77 Min Read

[The AI Show Episode 250]: Midterm Voters Reject Data Centers, AI Speeds Up Anthropic’s Research, Celebrating 250 & Nvidia’s Spy Probe

By Claire Prudhomme on November 10, 2026

Podcasts
Featured Image

Voters backed limits on data centers in 40 of 41 local ballot measures during last week's U.S. midterms, and the AI industry is already scrambling to respond.

Paul Roetzer and Mike Kaput break down what the backlash could mean for the build-out. That includes OpenAI's rebrand to "compute campuses" and why Paul expects your AI licenses to get more expensive.

They also dig into METR's independent assessment of AI-driven acceleration at Anthropic and mark the show's 250th episode with a look back. Rapid fire covers Nvidia's board investigation into Jensen Huang, a messy week for AI in education, and Menlo Ventures' $177 billion enterprise AI spending figure.

Listen or watch below—and see below for show notes and the transcript.

This Week's AI Pulse

Each week on The Artificial Intelligence Show with Paul Roetzer and Mike Kaput, we ask our audience questions about the hottest topics in AI via our weekly AI Pulse, a survey consisting of just a few questions to help us learn more about our audience and their perspectives on AI.

If you contribute, your input will be used to fuel one-of-a-kind research into AI that helps knowledge workers everywhere move their companies and careers forward.

Click here to take this week's AI Pulse.

Listen Now

 

Watch the Video

Timestamps

00:00:00 — Intro

00:07:53 — U.S. Voters Reject Data Centers During Midterms

00:24:33 — Has METR Found Proof of RSI Happening in the Labs?

00:40:21 — Episode 250 of The Artificial Intelligence Show

00:53:45 — Nvidia Opens Probe Into Alleged Spy Operation Targeting Jensen Huang

01:05:06 — AI in Education Is Getting Messier

01:16:46 — Enterprise AI Spending Hits $177 Billion

01:22:00 — Controlling AI Costs

01:31:58 — AI Use Case Spotlight

01:39:16 — AI Product and Funding Updates

  • Anthropic Brings Personal Finance into Claude
  • OpenAI Scales AI Bug Hunting as Replit and Cognition Push Stronger Agent Security

 

AI-Academy-POD250-promo

To mark our 250th episode, get $250 off a new AI Academy by SmarterX individual license through November 30 with code POD250. That’s 25% off your first year of the full Academy experience: courses, professional certificates, live sessions, and community.

Leading AI literacy for your team? Buy 25 or more licenses by November 30 and Paul Roetzer will run a private one-hour virtual session with your team. Buy 250 or more and Paul will come to you for a private in-person session in the first half of 2027 (US only). Existing business members adding 25 or more licenses qualify too. Join at academy.smarterx.ai, or contact our team at support@smarterx.ai about business licenses.

Read the Transcription

Disclaimer: This transcription was written by AI, thanks to Descript, and has not been edited for content.

[00:00:00] Paul Roetzer: Everything looks different to me right now. Every business, every person, every career path, it just looks fundamentally different based on what's available today and how good it's gonna get over the next like 12 months. I just can't look at 2027 the same right now. Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable.

[00:00:23] My name is Paul Roetzer. 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.

[00:00:51] Welcome to episode 250 of the Artificial Intelligence Show. That is what I said, 250. I'm your host, Paul Roetzer. I love my co-host [00:01:00] for 230 of those episodes, I believe. Mike Kaput. Welcome, Mike.

[00:01:05] Mike Kaput: All right, Paul. This is a big one.

[00:01:07] Paul Roetzer: Yeah, it's kind of nuts. We actually thought about it like, I think it was last week when I saw the 249.

[00:01:11] I said, oh. Next week's 250, that's a lot of episodes we should probably do something for that. So actually we are gonna talk about the 250th 50th episode in one of our topics today and give you a little background. But yeah, thanks for being with us today. Thanks for being with us. some of you have probably been with us since the first episode, I would imagine, but, we appreciate everybody being here for another week of covering AI news and wow.

[00:01:37] Has it gotten a little bit crazy? you know, we've, we've kind of joked, I guess there is a new CBS show about Silicon Valley and ai. I think the writers gotta go back to work because whatever is happening right now with Jensen Huang and Nvidia is like off the script. Like it is, it's beyond Hollywood.

[00:01:57] and we're gonna get into that one too. [00:02:00] 'cause that was one I didn't want to talk about on Friday. I literally shared it with Mike. I'm like, yeah, I, I don't think we're gonna cover this. And by Sunday night it's like, oh my gosh, we gotta cover this. So if you haven't heard the news on Jensen Huang and Nvidia, if you're wondering why your Nvidia stock is probably cater cratering, we will explain all that in a few minutes here.

[00:02:18] Okay. So this episode is brought to you by AI Academy by SmarterX. You hear us talk about AI Academy all the time. This is like the main focus of our business at SmarterX, in addition to our events in this podcast. but there has never been a better time to become a part of AI Academy by SmarterX because we're celebrating our 250th episode.

[00:02:38] With a special offer, which I will get to in a moment. So if you're new and aren't sure what AI Academy is. So AI Academy helps individuals and businesses accelerate their AI literacy and transformation through personalized learning journeys and an AI powered learning platform. In addition to our collection of on-demand professional certificate course series, there's 21 of them I [00:03:00] believe, Mike, and that's across AI foundations, AI for departments and AI for industry.

[00:03:04] So it's a collection of course series. Within those areas, new educational content is added every week, so you're always up to date with the latest AI trends and technologies. Plus, as will be highlighting during our November 19 AI Academy Day, we have a major focus moving forward on dynamic learning.

[00:03:22] Including more real-time content about AI agents, the inside stories about personal and business transformations and platform specific courses, which everyone's been asking for. Those are coming soon for Chad Chip t, Claude, copilot, and Gemini. So if you're learning a development team in a big enterprise is struggling to keep up with the pace of change, we've got you.

[00:03:43] We're gonna be launching all new course series tied to the platforms themselves. And so now onto the offer. So again, 250th episode only. This is gonna run through November 30th. So you have, what about 20 days or so to take advantage of this? There's two options here. So if you are an individual and you're buying a license [00:04:00] for yourself.

[00:04:01] You can get $250 off a new AI Academy individual license. So that's the full academy experience, courses, certificates, live sessions, community, everything we're building to help you, get ahead 25% off your first year. So again, $250 off your first year, with POD250. I'll give you that code again in a minute.

[00:04:21] So that's for individual learners. Now, if you're a business leader who's charged with driving AI literacy and transformation for your team or organization, we have two options here. You can buy 25 licenses or more by November 30th, and I will do a personal, private one hour virtual session. With your team.

[00:04:40] We can do AMAs, we can do a state of AI talk, we can just do a consulting session, whatever you wanna do. So, one hour of, of my time for 25 or more licenses by November 30th. if you're an existing member and you want to add 25 licenses, we'll extend that offer to you as well. And then the big kicker is for 250 or more licenses.

[00:04:59] So [00:05:00] if you want to, get a business account with 250 or more licenses, I will actually come to you and do a private session in the United States only. So it's gotta be in the us in the first half of 2027. I hope I regret this. Like this is one of those offers where I'm, I'm kind of thinking about my schedule and saying, how in the world would I do this?

[00:05:19] But like, I hope that what I have to solve for is how to make this many trips. 'cause I would love to come and meet with as many companies as we can, hear what's going on in your organization and try and help, you know, really drive that human-centered transformation. So again, if you're buying a business account.

[00:05:34] 25 or more licenses by November 30th, I will do a private one hour virtual session with your team. If you do 250 or more, I will come to you and we will do a private session, in the first half of 2027. So again, there's never been a better time to join AI Academy as an individual learner. Or get your organization started with a business account, you can go to academy.SmarterX.ai and you can join right there with the POD250 at [00:06:00] checkout for individuals.

[00:06:01] If you're looking at the team licenses, just reach out to our team and we'll get you hooked up there. and then the last note here, Mike, is we have that AI Academy Day I mentioned on November 19th. We'll put a link to it in the show notes. You can go to SmarterX.ai and register for that today.

[00:06:16] That is a free event on building an AI forward workforce. So while we're gonna talk about. All of our plans for dynamic learning and what's coming to academy. it's also gonna be sort of an, an educational session about how we think about learning and development within organizations. So you'll find value in it, whether you become an AI Academy member or an AI Academy member, or just want to be a part of that day.

[00:06:36] So again, academy.smart Rx ai, $250 off, individual memberships. And then either, an in-person session for 250 or more licenses. Or a virtual session for 25 or more licenses. So again, I hope my schedule's packed and I'm regretting that I made this offer, but we, we were trying to think like, what is something we could do that we've never done before?

[00:06:56] And we've never done either of these before. So, we hope you take [00:07:00] advantage of that. Alright, so onto this week's. episode the survey. Mike, give us the rundown of what's going on with AI Pulse. We always start off with that.

[00:07:09] Mike Kaput: Yeah, so we are running our AI pulse again this week that we've been running the last week or so.

[00:07:13] So this question is about how AI has expanded your role and if it's expanded it into work that typically belongs to adjacent roles. So it takes 15 seconds to fill out. If you have not filled it out yet, please go to SmarterX dot ai slash pulse. We'd love to hear from you.

[00:07:30] Paul Roetzer: All right, this feels like it was already a year ago, Mike, but last Tuesday was the midterms in the United States, and as we've said, we generally try and avoid talking politics on this show, but wow, there, there is no way to avoid discussing the implications of the midterms last week on the AI industry, which is where we are going to center our discussion starting today.

[00:07:51] Mike.

[00:07:53] U.S. 

Voters Reject Data Centers During Midterms

[00:07:52] Mike Kaput: All right, Paul. Yes. So this past week's US midterm elections delivered among other things, a potentially [00:08:00] major setback for data center expansion. A few threads to pull on here. So first, voters favored limits on data centers in 40 out of 41. That's 40 out of 41 local ballot measures. So they either backed restrictions that were on the ballot or rejected proposals supporting.

[00:08:20] Development. Now, voters also elected several prominent candidates who campaigned on slowing the data center build out. So there are incoming governors, including Gina Hinojosa in Texas, Keisha Lance Bottoms in Georgia, and Amy Acton in Ohio, all of whom in one way or another, supported pauses on new data center development.

[00:08:42] And Texas is especially important there. There are varying conditions in each state for allowing projects to proceed, but data centers took a center stage in those campaigns. This push also reached the Senate, Texas Senator Elect, James Talarico called for ending big tech tax breaks and making [00:09:00] data centers pay for their own infrastructure.

[00:09:02] Michigan Senator Elect Abdul El Sayed supported pausing approvals until federal protections are in place for data centers and for the communities where they're being built. So these positions are reflecting a lot of voter concerns about things like electricity, bills, water usage, how much say communities have over these projects.

[00:09:23] Now, the election results do not stop a lot of the construction already happening, but they are going to probably influence pretty heavily. How planning and considerations go for newer projects in 2027? So Paul, there's a lot of kind of threads to pull on here. openAI's is already scrambling to react to the changing political environment.

[00:09:46] I'm curious, as the electoral wind in the US have shifted so dramatically since we recorded our last episode, what do you see in your mind? what does it mean that the blue wave sweeping Congress and state level [00:10:00] races, what is that going to do to AI policy in the near term?

[00:10:04] Paul Roetzer: Yeah, I mean, at a high level, obviously there's gonna be a slowdown in the data center build out.

[00:10:07] There's gonna be lots of people being called before the Senate to, you know, explain what's been going on. And I think there's gonna be a lot of movement on regulation, probably most likely at the state level. Like, I think things are gonna start getting pushed through, so it's just gonna make it more challenging definitely for the AI labs to move forward at the pace that they were prior to this.

[00:10:28] I just, I I'll, what I'll do, Mike, is I'll go through a collection of, articles that I, I, I was kind of reading, catching up on all of this and, and just sort of pull some excerpts out. And then we will kind of close out this when we're talking about what you were alluding to with the compute campuses as openAI's is now calling them.

[00:10:45] Nothing like a rebrand to change perceptions. All right. So Axios, this is from, and again, we'll put links to all of these in the show notes. This is from November 4th, I think. voters show deep distrust of the AI race Tuesday, approving dozens of [00:11:00] data data center, moratoriums and electing a raft of candidates who favors slowing a $10 trillion build out, including in Texas.

[00:11:09] As you mentioned, Mike, the epicenter of the boom, which is also where Elon Musk was building all of his stuff. So why it matters in an election that hinged on an affordability crisis for many Americans, voters made it clear they, they view Silicon Valley's AI ambitions not as a technological miracle.

