66 Min Read

[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

By Claire Prudhomme on November 3, 2026

Podcasts

Anthropic's S-1 is public, and Paul Roetzer thinks there's a very strong probability the company comes to market worth more than Meta.

Roetzer and co-host Mike Kaput unpack the filing: $16.3 billion in first-half revenue, hundreds of billions in compute commitments, more than 90 pages of risk factors, and a vision of "virtual collaborators" working alongside knowledge workers.

They also critique Replit's prediction that every Fortune 500 company will be self-driving by 2030. They dig into Google DeepMind's new research on AGI, wages, and the meaning of work. And, in rapid fire, they cover Gemini 4 Carbon, Big Tech earnings, Microsoft's 60,000-employee buyout offer, and a run of AI security incidents.

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

00:06:57 — Anthropic’s IPO Filing

00:27:42 — The Rise of the Self-Driving Company?

00:41:35 — Google DeepMind and OpenAI Say AI Could Transform the Meaning of Work

00:56:12 — Google Unveils Gemini 4 Carbon and Genie 4 as OpenAI Releases GPT-6.1 Terra

01:01:44 — Google, Microsoft, and Amazon Report Earnings

01:06:13 — Microsoft Offers Buyouts to 60,000 Employees

Microsoft offers buyouts to 60,000 employees - The Verge

01:12:15 — Amazon Refocuses Its AI Strategy on Frontier Models and Business Results

01:16:49 — AI Is Forcing a New Security Agenda

01:21:47 — AI Use Case Spotlight

01:28:26 — AI Product and Funding Updates

OpenAI Updates the Model Spec for More Capable AI

Mercor Offers Businesses Payment for Workflow Data

How do you know where your organization actually stands with AI?

Join Paul Roetzer for Introduction to AI Score for Business Leaders on Friday, November 13 at noon ET. He’ll walk through AI Score by SmarterX, our new change management assessment built to make AI transformation more measurable and actionable. You’ll see how the free Business AI Score and Personal AI Score assessments reveal strengths and gaps, guide personalized learning plans, and shape your AI transformation roadmap. Register free for live and on-demand access at SmarterX.ai

This episode is also brought to you by the MAICON 2026 On-Demand Bundle. Couldn’t make it to Cleveland, or want to rewatch?

The MAICON 2026 On-Demand bundle gives you more than two dozen recorded main-stage and featured sessions (12+ hours) from this year’s Marketing AI Conference, ready to watch on your own schedule. Grab it at MAICON.AI

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: So there is a very, very strong probability that Anthropic will be worth more than meta. When they IPO. So that's probably the best way to think about this in terms of how much is this? It's like Meta is one of the most successful, profitable companies in the world. Yeah. With their ad business. And Anthropic could come to market as a five-year-old company and be worth more than them.

[00:00:24] Welcome to the artificial intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable. 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.

[00:00:44] 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.[00:01:00]

[00:01:00] Welcome to episode 249 of the Artificial Intelligence Show. I'm your host, Paul Roetzer, along with my co-host Mike Kaput. We are recording on Monday, November 3rd, 9:00 AM I am happy to say like it is election day in the United States if you're listening to this on the third, and I'm happy to say I don't think we have any political stuff to talk about today in the episode, if I'm not mistaken, Mike.

[00:01:21] Mike Kaput: Incredible.

[00:01:22] Paul Roetzer: Okay, so that's it. The\ midterms are happening, on November 3rd, and I'm sure we will have some, what does this mean to the AI industry on, on next episode, but you are political information free on today's episode. So that is a joy for all of us to hear. We do have a lot of stuff. I don't know what happened this week, but like.

[00:01:45] A lot of stuff related to the economy and jobs. Like I, I think our first three main topics, I'm not mistaken, I'll either directly address that or Yeah. and certainly in indirectly. Yeah. but lots, lots to get into some cool stuff, some good [00:02:00] research. big IPO forthcoming for Anthropic. So, I don't know, this is one of those, like, I actually, like Friday I was like, Hey, this could be a, this could be a fun episode.

[00:02:09] Like, there's nothing like too crazy happening. So, I was just looking forward to this one. Same. Alright, good. All right. So, this episode is brought to us by our upcoming webinar, introduction to AI Score for business Leaders. You don't wanna miss this one. This is gonna be pretty cool. I'm actually, probably, I gotta travel the next two days and then I'm gonna be working on this, most of the end of the week and into the weekend.

