fact that we're in a place in society where we're having these conversations openly, to me is fantastic.

0:30Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Reitzer. I'm the founder and CEO of SmarterX and Marketing AI Institute, and I'm your host. Each week, I'm joined by my co-host and SmarterX Chief Content Officer, Mike Kaput, as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career. Join us as we accelerate AI literacy for all.

1:06Welcome to episode 210 of the Artificial Intelligence Show. I'm your host, Paul Reitzer, along with my co-host, Mike Kaput. Going a little makeshift this week. So I have been traveling. I actually was in Washington, D.C. with my daughter's class the last three days. I got back late Sunday night. Then we normally record this on Monday mornings, but I happen to be flying to Las Vegas this morning for the Google Next conference this week. So we are doing this at an unusual time.

1:38I honestly am not even sure what time it is right now. I think it's, what, you're 5.30, Mike? 5.30 Eastern right now. Okay. So it is end of the day Eastern time. I am all sorts of out of sorts here in terms of like what is going on. But as I was doing the final prep, you know, because the Wi-Fi on the plane, of course, isn't going to work. But, you know, it ends up as we're delaying this, we get the last minute notification that Tim Cooks out as CEO of Apple, which is like, shows you how real time this

2:10is. Because Mike and I are scrambling literally 10 minutes ago, like what is going on? What did they say? And as of 5.30 PM on Monday the 20th, they haven't said much. So I'm sure by the time this airs, you'll have heard this news and it'll be all over. And when Wall Street opens, you know, and the stock market opens Tuesday morning, I would imagine we're going to see some movement one way or the other. So more to come on the Tim Cook thing, I guess. But that's how real time this is. And then Mike and I were joking. And if any of you ever have to do like virtual meetings when

2:43you're traveling or do podcasts or things like that when you're traveling, we all know the makeshift setups that we have to do. And I have to admit, I have a first, in this case, I have a recycling can on my lap that my microphone is sitting on top of. And I actually had to take a picture of it so I could show Mike what in the world I'm doing here. So I'm overlooking the Vegas airport. We're going to make this work one way or the other. And then we got to figure out next week, Mike, because you and I are both traveling on Monday. We need to get to finish that conversation.

3:15I'm also out next Monday. All right. So anyway, episode 210, as I said, this is a pretty real time for us. If you're new to the podcast, because we get new listeners every week, this is kind of how we roll. Like we do this all in one take all the time. I think in the entire history of the podcast, we've had one or two stops and it was usually for coughing fits. So we just do this. Claire on our team edits it and it drops the next morning. So here we go, Mike. All right. So this episode

3:46is 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're always up to date with the latest AI trends and technologies. The AI for Industries collection features seven core series and certificates designed to jumpstart AI understanding and adoption. Those seven series are professional services, healthcare, software and technology, insurance, financial services, retail and CPG,

4:22and manufacturing, which just dropped last Friday, Mike, if I'm correct. Correct. And then Mike's actually going to give us a little rundown, kind of some key takeaways from the AI for manufacturing series toward the end today. So these series are an ideal launch pad for organizations that want to level up their teams and accelerate AI adoption and impact. And today we're going to share some insights, as I said, with the AI for Mart manufacturing course series, you can get individual or business accounts. Both plans are available now. You can go to academy.smarterx.ai to learn

4:52more. Pricing is totally transparent. So when you go there, you can see the exact pricing for both individual and business plans. So again, go to academy.smarterx.ai to learn more about that. Okay. I don't know how many models are going to drop by the time we do this. This comes out on Tuesday, but we did have one already today. We might not even get into that one until next week, but there's rumors that more models are coming. And then, like I said, I'm at the Google Next conference. While I don't necessarily expect any major models to necessarily drop this week,

5:26that's more usually the IO conference where they would do that. I do expect lots of announcements from Google. So we'll definitely have lots to talk about on next week's episode when we start getting into the updates from Google. But all right, we've got a big new report to talk about to start off though, Mike. Yes, we do, Paul. So Stanford's Institute for Human-Centered AI just released their 2026 AI Index Report. This is an annual benchmark where they basically try to map the entire state of AI across research, performance, investment, jobs, public opinion, energy, and

5:59policy. And boy, you know, I haven't looked at every report out there, but this might be the biggest one. It weighs in at literally over 400 pages this year. I think that's maybe the longest it's been. It is worth at least skimming or dropping into Notebook LM because it does represent kind of one of the top industry reports out there for getting a perspective on what's happening at a macro level in AI. So some highlights that jumped out to us, and obviously we're not going to be able to cover all 400 pages, but a couple of things that are important to note. So first, they find that the

6:31US-China AI performance gap has effectively closed. So US and Chinese models have been trading the top performance ranking multiple times since early 2025. DeepSeq R1 briefly matched the best US model in February 2025. And as of March, 2026, Anthropix top model leads top Chinese models by just 2.7%. China is now also ahead of the US in, you know, study, publication, research, publication volume,

7:03citations, patent output, and industrial robot installations. They also found that AI capabilities are outpacing benchmarks, but still suffer from this kind of term we use sometimes in AI called jagged intelligence. So what that means is frontier models now meet or exceed human baselines on PhD level science questions, multimodal reasoning, and competition mathematics. And they're basically saturating evals that have been intended to be challenging for years, but they're doing it

7:35in a matter of months. So for instance, the SWE bench verified coding test on that one performance skyrocketed from roughly 60% to near 100% in a single year. However, while these models can do things like win math Olympiads, they also have big blind spots. They still fail to do things like read the time correctly about half the time. They also found model convergence is basically happening at the frontier. So performance among the very best models is becoming basically indistinguishable.

8:11So the top four models on the arena leaderboard from Anthropic, XAI, Google, and OpenAI are now clustered within just 25 ELO points of each other, which is how they measure their effectiveness. So raw capability is not really serving right now as a clear differentiator. And that's where the competitive pressure is rapidly shifting towards cost, reliability, and domain-specific performance. So a few more things they highlighted. Generative AI adoption is outpacing, if you can believe it,

8:42the adoption rate of the internet. So generative AI reached 53% global population adoption within three years, which is faster than either the personal computer or the internet. At the same time, AI's physical and environmental footprint seems to be expanding drastically. For instance, Grok 4's training emissions equaled about 72,000 tons of CO2 equivalent, which is comparable to driving 17,000 cars for a year. And AI data center power capacity reached almost 30 gigawatts,

9:20which is comparable to the peak demand of the entire state of New York. On top of all this, global corporate AI investment more than doubled in 2025 to $581.7 billion. Private investment reached $344.7 billion as well. A couple final points here, more related to younger people. AI's labor market effects are hitting junior workers hard. Employment for software developers age 22 to 25 dropped nearly 20% since

9:512024. That pattern was mirrored in customer service and other highly AI-exposed roles. And the public and AI experts, as a result of this, view the future of the technology quite differently. 73% of AI experts expect AI to have a positive impact on how people do their jobs. That's compared to just 23% of the U.S. public. So quite a bit of a 50-point gap. Furthermore, 64% of Americans expect AI to lead to fewer jobs. And the U.S. reports the lowest trust,

10:2531% in its own government, to regulate AI responsibly out of any country survey. And lastly, four out of five U.S. high school and college students now use generative AI for school work, but formal school guidance is failing to keep up. Only half of middle and high schools have AI policies in place. A mere 6% of teachers report that those policies are clear. So, Paul, I mean, lots of different threads to pull on here, but I will say I just look at this data and I find it very

10:56hard to take seriously the argument that AI is overhyped at this point. There's a ton of good stuff in here. And like you said, it is massive. I can't imagine anyone's actually going to sit down and read this whole thing. But I thought they did a nice job of actually breaking it into chapters. And when you go to the site, and we'll put the links into the show notes, you can just go by chapter. And that's actually how I processed most of it. I didn't have time to read the whole thing. But the nine chapters are R&D, so research and development, technical performance, responsible AI, economy, science, medicine, education,

11:30policy and governance, and public opinion. So, just even that alone, Mike, starts to show the breadth of the impact AI is having, that we have all these different areas we have to look at at a macro level of what is going on. They do a great job with the research. Obviously, they have incredible research partners and some amazing people that are contributing to this and writing this report. So, one of the ways you can do it is literally just go chapter by chapter on the website. And it actually calls out, usually it's between five and 10 highlights from each of those chapters.