[00:11:25] But as a resource straining threat to their wallets and communities. All told nearly two thirds of 2028 data center capacity could face delays. Although most political analysts don't see Democrats permanently blocking construction because they would then run the risk of impacting the economy negatively, which could then cost them.

[00:11:45] The 2028 election. So it's this like fine balance that Democrats not have to play. Axios continued AI had its official debut as the biggest issue in American politics. I don't know that I agree with that, but that's what, assembly member and RAISE Act author Alex [00:12:00] Boros told Axios AI is, this is, this is a quote from Boros.

[00:12:04] AI is reaching into every part of our society and economy, and it's now crystal clear. It's impacting every part of our politics too. I, I would not agree that it was, the biggest issue in American politics. It was certainly an issue. Yeah. I would agree that it has the possibility to be the biggest issue in the 2028 elections.

[00:12:22] Alright, so then the Washington Post said in Texas where industry estimates show a quarter of the tech industry's new data centers are planned to be built. So a quarter of all data centers planned to be built. Democrat, Gina, how did you pronounce her name? Mike. I don't wanna crush

[00:12:37] Mike Kaput: ina

[00:12:38] Paul Roetzer: Hosa. Okay.

[00:12:38] Ina Hosa, over incumbent governor Greg Abbott rattled the industry and its investors. She champions a moratorium on new projects along with regulations that could impose aggressive protections for the environment and electricity rate payers from the effects of data centers, immense energy consumption.

[00:12:56] Now, you alluded to this, Mike. There wa there was one place where data [00:13:00] centers won. I thought this was a fascinating side note. It said just one community appeared to give a thumbs up to data centers Harwood North Dakota, a town of 800, where three quarters of voters struck down proposed restrictions on industrial development.

[00:13:14] So North Dakota, we had a winner in the town of 800. All right. wall Street Journal, the Election Day Revolt that threatens America's ai boom is the, title of this one. This one says this is going to probably, this is a quote, I'll get to that in a second. So this is going to probably elongate the timelines for the data center build outs according to Sequoia capital partner, pat Grady.

[00:13:37] that's probably the most tangible impact of what happened with the elections yesterday. Now, Brad Gernstner, who we've talked about, the BG two podcast, very influential in the Trump administration. He said, now with this blue wave, I think it'll slow down even more. on the All In podcast where he's frequently a guest, adding that roughly 25 to 35% of next year's data [00:14:00] center build out in the US could be at risk quote.

[00:14:03] That's why I think it will be really challenging for the AI info trade. he said, referring to shares of data center companies, the article continued. SpaceX, CEO Elon Musk poured nearly a half a billion dollars into the midterms, while others in the sector donated tens of millions of dollars to help elect candidates who are more open to data centered development.

[00:14:23] Elon focused a lot of his money on the Texas races because his headquarters are in Texas, and that it just, it didn't go well. Now don't feel bad for Elon. Him spending 500 million on the midterms is like you and me, you know, spending $50 on dinner, but like still a lot of money. data centers in space are looking pretty appealing right now.

[00:14:45] Musk said on x early Wednesday. Referring to his idea of locating the facilities in orbit, he resides in Texas and SpaceX is based there. The data center Buildout has buoyed economic growth at a time when overall hiring has slowed. And other economic Inca [00:15:00] indicators have stagnated. The boom has flooded.

[00:15:03] Many rural communities with investments and property tax revenue. Data centers have been the primary driver of construction spending growth. set of chief economists at the Associated Builders and Contractors Association. The construction trade group said, one can imagine that the data center construction boom is going to really soften next year.

[00:15:22] Okay, so now onto openAI's. And they published a article that said, setting a new standard for responsible AI infrastructure. So this is obviously timed to knowing that this was going to happen. so he said, the promise of AI is that it gives people more power to shape their own futures. It can help someone learn a new skill, build a business, or solve a problem that was previously out of reach.

[00:15:46] Building more computing infrastructure capacity in the United States can help lower the cost of delivering ai. Put these tools in more people's hands and ensure America stays at the forefront of the technology. The communities that host this [00:16:00] infrastructure should have a clear say in shaping their futures too.

[00:16:03] Residents should be able to influence how a campus is developed. I'll come back to that word in a minute. Hold developers accountable for covering the cost of powering it and share in the jobs and business opportunities it creates. How we build these campuses is one of the first tests of whether the benefits of AI will be broadly shared.

[00:16:20] So they went on to say in this article. Today we are publishing our national Community Compact and a compute campus policy blueprint. So compute campus. the compact sets out what communities can expect from openAI's at future campuses that we own and operate. The blueprint proposes policies to help ensure that compute campus development gives communities a meaningful voice, protects families from new costs, and broadly shares the economic gain.

[00:16:49] So this is, they're basically taking the hard lessons of the last few years and saying, we understand it. You all don't like the data centers and how we've gone about building them. This is our plan for how [00:17:00] to make this work. So the, the thing that they link to the Compute Campus policy blueprint then takes you to A PDF that says, AI infrastructure that works for America are compute Campus Policy Blueprint.

[00:17:14] just grab a intro to there. It says, unlike previous waves of software driven innovation. AI requires enormous physical investment in American comput com communities in data centers, energy infrastructure, advanced manufacturing, and the skilled workforce needed to build and operate them. AI training data centers more accurately de described as compute campuses.

[00:17:37] There's our rebrand are unlike most data centers that power internet searches, email or phone calls, so they're trying to differentiate between what we used to think of as data centers and what we, we are actually building. The scale of operations creates massive demand for skilled construction and operations jobs.

[00:17:53] they give the example of in Pike County, Ohio, where the compute campus is expected to create 2,500 long-term [00:18:00] operating jobs with an average annual salary of $80,000 and 26,000 construction jobs. then they talk about how every future compute campus that they own will have local compact, they're gonna pay their own way on energy and infrastructure.

[00:18:15] They're gonna protect local resources and report on their impact, and they're gonna create opportunities for local workers and businesses. so big picture. One final article to site is Tech Billionaires Face Tricky Political Waters After Democratic Sweep. This is from the New York Times, and it said many donated money to President Trump's favored causes and became his allies.

[00:18:37] Now that political alignment may be a liability. tech leaders may find their business dealings under investigation quote. They are going to need to plant new forests for all the subpoenas will be printing. Said Senator Edward Markey, Democrat of Massachusetts, posted on social media on Tuesday. So yeah.

[00:18:56] voters don't like data centers. We [00:19:00] know now,

[00:19:00] Mike Kaput: so rename them.

[00:19:02] Paul Roetzer: yeah, so compute campuses, super intelligent compute campuses.

[00:19:06] Mike Kaput: Yeah. There you go. I was gonna ask, is compute campus the new super? I dunno, space intelligence,

[00:19:13] Paul Roetzer: I don't know, but yeah, it's, so there's going to be a lot of AI tech leaders before Senate starting in January.

[00:19:20] you can get used to that. I feel like it's gonna be a new weekly session as to like segment what, what, you know, this week in AI leaders visiting Washington. yeah. And things are gonna slow down a little bit, but like I've said before, like I'm, I'm not sure that's a bad thing, but yeah. What the way it could, I don't know, like try, I'm trying to make, how do you even make this like tangible to people?

[00:19:43] the cost of your licenses may go up real soon for ChatGPT and Claude and Gemini because it's just supply and demand. There's going to be less supply for the inference compute that we all need to use our AI tools. if they can't build their data [00:20:00] centers, they're going to raise the cost of your licenses.

[00:20:03] That, I hadn't really thought about that till this moment, but I can almost guarantee you price increases will be coming to the favorite AI system near you.

[00:20:12] Mike Kaput: Or, or possibly usage reductions, right? Massive.

[00:20:16] Paul Roetzer: Yes.

[00:20:16] Mike Kaput: Yeah. Yeah. And it also may, so they may keep the price anchored where it is maybe, and yeah,

[00:20:20] Paul Roetzer: especially on high inference cost solutions like video image generation, world models, agents, things that require a lot of.

[00:20:31] You might see like significant reductions in the lower priced models. And you could see, like I I, I would not be surprised at all, Mike actually, as I'm thinking about this, if you don't start seeing the first like thousand dollars a month, $2,000 a month, where they're now gonna say, listen, this is replacing the need to hire someone.

[00:20:48] Mike Kaput: Hmm.

[00:20:48] Paul Roetzer: It's gonna cost you $5,000 a month, but it's gonna do the work of five analysts and two marketers and whatever. So it could actually accelerate the positioning of this as an alternative to people [00:21:00] because they're gonna start charging you more because they, they need to limit the compute usage.

[00:21:04] Mike Kaput: It wouldn't surprise me if some of that messaging to centers around like, Hey, look at this business owner who couldn't hire someone to do this, so therefore they turn to the thousand dollar a month plan, you know, as a way

[00:21:16] To almost say, I'm sure behind the scenes, that's not the small business. A hundred percent that type of thing. Like, hey, so and so company needed an analyst on demand 24 7. They could not do that because people couldn't do it or didn't have the skills or whatever. Pick your angle. I wouldn't surprise me. I think

[00:21:34] Paul Roetzer: actually now you're saying that I saw an ad for ChatGPT last night and I was like, oh, this is better than their really bad ones that were running at the beginning of the year where they would like scan the text real fast.

[00:21:43] Like what does this ad even for? This was one that was more human interest and like helping people run a business or start a business. And so I agree. I think you could see a change in messaging where it might cost you more, but it's powering small business. Right, right across the world. And

[00:21:58] Mike Kaput: what one final from

[00:21:59] Paul Roetzer: market view [00:22:00] campuses?

[00:22:00] Mike Kaput: From our right. Yeah. We'll see how long that lasts. what, one final question for you. So obviously it's a super complicated issue in terms of the infrastructure build out, but it sounds like, so there's projects already greenlit for 2027. I think a lot of this gets pushed into 2028, but how likely do you think it is we actually see.

[00:22:20] A major blocking of data centers? Or is this gonna be a lot of bluster and maybe some local moratoriums, but, or do you expect like really existential issues here?

[00:22:32] Paul Roetzer: I think, well, I think there's like fundamental beliefs about they're not good. Like they, you have to deal with that. But I do think that more likely because the Democrats understand the importance of AI build out to the economy

[00:22:45] Mike Kaput: Mm.

[00:22:46] Paul Roetzer: And to national security and things like that, it's more of an en enormous leverage point now to get money funneling back into these communities. I see. And making a real positive impact. So the Democrats are gonna hold [00:23:00] a lot of negotiating power here and I would imagine they're, they're gonna use it to an extreme,

[00:23:06] But the AI labs are gonna have to play ball. Like there's, there's gonna be no way around it. So I think they'll probably arrive at it. I just feel like it's gonna be really messy. there's gonna be a lot of things going on, a lot of articles coming out, a lot of leaks happening, but at the end of the day, everybody knows that we can't just stop ai.

[00:23:25] Like it would not be good for the economy to do that. So. Yeah, I think it'll move slower, but generally the commitments are usually like two years out. So like 2026 we're building for commitments that were made in 2024.

[00:23:39] Mike Kaput: Right.

[00:23:39] Paul Roetzer: So most likely, you know, the first real impact you would see of this is gonna be stuff that would've been coming online in 28, 29 probably.

[00:23:47] Mike Kaput: Okay.

[00:23:48] Paul Roetzer: but anything that was committed in 25, 26 that's already broke ground or started, like, I, I don't think that stuff's stopping, but I haven't analyzed all, you know, 40 bills and stuff

[00:23:58] Mike Kaput: like that.

[00:23:59] Yeah.

[00:23:59] Mike Kaput: [00:24:00] It's like the AI generated viral video out there where they have all the AI generated tech leaders saying, we were always blue.

[00:24:07] Paul Roetzer: Dude,

[00:24:07] Mike Kaput: remember

[00:24:08] Paul Roetzer: that? Dude, we have. Okay, so if you don't know what Mike's talking about, there's one of the best AI slop videos like ever created. and we'll put a link to it in the show notes. I, the guy who did it just like creates AI videos online. It's. Amazing. It's like, it's

[00:24:22] Mike Kaput: so funny.

[00:24:23] Paul Roetzer: It's all the AI tech leaders, like singing like Kumbaya, like we are always like Democrats and this is gonna be great.

[00:24:29] And it's, it's entertaining.

[00:24:33] Has METR Found Proof of RSI Happening in the Labs?