[00:02:35] And then into next week. But, I'm gonna be spending a lot of time building this. This is a completely original presentation. It's all related to the new AI score platform we just, released into public beta. So if you're responsible for AI adoption in your organization or trying to move it forward in your organization in some way, how do you know where the organization stands?

[00:02:57] How do you measure that? And then what needs to [00:03:00] happen? And if you're making progress, these are the things we've tried to solve for. So Friday, November 13th at noon Eastern time. I'll be walking through AI score by SmarterX, our new change management platform that is designed to make AI transformation more measurable and actionable.

[00:03:17] You'll get a look at business AI score and personal AI score. So there's two assessments that are part of this platform. they establish a starting point for your organization and for your people or for yourself if you wanna take this individually. The personalized score, I'll explain how the assessments can reveal strengths and gaps, guide personalized learning plans, and inform an AI transformation roadmap.

[00:03:38] I'll also 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 or organizational transformation, this session is going to help you understand how to connect assessment.

[00:03:54] with your next steps. Register for free, for the live and on-demand [00:04:00] access is gonna go to SmarterX dot ai. and then you'll see on the option on the right, on the homepage to register for introduction to AI score for business leaders. There's also, you can go to the webinar section on the resources area and it'll take, is that where we have it?

[00:04:13] It's under education is webinars, is that where Yeah,

[00:04:15] Mike Kaput: yeah,

[00:04:16] Paul Roetzer: yeah. Okay. Now there's, you can see all of our past webinars and everything that's coming up. So SmarterX dot AI get registered for introduction to AI score for business leaders. Again, happening November 13th. There's a bunch of stuff like, we've only done a very soft public beta rollout of this product, so I'm gonna, for the first time sort of demonstrate the vision behind it.

[00:04:35] I may actually. Show some features and capabilities we did not release in this first version that we, we sort of have, that I'm pretty excited about. So, yeah, I, I'm not sure exactly how much I'll share, but what we're gonna basically do is go through with a focus on both of the assessments and then teach you how to use them.

[00:04:55] They're both free tools, so you could start using them to drive transformation for yourself, for your [00:05:00] business. So, definitely check that out again, Friday, November 13th, SmarterX dot ai to learn more. Alright, Mike, AI pulse survey. This is our informal poll we do each week with our listeners. We asked on the most recent one, when AI reduces the need for a particular role at a company.

[00:05:17] What should leaders do first? okay, so 42% said it depends on the role. Yeah. 39% retrain people and move them into new roles. 10%, redeploy them into roles that are already open and 10% let the role, shrink through attrition in hiring freezes and then reduce head count. Looks like no one said that, or if it did, it was a very, very small sliver.

[00:05:44] Mike Kaput: Yeah, yeah.

[00:05:45] Paul Roetzer: Okay. So we're, yeah, it's a mix of like, depends on the role. So it's unknown, which I, I, I don't know. I kind of agree with, it's like I, it's hard to tell. this survey is going to be very relevant to a couple of the topics we have today. 'cause there's just a lot, [00:06:00] like, as I alluded to earlier, there's a lot coming out right now that's trying to understand what was the impact of AI in the, we'll call it the assistant age.

[00:06:09] Mm-hmm. So for like the four years where AI largely functioned as an assistant, where it was all human led and now we're entering the AGI agentic era where it is possible for it to be. human in the loop or not at all. And that's changing the dynamics of how people really think about this. And I think that's kind of like a central theme of today's episode, Mike?

[00:06:32] Mike Kaput: Yeah. All right. We'll, we'll go ahead. We'll have a new, we'll have a new pulse up as well by the time you listen to this, and we'd love to get your perspective on that. it'll be about how ai, if at all, is expanding your role into work that typically belongs to other roles or adjacent roles. So super excited to hear results on that.

[00:06:52] Paul Roetzer: And that's SmarterX dot ai slash pulse. Take part in that.

[00:06:57] Anthropic’s IPO Filing

[00:06:57] Mike Kaput: All right, Paul, so let's dive into [00:07:00] this week's topics. First up, Anthropic's S one. So last week we had discussed a draft of Anthropic's IPO prospectus that was like obtained early by the Wall Street Journal this past week. The company publicly filed.