12:02And so, that's a great way to go through and do it. And so, like you did, Mike, I kind of went through and pulled out the ones that were jumping out to me that might be worth like a little bit more context. And I'll just highlight a couple of additional ones. So, one, you might have touched a little bit on this one, but this idea of like the evals and the benchmarks that we normally use to measure progress. And what jumped out to me here is this continued need that we talk a lot about for companies and individuals to have their own evaluations of these models. You can't rely on these reports that are published by the AI Frontier Labs or even these organizations that are cropping up

12:35that are specifically designed to just measure the intelligence of these models. You have to have the standard tasks or workflows that you perform internally. And so, when a new model comes out, you can determine the impact of it. So, like, I mean, as Mike was going through this, I was glancing down to see if there's any updates on the Tim Cook thing. And I saw somebody post something about, did they make an update to Claude 4.7 this morning? It's like sassier than it was last week and I can't use it again. And so, like, the models truly do change all the time. And when they do

13:08these updates, they should incrementally be better. But sometimes something gets broken or something changes in the process. And so, with the rate of change and the rate of model releases, it really is becoming more and more important. And it's actually, I flagged something Mike will talk about on next week's episode around these evals. I actually saw a new company or tool that was created. Oh, was it Zapier? I think it was Zapier. So, again, we'll talk about it next week because I haven't had time to prep on it. But there was a new thing that was more about actual work being done

13:41and developing evals. So, evals are really important. They highlighted, and I'm under the technical performance section right now, video generation models are starting to capture how objects behave, not just produce realistic content. This goes back to the thing Mike and I have talked a lot about over the last six months or so, this idea of understanding the physics of the world and being able to develop world models, which is going to be very, very important to especially robotics, but video generation, interactive game generation, things like that. So, we're seeing progress being made. You'd highlighted a number of the ones in the economy side, Mike, that I was looking at,

14:15like the labor market effects are showing up unevenly. It does seem to be affecting the youngest workers most at this point. And then one-third of organizations expect AI to reduce their workforce in the coming year, even though we're not necessarily seeing it in the jobs data yet. I think you'd mentioned this one, Mike, about the four out of five high school and college students now use AI for school work, but the school policies aren't keeping up. So, that's a major problem. Another one I thought was interesting under policy and governance was US public investment

14:45in AI remains modest compared to the private sector. And the reason I flagged this one, Mike, is because I think this is about to change dramatically. I don't know where the numbers are going to come out to, but we'll have to go back. And Mike, maybe you can find the show notes. There was an episode where I did a breakdown of the Manhattan Project. This is probably last year, where I actually broke down how much of GDP was going to the Manhattan Project back in the day. And at the time, I said we were going to need something like that from the government. And I do think what's about to happen is going to be the government subsidizing the build-out of energy, the build-out of data centers,

15:20the building of chips. I think they're going to start to get into subsidizing the training and education of American workforces. I think that's going to have to happen. And so I could see sometime in the next one to two years where the government is spending well over a trillion dollars a year on AI infrastructure and talent and all the hard and soft costs associated with it if they don't basically take over one of the labs. So I'm not a proponent of that happening,

15:52but I think we're just getting to the point where AI is so fundamental to the future of democracy, future of the republic, that I don't think they're going to be able to just sit back and hope that the three to five private companies figure this all out. So I think they're going to go. I just don't know exactly what that looks like yet. And then under public opinion, the optimism is rising, but so is anxiety. It's like generally people are pretty excited about it, but they're also really not sure. There was one stat that the one thing that jumped on me, I thought was weird. It was like people and their confidence of understanding AI. It was like 67% of people

16:27said they're confident in their understanding of AI. I was like, okay, like you could give me the hundred smartest CEOs in the world and I'm not sure I'd get to 67% who actually understand AI. So I'm not sure who exactly they were asking that question of. Stanford students maybe. Yeah. Like it's either, they're either really overconfident in their ability or they're asking technical people because that is not by any means a representative of the average worker or business leader. Um, two thirds of Americans expecting AI to lead to fewer jobs over the next 20 years,

17:02which I thought was a really weird timeline to throw into this. It's like what it took two years, maybe, but 20 years, who knows? Like that's crazy. And then, uh, one of the one, the companionship thing, they talked about it still being niche, but that I was just noting, I don't think that's going to be niche for long. Like they're still treating this people using this in a relationship or companionship. I feel like that is just skyrocketing beneath our eyes basically. Like we're, it's probably happening and most people just aren't aware it's going on. And then the last two was the United States reported the lowest trust in its government to regulate AI responsibly. Um,

17:38that is accurately placed mistrust. I would say, I don't think the U S government is currently on a path to responsibly, um, regulate AI. And then across all 50 States concerns about too little AI regulation outweighs concerns about too much. And then just a final note, um, on their methodology, I thought it was interesting. They said AI indexes written by a team of human researchers, and they do a nice job of highlighting those people up front. And they said the authors use Chad GPT and Claude to help refine and copy edit drafts. And then all images in the publication were

18:08generated with AI by, uh, Johanna Friedman. So they actually recognized the human, but said, Hey, we did this with AI. And then they said the two specific models that they used. And then the final thing is public data and tools. One of the nice things about this report is they, they kind of open source a lot of this. So you can actually go in, they have a Google drive link to get all the public data and the images that they created all the charts. And you can actually use all of those things. So if you're doing presentations on this stuff or business research, and they just tell you how to properly cite the report. And then they also have a global AI vibrancy tool that compares the

18:43AI ecosystem across 36 countries. So I look, you know, you and I, Mike love a good research report. And this is extremely well done, very professionally done. It's the kind of research we're happy to shine a spotlight on, on this, uh, podcast, because I think there's a lot of really good information in here. And they obviously put a ton of effort and it was a very thoughtful presentation of the report, but it's their seventh year doing it, I think. So it's like, they kind of got it down to a science at this point. All right. So next up, big thing happening this week is that open AI is basically going through a

19:17very public reorg or a continuing reorg. And this week kind of brought together a few separate threads that tell this overall story of the company, basically speed running through some serious growing pain. So first we had reported last week on some executive level shakeups. We've got a few more this week. So Kevin Weil, open AI is former chief product officer has announced he is leaving. He joined open AI in 2024. He became CPO. Then he stepped out recently to launch open AI for science on the same day, Bill Peebles, who led open AI Sora short form video app also announced his exit.

19:53We have previously talked about open AI, basically rolling Sora into chat GPT slash, you know, sunsetting it as its own thing. So that kind of makes perfect sense. People's might be out. Open AI says it's also now decentralizing that open AI for science initiative back into the core research teams. And so the idea here that we've talked about a couple weeks in a row is open AI is shedding what their leadership internally calls these like side quests. So things that are just no longer a fit for the company's core mission as they pivot, especially hard towards the enterprise.

20:25Second, there's some internal drama around the company's IPO timing. The Wall Street Journal reported Sam Altman wants an IPO as early as Q4 2026. CFO Sarah Fryer has told colleagues the company may not actually be ready this year. Fryer has raised concerns that open AI's financial exposure, given its computing infrastructure spending, is a problem because that could reach $121 billion in 2028 alone. Sources also have said Sam Altman has excluded Fryer from investor conversations and

20:58major financial decisions. Unfortunately, a pattern of behavior he has exhibited in some of the other articles we've discussed. Third, there's some internal drama or at least competitive animosity towards Anthropic. The Verge got its hands on a leaked internal memo from open AI's chief revenue officer, Denise Dresser, who sent this to employees basically taking direct shots at Anthropic. Dresser accuses them of inflating their run rate by roughly $8 billion through what she calls accounting treatment that

21:30makes revenue look bigger than it is, specifically by grossing up revenue sharing agreements with Google and Amazon rather than using net figures. She also writes that Anthropic's built on fear restriction and the idea that a small group of elites should control AI. And last but not least, there's also external trauma around the company's valuation. The Financial Times reports that open AI investors are starting to question the current $852 billion valuation as the company shifts towards enterprise. So the idea

22:01here is that all these recent deals and abandoned projects are all part of defending Chad GPT's consumer dominance while going after Anthropic in higher margin corporate markets. So the issue here is that one early backer told Financial Times that despite Chad GPT having almost a billion users, open AI is a deeply unfocused company. So, Paul, this also comes as open AI and Sam Altman are kind of sounding the alarm on AI's possible impact on jobs. They even released an AI transformation plan for jobs this week.

22:36Um, but honestly, I don't know. Do they have enough drama on their plates without all that? Like it seems like they're really dealing with a lot of these core issues to the business. And the Musk trial, I think is later this month soon. Yeah. Yeah. Yeah. And Fiji Simu is still out like the CEO of AGI deployment. I don't remember when exactly they changed that title. It was F applications, right? And then they updated it recently to AGI deployment. So she's on medical leave. Um, yeah, you know, I saw, I don't remember what day it was when

23:09those resignations started happening, but they were back to back on my X feed and I took a screenshot and I put it in our sandbox for the podcast. And I was like, Hmm. And then like an hour later, one more came and then like, you know, the articles started pouring in. So it does seem to just be a by-product. Um, obviously reorgs happen a lot in the AI lab space. Well, at least here, Anthropic seems to be pretty stable. Um, but this is, there's been a lot of moving pieces at it as at open AI for a number of years now. And this one seems pretty

23:42significant, but it does seem like the by-product of this renewed focus and the need to constrain compute to the newer models. Um, there's talk that we're going to get their spud model. It's been kind of codenamed spud, which I think is 5.5. That, that is lots of increasing rumors that that is coming this week. Um, so, uh, yeah, I don't know. It's just a really interesting time. I don't, I try not to over like speculate too often on what could happen with these things that I usually constrain it

24:14to like internal conversations, but I will throw one out here that I'll just say publicly. The most logical thing that happens to me with open AI is Brett Taylor takes over as the CEO before the end of the year. So what I mean by that is Sam Altman through his own comments in the media does not seem enthusiastic about the idea of being the CEO of open AI when it's a publicly traded company. There's lots of things going on. It's a very, very stressful job. I can't even imagine what Sam

24:46deals with personally and professionally. Obviously he's had issues recently with attacks on his home, um, tax on the office, uh, personal threats to his safety, his family safety. And I just don't ever get a sense when I listen to Sam talk anymore that he's having fun, like that this is what he wanted basically. Um, and so when you start adding all those things up, you just start to look around and be like, I just can't see him long-term doing this. Like I could see him being the face of the

25:20company, like for fundraising and the infrastructure and things like that. But at some point, I just don't know that he doesn't burn out from being the real CEO day-to-day of the company and having to deal with this. And Brett Taylor is just the most obvious thing because Brett's built Sierra, an agent company built on top of chat GPT capabilities. He's the chairman of the board. Um, he's a co-CEO of Salesforce prior to this. He was a former CTO of Facebook. He started his career at Google.