[00:24:33] Mike Kaput: All right, so our second big topic this week, this past week, the AI safety nonprofit METR, METR, which we have talked about at length on this show, published an assessment. Of how much AI is accelerating research at Anthropic. So according to METR, this is directly from their post quote.

[00:24:52] From August 12th to October 29th, 2026 METR staff performed a high access [00:25:00] evaluation at Anthropic to assess the degree to which AI is accelerating ai r and d end quote. So this included examining research data and interviewing Anthropic staff. And basically what's happening here is METR is searching for evidence of what we've talked about in the past, recursive self-improvement or RSI within a major lab.

[00:25:22] So RSI is that concept of AI getting good enough to be creating and improving itself. So. AI creating the next generation of models, which some theorize would lead to basically a runaway explosion of AI development. So METR did not find proof that RSI is happening at Anthropic, at least from what they disclosed.

[00:25:45] We'll come back to that in a second. But they did find some interesting traces of acceleration at Anthropic. They say the company has made roughly two years of progress in the past year alone, and that most of the [00:26:00] acceleration comes from using AI for AI development. However, they do say this is still primarily AI helping human researchers not doing all the research itself.

[00:26:12] The report does not show AI independently building successive generations of more powerful ai. Now there's a big, big caveat here. Anthropic redacted much of the evidence from the public report that METR published. So the conclusions in this whole thing are pretty difficult to check. So METR actually mentions this frequently in the report that under this arrangement basically Anthropic was able to redact things and that the redactions make explaining certain conclusions quite difficult.

[00:26:44] Now, Anthropic said these redactions were quote aimed at removing confidential information that could accelerate other AI developers. So Paul, I'm curious, like, can you walk us through the big picture here? Your thoughts on it? I'm just curious too, why does this matter for your [00:27:00] average business leader who's trying to figure out how to better approach this technology?

[00:27:05] Paul Roetzer: Your company can get disrupted way faster is like the, like. So at a very high level, the way I interpret what's going on is things are moving faster than expected, even within the labs themselves. And even faster than the most aggressive models that we talked about last year. Like, like the models of growth, in the AI 2027 report.

[00:27:28] So if you remember, I don't remember what episode we talked about that on. We've talked about it multiple times. Couple

[00:27:33] Mike Kaput: times, yeah.

[00:27:33] Paul Roetzer: Yeah. It's like basically AI researchers playing out, like the scaling laws within AI and how fast things were gonna go and if they got to recursive self-improvement, how it could accelerate.

[00:27:45] And so the most aggressive versions of that, it's actually pacing ahead of that. And at the time, people were laughing at their concepts of how quick things could emerge

[00:27:55] Mike Kaput: because their prediction got very weird. Very quick. Yes. If you have [00:28:00] this runaway improvement, I won't get into the details. Yes. But it's like you're dealing with like super intelligence even.

[00:28:05] Paul Roetzer: Correct. Like if you play this out, like what could happen to businesses, to jobs, and to the economy if the models actually get this good this fast. So yeah, like a super high level, and I'll unpack it in a second. It's just like. If you didn't have a sense of urgency to figure this stuff out before you sh you really, really need to have a sense of urgency to figure this stuff out either for yourself, for your, your coworkers, for your organization, for your industry.

[00:28:32] if you think it's moving fast and hard to keep up with now, it's only gonna get harder. Mm. And so like. We, we just need to be honest with ourselves about where we are right now. So, METR, you kind of alluded to this, Mike. They, they said in a, their ex-post when they shared this research, which by the way, I think we mentioned this last way, this research like blew up in AI circles.

[00:28:53] Yeah. Like, and again, our litmus test is always do the people we trust and follow care [00:29:00] about something. And every person I follow in this space was talking about this article. So like, it's a big deal. So they said, we submitted a 100 page full report of our assessment to Anthropic for the public version.

[00:29:12] Anthropic redacted a majority of the evidence we found most informative. As a consequence, the public must rely on our judgment more than we consider ideal meaning. They're saying, trust us, we saw things. Right. It is moving really fast, but we're not allowed to tell you exactly how we know that Now, Sam McCandlish, who I've never heard of before, this is a co-founder of Anthropic.

[00:29:38] I find myself, I was like, how many co-founders of Anthropic are there? Like how many more have we've never talked about actually, like are listed as co-founders in Anthropic.

[00:29:45] Mike Kaput: Well, he, he posted, this is his first post on X. Okay. He created, he created his account in November of 2026. I had to go look this up because

[00:29:54] Paul Roetzer: someone said

[00:29:55] Paul Roetzer: It was his first.

[00:29:55] Mike Kaput: It’s real. Yeah. Right.

[00:29:56] Paul Roetzer: Yeah. Going, yeah. He was on, he's a co-author on the [00:30:00] 2020 scaling laws research that Dario Amodei did that I've

[00:30:04] Mike Kaput: never even heard of this guy.

[00:30:05] Paul Roetzer: Yeah. That projected all of this. So yes. So he was apparently the CTO of Anthropic and now he's the chief architect. So it just shows you like Mike and I are as tapped in as you can be, and like there are influencers lurking within these companies that just like show up every once in a while you're like, wow, okay.

[00:30:21] Who's this person? Okay, so co-founder, chief architect at Anthropic, he's one of the researchers who helped establish the silence of scaling ai. Again, going back to the 2020, research with Dario Amodei and now helps lead the technical work of training Claude. So this is his post. I'm Sam, one of Anthropics co-founders, and I lead much of our AI safety work.

[00:30:43] I'm going to have more to share on the work over the coming months. Many of METR's conclusion conclusions match my own. So again, this is now the guy who basically is probably in charge of this METR research. Yeah. He's saying yes, we, we extracted most of the useful [00:31:00] information, but I'm just going to tell you what they're telling you is true, basically.

[00:31:05] So their conclusions match my own AI is speeding up our research, but it's hard to say exactly how much while they conclude. AI can't yet make the most important contributions to model training on its own. They think the pace of development could, by default accelerate further. We determined that sharing exactly what is driving the speed up isn't the right thing to do, both because it's commercially sensitive and because it risks handing out a roadmap to others to go faster.

[00:31:37] So he's saying we made the right decision. Like yes, we extracted a whole bunch of stuff from METR, but had we put that out into the world this, this could go sideways really fast. So he doesn't, he doesn't regret the decision that they've made. He then said, in the coming weeks, we plan to share a report with our own analysis of the speed of AI progress within Anthropic, including where we stand against our [00:32:00] responsible scaling policy, which you just alluded to, Mike.

[00:32:02] They had to update in the S one. Plus new steps we're taking to strengthen our safety and security practices as AI keeps improving quickly. So then I was like, okay. How do we sort of encapsulate, like what are the implications of this? So we'll go back to Sergey Brin 'cause it's been talked about now for about six to 12 months that Sergey is pushing hard on recursive self-improvement at Google.

[00:32:23] That he sees this as the path forward. Which I don't, well, I don't really wanna speculate too much, but I, my guess is Demis Hassabis would not be in the camp of accelerating RSI. I think Demis sees the risks more than others. Hmm. And I, I don't know that, maybe that's not part of the reason why Demis is no longer like.

[00:32:44] Leading DeepMind is maybe the research path took a a different direction than he would maybe be comfortable taking it.

[00:32:51] Mike Kaput: Mm.

[00:32:51] Paul Roetzer: okay. So, October 15th of this year, we have a Business Insider article that was a prelude to Argon, which is the newest model. Its said in a [00:33:00] round table discussion about the creation of Google's new Gemini four model.

[00:33:03] Argon, company co-founder Sergey Brin said that AI development was accelerating internally and that models are now helping build their successor. So that's RSI, early forms of it. Brynn pointed to the release of Google's flash models before Gemini four as evidence that this was happening and said the company was quote, starting to get onto the RSI treadmill.

[00:33:23] Treadmill, continued and now our researchers and the model work hand in hand to make the next version of the model. So again, that is. Early forms of RSI. Others at Google appear to share Brynn's perspective in a post on x Vihan Jain, RSI leader for Gemini at Google DeepMind remarked how embedded the Argon model had become in its own evolution.

[00:33:47] He said, when your post training demand shifts from asking for raw ships to asking for argon quota, you know, the RSI fly reel has reached a new gear. Hmm. So not even hiding the fact this is [00:34:00] happening now, the potentials here. So this is a tweet or a post from, I think this was just the other day. Pushmeet Kohli, or Pushmeet Kohli, the chief scientist, Google Cloud, VP of DeepMind.

[00:34:14] He said, across our projects at Google DeepMind, we are witnessing a genuine phrase, phase transition in the pace of scientific discovery. Yeah, I think I saw this on Sunday. you'll see a glimpse of this in our releases in the next few weeks. As this AI powered acceleration unfolds, our focus will be on empowering the broader scientific community to harness these tools responsibly.

[00:34:37] Greg Brockman posted over the weekend, frontier models with lots of test time compute are becoming the engine of scientific discovery. So they're all saying is like, we are all seeing something you aren't all seeing yet, but we are about to unveil it to you. So imagine what happened in mathematics now, like imagine that being applied to all areas of scientific discovery, which is [00:35:00] super exciting, scary, but like exciting at the same time.

[00:35:04] And then one other one to throw in the mix says like. Again, the whole, you have to be aware because things could get disrupted real fast. So it always probably weird to people when we quote Pliny the Liberator. this is an ex-US an anonymous account who is, definitely followed closely by the AI labs.

[00:35:23] This is someone that is, you know, a verified insider, AI researcher, kind of a hacker, sort of like takes the models, reverse engineers, releases, system instructions, things like that. but has a, a, a, a trusted history, I would say on X as you know, some, someone people follow. So on November 5th, he, he, she posts, my Spidey senses are telling me the labs are, as we speak, using on guard railed versions of their unreleased models to systematically reverse engineer all in caps of the internet software, [00:36:00] apps, and games.

[00:36:01] Then feeding the reconstructed source code straight into their training pipelines. While, while simultaneously building guardrails to ladder pull the rest of us. Hmm. So what he's saying is they're basically taking these, the most powerful versions of their models without the guardrails, letting them loose on the internet, stealing not just the text and video and images, which is how they train the original models.

[00:36:26] They're stealing the underlying code of how everything is created and then using that code to build the next generation of everything where they don't need the original sources of that software. It's a, it's a wild concept that I hadn't really like, deeply thought about, but, wouldn't surprise me.

[00:36:48] Right, right. If that's a new source of training, data is all the underlying code for how the internet was built, how software was built, how businesses are built.

[00:36:59] Mike Kaput: Yeah. [00:37:00] Wow.

[00:37:01] Paul Roetzer: So again, as I started off and we didn't even get to the Jensen topic yet.

[00:37:04] Mike Kaput: Yeah, right.

[00:37:05] Paul Roetzer: Some things this week are like wilder than what we would write in Hollywood.

[00:37:12] Mike Kaput: Well, you know, one thing that jumped out of, of that METR, profile that they did, or a recap they did rather, is both METR and Sergei Bry are still talking about humans working with AI to accelerate this. Yeah. So that's kind of like his first step onto the RSI treadmill. It's not even AI itself. improving things autonomously.

[00:37:34] And so in the METR report, they said that when they had talked about this two years of progress in one year happening at Anthropic, they said, this should make you feel old. Basically listening to this, they said this is roughly the same level of progress from Claude 2.1 to Claude Sonnet 4.5, which has already happened in a crazy small window in half.

[00:37:59] The time. So if [00:38:00] you think of even, I don't even think of the last time I said the words, Claude two point, whatever, right? If I even ever have, frankly. So that level of progress in half the time, that's how far they're going, how fast they're going.

[00:38:15] Paul Roetzer: All right. So let's, I just thought of a way to put this in perspective, Mike.

[00:38:18] And this actually is, tees up a talk, a session we, or a segment we have coming up later about, education.

[00:38:24] Mike Kaput:

[00:38:25] Paul Roetzer: Chad, GPT was released when the current college seniors were freshmen.

[00:38:32] Mike Kaput:

[00:38:33] Paul Roetzer: So November, 2022, which would've been four years ago, right? So 23, 24, 25, 26. So if you were a college freshman in November, 2022, you are now a college senior and ChatGPT just got introduced into the world when you were a freshman.

[00:38:53] And so imagine. The level of change we've seen in the blink of an eye. I mean, think if you have kids [00:39:00] in high school or college, you know how fast those four years go?

[00:39:02] Mike Kaput:

[00:39:03] Paul Roetzer: In the matter of time that your child could have gone through high school or college. Same with high school. They were freshmen high school, now they're senior high school.