[00:07:15] Its IPO paperwork, which is known as an S one with the SEC. So a few details that are worth noting. This document is dozens of pages listing out. The specifics of their business. So Anthropic has applied to list on the Nasdaq under the ticker ant a NT. It has not yet disclosed a share price, how many shares it will sell, or how much it expects to raise.

[00:07:37] The filing shows how quickly this business has grown. Revenue was $16.3 billion in the first half of 2026. That's compared to $1.2 billion during the same period last year. So from 1.2 to 16.3 billion. Now this filing reports this massive $80.7 [00:08:00] billion net loss for the first half of the year. But like we had.

[00:08:03] Referenced on a previous segment about 72.4 billion of that was a non-cash accounting charge. This was driven mainly by an increase in the value of convertible notes as Anthropic's valuation Roetzer. So basically these are notes that are just loans that can turn into shares. So as Anthropic's valuation has risen, the estimated value of those obligations has increased.

[00:08:25] The company recorded that change as a loss. This is not them spending $72 billion in cash to run the business. In fact, Anthropic recorded a $974 million operating profit in the second quarter, but for the first half of the year as a whole, it reported a $693 million operating loss. They reported $8.9 billion spent on compute and infrastructure expenses in the first half of the year.

[00:08:53] Nearly half the company's revenue in this period came from customers buying cloud through cloud [00:09:00] marketplaces. So some of those companies, like Amazon's of the world also supply Anthropic's computing capacity and compete with it in ai. That also includes Google. Microsoft, et cetera. it's worth noting in this S one Paul that Anthropic's future compute commitments are insane.

[00:09:17] It has agreed to spend at least $114.5 billion with Google through 2033. $110 billion with Amazon through 20 36, 44 0.9 billion with Microsoft through 2033, and those minimum spends apply apparently regardless of usage, though they are subject to limited exceptions. Now, it also has about $161.2 billion in payments, excluding interest under a five year equipment lease with Broadcom.

[00:09:49] This is for AI compute hardware. There's $120.5 billion with SpaceX AI for computing capacity through 2029, though that's a more flexible agreement, it sounds [00:10:00] like after the initial lock in periods ending in January and March, 2027, either of those parties can cancel with 90 days notice. So the idea here, Paul, is that actually once you take away that 72 billion or so with the convertible notes, this business is.

[00:10:15] Not profitable yet, but had a profitable second quarter, but has this looming compute spend, like if you wanna bet on something, bet on, I guess the cloud providers.

[00:10:27] Paul Roetzer: Yeah. It's, it's hard to comprehend those numbers honestly, because you, it's like, it's unreal. Yeah. You're going through, like, there's an all of a sudden, like 8 billion sounds small.

[00:10:35] It's like, oh, okay. Well that's, yeah. It is hard to put in context mentally, like the scope of what we're talking about, how big these commitments really are to play down on the stream. Like what if they can't fulfill the commitments kind of thing. Mm. the, like, one helpful way I think about this is like, how big, when they go through with this IPO in a couple weeks, how, how big are they relative [00:11:00] to the other largest companies in the world?

[00:11:02] So obvi, you know, the primary measurement we would look at here would be the market cap. so Apple, as of. Morning of November 2nd is the most valuable company in the world at just over $5 trillion. Nvidia, they, they trade places each day, depending on certain things, but Nvidia is at 4.9 trillion. Then you have Microsoft down at 3.9 trillion.

[00:11:24] Then alphabet slash Google 3.4, Amazon 2.6 TSMC, which is, you know, basically is the supply chain for the chips, for all of this is at 2.1 trillion. Then we have at number seven is SpaceX at 1.8 trillion. So the current. Word on the street is that the IPO will value them somewhere between 1.5 and $2 trillion.

[00:11:50] So they are potentially in the top 10 most valuable companies in the world, somewhere around that TSMC to SpaceX [00:12:00] range. Likely number eight is Saudi Aramco at 1.6 you have Tesla at 1.6, meta at 1.5. So there is a very, very strong probability that Anthropic will be worth more than meta. When they IPO. So that's probably like the best way to think about this in terms of how much is this?

[00:12:22] It's like meta is one of the most like successful, profitable companies in the world. Yeah. With their ad business. And Anthropic could come to market as a five-year-old company and be worth more than them. Some of the other ones, just for relative perspective, Samsung is worth 1 trillion. Walmart is at number 16, worth less than a trillion.