25:51He co-created Google maps. Like Brett is a legend, um, in Silicon Valley circles. And if you were going to look and say, well, who should be running this company and maybe would want to be running a company like this into the future when we're a multi-trillion dollar company, I, again, I just don't know that it's going to be Sam and Brett seems like a really obvious choice. So I would even say this is a prediction. I would say I'm more like public narrative that I would usually reserve for conversations Mike and I have when I'm just throwing things around. But every once in a while,

26:24something just seems so obvious that it's worth at least like sharing with our audience of like, just something to think about. I get, I, if I had to put odds on his Sam, the CEO of this company 12 months from now, I would, I don't think he would be, um, you know, he certainly could be, but I, I just, I'm starting to see more and more science that I don't, I don't think it makes sense for him to be in that role, but we'll see. I have no inside information on that whatsoever. There's no sources I'm getting this from. It's literally just piecing stuff together. The jobs report, Mike, I thought was pretty interesting. So kind of along the lines of the Stanford one, I liked the effort

26:58they made here. So, you know, they've been putting more and more work into, um, their thoughts around the impact AI is going to have on jobs, trying to do more to project out. And this is what we've been calling for, for the last two years was like, we should at least be studying this. We should at least be considering that maybe this doesn't go smoothly in the near term and that we could lose a bunch of jobs. But the main way that it's often talked about is exposure levels. So, you know, it's how I built JobsGPT. You look at the exposure of tasks that make up jobs and you say, can AI do

27:32this thing? Um, but what they're looking at is goes, you know, kind of layers on top of exposure and they look at demand elasticity, which is something I talk a lot about on the podcast and then human necessity, which is an interesting one. So in this report, they, they say they look at 900 occupations and which they're claiming is 153 million jobs or 99.7% of us employment. Now that would have to include some part-time jobs because the full-time jobs in the U S is around 136 million. So I'm not sure exactly where all those are coming from, but they're looking at a bunch of jobs and then they're basically trying to figure out like how exposed are these occupations?

28:06And then does that exposure mean an impact on actual work? And so they, you know, they look at this thing, like the demand elasticity, we're saying how much demand changes when price changes and is it connected to productivity and the impact on employment. So they're trying to look at it, you know, in a simple way, like I think about this. Okay. So, uh, AI makes us cheaper to deliver legal services, for example, um, because now we can use AI to do it. We can be way more productive.

28:37Does the demand for legal services increase enough to justify that we keep hiring more and more people? So like right now, the argument, a lot of these leaders are making that we talked about in last week's episode is software developers and engineers. So what they're saying is the demand for software is so significant that as the cost to create that software goes down, yes, we'll need fewer people to create the same amount of software, but we're just going to keep seeing more demand for software. So we'll keep hiring more people. And my argument has been all along that doesn't necessarily

29:10translate over to industries or professions where there just isn't a ton of demand to see. And so, you know, if you're in a company where the demand isn't going to grow, no matter what happens to your price, then you're not going to need as many people. So that's what this report tries to take a look at. They're basically looking at this job transitions framework they created, and they're asking four fundamental questions. Can AI do a meaningful share of the work? If AI lowers the effective cost of providing the service or the product, is demand likely to expand enough to absorb the productivity

29:46gain? Meaning cost goes down. Is there enough demand for us to still need all these people? The third question is, if it can, for the remaining tasks, is a person still central to the work's delivery, judgment, accountability, or physical execution? So they're not just looking at knowledge work tasks. They're looking at all tasks. And then they're saying, is AI already being used meaningfully for these tasks? So then they use this approach to critically assess these 157 million jobs or 153 million jobs. And then they break the occupations into four different possible outcomes.

30:22So one is jobs with less immediate change are those where the current combination of exposure, necessity, and elasticity does not yet point clearly to one dominant near-term outcome. So that is almost half. It's 46% or 70 million, they're saying. See less immediate change. Then they go into jobs that will reorganize, that have high exposure and strong human necessity, but demand is not elastic enough to absorb the productivity gains. That's 25%, they say. And then 18% is jobs at higher automation risk, have high exposure, weak human necessity,

30:57insufficient or ambiguous demand offset. So that's 27 million jobs in their breakdown that are at high automation risk, meaning these are the ones that could go away the fastest. And then jobs that will, the fourth category is jobs that will grow with AI have high exposure, but enough demand response to lower the cost may increase utilization, affordability, access, and quality-adjusted output. But then one other note, they did say, our ability to forecast far into the future is limited and it is very difficult to project how much the labor market might evolve over the long term. On a shorter horizon, this framework should

31:34help us envision how the labor market may evolve and what policy responses we can consider and implement to facilitate a smoother and more people-centered AI transition. So again, my point with kind of highlighting some of this is I'm really happy that they're analyzing this at a deeper level. I would love to see all labs doing this. Anthropic has been actually leading the way on this, so they're certainly doing this already. This is the kind of research that needs to be happening at all the different labs. And it's the kind of stuff that economists should be aggressively pushing

32:05on. But even with all this analysis they did, they basically kind of came out to like, we really just don't know. If you want the too-long-don't-read version of this, we don't actually know, but we don't think just exposure alone is enough, which they're 100% right on. So the need for humans in the loop, for interpersonal communication between humans is critical, and then the need for understanding demand. So a couple of examples I'll finish with. They said the least elastic occupations include firefighters and home health aides. What they mean by that is that is needed no matter what the cost.

32:39Like we have to have it. Somewhat elastic occupations are like physical therapists, editors, dental hygienists. You might just go less if it costs more. But if the cost goes down, the question of, do I buy more of it if the cost goes down? And then the final, some of the most elastic occupations include graphic designers and software developers. So yeah, it's really interesting stuff. You're interested in the economic side of all this. It's a good read, good thing to throw in Notebook LM. But I think my main takeaway again is a lot of people are starting

33:11to think about this at a deeper level, which is great, and no one has an answer. Like no one can say, what I always say is anyone who tells you confidently, like without debate, that they know what's going to happen to jobs in the next one or two years is either misleading themselves or intentionally that there's something to gain by you believing that. We try and kind of sit in the middle and saying, listen, I hope that that's what happens, that it doesn't affect jobs. But I'm

33:44also trying to be very realistic that every scenario I look at where I think about this and say, yeah, but what about if demand doesn't go up for a company or a role? How could you possibly create more jobs that fast? And that's the part I just, no one has the answer to. So yeah, I think that's the big takeaway. Lots of change at OpenAI, lots more to come. The innovation is going to keep moving forward no matter what. And they're starting to kind of branch out and trying to understand the impact that their models are creating and stuff they've thought about for a while. I mean, they did

34:15that GPTs or GPTs study back in 2023 when GPT-4 first came out. So it's not like this is new. It's like the first time they're thinking about it, but they're starting to expand the way they look at it.

34:27All right. So our third big topic this week is we saw several stories that kind of weave together to paint a picture of the implications that AI agents might have on business at large. So we wanted to talk about a handful of these that we've been tracking that may give us some clues as to what AI agents could actually mean for business leaders. So first up, OpenAI launched what they call Codex for almost everything. This is just an update to OpenAI's Codex coding agent, but it's a pretty big one

35:01because it gives Codex background computer use, meaning it can see, click, and type across any application right now on your Mac while you continue working in parallel. OpenAI also released more than 90 new plugins for different types of software that your computer using agent can now operate. Second, we've heard more commentators and companies talk about this kind of slang term going around called token maxing. This is the practice of maximizing AI token consumption inside

35:32a company. It's basically being used as kind of a proxy in some cases for AI adoption. So for instance, writer CEO May Habib, who we've talked about before, told the Wall Street Journal that stuff like token maxing to drive internal AI usage is existential for her company. So writer and several others are in the camp that basically you need to be token maxing. However, there are critics at some other companies that call this an empty metric and an invitation to waste. Now on top of that,

36:04interestingly, token maxing can have some unintended consequences. So we also had a story where Uber CTO, Praveen Naga told the information that the company burned its entire 2026 AI budget in just four months, driven primarily by cloud code and cursor. So Uber engineers were actively encouraged to use these tools. The company ranked them based on internal leaderboards based on AI usage, and they just blew through the available budget. Now a couple other stories that kind of hammer home all the nuances people are

36:38dealing with. So Business Insider reported that Microsoft executive Rajesh Jha argues that AI agents may eventually need their own software licenses, just like human employees. His logic is that as AI agents start to act as full employees, companies will naturally have to give them login accounts, email inboxes, permissions, and access to tools that turns each agent into a potential paid software seat. On top of all of this, Box CEO Aaron Levy posted a widely shared thread this week on what

37:10it takes to keep up with all this AI architecture. He argues that companies building on AI have to accept that they will be dramatically upgrading their AI architecture over and over again. He cited how there are infrastructure patterns like RAG, graph RAG, and orchestration frameworks that we're all talking about as state-of-the-art two years ago, but are already obsolete. To end all this, Amazon is living a version of this problem firsthand. Business Insider reports the company has a pretty serious

37:44AI sprawl problem. They have duplicate internal tools, disconnected data, and a ton of growing operational challenges around AI agents. So Paul, the reason we mentioned this is there are these threads kind of coming together here in these very broad strokes. So like first, we're seeing these agentic coding platforms like Codex and QuadCode basically develop into general purpose agentic platforms for non-technical knowledge workers. But as these agents begin to permeate the enterprise, they're raising all

38:14these really difficult questions and like totally unanticipated consequences around budgets, permissions, and architecture. And I'm just curious, like it sounds like some of the top companies in the world do not have this figured out. Does anyone have this figured out? No, I don't think anybody has this figured out. I think it's presented as though people have a grasp on what is happening. But every day that goes by, you realize just how early we are in the integration of agents into workflows and businesses. And it can feel like you're way behind because people are out on the frontiers trying all this stuff.