[00:39:11] All of this from a simple chat bot that you could text in text out to talking about recursive self-improvement, where the models are writing themselves and taking down fields like mathematics and biology four years. And what these labs believe is that level of change will start to happen every like six to 12 months.

[00:39:32] Hmm. Like that amount of change compressed into like months instead of years. Yeah.

[00:39:38] Mike Kaput: And to Pliny's point, it may be happening completely behind closed doors. The things that are possible, we're not even experiencing. We're experiencing barely a fraction of it, and it's still incredible.

[00:39:49] Paul Roetzer: Right. And the only way we might hear about it moving forward is if they've called these tech leaders before the Senate, because they're still gonna keep developing the technology.

[00:39:57] They just won't release it.

[00:39:59] Mike Kaput: Yeah. [00:40:00] Funny to think in that four year period too with those college students, the, how they voted in 2022 versus voting in 2026 now with, some of these issues is wild.

[00:40:10] Paul Roetzer: Yeah. And I guess to tee up the next topic, Mike, what also has happened in the last four years

[00:40:14] Mike Kaput: is this, is, this pulls everything together very well because this episode

[00:40:18] Paul Roetzer: unintentional, by the way, I did not skip any of that.

[00:40:21] Episode 250 of The Artificial Intelligence Show

[00:40:21] Mike Kaput: Yeah. I appreciate this, major tee up here. So our third big topic, Paul, is this episode is a big milestone. It is episode 250 and you know, we could pat ourselves in the back. It, we have plenty to celebrate. It's no small feat getting to 250 episodes of a podcast. So we wanted to maybe take a few minutes to kind of look back at where the show started, comment on a few things related to that and kind of just overall take a quick little victory lap and then move on to weightier topics.

[00:40:52] But. I wanna kind of just really quickly walk us down memory lane, because actually though it is, it's [00:41:00] 250, but four years ago almost exactly to this date is when the show really kicked off in a big way. So, as a reminder for folks, the podcast launched in January, 2021 as the marketing artificial intelligence show.

[00:41:14] The first 19 episodes were in a completely different format. They came out periodically. They involved Paul interviewing guests,

[00:41:21] Paul Roetzer: so they never happened 'cause it was dependent upon me by myself.

[00:41:24] Mike Kaput: there was, you know, you got nine, there were 19 episodes. They were great, but it was not until October, 2022, episode 20.

[00:41:34] We switched to the weekly news format you all know and love. Episode 20 actually aired on October 14th, 2022. It covered, this is just a wild walk down memory lane. These topics, it covered the Biden White House's AI Bill of Rights, which I clicked on the link for and no longer exists. So not surprising.

[00:41:51] It's shocking how to build an AI driven company and meta's early text to video technology. I'm gonna go through a few more episodes because we're hit. We're about [00:42:00] to hit a milestone here. So episode 21 talked about AI's impact on sales funding rounds for Jasper and stability ai, whether AI would replace workers or help them do their jobs.

[00:42:10] Episode 22, Adobe's plans for AI generated images using DALL-E 2 for product design and sequoia's bet Sequoia Capital's bet on generative ai. 23 was Google's treatment of AI generated content, responsible AI guidelines and what AI can mean for local news. Finally, this is the column before the storm, episode 24.

[00:42:31] It was on ai, improving workplace relationships, openAI's funding entrepreneurs, and Google's latest AI research projects. Then we hit episode 25, which aired, oh, it's funny. It's really interesting. Aired on December 8th, 2022. This was the launch of ChatGPT and looking back at the conversation, I'm gonna share some quotes in a bit, Paul, but we were already asking some really interesting questions.

[00:42:55] Just wild to see the, that progression of topics. Also, I thought I was [00:43:00] missing something. Those first episodes were like 30 minutes long. I was like, where are all the topics? There were like 30 to 40 minutes long. I was like, no way. There's literally three to four to five topics per episode, so I'll stop there.

[00:43:12] I'm curious, like what's top of mind for you as you think about where we started, where we're at for episode 250?

[00:43:18] Paul Roetzer: Yeah, I mean, I think it's like, you know, it's interesting a lot of times in business, you know, I was at, I was at Ohio University last week talking with students and we'll get into that a little bit later on the education topic.

[00:43:30] You know, one of the things I stress to them is like, you gotta play the long game and you gotta have conviction about something. And, and so, you know, in the early days of this podcast, I've told this story before, but the only reason we did this initially was so that it forced me and Mike to talk about this once a week and figure out what it means for our business.

[00:43:48] Like we had no idea if anyone would listen. So me and Mike would stand around the office and like we'd share links with each other each week, and then you'd never read 'em, like you wouldn't have time to go through it. We were just going about building the business and then [00:44:00] I went to Mike and I was like, listen man, like we're just, our blog is becoming stale.

[00:44:04] We're just republishing old stuff. There's all this stuff happening every week. You and I know what's going on. We talk about like, why don't we just. Start doing a weekly where you and I just cover like three big things and then we'll like hit some rapid fire items. And that was the origin of the weekly podcast.

[00:44:19] We had no idea how to track performance. We didn't know anything about it. We, we had the podcast set up because I'd done these like, you know, 19 or whatever it was at the time. but it's like, let's just start talking about AI each week and it'll be a forcing function for us. So in 2022, so we started this in October of 22.

[00:44:37] we had 4,800 downloads for the year. So we did 18 episodes, 4,800 downloads,

[00:44:43] Mike Kaput: man.

[00:44:44] Paul Roetzer: 2023, we hit 5000% gross. So we had 22, 260 2000 downloads across 50 episodes. 2024, we hit 400,000 downloads across 51 episodes. [00:45:00] 2025, you start to see the compounding value. We grow 137% in 2025 to 953,000 downloads. We did 60 episodes last year and then this year.

[00:45:11] We're projected to hit about 1.7 million downloads, so it'll be like, you know, 80 ish percent growth. and we've done 56 episodes so far. So in addition to the Weekly, we've obviously been doing the AI transformation series that Mike does. We do the Ask Me Anything sessions. so we, we do AI answers, so we've started mixing some stuff in, but like the Weekly is definitely the core focus of what we've done, and it's just all organic and it's all because the interest in the topic has expanded.

[00:45:37] And, you know, we're, you know, to us it's like an amazing privilege to get to do this every week. Like I've said before, like Mike and I would do this if no one was listening. Like, it's, it's literally just like it's therapy for me to like wrap my brain around it. Get a sense of what's actually going on, figure out the most important things.

[00:45:53] So I can start my week each week with like, okay, I, I know what happened last week. I know what it means to us. You know, we try and share [00:46:00] what it means to all of you. And like, that's it. Like it there, there's not like, some crazy strategy other than just tell the story of AI each week and try and pick the things that are most relevant to people.

[00:46:10] Mm. And somehow that's gotten us into the top 0.5% of all podcasts in the world. Like, it's just, so yeah, I, I don't know, like, it, like I'm really grateful for people to listen. I think it's amazing that when we're in person, you know, we get to hear the stories of our listeners because podcasts, like, you have no idea who's listening.

[00:46:26] You don't get data on them. We don't track anything other than like, you know, downloads and streams and subscribers. I think our YouTube channel has like 35,000 subscribers now. And then I think between Spotify and apple Podcasts we're somewhere around like 55 to 60,000 subscribers just through there.

[00:46:44] I don't even look at that data. Like I, I, it's like every once in a while I'll pop and like, oh, cool, we got, you know, 35,000 subscribers over here.

[00:46:50] Mike Kaput: Yeah.

[00:46:50] Paul Roetzer: So, yeah, we're just grateful to everybody for being a part of it each week and we love hearing the stories. And the one thing I'll ask, I don't really ever ask for anything on this, is leave us a [00:47:00] review.

[00:47:00] Like if you're a regular listener, if even if you just found it for the first time and you find it valuable, that's what keeps the growth going is the reviews on Apple Podcasts and Spotify. so we would really appreciate that if you do have time, you know, drop a, a rating and a review in there and, help other people discover it.

[00:47:16] So that was cool though, the walk down memory lane, Mike was pretty fascinating

[00:47:19] Mike Kaput: the way. Unreal. It's unreal to revisit these, you know, one quick thing I did wanna share, because I'm sure we'll talk about in a whole segment, kind of like the fourth anniversary of ChatGPT, but I want to go back. I, I pulled some quotes.

[00:47:32] I was just like, what did we say

[00:47:34] Paul Roetzer: Yeah,

[00:47:34] Mike Kaput: about ChatGPT on episode 25? And like, I'll be honest, I was like, I'm always, the no Ps like, I'm genuinely always impressed by like the takes and stuff you've got. But like this one I was reading, I was like, god damn, like we predicted the future here because like, I'm gonna read just a couple quick highlights.

[00:47:55] So interesting observations from episode 25. This is where we did an extended discussion of [00:48:00] like, what does ChatGPT mean? What is this thing? These are like paraphrased from some of the quotes. So first, Paul compared the moment to the iPhone releasing and the app ecosystem. Basically you said quote, we're talking about trillions of dollars in value and wealth creation over the next decade.

[00:48:16] The real key to me here is the barriers to entry. Think about the tech. Anyone can try it now. Technical skill, no technical skills required to use it. No code. Very interesting observation that I think that I talked about it as an answer engine and asked what improving answers would mean for people who sell expert knowledge.

[00:48:35] Like it's not just about the writing, it's about like how many careers, jobs, or functions are essentially being a human answer engine. You talked about how schools should respond. You talked about your daughter's songwriting, homework at the time. I didn't revisit the exact scenario, but you were like. If AI can, you're like, I realized ChatGPT could probably do this after we completed this assignment.

[00:48:56] Should I teach her how to use it? She was 10 at the

[00:48:59] Paul Roetzer: time. [00:49:00]

[00:49:00] Mike Kaput: Yeah, exactly right. We both talked about what happens to SEO and Google's model when people can get answers directly instead of navigating search results. And then finally, Paul, I want to actually mention the final thing you said here. You explained in this segment a shift in the show's tone.

[00:49:18] So you said, look, we've been very cautious to not create fear. We've made every effort in our messaging and our content to make AI aspirational, inspirational, drive curiosity, let people explore it. But what I would say now is if you haven't picked up on it, we have an increased sense of urgency. It is very, very important that we solve for this.

[00:49:38] You have to make AI, understanding AI and figuring out how it's going to impact you and your company. A top priority in Q1. Of 2023, there's no waiting anymore. I think that's the biggest change with what's happened in the last few months.

[00:49:53] Paul Roetzer: And here we are again saying the same thing at the start. Yeah.

[00:49:56] It's funny when you look out at, like, if I look back at the things we've gotten [00:50:00] kind of right through the last 15 years or so that I've been in this space, it always just seems so obvious at the time. Like I don't feel like I'm predicting the future in any way when we say the things we say on the podcast because when you just live in in this stuff every day, you can just see around the corner and you all can do it too.

[00:50:16] Like there's nothing uniquely special about the ability to see around the corner if you immerse yourself in a topic. And that's just what we've done like for 15 years. Like I've thought deeply about this stuff. And when you do that. it just seems very obvious and it sometimes it takes, you know, the world or different businesses stuff, time to catch up and see it too.

[00:50:36] But it happens. you know, I, I remember going back to, you know, 2018, 2019, talking to leaders of journalism and communications schools and I'm like, Hey, the AI's gonna be able to write like humans. You should start preparing for this. You should bring in your alumni and have conversation about what this is gonna mean when AI can do this.

[00:50:52] And, you know, they wanted nothing to do with those conversations. Back then in our book in 2022 that came out right before we started this, the weekly [00:51:00] model, there was a literally a section of the chapter says, what happens when AI can write like humans? So it was like, it was obvious then like, but not to everybody 'cause not everybody was paying attention.

[00:51:09] And so, yeah, hopefully, you know, that's what this podcast gives you the power to do is see around the corner. It's what I always tell people, like, your best chance to thrive in your career and to, you know, build a successful business. Is to always see around the corner. Like you gotta get there first. You gotta be ready for what's coming.

[00:51:27] And hopefully that's like each week, you know, we help you do that a little bit.

[00:51:32] Mike Kaput: I love that. Alright, before we dive into rapid fire, one more thing. In addition to the offers that Paul mentioned at the top of the episode, this week's episode is also brought to you by our upcoming webinar, introduction to AI Score for Business Leaders.

[00:51:46] So if you are responsible for any. Facet of AI adoption. The question comes up, how do you know where you actually stand, what your people need, and whether you're making progress. So on Friday, November 13th at [00:52:00] noon Eastern, Paul will be walking through AI score, our new change management platform designed to make AI transformation more measurable and actionable.