[00:12:40] So again, just gives you a little perspective as to how much we're talking about here. okay. So what did we learn from this S one? It is a, it's a big document. It's actually pretty readable though. Like it's a, yeah, it's good. If you want to understand just their overall philosophy, what they think their market factors are, they're gonna influence them, where they're gonna grow, things like that.

[00:12:59] It gets a little [00:13:00] technical. You know, a little deep financially, but like overall, you, you can scan this thing you throw into the, you know, Google notebook and like, you know, have it analyze it for you. one of the things interesting it does say, so this is from the information, one of the takeaways is that it looks like week of November 16th.

[00:13:16] So if people are trying to figure out, you know, when this is happening right now, it's looking like about two weeks from now, that can change, obviously it could change the day before, but like that's roughly what we're looking at. like I mentioned about 1.5 trillion is what they're expected now.

[00:13:31] That's just what they come to market at. It could go up 15% that day and we could skyrocket.

[00:13:36] Mike Kaput: Right?

[00:13:36] Paul Roetzer: or it could go down like just not giving investing advice. there was over 90 pages of risk factors. Yes. So like usually when you IPO, it's like, oh, this could happen, this could happen. You know, that would be bad for us.

[00:13:50] They have 90 pages of that. Google and Amazon are among the biggest shareholders. We knew that. they do not disclose their [00:14:00] current ownership stake, but it does say that Amazon can't hold more than 12% of Anthropic stock, and Google can't hold more than 14.99. I believe both of them were somewhere around those numbers.

[00:14:13] If I'm not mistaken. When we last heard it was like Google owned somewhere around between 12 and 14%, and I think Amazon owned somewhere between 10 and 12% roughly of the company. Bloomberg had a good summary of what are some of the risks, so they do get into the catastrophic harms. So they said, Anthropic highlighted the potential of its AI to behave in unintended, misaligned, or autonomous ways.

[00:14:36] Companies said that if its technology is not developed with adequate safeguards and oversight, it could threaten humanity. Particularly if su such systems were misused, replicated, weaponized, or allowed to operate with a high degree of autonomy. Again, things we've been talking a lot about, insurance problems I thought was a really interesting one.

[00:14:53] They said to complicate matters, Anthropic warrants that existing insurance policies may prove is quote may [00:15:00] prove insufficient to help AI firms insulate themselves from legal liabilities that arise from their technologies causing real world harm, especially catastrophic AI related incidents, computing supply.

[00:15:12] So another major risk they flagged is the potential inability to secure enough computing power, which seems crazy, Mike, given the obligations they have through 2033.

[00:15:22] Mike Kaput: Yeah.

[00:15:22] Paul Roetzer: But when they buy this compute, you're basically hedging against the next two years. And I have now listened in the last seven days to interviews with Mark Zuckerberg and Sam Altman.

[00:15:32] Who both said they didn't buy enough compute in 2024. Wow. So they all got reamed in 2024 for over betting on the CapEx spend to buy that compute. And both of them, if they could do it again, would buy more than they bought. Hmm. And so that's like this interesting dynamic and Dario has been on the record saying like, listen, this is maybe the biggest issue.

[00:15:51] We never buy enough, but if we ever did, we're bankrupt, basically. Yeah. So they're constantly playing this game of like, do they have enough compute for [00:16:00] what's gonna happen? And to do that, they basically have to now make bets on. The adoption and enterprise, the two, two main factors I think about adoption of AI and enterprises.

[00:16:08] 'cause I still believe it's extremely low and how widespread the use of agents are going to be. Yeah. Because those are like two major factors in how much compute is gonna be needed. So you're trying to like make these bets that like, and it's tough. couple other government scrutiny. Obviously the government has had its issues with Anthropic and that could get worse.

[00:16:27] Data center opposition. So societal pushback to the data centers that are needed. Hackers, which we've heard a lot about recently. TSMC and the chip supply chain. So that's the company I just referenced. Like there could be a breakdown in the chip supply such as like there could be a war between China and Taiwan like that.

[00:16:43] Things like that would've a massive impact. company culture I thought was an interesting one. So what they're basically saying is we have had people for the first five years who were all in on our mission. Now we're gonna run into an issue where to hire enough people, like can we [00:17:00] scale that same thing?