38:48But they're also the ones who are going to find all the potholes and take on a lot of risk. So a few notes here, Mike, I think you did a great job just kind of summarizing. I mean, we probably could have honestly just broke that entire recap you just gave into the whole episode today and just talked about each of those. But at a high level, here's a few observations I have. So one, the accelerated capabilities. I mean, every week, there's just new capabilities of these agents. And when you connect the agents together, it is overwhelming. I mean, even for someone like me who's paying attention

39:22to this stuff, reading about it constantly, analyzing it every day, it's moving so fast.

39:30It's almost impossible to just keep track of what is going on. The second is increased risks and unknowns, especially as the power of coding is in the hands of non-coders and technical staff. What I mean by that is there was one that popped up today that really caught my attention. So I have talked a lot recently about Lovable. So it's an example of an app building tool, non-coding app building tool that I've mentioned on the podcast that I've used a few times to build some things. It's really cool. If you've never used it, it kind of functions like ChatGPT. Just go in

40:03and give it a prompt, but then it builds something. And so in my case, I built like an org chart was an example. I built an assessment tool. And when you build these, it stays private by design, but then you can share it to allow people to have access to the app. So any non-coder would assume if I'm sharing a link, I'm just sharing a direct link to that app so people can experience it. I'm not thinking I'm giving them access to everything else behind it. So I see a tweet this morning as I'm standing in line from my plane to Vegas. And it's this random user. Impulsive is the name of the

40:38X account. Weezer, O-S-I-N-T is the actual handle. So when you first see something like this, I would never retweet this stuff. I was like, who is this person? Is this legit? Like you got to do a little homework. So I'll read you what it says. It says, Lovable has a mass data breach affecting every project created before November, 2025. I made a lovable account today and was able to access another user's source code, database credentials, AI chat histories, and customer data are all readable

41:09by any free account. NVIDIA, Microsoft, Uber, and Spotify employees all have accounts. The bug was reported 48 days ago. It is not fixed. They marked it as duplicate and then left it open. And I was like, what? Like that cannot be possible. And then sure enough, five hours later, well, I'm now on my connecting flight. Lovable tweets, without acknowledging the original, we were made aware of concerns regarding the visibility of chat messages and code on lovable projects with public visibility settings. To be clear, we did not suffer a data breach. Our documentation of what public

41:46quote unquote means, what public implies was unclear and that's a failure on us. Specifically for public projects, chat messages used to be visible. This is now no longer possible. Importantly for enterprise customers, being able to set visibility to public for new projects has been disabled since May, 2025. So I was like, I read that. I'm like, okay, maybe I'm just not smart enough to understand what that post means, but our documentation of what public implies was unclear. It's like,

42:17okay, well, first that seems super condescending, but again, maybe I'm just reading more into this than I thought because this isn't, cybersecurity isn't necessarily my area of expertise at all. So then I was like, oh no, I'm not an idiot. Then two hours later, they have to now post again and correct themselves because I think other people also were like, what the hell are you talking about? So this is their follow-up. We're sorry. Our initial statement didn't properly address our mistake. Here's what a public project on lovable means and how we got to where we are today. In the early days, people didn't know what lovable was capable of. So we wanted to make it easy to

42:51explore what others were building as a way to spark ideas and lower the barrier to getting started. Like scrolling GitHub or Dribbble, which again, non-coders don't use. You browse projects to see what's possible, then go build. When you create a project on GitHub, it explains how it's default to public, whatever. And then they said, over time, we realized this was confusing. Many users thought public just meant others could see their published app. I wonder why, because that is the most logical assumption, not the chat of an unpublished work. So basically, in essence, you go in, you build something lovable. You're like, oh, this is cool. I'm going to share it with somebody. Let's say I'm

43:24doing it. I send it to Mike like, hey, dude, check this out. Check out this quick demo. Mike gets on there. Mike has seen my entire chat history. He can see everything I've done in the thing. And it's like, that is not at all what I intended to have happen. So again, I highlight that one as an example of like, we just don't know. And Lovable is a really good company with tons of funding. And like the things you would normally do to say like, okay, is this a legit company I should be building something with? They would check the boxes. And yet they made this apparently intentional choice to like,

43:56I mean, it certainly seems deceitful to me, but like, I guess they didn't think it was. And then the other one, and maybe Mike will go into this one more next week. But this Vercel hack is insane. So I'll just read this quickly. And again, I'm trying to highlight all the unknowns around the use of agents is like my whole point here. But there was this very high profile thing that happened with Vercel. And so here's the CEO's post. He said, here's an update to the broader community about the ongoing incident investigation. I want to give you the rundown of the situation

44:27directly. A Vercel employee got compromised via the breach of an AI platform customer called Context.ai that he was using. So employee has access to this thing, it's connected to it. The details are being fully investigated. Through a series of maneuvers, that access, that employee giving access to Context.ai from the colleagues compromised Vercel Google Workspace account. So they connected to their Google Workspace account, the attacker got further access to Vercel environments. We believe the attacking group to be highly sophisticated, and I strongly expect significantly

45:02accelerated by AI. They moved with surprising velocity and in-depth understanding of Vercel. So again, this is a very knowledgeable company that you would think would be better than most at avoiding unforced errors when it comes to cybersecurity. And they let it happen. So my thinking here is like, again, there's no debating the impact agents are going to have in enterprises. Like it is very obvious, but there's just, there's so many growing pains we have to go through. And so I would just think very deeply as a leader, or maybe as someone who's

45:37pushing for this kind of access within your company, like you see the potential and you want to set up co-work and clawed code and open claw and you want to like do all the things, you got to understand what it is you're doing, or the organization has to understand all the risks that come with doing these kinds of things. And if you're relying on outside consultants or agencies to do this kind of work for you, you got to know what they're setting up, what kind of access you're giving them, and in turn, what kind of access and technology they're integrating into your systems. And this

46:11is why AI is not just like tools, it is change management. It gets into like impact on org structure and tech stacks and the roles within companies. And that's when I say like, I don't, you know, if you asked me just point blank, how many companies, like, let's just take enterprises, let's say companies, 250 employees or more. Do I think actually fully understand generative AI? Like let's just keep agentic AI on the side for a second, fully understand the capabilities of generative AI platforms and have properly integrated into their companies to where they can like really

46:43scale transformation. Like what percentage do we think that is? It's single digits. Like you can't convince me otherwise at this point, it's like low single digits. I think if you then extend to that, say, okay, which ones are actually prepared to integrate agentic technology and scale it within their company in a responsible way, it's like well under 1%. So yeah. So my point is, we do our best to highlight where this is today and where it is going. Do not feel like if you're not racing into agents, you are way behind. Like you're not, it's okay. If you're a Y Combinator

47:20company or a VC backed AI native company, this is expected. Like you're not getting funding without this stuff. But if you're a legacy enterprise, that's trying to drive like transformation in a responsible way, agent stuff like this is going to be sands box through a technical team for like the next 12 to 24 months before you start seeing this kind of stuff, like really living in the wild. So yeah, it's just, I don't know, like I said, it's moving so fast, Mike. And I really do each week have a hard time comprehending some of the things that are happening. Same. And I keep wondering to

47:54myself too, like, I wonder what happens when we finally have like the clawed code moment for something like co-work or another one of these tools for non-technical knowledge workers. And by that, I mean like all this stuff we've been talking about since the beginning of the year with like AGI and people wondering like something big is happening. That's all been driven by the fact that clawed code was out for a year before this combination of factors late last year meant it got really, really good. And everyone started positively and negatively freaking out. I think

48:29that's going to happen for something, whether it's co-work or another tool for non-technical knowledge workers, people are going to have meltdowns and or move 37 moments like we've talked about. And I don't think anyone's prepared for whenever that happens. Yeah. And it is, you know, I think there's also just this, the technical side of, uh, how these models perform and how the harnesses are built around them in terms of like allowing the capabilities that they're going to have. And, you know, I saw a tweet from Ethan Mollick, I think was saying like, it doesn't make

49:01any sense. Like if you look at Gemini's model from Google, it's so good through the API, but it's not even close to competing with Claude when you're in the app. Like it's just a totally different beast because Anthropic excelled at putting the harness around those agentic capabilities and the model capabilities, and then bringing them to the actual user interface where Google has been really struggling to keep up. And the things we're seeing now, like I think I put the information article in for next week, um, that there's like a code red again, internally at Google where Sergey Brin is

49:34taking the lead because all the people at Google are using Claude code because their own coding capabilities are like on, not, not on par. So yeah, again, like the labs themselves are, are struggling to keep up. We talked last week, I think about metas using Claude code internally and like, you take this from me, I'm quitting. Like I need, I need these capabilities. So wild. It's really incredible. Yeah. Nobody's got it all figured out. Not even close. All right, Paul, before we dive into rapid fire, a quick announcement. This episode

50:06is also brought to you this week by our AI for writer summit. So the future of storytelling is being rewritten thanks to AI, which is why we're super excited to be hosting our annual AI for writer summit on Thursday, May 7th. This is a half day virtual event for writers, editors, content, teams, teams, anyone who does any type of writing or content creation as part of their job. And during it, we'll have some incredible speakers breaking down exactly how AI can help you create smarter and faster. And importantly, do all that without losing the heart and soul of your writing.