[00:52:10] Paul will give you a look at business AI score and personal AI score, which help establish starting points for your organization and its people. Paul will walk through and explain how assessments in AI score can reveal strengths and gaps, guide personalized learning plans, and inform an AI transformation roadmap.

[00:52:28] Paul is also going to explore how to organize employees into learning cohorts, prioritize education and change management efforts and track. Progress over time. So if you lead AI strategy, learning and development, hr, any type of organizational transformation, do not miss this session. If you have not taken AI score yet, no worries, we'll walk you through it.

[00:52:48] This will help you understand how to connect this free assessment with your own work. So register for free for register for free for live or on demand access by going to [00:53:00] SmarterX dot ai. You'll see an option right there on the homepage to register. For that webinar.

[00:53:06] Paul Roetzer: We're hear, hearing you explain that.

[00:53:07] It's like, man, I got a lot of work to do before Friday.

[00:53:09] Mike Kaput: Yeah. I was reading three this. I was like, am I, am I giving you more to do here? This is

[00:53:13] Paul Roetzer: the I know, but seriously, no, it'll be really cool. It's based on

[00:53:15] Mike Kaput: our public description of the webinar,

[00:53:17] Paul Roetzer: so I No, it'll be really good. I'm, yeah, I'm gonna basically walk through, do some demos.

[00:53:21] I may actually show some stuff that's coming too. I, I haven't publicly talked about some of the features we've, we, we built but didn't release. And then we may, we may, I, I have to make some decisions by Friday, but we may show some things that are gonna be coming as well to really, you know, increase the value that this, these tools, you know, give to individuals and organizations.

[00:53:41] So, yeah, I'm, I'm really excited about it. I just got a lot of work to do before Friday.

[00:53:45] Nvidia Opens Probe Into Alleged Spy Operation Targeting Jensen Huang

[00:53:45] Mike Kaput: All right, so let's dive into some rapid fire topics, Paul. So, you know, we've, we've had some pretty positive, for the most part, even the weighty but interesting topics, it would not be a weak in AI without some weird and unfortunately kind of negative stuff going on.

[00:53:58] So this [00:54:00] topic is fast developing Nvidia, CEO, Jensen Huang, as admitted to having a romantic relationship with Alice Huang. Also, same last name, not related,

[00:54:11] Paul Roetzer: but also maybe not her actual name.

[00:54:13] Mike Kaput: Not her actual name, possibly because she is an executive at Singapore based Nvidia, customer Megaspeed. But US intelligence officials also believe.

[00:54:23] That she is a Chinese spy. So NVIDIA's board has opened up an investigation, independent investigation into this incident. The New York Times reports the, this was an alleged intelligence operation in which Jensen was targeted as they were seeking information about American AI chips and potentially Chinese, what you would call influence.

[00:54:44] Paul Roetzer: Call in. What you would call in this world a honeypot where

[00:54:48] Mike Kaput: you a honeypot like a Cold War espionage. Yes, exactly. Yes. You

[00:54:50] Paul Roetzer: attract someone to capture them. Yes.

[00:54:53] Mike Kaput: So this is linked to Chinese intelligence. They're trying to get information about American ai. Chips allegedly potentially [00:55:00] influence NVIDIA's effort to sell advanced chips to China.

[00:55:04] The Times also reports that Secretary of State, Marco Rubio, apparently warned senior Trump officials about this relationship. In a classified memo this spring, Nvidia says, independent directors and outside lawyers will oversee its investigation. The board says in a statement that it currently has no reason to believe national security or NVIDIA secrets were compromised and continues, at least for right now, to back Huang's leadership.

[00:55:29] Huang apologized in a message to the company. He said, I'm just gonna read a very short excerpt. He said, plenty of things in it. I owe my family and all of you an apologies. I was unfaithful to my wife. My actions hurt her and our children. There's no excuse for what I did. I'm deeply sorry. Throughout my life, my family has given me love and support.

[00:55:47] They deserve my honesty and loyalty. Instead, I let them down and damage the life we have built together. I take full responsibility for what I did and the hurt I caused, and I accept the criticism that comes with it. So Paul, I'll be honest, I don't know what my bingo [00:56:00] card looks like in ai. This was not on it.

[00:56:02] Really, like few companies are more important to AI and the AI economy than Nvidia and Jensen is like one of the most influential figures in ai. Like. How bad is this and how is it bad? Is it gonna get

[00:56:18] Paul Roetzer: worse than people can imagine? I would. I, I think so. I'll again, I'll, I'll try and objectively provide some context here.

[00:56:26] I will tell you right now, like, if someone doesn't have a, a, the rights to a book on this in the next like week, I'll be shocked like somebody's writing this book and movie writes to follow because it's gonna get way more bizarre before it gets cleaned up. Hmm. So you mentioned Rubio. How about we go back, Mike to October 9th, 2025?

[00:56:49] this is 13 months ago, the New York Times. On a humid June night last year. Jensen Huang, the chief executive of Nvidia. [00:57:00] Held court with several of his company's major Asian customers at a bar with sweeping views of Taipei. They stood and toasted the booming artificial intelligence industry. Next to the Nvidia Chief was a woman named Alice Huang or Alice Huang, an executive of a Singapore based data center company called Megaspeed, which was poised to buy $2 billion of Nvidia AI technology over the next year.

[00:57:24] Though Ms. Huang and Megaspeed are little known players in the AI industry, their association with NVIDIA and its chief executive has recently become a preoccupation in Washington Commerce. Department Officials have been investigating whether Megaspeed, which has close ties to Chinese tech firms is helping companies in China sidestep American export restrictions according to more than a half a dozen current and former officials and other people familiar with the companies who spoke on the condition of anonymity to discuss an examination that is not public.

[00:57:57] The inquiry, which is active [00:58:00] calls into question how closely NVIDIA is tracking where its AI chips end up and highlights the possibility of American export laws easily being sidestepped. Megaspeed is also facing scrutiny from Singaporean police who told the New York Times in a statement that they're investigating the company for breaching local law.

[00:58:21] okay, we'll come back to that later. Okay. So that was October of 25.

[00:58:26] Mike Kaput: If they were already tracking this kind of relationship, whether it was romantic or not, at the time they were

[00:58:32] Paul Roetzer: entangled, they knew exactly who Megaspeed was and who this lady was. And that was the New York Times with the story, which means people inside the government knew longer.

[00:58:43] here's one I found interesting, and then I'll kind of come back to the pho. And this is gonna end up going a little longer than a normal rapid fire, but I think this is very important. so MG Siegler, who's a, a, a journalist, he tweeted something this morning that like really piqued my interest. He said the whole Jensen [00:59:00] left behind for the big Trump China meeting, only to hop on the plane last minute after everyone pointed out how weird it was, is now about 10,000 times more wild was he left off that trip.

[00:59:13] Due to the Rubio memo, and did the president override it? Keep in mind, the Democrats take power of Congress starting January 1st or sixth or whatever it is. There is no way Rubio isn't like first up, like, what did you know? When did you know it? What is happening? and what we're referring to here is a couple months back, Trump went to, visit Xi in, in China and it was like, where, why isn't Jensen on this flight like the day before, Jensen's not on the list, and then all of a sudden morning of Jensen's, like on the plane with everybody else.

[00:59:49] Okay, so then Semafor, has an article. This is from Sunday. There's little precedent in US corporate history for the crisis facing Nvidia. After reports that [01:00:00] Jensen Huang was targeted by a Chinese honey trap, operation. The national security implications of the allegation, the chip maker's founder CEO's centrality to his company's growth story and NVIDIA's pivotal place in the AI ecosystem all mean.

[01:00:14] The fall it is likely to stretch further than other cases of executive infidelity, is in, has been inseparable from commercial success. Sorry, Jen Jensen has been helming its Purdue product presentations, negotiating the loans to customers that have made Nvidia the Central Bank of AI and deal dealing directly with President Donald Trump and Chinese leader Xi Jinping.

[01:00:36] At a moment when access to the world's most advanced AI chips has been called the de defining strategic question, facing the two superpowers on Sunday, Sequoia's Shaun Maguire, A Huang supporter called him irreplaceable. For America's AI leadership and basically said, we have to forgive this. Like we cannot let Jensen go down for this.

[01:00:56] it sucks that he had an extramarital affair maybe with a Chinese [01:01:00] spy, but like all is forgiven, don't let Jensen go down, is basically what people supporting Jensen are, are saying right now on X. The company can also expect litigation from shareholders and possible regulatory action if Huang or the board held back material information from investors.

[01:01:17] Nvidia previously denied the allegations as recently as like last Friday of an affair to media outlets and Huang is likely to avoid high profile investigations from oh, unlikely to avoid high profile investigations from Congress whose incoming Democratic led majority is suspicious of big tech CEOs and is already preparing a rash of subpoenas for Trump allies.

[01:01:37] Huang, unlike many other Silicon Valley founders, is not in control of his own fate. This is a really interesting and important point. Hmm. He has run Nvidia as if he fully controls it, but he lacks the dual class stock that Shields tech founders like Meta's, mark Zuckerberg and Alphabet co-founders Larry Page and Sergey Brin, while his near 4% stake gave him a net worth of [01:02:00] $155 billion at Friday's close, it does not protect him should the border investors try to oust him.

[01:02:06] For now that outcome looks unlikely. There is no obvious successor waiting in the wings. Lawmakers are fully aware of NVIDIA's strategic significance to the US and Wall Street knows Huang is a brilliant founder who has made fortunes for his shareholders. All I'll end with saying Mike is like. Someone in Washington wanted this out.

[01:02:27] Like if you look at the timing, not only of this, but of some of the other leaks that came out over the weekend. This was a, this was a, coordinated effort by someone in the US government who had this information to put it out to the Wall Street Journal of the New York Times, who had the story simultaneously.

[01:02:44] Jensen has made a lot of enemies while Jensen is like critical to the us the economy, everything, Jensen has been the biggest supporter of chip exports to China, which is not a commonly held position by most [01:03:00] people in power. And he has been the guy pushing for exports to China. He has lost that leverage.

[01:03:07] So like the immediate implication of this, outside of all the obvious stuff, is like the most vocal person pushing for coordination with China. Cannot be in those meetings anymore. So he is going to have to take a backseat, not only for that with China, but as an influential figure in Washington. He's gonna have to get out of the room right now.

[01:03:27] And the downfall of that, like I, so I just, my guess is over the coming months we're gonna hear a whole lot of like, really, really wild stuff because the people who know what's been happening for over the last year may not be his allies. And, and if Jensen has people who are not his allies in the US government right now, including the CIA State Department, it's, it's gonna get real uncomfortable for him very soon.

[01:03:55] Mike Kaput: You know, in the reporting you had mentioned that he doesn't control stock. I found [01:04:00] this interesting as well from one of the New York Times articles. It said, Mr. Huang 63 will continue to lead the company during the investigation. The board said, as a co-founder, he has set up a unique structure at Nvidia where he dictates all aspects of its business while overseeing some 60 people who directly report to him.

[01:04:19] So not exactly someone you could totally rip out tomorrow.

[01:04:22] Paul Roetzer: No, not at

[01:04:23] Mike Kaput: all. The company, I mean, unreal. it would be chaotic at say, yeah,

[01:04:27] Paul Roetzer: just real quick for impact. So their stock started the day down at 8%. So I'm now at 10:15 AM Eastern Time on Monday the ninth. So their stock started down 8% while all their tech stops are tech stocks were up to start the day.

[01:04:41] They're now been wa sitting here watching it as we've been talking, it's now 3.7%. So like initially. Cratered. Now it's gonna come back to to earth. Google's up 3%, meta's up 5%, Amazon's up 5%. SpaceX up 5%. So everybody else is up 5%. They're down [01:05:00] three. Yeah. So just to give you a sense of initial reaction could change by the time you listen to this, this episode.

[01:05:06] AI in Education Is Getting Messier

[01:05:06] Mike Kaput: All right. So next step. There's a few items in AI and education this week that have us thinking about maybe how complicated or how messy things are getting in the space. So none of these are directly related to each other, but we thought these are kind of worth threading together and mentioning. So first, Google is changing its AI policies for schools after months of pressure from parents, educators, and lawmakers about how it has pushed the technology into classrooms.

[01:05:33] So this backlash started to really intensify in August of this year when Google made Gemini available in classroom to students of all ages at schools that had already permit permitted its AI tools. Educators complained they were left scrambling to vet the technology and set rules for it. Last month, a coalition of 13 organizations and experts urge the 150 largest US school districts to disable Gemini on school issued [01:06:00] devices.

[01:06:01] Google is now saying it will turn Gemini and search off by default for school managed K to 12 student accounts until administrators enable them. Schools can then turn those services on separately. But Google still is not offering search without, its built-in AI features. So districts are pushing back against kind of giving, being, having AI being given as a choice, but allowing Google search with AI is mandated or they have to leave search totally unavailable.

[01:06:30] So there's all these kind of complications here, Paul, with Google and how they're rolling AI out in schools. Meanwhile, we saw an interesting essay in education in our own community that was published by the president of St. Edward High School in Lakewood, Ohio, which was titled, what are We Even Doing here?

[01:06:49] And it basically posed this big question about high school education, which is what should a high school education provide as AI becomes more capable? So, [01:07:00] Casey McKenna, who is the president, said, knowledge and deep understanding still matter. But the central question of education becomes what can you do with what you know he goes through in this essay, 11 questions for judging whether students are prepared.

[01:07:14] Things like can they identify worthwhile problems, evaluate evidence, make difficult decisions, create something original, collaborate, teach themselves new skills and learn from failure. But he also asks whether students can decide what is worth doing and discovery purpose beyond personal achievement.

[01:07:31] And he ends by saying, this means we need to rethink homework, grades, testing, and classroom instruction based on what AI is now changing in education. And last but not least, Paul, you just spoke at your alma mater, Ohio University, so you're fresh off speaking directly to an educational institution. I'd love to maybe hear your thoughts on these different threads, how you're thinking about some of the, for lack of a better word, messiness of AI and education right now.

[01:07:58] Paul Roetzer: Yeah, it's interesting. So [01:08:00] Casey's actually a friend of mine. yeah, we've played a lot of basketball together. That's nice. Ironically, we haven't really talked AI together, which is unusual, like normally at basketball, like everybody's asking me AI questions and stuff. But no, I mean, I've known Casey for a long time.

[01:08:11] Our kids actually graduated from grade school together. so I'd say that just as like, I'm not biased in any way about it. It's a really good essay. but I do know him and St. Edward, if I'm familiar, is an all boys school. So just like, as we're saying, like young men and stuff, just know it's, it's an all boys school in, in the Cleveland area.

[01:08:28] So I thought it was really good. I thought it was, a nice reflection of where a lot of educational leaders are at. so he said for generations education has been organized a a relatively straightforward premise. Knowledge is valuable because knowledge is scarce. Schools existed to provide access to information, help students retain that information and measure how successfully they could demonstrate what they had learned.

[01:08:51] But what happens when knowledge is no longer scarce? What happens when artificial intelligence can perform many of the tasks? We have spent decades teaching students to perform. And then towards [01:09:00] the end he said, we should be willing. To challenge our assumptions about homework assessment, grades, standardized testing, classroom instruction, and even the traditional boundaries between academic subjects.

[01:09:10] We should ask whether our students are spending enough time creating, collaborating, experimenting, questioning, and solving meaningful problems. We can help them become young men who are intellectually curious, morally grounded, creatively ambitious, emotionally resilient, and capable of meaningful relationships.

[01:09:26] We can teach them to use technology as a tool rather than allow themselves to become tools of technology. We can challenge them to think independently, act courageously, and recognize their obligations to people around them. Yeah, that was really well said. Yeah, I think it's, You know, again, it's like the very, specific like 11 questions we should be asking.

[01:09:43] I think that's very actionable information and I'm just like, I'm happy to see the conversation moving forward there. The college campus experience. Wow. Okay. So, so that was, it was an amazing opportunity first. there's a wonderful lady, Betsy Lancaster, who, created [01:10:00] an endowment in honor of her husband.

[01:10:01] They had met at Ohio University, and so she created an endowment to bring executive speakers to campus. And so I had the privilege to go back and, and do that as part of like a business school program. And so we had about 150 or so students in the room. I did a, a talk and then we did a q and a that was student led and they did a great job with that.

[01:10:17] A couple of students from the Ohio AI Council. Ohio AI group. And so it was amazing and really inspiring and I had a, a bunch of other separate meetings throughout the day with everybody. That being said, you also get exposed to the other side of the angst around ai. So, like a, a dry erase board was shared with me, I guess like pictures of a dry erase board where they had put somewhere on the college campus.

[01:10:42] I think it was in the library maybe. it just said, what is your approach to ai? How is it different from, for coursework, class research versus personal work? And people would just walk up and write whatever they wanted on these boards. I shared these with Mike. So he's, he's seen this. There are probably some positive statements on those boards, [01:11:00] but they are heavily weighted against, and I'll just read you a few exact quotes from these boards.

[01:11:07] never need it. Never use it under any circumstance. Why would I take a forklift to a gym? Meaning why would I let the, you know, something else do the work for me? another AI is an offense to humanity. Another, I hate ai. Another f ai, another AI magnifies inequality and bias another down with the machine.

[01:11:28] It's stealing our water and making people stupid. So,

[01:11:33] Mike Kaput: so this is all written,

[01:11:34] Paul Roetzer: these are students wrote these

[01:11:35] Mike Kaput: and it's written quite largely often in the image you showed me on a public board where I think they probably didn't know what they were getting into asking opinion. That's a pretty benign question about ai.

[01:11:46] Yes. And it's overwhelmingly negative on that

[01:11:49] Paul Roetzer: 95%. Yeah. Like negative. Yeah. So I, knowing that these feelings exist, Because I even asked like, so the one guy was like the [01:12:00] co-founder or founder of the Ohio AI group and I asked him this, you know, a student, like how many people show up to these meetings each month?

[01:12:06] And he said, oh, you know, like about 20. And I was like, 20. You should, every kid from the business school should be at these meetings every, like, what are we doing here about you? Now you understand why It's like they don't like it. So the thing I said to the students on stage, I did like a state of AI kind of talk and like leading an AI forward future.

[01:12:23] And the thing I tried to. Convey to them is like, listen, I understand it. Like I have tremendous empathy for people who don't like it, who are afraid of it, who fear it for different reasons, who think it does, you know, bad things to the environment. I completely understand all of that, and I, I empathize with all of that.

[01:12:40] I wouldn't be doing my job here today if I didn't tell you that. You have to figure out how to balance those feelings with becoming good at using the technology, because if you don't, you won't have a job in three years. Like there is nowhere you're going to go work when you leave this school in one year, or five years or four [01:13:00] years, whatever, where AI isn't infused into that job.

[01:13:03] And so you have to find a way to understand the technology more deeply. You have to champion the use of it in a responsible way, but you have to, for your own career potential. You have to find ways to embrace it and get really good at it because the best job security you're going to have is being the best at AI in whatever it is you're majoring in.

[01:13:25] Like whatever you come outta school with, be the best in AI edit. So one other quick note on the college campus thing. Mike was this Princeton? So the Daily Princetonian, this is from the editorial board, they said this is November 6th. So just this on Friday, I think. last week Princeton announced a plan to provide paid versions of ChatGPT to all faculty, staff and students who request it.

[01:13:48] However, the university's broad allocation of the more basic plus version lacks pedagogic grounding and takes a fundamental reactive approach to technology with uncertain effects. [01:14:00] We think it is clear that AI is not academically necessary to a large portion of the student body. Hmm. And may have destructive consequences since that distinction largely depends on what students are studying.

[01:14:11] We believe that Princeton should implement a department specific approval process. The research on how AI affects learning is still an emerging field, but there is some evidence to suggest that in a university setting, the use of chatbots can be detrimental to students. Eric Klopfer, an MIT professor, who chaired a committee studying the use of AI at the university, told the Washington Post that AI can lead to quote the illusion of learning through providing universal access to ChatGPT.

[01:14:37] Plus, the university grants the same baseline direction to computer science majors who often require AI for coding as English majors, whose subjective analysis is largely dependent upon their own critical thinking. Although the university is DO does not describe the pilot as an endorsement of ChatGPT.

[01:14:53] Or use of generative AI tools. Indeed, as one of our colleagues pointed out, it seems to run counter to Dean [01:15:00] of the college, Michael Gordin's suggested teaching philosophy, it's choice to provide access to all students incorrectly treats AI as a universal student need.

[01:15:12] Mike Kaput: I got call back to the title of the St.

[01:15:15] Ed's a what are we even doing here?

[01:15:17] Paul Roetzer: Yeah.

[01:15:17] Mike Kaput: Like this is crazy.

[01:15:19] Paul Roetzer: Like you don't have an obligation as a university to prepare them for a future of work. English major, not, I don't care. You don't have, you don't have to use eng, you don't have to use AI right. In your English writing. That's fine. But to pretend like we don't have an obligation to teach students what it is and how to use it responsibly is a, a, a lack of fulfillment of the duties that are expected by the parents and the teacher.

[01:15:40] Like, or the students. Like we, we have to prepare them. I don't care if it's just an Intro to AI Class, like it has to be something we can't just like present it as the evil empire, which is basically how this reads. It's like. Yeah,

[01:15:53] Mike Kaput: I get these, these questions are very challenging and college is a big thing, but I'd be yanking my kid after reading that

[01:15:59] Paul Roetzer: [01:16:00] it's bad.

[01:16:00] Yeah. God, it's not a good look at all.

[01:16:02] Mike Kaput: It's just like, I just to quickly call back though to what you were saying in the episode 25 at the end about ChatGPT, like we, we don't have any time anymore. Yeah. Like, this is not, this is not like, oh, hey guys, let's get our act together. It's like we, there's no time left here to, this feels like this could have been written four years ago.

[01:16:22] Paul Roetzer: Yeah. I mean, go back to the pace of change thing. So there was, in the room I spoke at last week, there were freshmen. So four years from now, we'll be November what, 2030 or whatever. Imagine how much will have changed by the time they are seeing you Were

[01:16:37] Mike Kaput: in college when RSI happened.

[01:16:39] Paul Roetzer: Yeah, yeah.

[01:16:39] Mike Kaput: Right.

[01:16:40] Paul Roetzer: When agents took off.

[01:16:41] Mike Kaput: Right? Yeah. Right, right. Unbelievable to think about. Yeah.

[01:16:46] Enterprise AI Spending Hits $177 Billion

Alright, so our next topic this week, venture firm, Menlo Ventures, released its fourth annual state of enterprise AI report this past week. So this report draws on data they collected via survey in September of this year from nearly 400 [01:17:00] US enterprise decision makers that Menlo says are responsible for more than $4 billion in AI spending.

[01:17:07] So some highlights from this report include, I'm just gonna read off some things that jumped out to me. Menlo found that enterprise AI spending has exploded this year. It reached $177 billion annualized in an annualized pace versus $33 billion last year since late last year's. Agent inflection point, which you talked about on this podcast this year.

[01:17:29] Menlo said that agents quote have now grown to become the biggest consumers of AI models. They account for 80% of all tokens on OpenRouter, which is a popular platform to use a bunch of different AI models. So it's one proxy indicator that's roughly seven x the human driven usage on the platform. Menlo estimates Anthropic, openAI's, and Google Command.

[01:17:51] 69% of all deployment dollars. That's up from them commanding 50% last year. That's 2.3 x The rest [01:18:00] of the market combined. Anthropic and openAI's account for 86% of enterprise L-L-M-A-P-I spend. Again, according to the survey of this cohort, this is up from 67% last year. I found this super interesting Anthropic share grew from 40 to 55%.

[01:18:18] openAI's share grew from 27 to 31%. Google's fell from 21% to 9%, which is rough. I

[01:18:26] Paul Roetzer: mean, Google's models have sucked this year, so this, it'll be interesting to watch the next three months.

[01:18:31] Mike Kaput: Now they have newer models. Yeah. This can change very quickly, I think. Yep. Companies are also spending far more than they planned, so average AI spending reach four points, seven times last year's level.

[01:18:42] They had actually planned on spending reaching 1.9 times. The previous level. So a bit more than expected. Coding and Enterprise Chatbots account for 76% of AI application spending in this group. Finally, 55% of incremental AI spending comes from new budgets [01:19:00] versus 22% redirected from existing software budgets.

[01:19:04] So Menlo, which you know, has many dogs in the fight because they invest in software companies, so take it with a grain of salt. They did find pretty limited evidence. Companies were replacing core systems like Salesforce or Workday, for instance, with ai. So Paul, some pretty interesting data here.

[01:19:21] Obviously this is just one snapshot of a specific segment of companies, but among them, my gosh, like enterprise AI spend is on a tear.

[01:19:30] Paul Roetzer: We talked last week about the self-driving company concept.

[01:19:33] Mike Kaput: Yeah.