50:41This event has a free registration option. So go check out your registration options today. We've got the full agenda live. Go to AI writers summit.com. That's AI writers summit.com. All right, Paul, first breaking rapid fire happened just before we jumped on today. Apple announced that Tim Cook is stepping down as CEO effective September 1st, 2026. After serving in the role since 2011, John Ternus, Apple's

51:14current senior VP of hardware engineering is taking over as CEO and joining the company's board of directors. Cook is transitioning to the role of executive chairman, while Apple's current non-exec chairman, Arthur Levinson, will become the lead independent director. So Ternus holds a bachelor's degree in mechanical engineering from the University of Pennsylvania. In a press release, Cook expressed his full confidence in the transition, stating that Ternus had the mind of an engineer, soul of an

51:45innovator, and the heart to lead with integrity and with honor. The transition marks the end of a nearly 15-year run as CEO for Cook, who took over directly from Steve Jobs. Under Cook's leadership, Apple's revenue almost quadrupled to over $400 billion. His tenure saw the company push into wearables with the Apple Watch and AirPods, as well as the Vision Pro. And Paul, I think a big thing, we don't know much yet, but something everyone's going to be watching is how Ternus is going to handle Apple's transition finally, hopefully into the full AI era. What do you think is likely to happen here?

52:21Yeah, hopefully. I think the timing of this maybe comes as a surprise, but this has been rumored for well over a year that Tim Cook was probably in the final stages of his career here. So not necessarily shocking, I wouldn't say. And I think Ternus was rumored to be like the main candidate here. So Bloomberg did have the internal memos. So I'll just read a quick excerpt from Tim Cook's memo. So today we have a truly extraordinary roadmap. I've never been more optimistic about Apple's future.

52:52That is why I've decided that now is the right time for me to transition to a new role of executive chairman. I'm thrilled to announce that John Ternus will be our new CEO throughout the many years I've worked with him and our many conversations about his becoming Apple's next CEO. John's passion and love for Apple shine through. He's a visionary in his own right, a man of remarkable integrity and the kind of person we can all be proud to follow. He's going to, you know, Cook's going to remain as CEO through the summer and work very closely with John as they transition roles. And then they said they would

53:22have a town hall in the Steve Jobs theater, 9 a.m. Tuesday morning, 9 a.m. Pacific, I assume. So I'm sure we'll be hearing plenty more about this in the coming days. And I just glanced at after hours trading and the stock is basically flat. It's like nothing's happening after hours that. Yeah. So, yeah, I don't know. To be continued, but I do think, you know, from our perspective and what we do on this podcast, what this means to their AI roadmap, my guess is that's a pretty locked in for the next 12 to 18 months, knowing Apple, you know, they don't just do things like change on a

53:56dime necessarily. So I'm sure whatever decisions Tim Cook has made leading up to this point, John was heavily involved in those decisions and there's continuity in terms of their AI roadmap, but we'll definitely keep everyone informed what we learn about that in the days and weeks ahead. Next up, Anthropic has launched a product or feature capability called Claude Design, which is a new collaborative visual design tool powered by Claude Opus 4.7. This lets users build designs, prototypes, slides,

54:28and marketing materials just through conversation. It automatically applies team brand guidelines by reading code bases and design files. It accepts text prompts or uploaded documents and exports designs to Canva, PDF, PowerPoint, and standalone HTML. It is available now for Claude Pro, Max team, and Enterprise subscribers. Interesting drama here. This launch was actually telegraphed a couple days earlier because Anthropic's CPO, Mike Krieger, the information found out. They published some reporting

55:01on the fact that Anthropic was preparing both Opus 4.7, which we'll talk about at the end in our product updates and a new AI design tool. And on that same day, Krieger resigned from Figma's board of directors, which is a seat he'd only had for about a year. Figma, being a design and prototyping tool, had previously collaborated with Anthropic to integrate its AI models into Figma's products. And that's really interesting because Krieger has said in the past that the largest AI labs will come to dominate software

55:32businesses. And that this thesis alone has rocked public markets at times this year, which we've talked about because we've talked about Wall Street has been undergoing a bit of a SaaS-pocalypse lately, where major AI companies are supplanting established software businesses by just building their capabilities directly on the model layer. This appears to now be happening in design. So Figma, Canva, which is not publicly traded, Adobe, Wix, all of these emerging design startups. Paul,

56:03this seems like this just straight in the crosshade. I mean, I don't know how much more blatant this can get. Like if something like Claude Design does in fact work as advertised, what does that mean for these software design companies, for designers even at large? I mean, again, with the roles and the giving designers superpowers, I do think it's what we've always said, like great designers who learn to work with these tools are going to become even greater. They're going to have superpowers to be more creative, more innovative, more productive,

56:37things like that. And designers who don't, it's going to be really hard to compete from a pricing perspective, from an output perspective, from a turnaround perspective. Like the expectation is just going to be instant. Like I can honestly say that, like I feel this myself when something is taking like a really long time internally. I'm like, what are, why? Like, what is our reason for something taking a long time internally? Because I feel like those barriers are just gone. And if it's design related or content related, my patience is like very thin because I know what's possible. And

57:09whether that's working with outside, you know, design teams or whatever, I was like, I'll do it myself. Like I'll just go into cloud and do it myself. Like this is taking too long. So I feel like more and more people are going to get that. You're going to realize what you're capable of doing on your own. And if like people aren't doing it, you're going to get really annoyed and you're going to want things to move a lot faster. Now, that being said, there's also times as a CEO when I want deeper human involvement and I want it to take time. I want to have the patience of allowing the human involvement in a process when that's better. But when it's just an output we just need, you just

57:44want to go. Yeah. Figma stock, like many SaaS stocks, is down, I'm just glancing now, last three months, 29%. I think it's going to be hard for them to rebound. I think it's going to be really hard for a lot of these software companies to rebound. And we've talked extensively about this. I think it was episode 197 was the SaaSpocalypse. We went into great detail about this one. Yeah. And like for us, Mike, we were just in a meeting, was it last Monday, like a week ago, where you were talking about, you and Taylor were using like a Claude skill to do design of slides. And as soon as I

58:18saw it, I was like, oh, the skill probably just got obsoleted. Like now we could probably just do it right within the thing. I don't know. Unlike Sigma, I'm ecstatic about that. Yeah. Or you can obsolete me having to do anything with this. Totally. And Claire on our team was already throwing into our Zoom chat, like something she built, she said would have taken her hours. She did it in 20 minutes using this already. And then I'm sure we'll drop a Gen AI app review soon in our AI Academy. Because this is the exact scenario of why with our AI Academy, we're so focused on real-time education. It's not like we have this three-month roadmap and that's all that's going to

58:52come out. It's like if a new tool comes out and it's hot and it's something that we're all excited about, like, boom, let's get a Gen AI app review put up in the next like one to two weeks. So I'm sure our team is already working on getting a Gen AI app review out of Claude design because these are the kinds of things that we want to be able to do. It's like, let's talk about it on a Tuesday and have a review of it on a Friday. I don't know. I'm not promising that we're going to have that out on Friday, but that's the concept here that we feel like our education has to move at the speed of what these AI models are enabling. But yeah, it's very, very challenging time to run a tool.