[01:19:33] Paul Roetzer: From Amjad at Replit. when you understand the potential of that idea of like agents, you know, doing a lot of the work in humans, kind of orchestrating those agents, you, you can start to see this path.

[01:19:45] Like I think even in the last seven days for me, based on some things maybe we'll touch on the use case spotlight. I think we'll talk about controlling AI costs. I see. As next step. I could see it like, I, I, if, if I had a week right now to just automate all the things that [01:20:00] I now realize I can automate with agents and some of the capabilities we could put in place that we just don't have right now, or there's gaps in human capability within our company.

[01:20:08] Mike Kaput: Yeah.

[01:20:09] Paul Roetzer: I I could see 10 xing our spend in 30 days. Like it's, and they're

[01:20:13] Mike Kaput: being worth it.

[01:20:14] Paul Roetzer: Totally. Yeah. Like if, if I, I, I mean I, these are crazy numbers, but like. 50 x, a hundred x like I, there, there's almost like now that I have a, a vision for how to apply agents across different roles, like a clear vision, there's almost no limitation to what you could apply them to.

[01:20:30] And they're very token, they take a lot of tokens to, to perform their tasks. And it, you can really start to see how Anthropic and others could be spending more on tokens than they are on human capital. Mm. and I'm not saying it's a good thing, honestly, like in some cases, like I'm, there's some things that I'm thinking right now and like seeing what's possible where I'm honestly almost like hesitant to even start talking about it publicly because [01:21:00] it's starting to feel, to me, and I, I don't know if I said this last week or not, Mike, but it's kind of starting to feel to me, like within the AI labs where you have sort of these on guardrail models.

[01:21:10] Hmm. And you have like more powerful versions of something that you're just not releasing publicly. There, there's like, there's a part of me that's starting to like really understand how to build that kind of company, and I'm not so sure I really want to talk about it because I'm not a hundred percent sure how to do it in the most responsible way yet.

[01:21:30] Um hmm. So, I don't know. I think, I think it's kinda like an iterative deployment of like knowledge at the moment and not 'cause I want to like not share it with everybody. I just don't know how to share it responsibly. Like, I, I just, I think it's very possible right now to build a very different kind of company than I thought 30 days ago.

[01:21:49] Mike Kaput: Hmm.

[01:21:50] Paul Roetzer: And like, it kind of like, it's exciting, but I actually also find it kind of personally scary to, to know how to do it. [01:22:00]

[01:22:00] Controlling AI Costs

[01:22:00] Mike Kaput: All right. Our next topic is we have a few stories. Again, we're gonna kind of pull some threads here together about how companies are navigating AI costs. And again, while these aren't each related, I think each of these highlights an interesting wrinkle in how we're all trying to kind of navigate.

[01:22:15] AI budget. So first the information reports. LinkedIn has flattened the cost of running. Its AI powered hiring assistant, which uses large language models to automate candidate search and outreach for recruiters, even as usage of the tool has steadily increased in the past year end quote. So basically, to do this, LinkedIn has replaced models from openAI's with smaller models it trained and developed in-house that have been vastly cheaper than these off the shelf ones.

[01:22:43] So basically subbing in a different model into an existing product has kept costs flat while usage has increased. Second, we've had a post from Zapier, CEO, Wade Foster, talking about how companies can save a ton of money. By using AI only [01:23:00] where it's necessary and automating the rest of a business process with deterministic code or steps.

[01:23:05] He says, as part of this post, we'll link to it in the show notes. Go read the whole thing. He says, quote with Zapier, your agent builds the automation and Zapier runs it. The agent does the part that needs judgment. It reads your request, figures out the steps, and sets it up. Then the automation runs on a trigger like a new email with the same steps every time.

[01:23:24] If a step breaks, Zapier switches back to an agent to figure out the fix. Now, obviously, he's talking his own book here, but he does claim this approach of understanding where agents probab and probabilistic superpowers kind of fit in, and where deterministic code and automation fit in this approach can cut costs by up to 71%.

[01:23:46] Now finally. HubSpot's response to a rough earnings report recently includes that they are pursuing or considering a quote headless strategy that's basically letting customers use ChatGPT or Claude [01:24:00] to actually analyze their HubSpot data, instead of native HubSpot tooling. So HubSpot basically becomes headless as this data repository.

[01:24:09] so AI becomes the way you operate the software, where HubSpot still holds the customer data and powers the work. They're also testing a pricing model that replaces separate subscriptions to different products with a combo of user seats and credits for usage. Credits cover work, like sending marketing emails, resolving tickets, enriching customer records, so the bill can increasingly ref reflect how much work the software performs.

[01:24:33] So we're not gonna get deep into the HubSpot stuff necessarily, Paul, but the point here is like, okay, LinkedIn is cutting costs by tailoring models better for its core use cases. Zapier says you can cut costs using deterministic code versus probabilistic agents. And third, HubSpot may be becoming headless as a response to AI and maybe potentially.

[01:24:55] Charging by how much work gets done in the platform. So like, where is the signal [01:25:00] in this noise? Because like, how am I supposed to be thinking about saving money and not breaking the bank using AI tools?

[01:25:07] Paul Roetzer: Yeah, it's interesting. Like I am, I'm using HubSpot in this way, like in increasingly I don't go into HubSpot.

[01:25:11] Yeah. I, because it's connected to, for us, same

[01:25:14] Mike Kaput: to our

[01:25:15] Paul Roetzer: lms. Like this morning as an example, I'm guess I'm like preempting my use case spotlight. We have these couple of webinars coming up and I was like, I don't, I don't even know how many people are registered. So usually I would either message the team how, what is it looking like?

[01:25:26] How many people are registered? Keep me updated, or I would like, and I get a chance, I would go in and try and find that chart somewhere in HubSpot in the app. Right. And I would get annoyed and like it just wouldn't happen. Yeah. So instead I go to ChatGPT while I'm making my coffee this morning. Hey, at HubSpot, how many people are registered for the November 13th webinar?

[01:25:43] Gives me the thing. Great. Okay. How many about, how about the November 19th gives me the summary. Great. And it links to the list, so I wanna go look at it. I said, you know what, this is perfect. Can you send me, a report each morning of registration totals and a breakdown of like where they're coming from.

[01:25:56] Hmm. So it's great. Scheduled for you took about 30 seconds. Yeah. Not only. Now I have [01:26:00] the data. I just scheduled, an ongoing report and it knew to say, I'll run this through November 19th. Let me know if you want me to extend it beyond that kind of thing. That

[01:26:08] Mike Kaput: like, what else do you need in that scenario?

[01:26:10] What else do you need HubSpot for? Nothing. Nothing. We don't need dashboards. We don't need reports. We don't need Bri, no offense. HubSpot. This is just an example we use, but like, what do you need it for?

[01:26:19] Paul Roetzer: Correct. And then it like, you carry this out. Hey, that's looking a little behind goal. Let's send an email to anybody who took AI score in the last three weeks and like, draft that for me.

[01:26:34] Yeah. Sit there. You're making your coffee some more. Right? Okay. That draft looks really good. Let's get that scheduled and like ping, you know, the marketing team and tell 'em the draft is ready for review. I am like orchestrating

[01:26:45] Mike Kaput: Yep.

[01:26:45] Paul Roetzer: A, a campaign while I'm making my coffee. So when I was saying before, like I just have a different vision right now for what a company looks like.

[01:26:54] Very different from 30 days ago. And when you think about the cost, like now it's this like, okay, [01:27:00] now I get it now. We're gonna build analysts for everything. Hmm. And then we're gonna build producers that like work with our teams to like create everything. What the hell is that gonna cost us? So that leads me to a new role.

[01:27:14] So again, like a positive here is like what roles are gonna exist? AI value engineer, like cost and whatever you wanna call it. We're working on a home project right now, and the project manager for the builder, there's a value engineering component to it. Mm. It's like, how do we, we know what we want to create.

[01:27:29] How do we maximize the value creation of this process? Reduce the cost to get the same output we want to create. So if I think about the future of SmarterX or other companies like us, and I can see this future where all this. Okay, if I use ChatGPT, what would it cost? ChatGPT plus HubSpot. Okay. Now what if I took Zapier and I did all the deterministic steps outside of the LLM for a hundredth of the cost of the LLM.

[01:27:54] That's the value engineer role. Somebody needs to do that. Like, and so I, I mean, and that's like an [01:28:00] immediate thing like that, that role should exist on every, in every company right now. And I don't, I don't know of it. Like I don't, I don't know where that role is, but I might be hiring for that role in the near future.

[01:28:10] Mike Kaput: Yeah. So if you're, I, I'm always curious, like if your HubSpot, for instance, sorry, HubSpot's like catch and strays over here. It's this, you could say it's a venue software company. Like if you're HubSpot. Are you gonna try to start charging me to use my data more aggressively? Or even if you do, like, can I just use browser usage ultimately and tokens go get cheaper over time and like, it doesn't matter.

[01:28:34] Or I have, or just pay openAI's to get enough usage or whatever. Like

[01:28:39] Paul Roetzer: Yeah. So that's an interesting one. The browser usage might be a workaround in the near term, a workaround. So they're, they're absolute absolutely gonna charge you for the headless software that, that you're gonna be accessing HubSpot data in ChatGPT or Claude, like they're going, they're gonna charge you for that.

[01:28:52] They're gonna METR or that, but if you turn a computer use agent loose and HubSpot is on, doesn't know [01:29:00] that a human isn't the one clicking around using it.

[01:29:02] Mike Kaput: Yeah.

[01:29:03] Paul Roetzer: And you're not using HubSpot's AI tool, you're just using HubSpot standard data.

[01:29:08] Mike Kaput: Yeah.

[01:29:08] Paul Roetzer: Then how would they ever know that it's not a human doing it?

[01:29:12] Mike Kaput: I have this question for every software product, by the way, too, like every, every

[01:29:17] Paul Roetzer: software company, right? And like, how would HubSpot know? Geez. Oh man, how so? Okay, so if I have 10, okay, this, hold on a second. I gotta think about this one. All right. Screw it. I'll say, okay, so if I have 10 licenses to HubSpot for 10 different people doing 10 different roles,

[01:29:33] Mike Kaput: yeah.

[01:29:33] Paul Roetzer: But with computer use, I can have a single agent or string of agents that function in those roles, but they're using computer use. So they're logging in under a single credential, but then doing all the work of the 10 people.

[01:29:45] Mike Kaput:

[01:29:45] Paul Roetzer: How do you, how do you charge for that? How do you even know to what? To how do you know what's happening?

[01:29:51] Mike Kaput: Yeah. There may be technical ways to detect if it's like a agent person bot activity maybe. But even then, like I've found that so [01:30:00] far, I can tell you from personal experiences, not reliable, because I've done this plenty in certain experiments to see what's doable with software platforms and websites.

[01:30:10] Like there's no like universal ban that they have figured out to figure out to stop this. Maybe that will change, I don't know. And if it changes, it's like. Okay, we're probably gonna have to evaluate something different because how do you, it's ole game. How do you say we're not gonna use computer usage?

[01:30:27] If your company is allowing it or something and you're just like, wait, we're not gonna take advantage of this enormous tool if you're okay with it because of our software wants to make more money.

[01:30:39] Paul Roetzer: So then you have to charge by outcome or work product, which, which love

[01:30:43] Mike Kaput: it sounds like that's where they're going in some of this commentary, but I don't understand like, are we suddenly then gonna be charged, but based on how many emails we create in HubSpot, it sounds like potentially and okay, but this is

[01:30:56] Paul Roetzer: to variable cost hell,

[01:30:58] Mike Kaput: right.

[01:30:59] Nobody's wanna do that. [01:31:00] Right, right, right.

[01:31:01] Paul Roetzer: And then you actually may ac may just accelerate someone building the alternative themselves. It's like, screw it. Like this is ridiculous. Like it's. Wow.

[01:31:08] Mike Kaput: And, and at, no, this is

[01:31:09] Paul Roetzer: messier than I thought actually.

[01:31:11] Mike Kaput: And at no point is this company, well it may be today, but at no point has this company been invested in or priced as a headless data repository.

[01:31:20] It's something pre ai.

[01:31:22] Paul Roetzer: Right. And we're just talking about hubs, but this is every software.

[01:31:24] Mike Kaput: Yeah. This is not HubSpot, like me coming at HubSpot specific. The

[01:31:27] Paul Roetzer: Salesforce,

[01:31:28] Mike Kaput: yeah. A hundred

[01:31:28] Paul Roetzer: percent Workday. Like

[01:31:30] Mike Kaput: yeah,

[01:31:30] Paul Roetzer: project man. Everything, every SaaS product you use. This goes back to the SaaS pocalypse, which ironically I was looking at like what are the most popular episodes ever on the podcast?