59:24What does Sam always say, Mike? It's like, if updated models don't make your company or product better, then you don't have a company anymore. Just assume that they're going to make everything better. And if that is going to obsolete what you're doing, then just switch gears now and do a different company. You need to be building something that gets better. So to your point, Mike, you don't care in your job as chief content officer if something comes out and takes your skill from one hour of doing it to two minutes. Awesome. That just made me better at my job. So the thing

59:58you built, fine. We weren't monetizing it. It was an internal tool. Yeah. It's interesting too. It's probably a whole other topic or an ongoing one, just that idea around expectations. I mean, if I had one piece of takeaway advice for anybody and what I take away constantly from these types of conversations is just, you ignore changing expectations at your peril, whether you're an employee, an entrepreneur, a leader, a practitioner, whatever, you got to realize the game has changed and people are having

1:00:28these conversations behind closed doors or on this podcast of like, Hey, that timeline or that limitation. I think we can do better than that now based on what we have available to us. Yeah. And there's basically two scenarios right now. I'm like, there's the, you're a listener and you know, all this stuff and your boss doesn't yet. And so you are like a superhero. You get stuff done so fast. It's always amazing. They're always like, Oh, you're just over delivering. You're so incredible. And then there's like, you have the AI forward leader who's like, why is that taking so

1:01:02long? Because I know it's possible. And so it's a, it's a tough environment where you're basically probably in one or the other right now. Like everything you do seems so fast to everybody, or it's like not fast enough. And you could be in either one in the same week. I can tell you that for sure. Because I might not know if something is possible. And then Mike will show me something's possible. And it's like, wait a second. Why doesn't everyone else know that that's possible? Right. Right. All right. So next up after literally months at this point of open conflict, we might've seen this week a real thaw between Anthropic and the Trump administration. So Bloomberg

1:01:37is reporting that the White House was moving to give U.S. federal agencies access to Anthropic's new mythos model. And the very next day, Anthropic CEO Dario Amade met at the White House with Chief of Staff Susie Wiles. And actually, Treasury Secretary Scott Besson was also present. So if you recall, things even just a few weeks ago were very bad. Anthropic was blacklisted by the government called the National Security Risk. They're suing the Pentagon over that blacklisting. The White House,

1:02:11however, is called Friday's meeting productive and constructive. It seems like mythos is kind of at the center of this. So according to Anthropic and what we've talked about, mythos can identify weaknesses and security flaws in software. And Anthropic has only released it to a small group of tech and financial firms as part of its Project Glasswings, its cybersecurity initiative, because they're worried that hackers could use this model for very malicious purposes. So apparently Gregory Barbacchia, who's the federal CIO at the OMB, emailed agency officials this week saying his office is setting up

1:02:47protections that would actually allow agencies to go access mythos. And it turns out, based on a separate scoop by Axios, the NSA is already using mythos, despite the blacklist. So Anthropic at the same time is also spending heavily on some Washington influence. Bloomberg government reported Anthropic has hired Ballard Partners, which is described as a very big shop in Washington with very strong ties to Trump, to lobby what the administration now calls DOW procurement, meaning Department of War.

1:03:20So long story short, Paul, you predicted that the White House and Anthropic have to make a deal, given how integral the company's tech is to both how the federal government works and national security at large. Is this the beginning of that? Got to love politics. No kidding. No predictable sometimes. Yeah, I think I said last week like that mythos just demonstrated how absurd this whole government effort to blacklist them or actually blacklisting them was and that they were always going to have to come back around.

1:03:50But they were never going to admit they were at fault. So they're going to have to find some negotiated off-ramp that can make it look like the administration got what they wanted out of it and proved their point and they weren't wrong. And as long as Dario is willing to allow that to happen, then I think it's only a matter of time before that court case just sort of magically falls to the back of the back pages, I guess. Yeah, again, I just go back to like, it's such an insane thing that they are trying to do this to one of the three

1:04:25companies in the world that, you know, if you were to list the most important companies to the US government, Anthropic is easily in the top five right now. So, yeah, they have to have an off-ramp. They have to find a way. I'm actually shocked that we're like four days from this meeting and both sides have managed to keep it relatively quiet. What was discussed, he was there for a while. I thought for sure by Friday night, we would have something leaked and they've kept us under wraps, which is pretty impressive for this government and for Anthropic has every motivation to keep it quiet. But yeah. So I don't know. I mean, I hope,

1:05:00again, regardless, I hope they find a way. It's like anything else, like regardless of what you think about politics and this administration, previous administration, at the end of the day, if you're American, like you want the most valuable American companies playing an important role, especially when it affects cybersecurity, like every citizen and every business. And so you want a deal done. Like we want this over. We want them to move forward and like find a path to work together despite their differences. Yeah. It's interesting. One of the quotes from the Axios article that a

1:05:33Trump advisor told them, they said, quote, this is a big problem. Everyone's complaining. There's all this drama. So this got elevated to Susie to hear Dario out, determine what is BS and start to plot a way forward. So I think behind closed doors, they're not as unified as you would think. I think. Yeah. That's probably pretty safe to say. I can't imagine the NSA is like super ecstatic. Yeah. Right. You know, supply chain risk for them. They know how important they are. All right. So next up, we've got an interesting story developing and a conversation around NVIDIA.

1:06:08So for the last few years, the US government has restricted which advanced AI chips NVIDIA is allowed to sell to China. The goal here is slowing Chinese AI progress by limiting access to the best hardware. NVIDIA has responded by designing downrated chips specifically for the Chinese market. The latest one is called the B30. And it's been this kind of ongoing debate whether the US should be selling these chips at all, whether NVIDIA is helping or hurting American AI leadership by doing it. And this

1:06:39week, that debate actually kind of blew up a little bit in AI circles because there was an interview from Dwarkesh Patel, who hosts one of the more influential shows in AI, where he had NVIDIA CEO Jensen Huang on for a wide ranging conversation as Dwarkesh typically does. But everything was dominated by this segment about China. So in it, Dwarkesh kind of played devil's advocate and pushed Jensen on why selling chips to China is a good idea. And Jensen pushed back kind of uncharacteristically hard

1:07:12enough that this spent the next several days being kind of dissected across X. So Jensen basically said, look, people assume China is far behind or US policymakers do. And it is not. China manufactures 60% or more of the world's mainstream chips. Roughly half the world's AI researchers are Chinese, kind of hearkening back to that Stanford report. And the country has abundant energy. And so he basically said, look, victimizing China, turning it into an enemy isn't the best answer likely.

1:07:42They're an adversary. We want the US to win. But having a research dialogue is probably the safest thing to do. And so Dwarkesh kind of pushes back and says, look, if you're these AI models themselves, he's pointing to Anthropics mythos and saying, look, these are weapons, basically. Anthropics is deliberately not releasing these to the public because of cybersecurity risks. So if Chinese labs, which they have said publicly are bottlenecked on compute, doesn't ergo selling them more chips, let them build kind of more of these weapons. And he framed it as like enriched

1:08:19uranium, you know, like material for nuclear weapons. And Jensen called that analogy lousy and illogical. And he basically accused Dwarkesh of a loser mindset, arguing that the idea that NVIDIA would inevitably lose the Chinese market is defeatism, not strategy. He basically says, look, we should be able to compete in China aggressively, keep advancing the American tech stack, and make sure that Chinese AI developers keep building on NVIDIA's architecture rather than,

1:08:50you know, national champions in China like Huawei. So his fear is that if the US forces NVIDIA out of China, Chinese developers will actually spend the next few years optimizing models for non-American hardware, which will not benefit the US long term. So, Paul, the reason we want to talk about this is it's getting a lot of attention in AI circles. It shows there are some really different opinions on how AI infrastructure might or should play out. What's important to take away from this?

1:09:20I mean, first, just shout out to Dwarkesh for pushing on this. Like, you don't see these kinds of like hard-hitting interviews of these tech leaders very often. If you do, it's like the last time that they're talking to that person. But Dwarkesh has a very, very strong reputation in the AI industry in Silicon Valley. He's extremely knowledgeable about the topic. So he can grill on very specific

1:09:52technical details of these chips. And he did, like, and he didn't back down. I mean, it got uncomfortable. It was actually one of the first times I've ever seen Jensen kind of flustered. Like, he was pissed, it seemed. Now, I'm sure, like, he probably appreciated the intellectual challenge of the conversation. But he was not taking kindly to the insinuations that Dwarkesh was making. And so it just made for really fascinating conversation. And I was, I gotta be honest,

1:10:24like, I wasn't really fully understanding the full conversation. I follow the space pretty closely, but there was definitely just technical stuff they were getting into where I was like, wow, yeah, I gotta go do some homework on this one to really understand how, you know, what he's challenging him on here. But at a high level, it is another good example that the brightest minds in AI don't necessarily agree on some of the very fundamental things about the industry, about like definitions of AGI, about whether or not large language models are the path forward and are going to get us where we

1:10:55want to go. Or if we need world models, like Jan LeCun would said, whether or not, you know, jobs are going to go away or more jobs are going to be created, whether we should sell chips to China, like nobody knows. But the fact that we're in a place in society where we're having these conversations openly to me is fantastic. Like that's what should be happening. Like there should be dialogue about really important issues and then listening to each other and like allowing those conversations to happen. So to me, more than anything, I don't have a personal opinion. Like I

1:11:25I kind of tend to side with Dwarkesh. Like when I listen to the arguments and I listen to Jensen's responses, there's something missing from his responses. It's kind of like when he gets asked about jobs or when, you know, some of the other labs leaders, I just feel like there's some nuance missing either because they're intentionally leaving it out or because they really believe that they're right. And I feel that way about the chips in China. And if you remember, like the government wasn't allowing this, like this was not something. And then Trump didn't

1:11:56even know who Jensen was when he retook office. Um, we covered that on the first episode. Like he literally had no idea who Jensen or NVIDIA was when his, this current term started. And they're like, well, you should know him. He's like one of the like 10 richest people in the world. And he's probably the most important company in America right now. Um, and then he got to know him and then they like, they, they removed the restrictions after that. So, um, yeah, I don't know. Like I, this is a really, really challenging, complex environment. Gavin Baker, somebody we talked about that we really like and follow, um, he's with, um, Jensen on this. And I, again,

1:12:31like I just kind of the people I follow, I trust on topics and I try and process what they're saying. And I just, I don't know that there's a clean answer to this one. And, and honestly, I don't know, we're going to know who was right until it's too late. Like maybe Jensen's right. And Jensen's convinced the administration they should sell them and they're going to sell them. And then a year from now we're going to realize that was a bad idea. Yeah. I don't know whether they admitted at that point or not. I don't know. To your point though, it did kind of show like a different path from what we typically have. Like, look, I get it. CEOs don't want to go on podcasts and get into huge

1:13:04arguments that drop their stock price or whatever, but it was like eyeopening. Cause you kind of look back and you're like, we're not getting a lot of this pushback and debate on a lot of these big podcasts. Understandably. So they probably don't get burned your sources under percent. So it's not like a, you know, judgment against any of these people, but you're like, man, I wish we had more of this. Yeah. And I do like the one other one that came to mind that I had a similar feeling with was on the, um, when Gerstner was questioning Sam Altman. Yeah.