[01:31:38] Yeah. For the 250, if I'm not mistaken. I looked, I did it, the SaaSpocalypse episode. I think it was like 1 91 maybe is the most popular episode we ever did, and that was like February of this year, I think.

[01:31:48] Mike Kaput: Nice. That's funny.

[01:31:49] Paul Roetzer: May have to, may have to do part two on that one now that we're thinking, Hmm. Oh, this

[01:31:54] Mike Kaput: sounds like we'll have more to say on it at some point.

[01:31:57] Yeah. All right.

[01:31:58] AI Use Case Spotlight

Well let's [01:32:00] move on. This actually kind of related, at least on my end here, it sounds like on yours as well, we have our AI use case spotlight segment where every week we give you a quick look under the hood at some real AI use cases we're exploring. So Paul, I'm gonna share one real quick, then I want to hear maybe a little bit more about how you are approaching things.

[01:32:17] So I have talked about in the past, using AI to collect voice memo transcripts. I record a fair amount of voice memos on walks and things. I've talked about building a personal context system that works across different AI tools. So this week I have been building on those ideas by building a more complete, more code heavy to Zapier's point second brain.

[01:32:39] So. Here's basically the overall problem. I do a lot of thinking that happens in different places and mediums. So I record voice memos, I talk through ideas with ChatGPT and Claude, both typing and by voice. I write notes in Google Docs. I save a lot of bookmarks on X of ideas I'm interested in. So I wanted to bring all this together better [01:33:00] so AI could help me search and make use of it.

[01:33:02] So I actually used Codex to build a searchable database on one of my computers. So it uses SQL Light, which is basically just a way to create a structured database in a single file. So each record keeps the original text of a voice memo, a chat where it came from, all the relevant stuff. Basically a big old database.

[01:33:22] And for AI conversations, it also preserves who said what, like me or the ai, which platform I was on. So I started basically with one source. Voice memos were the kind of biggest immediate need. So I'd Codex write code to pull apple's existing transcripts into the database. We set up automatic updates. I recorded a test memo to check that this works.

[01:33:42] So that's all code. None of that has anything to do outside of using Codex to build it with ai. So there's no usage limits, just local code running on your computer. Then I added sources one at a time, X bookmarks ChatGPT, and Claude Conversations, selected personal [01:34:00] writing and Google Docs, even photographs on my phone of handwritten notes that AI helps transcribe.

[01:34:05] So there's some AI getting injected, but still a lot of this is just scheduled. Code. So I wanted all the repeatable collection of this material handled by code wherever possible. I literally screenshotted that Zapier post and dropped it and was like, I wanna do this, basically. so, you know, AI handles things still, like reading handwritten pages and reviewing which parts of a document belong in the database.

[01:34:29] And then finally I can have Codex search that archive pull up relevant source material. I then can use AI on top of this and just point it at this great database that's auto updated and maintained. Say like, Hey, like what are some themes that are coming out recently in my writing? What about that idea I had like two weeks ago, whatever.

[01:34:47] So this is still a work in progress, but it's pretty cool to be able to have. This actual database. There's code in a private GitHub repository now that I finally have, which, you know, years too late now, [01:35:00] finally here, you could certainly point AI at this stuff and just ask questions, right? Like there's no problem trying to do that to start, which is what I've kind of done in the past.

[01:35:09] I just wanted to challenge myself again, kind of like what we've been talking about, to incorporate far more deterministic code into the process for reliability, efficiency, and cost reasons. So I'm not just ripping through usage, trying to point code X on like Astro Max at like my Google Docs, right?

[01:35:26] So super informative, super cool exercise to go through and felt very empowering. It works really well so far.

[01:35:33] Paul Roetzer: That's wild. Yeah, mine was pretty straightforward. I feel like I'd say this every week. Like, you give this like insane, like, wow, okay. I was an advanced masterclass in something and then, and I'm like, yeah, I asked Dots what to do with it.

[01:35:45] no, I actually did. So DOTS is in my personal ChatGPT account and I started playing around with it. The new, a personal agent, if you haven't, you know, been following along with dots, named mine, bluey love, cute little circle. Love it. And so [01:36:00] what I found was like, there was a specific thing I used it for in my, in my personal life actually.

[01:36:04] but then it started like proactively saying, Hey, what can I help you with? Like, you were looking into this last week. Is this, you still, you still need help with this? And you just immediately start to see like, oh my gosh, like once there's proactive personal agents that are allowed to have access to your business data.

[01:36:20] Mike Kaput: Yeah.

[01:36:20] Paul Roetzer: The amount of like, you know, this, imagine like a, I program or create an analyst that goes and deterministically does this thing every day for me. Analyze the segment of people who've registered this webinar, gimme that report, tell me you know, what they're doing. Just kind of like it does a thing I've asked it to do.

[01:36:38] But instead imagine like agents that are just always watching sets of data. Or business units or production, whatever. And they can be proactively engaging with people throughout the day saying, Hey, you want me to look into this, or, I noticed this fell, or like, whatever. It's just like the idea of proactive agents is, is where my brain is like really starting to think very aggressively [01:37:00] about and how much that would change the way jobs function today.

[01:37:05] Mike Kaput: Yeah.

[01:37:06] Paul Roetzer: And that's, it gets into a little bit of the stuff I was saying, like I'm a little actually afraid of the direction because you can really start to see more and more how. A lot of things that people do today won't, won't be needed. and what does that mean? What does that mean for us?

[01:37:21] What does it mean for other companies? And I think this gets to that seeing around the corner thing, like you can't unsee it. And I, I kind of feel like I do today with maybe how I did with Chet in 2022. You just, everything looks different to me right now. Every business, every person, every career path, like, it just looks fundamentally different, based on what's available today and how good it's gonna get over the next like 12 months.

[01:37:47] I, I just can't look at 2027 the same right now. so yeah. And, and it's from innocuous, like use cases. Yeah. Right. It's like I'm trying to find this thing and then you're like, it finds the thing and [01:38:00] you're like, oh. H Okay. Like there's a whole lot of things, like this one little thing I just asked you to do that would fundamentally change how I run my company and how I staff my company and how, what those people do within my company.

[01:38:14] And yeah, just like I said, you can't unsee the future once you see it. And I just, it's, it's running through my brain a lot right now.

[01:38:24] Mike Kaput: Yeah. I know some of the things you're talking about, I suspect others, but it feels to me, I don't think it's crazy to say this is far bigger than ChatGPT

[01:38:33] Paul Roetzer: Oh,

[01:38:33] Mike Kaput: the ChatGPT moment, actually totally in kind of, not, not like one technology per se, but, the implications of, I think what you're talking

[01:38:42] Paul Roetzer: about.

[01:38:42] Well, and I think like for four years we've let individuals figure out AI systems.

[01:38:48] Mike Kaput: Yeah.

[01:38:48] Paul Roetzer: My concern slash excitement is if you centralize this and you just go do this for people. Yep. Then the self-driving companies that Amjad. [01:39:00] Proposed could happen way, way sooner. And I, that's why I'm like, I'm not sure that the economy is ready for, if you have the right people within an organization centralizing this process, things could move really fast.

[01:39:16] AI Product and Funding Updates

[01:39:16] Mike Kaput: Yeah. Alright. Our final segment this week is our AI product and funding update. So I'm gonna go through very quickly a handful of updates we've got here and we'll wrap up episode 250. So first up, thinking Machines Lab. The AI startup led by former openAI's CTO Mira Murati is nearing a $5 billion round led by Accel at 40, a $40 billion valuation.

[01:39:38] According to Business Insider, openAI's has a few updates. First, they launched images 2.6 across all ChatGPT and Codex subscription tiers in its API, adding animated gifs, more consistent image at its and generation, the company says is 50% faster. openAI's says its experimental AI security pipeline has [01:40:00] developed exploits and candidate fixes for hundreds of unique Linux flaws that could let a local user gain administrator access.

[01:40:07] So basically, they're using AI to fix cybersecurity issues. This cost, this particular project cost about $1.2 million in API usage. openAI's also introduced something called LifeVal LifeVal Benchmark Measuring how well AI completes everyday tasks, such as budgeting, shopping, and finding childcare built, using time, use studies, and a survey of nearly 4,000 people.

[01:40:33] Anthropic began rolling out personal finance in Claude in beta for U.S., pro Max and Ultra subscribers. This provides read-only access to connect bank and card accounts so users can analyze spending flag, unusual changes, and build financial dashboards. Apple has opened Siri AI to all developers in its latest operating system, betas, allowing apps to expose actions such as booking reservations, and combine multiple [01:41:00] actions into skills.

[01:41:01] This is ahead of a planned later customer rollout.

[01:41:04] Paul Roetzer: I have been testing the new Siri AI on my phone. Yes, it has gotten smarter.

[01:41:08] Mike Kaput: Great.

[01:41:08] Paul Roetzer: it is not computer use yet, but that is also where they're going. There's an article that just came out about that we'll talk about it next week, maybe more in detail.

[01:41:17] They know they're falling behind on personal agents and it sounds like the CEO has a major focus on that and you can start to see it in sir now.

[01:41:25] Mike Kaput: Cool. That's exciting. And then finally, replit and cognition, published security guidance recommending controlled cloud environments for AI agents, permissions limited to each task and removing production credentials from developer's laptops.

[01:41:40] So like everything in here, we will link this, but basically they have pushed stronger agent security and are offering some guidance around that from some companies on the frontier of all this. So. One final reminder, go to SmarterX dot ai slash pulse, take this week's survey. We would love to hear from you.

[01:41:58] And with that, [01:42:00] Paul, thank you. Happy episode 250.

[01:42:02] Paul Roetzer: Yeah, you too man. Great work and keep, keep going. We have another 250 in the seven least I think I

[01:42:07] Mike Kaput: right? I think so,

[01:42:08] Paul Roetzer: yeah. And that again, academy, do SmarterX dot ai use that POD250. I think a lot of the things we talked about today, Mike, is sort of where I envision Academy going.

[01:42:17] Yeah. It's like this AGI agentic frontier and we gotta figure out what this looks like together. which means we kind of, kind of probably build it together, and try and figure out how to do it in the most responsible way possible. So yeah, I, if you wanna be a part of all that, join now is a great time to join AI Academy.

[01:42:33] Alright, thanks everyone for being with us. We really appreciate, it's a privilege to get to do this with you all each week and, Yeah, we'll see what next week brings us. It's gonna be an interesting one. Thanks, Mike.

[01:42:45] Mike Kaput: Thanks Paul.

[01:42:46] Paul Roetzer: Thanks for listening to the Artificial Intelligence show. Visit SmarterX.AI to continue on your AI learning journey and join more than 100,000 professionals and business leaders who have subscribed to our weekly newsletters, downloaded AI blueprints, attended virtual and in-person events, taken online AI courses and earned professional certificates from our AI academy, and engaged in a SmarterX slack community. Until next time, stay curious and explore ai.
Claire Prudhomme

Claire Prudhomme is the Marketing Manager of Media and Content at the Marketing AI Institute. With a background in content marketing, video production and a deep interest in AI public policy, Claire brings a broad skill set to her role. Claire combines her skills, passion for storytelling, and dedication to lifelong learning to drive the Marketing AI Institute's mission forward.

Recent Posts

[The AI Show Episode 250]: Midterm Voters Reject Data Centers, AI Speeds Up Anthropic’s Research, Celebrating 250 & Nvidia’s Spy Probe

Claire Prudhomme | November 10, 2026

Paul Roetzer and Mike Kaput unpack the data center backlash, METR's Anthropic report, and the Jensen Huang probe.

Podcasts

[The AI Show Episode 249]: Anthropic Files for IPO, The Self-Driving Company, AI and the Meaning of Work & Microsoft’s 60,000-Employee Buyout Offer

Claire Prudhomme | November 3, 2026

Anthropic's S-1, Microsoft's 60,000-employee buyout offer, and Amazon's tokenmaxxing memo show how AI is reshaping work.

Podcasts

[The AI Show Episode 248]: The AI Cybersecurity Crisis, Dario Fights Back on Safety, OpenAI’s Monitoring Blind Spots & Claude Duet

Claire Prudhomme | October 27, 2026

Paul Roetzer and Mike Kaput break down the AI cybersecurity crisis, Anthropic's safety defense, and its $2T IPO filing.

Podcasts