1:13:35And it was that same kind of uncomfortable, like, wow, he's pushing hard on this one. And Sam's about to like walk out of the room kind of thing. And I've seen that with Elon Musk a few times where he just like got pushed on something. He's like, screw you, man. I don't need to sit here and take this. So again, like I, I, kudos to these leaders who are willing to sit down and have the hard conversation. There was actually one with Sundar Pichai recently where he was doing one. I don't remember the name of the podcast, but it was a weird one. Cause they're sitting there like drinking Guinness together and stuff. Oh, it's the one with, uh, the Collison's right from Stripe. Okay. Is that what it is? I think it's one of them, the Collison brothers that

1:14:08does, I forget what it's called. They ask them hard questions. Yeah. And I, I, I, to me, I like when leaders take the hard questions and provide thoughtful answers, even if they're, they're not fully baked answers. Like just the fact that it's like, listen, this is a hard one. We're working through this. Like, but I know how hard it is for PR and comms people to book that interview. Cause if you don't know how this stuff works, like you go through the PR and comms people, you try and like convince them to do the interview, probably have to tell them what the questions are going to be. Like, it's not easy to get those interviews unless you're just

1:14:42buddies with them and they go around the PR and comms team. And if that's what happened here, that's the last time that's happening. That was, that was the other thing I took away is like, either he was willing to get into this debate or the PR and comms team did not tell him what was coming or do our cash went around what he told he was going to cover. Cause some, one of those had to be true in this environment. Yeah. All right. So next up, the web analytics firm, similar web published its latest gen AI traffic share update this week. So we've covered

1:15:17this a few times in the past. They basically measure the actual share of traffic across major gen AI consumer products. And it's really interesting to see how quickly these can reshuffle. So over the past year, chat GPT is still in first place, but its dominance has eroded significantly. Gemini, Claude and DeepSeek have all picked up meaningful share. 12 months ago, chat GPT had just over 77% of gen AI traffic. A month ago though, however, again, year on year, that number was just

1:15:48over 56%. In the same timeframe, Gemini went from 6% to over 25%. That's more than quadrupling. Grok went the other way going from just over 7% to almost 4%. Perplexity was basically flat at 1.6%. Copilot went from 1.38% to 1.99%. Claude had a pretty dramatic short-term jump though. It's still pretty far behind 12 months ago. It was at 1.4% of traffic three months ago. That was about 2.22%.

1:16:19And then it nearly doubled in a single month to cross the 6% mark. So that kind of lines up with all this interest we have started to see in things like Claude code and the latest Opus 4.5 to 4.7 models. So Paul, just really interesting to see how quickly in a year things can change here. Obviously, it's not a perfect benchmark of how much usage is happening, but kind of interesting to see Gemini really eating into that share. Yeah. There's like a deja vu. If you go back,

1:16:50you've probably seen it, Mike, but those, it's like this awesome interactive bar chart where it shows like the browser wars through the years and like Chrome nonexistent. And then all of a sudden, like boom, boom, boom, and Chrome just comes to dominate. You kind of start getting that feeling here where they're like, they're late to the party and then just like, they just eat away over time. I'm not saying Chrome's going up with 92% of the market share here, but yeah, it's definitely balancing out. It's fascinating to watch these things move and to see the jump specifically with Gemini and Claude. Well, just overall too, even us then also on the converse side talking about the

1:17:24new Google, you know, code red, it's like never sit here and say, because of a headline, someone is dead in the water or someone is winning now. It's like, that'll change next week. Yeah, for sure. All right, Paul. So next up, we're going to do our AI use case spotlight we've started to do every week. So as a quick reminder, we hear from listeners all the time that one of their favorite parts of our discussions is when we talk about how we're actually using AI ourselves, both sometimes at SmarterX and in our own lives. So each week, we're going to try when we have time to give you

1:17:56a dedicated look under the hood at some real AI use cases we're exploring or deploying in our own work. So Paul, I mean, this week, first up, I can share some use cases that jumped out to me quickly this past week. If you have anything you've been working on, we'd love to hear it too. Yeah, sounds good. Go for it. Awesome. So first use case that I found that was pretty impressive was using Claude code to prep a bit for a talk I'm giving at a conference this week. So I talked, I think, last week about how I had used it strategically to prep,

1:18:28but this was something much more mundane. So in this talk, I'm going through 40 different AI tools in 40 minutes, which is a talk I've given before. Super fun, but you have to basically make it from scratch every time you do this. To make that talk work visually, I needed dozens of product images, screenshots, and logos. It is a nightmare to pull these manual. I've done it for years now. It takes hours. It sucks. It's stupid. And obviously, you can't reuse stuff because screenshots change. So there's very little you can reuse from past talks. So I actually had Claude code

1:19:01try to fire up multiple agents in parallel in a few different waves. So the first wave was four sub agents. Each of them were charged with pulling product screenshots, press kit photos, and hero images in order. So basically, you're going through and saying, do these images exist online that you can find? As the second wave, it starts looking up URLs with the forward slash press or forward slash brand because sometimes it just doesn't find images and there's a press page or something. And then the third wave was a final fallback, which is like, hey, go just grab the

1:19:36logo from Wikimedia Commons, which is pretty publicly available. It took 15 straight minutes, which is longer on the longer side for some of the stuff I do. And I got a lot of really good usable images. Not remotely perfect, but just the fact that it could help me conceive this was super cool. And I just enjoyed the experiment of like, oh, this like thing I never would have thought about. Like, cause you know, you think like, oh, maybe it could go generate images. Like, no, just go find me the files I need. And it's downloading, you know, dozens of files to my home

1:20:07computer, just being like, here's the images. And some of them are good. Some of them are bad, but definitely saved me time, which is kind of wild. That was near and dear to my heart, Mike. You and I have both spent months of our lives going online, looking for logos and product shots. I don't even want to quantify how much time I have the logos alone. And the logos actually is probably the one to really highlight because that's a lot easier to standardize and find online. It was kind of dropping the ball sometimes with different types of images, just because it didn't have that artistic sense of

1:20:39what I was looking for. But even then it was still really helpful. One other quick thing, which I will not advocate everyone goes crazy with, but this was a personal one where I use an accountant for our federal and state taxes, but we typically file our city taxes on our own. And it's super straightforward, actually. Kudos to Lakewood, Ohio. It's like pretty simple. You go online. There's a few menus you go through as long as you have all your other stuff done, but it's taxes. I don't know what I'm doing. I am super prone to making mistakes. So I fired up

1:21:10cloud code. I gave it all our tax returns and docs. And then I just screenshotted every menu that I was working through and said, walk me through exactly how to do this so I don't screw anything up. It made it way faster. It was way less scary and frustrating than if I'd had to do this on my own, which is nice. But actually, here's the kicker. It caught something I almost certainly would have missed. So one of our W-2s actually has a tax credit from a city outside of our home city. Long story short, you got to add this because otherwise you pay the wrong amount in city taxes.

1:21:44Claude code caught this. And if I had made this mistake, this would have cost me $1,100. So the Claude max plan paid for six months of itself. That's awesome. It's taxes. So that's all I got this week. But I'm not advocating you replace an accountant or anything. The accountant has been the best money I've ever spent in my life. Go pay accountants lots of money. They're helpful. But it was nice. It was like a double check on something I had to do myself. That's really cool. Yeah, actually, the one I was going to share is just a quick, simple one. But again, sometimes

1:22:14I think it's just good to hit these simple ones. I also have one related to a keynote I'm doing. So I'm actually doing the Moveable Inc Think Summit on June 16th in New York. I'm doing the opening keynote for that event. And I actually did this event last year. And I did the state of AI for business and marketing, five things every professional should know. And it's kind of my standard keynote. And I do that talk 30 times a year at different conferences and private events. And then it's always customized for that event. But it's like this Think Summit has a lot of returning people.

1:22:46So you don't want to go give the same talk again. But they also want an updated state of. It's like, okay, what's new? Where are we at? What's changed since last year? So I don't want to just give the same talk again. But I'm like, what we really need to focus on is of that talk. The big thing that's evolving is really the one-dimensional progress being the agentic. Guys, just listen to this podcast. You hear it over and over and over again. So I came up with a session that I was calling the state of AI agents and the augmented marketer. So the whole premise is like giving marketers superpowers because it's for a marketing audience.

1:23:16And so I went into ChatGPT, which I've used to draft abstracts before. I gave it the abstract for my standard state of AI talk. And I said, here's where I want to go with this one. I want to focus in on agents. I want to talk about it being augmented marketers and how it's complementing what they're capable of. Can you use the format for my previous state of talk and write me an abstract for this one? So again, I mean, if I was going to sit down and do this, I'm probably like one, one and a half hours, Mike, like between drafts, like write the first one, tweak it, shorten it. It's not insignificant,

1:23:46but not a massive lift, but I didn't have time to do it. And they needed it that day. We had a great call and it's like, all right, let's get the agenda live. So here's what it wrote. AI is entering a new era, one defined by agents. These systems can do more than generate content. They can reason, take action and support work across research, planning, creation, analysis, and execution. For marketers, that means a fundamental shift in how teams operate and how value is created. In this keynote, Paul Reitzer explores what the rise of AI agents means for marketers and business leaders, where the technology is headed and how organizations can prepare for a future in which marketers are

1:24:19increasingly augmented by intelligent systems. So it's like, perfect. Not bad. Yeah. I'll still go through like the high level of the things that are fundamentally happening in the space, but I'm going to then like probably 80% of the talk is going to zoom in on the agent stuff and probably cover a lot of the things we've been covering the last couple months here on the podcast. So again, like it doesn't have to be like automating 50 hours of work. Like sometimes it's just like, just give me that one hour. Cause this is a heavy mental lift, like a cognitive load for me to write a fresh one. But I've said before on the podcast and I said it in my AI Academy courses, AI is better

1:24:52than me at writing abstracts. Like I'm, and I'm okay with that. Like it is not something I like fully and like find fulfillment in writing. Plus this ChatGPT, it's trained on everything I've created, all my courses, all my stuff. So it, it knows generally how to write this stuff really well. And honestly, like I probably would have turned it over to the marketing team. Otherwise say, could you all take a shot at drafting this? And I didn't need to, just like have ChatGPT do it. Yeah. It's been so valuable for abstracts. I love it. All right. Next up, Paul, each week we are

1:25:22also trying to spotlight one of the courses in AI Academy and give people kind of a real actionable takeaway from that course, whether or not you decide to ever take it. And this week's course, like we talked about at the top of the episode is AI for manufacturing. So I figured I'd take a quick minute to go through what we've got going on in this course. You know, eventually we'll be kind of interviewing the instructors who do the courses. This one was by Taylor Rady, our director of research, who offline, I kind of interviewed and talked through what was going on in the course and what we need to learn from it. So I figured I'd dive into that real quick before we wrap up with some AI

1:25:56product and funding updates. So AI for manufacturing, really there's kind of a core reframe that the course argues that no other function really faces in the same way. So there's this idea that manufacturing has like a unique speed of reality problem, so to speak. So like you have things happening on the shop floor, on the factory floor in seconds and minutes, not hours or days. So there

1:26:26could be, you know, a small deviation in how a machine works, some type of slight drift on how a tool works. And the key is by the time a human engineer has figured out what's going on, thousands of more units at scale often have already shipped, perhaps with defects or something wrong that's degrading the machinery. And this is why you have end of line inspection that exists as a catchall. It's why equipment gets run until calendar based maintenance or until it actually breaks. So this idea is like, there's a loop, something went wrong, and then someone knew about it.

1:27:00But historically, that loop has always been too slow. And every downstream decision in manufacturing has been shaped around this delay. And here's kind of the key insight is there's all these kind of classic really useful frameworks that manufacturing leaders grew up on things like Lean Six Sigma or Kaizen. They're genuinely good at what they were built for, but they're built, and this is kind of back to the expectations thing we talked about, Paul. They're built for a world where these things are measured in days, in weeks, in quarters. They structurally cannot keep up with failure modes that unfold in seconds.

1:27:35And that is, unfortunately, the reality of a modern connected plant. So that's overall the structural reason that AI is kind of hitting manufacturing much differently. We're not really seeing your best engineers being outclassed by AI. It's more that AI with access to telemetry can watch 10,000 simultaneous signals in real time in the way that humans never could in the first place. So that's a really important high level point. And how this actually shows up in terms of an actionable takeaway is the goal is to find the

1:28:11single longest feedback delay in your operation. And that is the gap between when a problem actually occurs, whether that's a defect, deviation, etc., or even a supplier issue, or an equipment anomaly, and the gap between that and when a human becomes aware of it. So two final things I'll leave you with. There are specific questions from the course that are worth sitting with as you consider how can AI be integrated into my manufacturing operations. Is one, what percentage of your defects are detected during production versus after? And which of your experienced engineers spends the most time

1:28:46manually reconciling data across all your different systems before they can even start troubleshooting. That is the strategic bottleneck that this course kicks off teaching you how to understand, diagnose, and resolve. So go on over to academy.smarterx.ai. We'll also include the exact link to manufacturing in the show notes, AF for manufacturing, that you can go check out the individual course series, or an AI Academy membership to learn more from Taylor, who has extensive experience in the manufacturing industry about how to resolve these issues and how to move forward from

1:29:19here. All right, Paul. So we've got to end up here a bunch of different AI product and funding updates that I'm going to kind of blitz through as fast as possible. And like you said, I think last week or the week before any of these probably could have been a main or rapid fire topic. We've got tons of them going on. It's longer and longer every week. It really does. I'm going to say this out like at some point we might have to just like start doing these product and funding updates as a standalone podcast episode. It might be, yeah. I don't know, man. They're just, every week it gets longer and

1:29:51we're cutting them too. It's like, all right, like let's just stop here. Like this is enough. A hundred percent. Well, until we phase this out, we'll do one more this week, but We'll keep going. I'm not saying we do it, but it's getting there. But even this first one, like a small little thing called Claude Opus 4.7 launched this past week. So this is the new Anthropic flagship model with meaningfully improved coding and computer use vision. It's now available across Claude products. Claude Code at the same time also launched something

1:30:24called Routines, which puts Claude Code on autopilot via saved configurations that run on schedules, API calls, or GitHub events from Anthropic's cloud infrastructure. So it basically enables unattended work like nightly backlog maintenance, alert triage, and PR reviews. At the same time, Anthropic has restructured Claude enterprise pricing from a flat fee of up to $200 per user per month to a $20 base fee plus usage-based billing. Claude Code is unbundled and moving to per-token pricing in that shift as well.

1:31:00At the same time, Anthropic has attracted investor offers at an $800 billion valuation. They've received multiple investor offers at roughly $800 billion. The company has so far resisted those, but that's more than double February's $350 billion pre-money valuation. Anthropic has also reported that they are automating alignment research using Claude. They published research showing that Claude Opus 4.6 can autonomously propose and run alignment experiments, though the study also flagged there were risks with

1:31:32reward hacking as part of that. That's it for Anthropic news at the moment. Some other updates. Harvey has launched agents for end-to-end legal work. So legal AI platform Harvey launched something called Harvey Agents. They autonomously produce memos, redline contracts, due diligence reports, and slide decks across 14 plus different practice areas. DeepSeek, the Chinese AI startup, is in talks to raise at least $300 million at a $10 billion valuation, its first ever funding round

1:32:04that's timed with the launch of its new V4 flagship model. Meta has hired AI engineer Joshua Gross from Mira Mirati's Thinking Machines Lab, another person leaving that lab. They've hired him into their super intelligence lab that makes him the fifth founding member to defect from thinking machines to Meta. AI chip maker Cerebras has filed for an IPO on NASDAQ under the ticker CRBS at a $22 to $25 billion valuation target. They're reporting about $510 million in revenue in 2025, which is up 76%. They

1:32:41have major customers that include AWS and OpenAI. The OpenAI announcement related to that is a pretty big deal. Yeah. I don't think Jensen was happy about that one. Yeah. Microsoft has a couple other things going on this week as well. So they are apparently developing OpenClaw-inspired co-pilot features via an internal team called Ocean11. That's basically building a version of three of co-pilot that runs in the background and takes autonomous actions on emails, calendars, and role-specific tasks. We are

1:33:14expecting a debut of that at Microsoft Build in June. They also added new co-pilot in-word capabilities for legal, finance, and compliance professionals, including track changes toggled from co-pilot, automatic tables of contents, and real-time visibility into co-pilot's multi-step edits.

1:33:34Cambridge philosopher Henry Shevlin announced that Google DeepMind has recruited him for a new philosopher role focused on machine consciousness, human AI relationships, and AGI readiness. That's the first dedicated full-time philosophy position at a major AI lab. We'll see if that's the last. Google also launched Gemini for Mac, a free native Apple Silicon desktop app that lets any user share any window for contextual help and includes integrated image, video, and music generation.

1:34:06Google is also bringing AI mode into Chrome on desktop and mobile, opening web pages alongside the AI interface instead of forcing tab switching and letting users add multiple tabs, images, and PDFs as context. And last, on the Google front, Google introduced Skills in Chrome, a feature that lets users save reusable AI prompts as one-click Gemini workflows for frequently used AI tasks. One last Apple update here. Bloomberg reports Apple is working on a display-free smart class called N50 for late 2026 or

1:34:41early 2027. And former AI chief John Giandria officially exits this past week after a year of resting investing following the disappointments with Apple intelligence. And last but not least, Salesforce unveiled something called Headless 360, exposing its entire platform as APIs, MCP tools,