[00:17:00] Are we gonna have people who are just kind of coming in for the money? The other thing is, once the hold period ends, you're gonna have a whole bunch of people, like hundreds if not thousands of employees at Anthropic, who are now worth tens of millions of dollars Yeah. And have no need to continue working at Anthropic.

[00:17:17] And so when they can start selling some of their shares. You're gonna have people just go and do their own startups, maybe compete with Anthropic, like, I don't know, but like that becomes an interesting one. And then when you actually dig into the S one itself, I'll just highlight a couple of the key themes because I think it's important how they view we, we actually had this topic last week, Mike, about what is Anthropic, like, how do they view themselves?

[00:17:40] You get a little bit of insight in how they sort of write the introduction to this thing. So it says, for the first time in history, intelligence is something humans can build. The cognitive capability to reason, analyze, synthesize, decide, and create is producible at scale and accessible on demand. We believe this abundance of AI will rapidly transform [00:18:00] every sector of the global economy in ways greater than the impact of industrialization, electricity, semiconductors, and the internet.

[00:18:08] AI that delivers on its potential with the risks well managed, could significantly accelerate. Global growth and dramatically improved quality of life for people around the world. Then they had something they call virtual collaborator, which is the first I've heard this term. Hmm. it's basically like a AI coworker says, we believe we have reached an inflection point where AI systems can function as genuine collaborators and professional work.

[00:18:30] They quickly complete substantive tasks that might otherwise occupy a skilled professional for several, several hours, a full day or longer. AI systems also become increasingly capable of performing tasks that were previously too difficult or costly to complete, or surfacing original insights and analyses.

[00:18:47] We refer to these systems that work alongside humans to perform complex multi-step work at professional levels as virtual collaborators. Proper noun, capital VC, virtual collaborators carry the [00:19:00] access rights, and this is important what they call this. So virtual collaborators. So again, think of like digital coworkers, AI coworkers.

[00:19:06] They carry access rights. System connections and context that their organizations grant them, which allows them to function more accurately and reliably. So I'll just pause there for a second, Mike, because this gets into like the future. Anthropic envisions to justify this multi-trillion dollar valuation is that their technology when given access to the right systems, the right data, when they're connected to the systems of record, like your CRM system, your knowledge base, your financial systems, and when they have the context of what happens in the organization, meaning how decisions are made, who the key people are, what are the customers like, everything it needs to know.

[00:19:47] Then they can start to function very accurately and reliably. And so like I don't, Mike, if you had any thoughts on that, but like that to me is like a really important thing for people to understand where this is all going.

[00:19:57] Mike Kaput: Yeah, that's super interesting. And also an [00:20:00] interesting note here, which is quite obvious, but might be useful for people, especially if you're someone that is on typically the subscription plan, like 80% of their revenue based on the S one is consumption based.

[00:20:12] So this is all just a bet that people are going to use more and more and more and more tokens Exactly. In the way you described, right? Yep. By millions of different virtual collaborators spinning up. That's, I think, the bet and where they're putting their money, where their mouth is in hundreds of billions in compute commitments.

[00:20:28] Paul Roetzer: Yeah, and I'll, I'll actually address an issue a little later on as to why adoption is so low on these and why they need this to change. they then got in right after virtual collaborators into Claude Cowork. So all digital work is ultimately an expression of code they say and Claude's mastery of its powers, mastery of it.

[00:20:45] The code powers virtual collaborators capable of handling complex work far beyond software development. So, again, keep in mind the vast majority of Anthropic's growth and the driving of the use of the tokens and things that they have [00:21:00] to, like conti consistently limit is largely for. Research and development of software.

[00:21:06] Mm-hmm. Like we haven't really crossed over to where there's massive demand on marketing and sales and operations and HR and finance and legal and all the other domains. There's usage, but nowhere near the level is there that there is for coding and software development. So Claude Cowork extends that same AGI agentic architecture of Claude Code to knowledge workers more broadly by translating a user's intent into completed work autonomously, reading files, navigating applications, and executing multi-step workflows on a user's desktop desktop while the underlying code is abstracted away.

[00:21:39] By extending Beyond Claude Code with applications purpose-built for specific professions and industry verticals, we accelerate Claude's diffusion into new categories of knowledge, work and deepen our presence across the enterprise. So again, I'll pause here. What this means is for you, the the listener, to use Claude to create a [00:22:00] presentation or to write a report, like to them build an application, a prototype of an application, it's coding.

[00:22:06] Like coding is what allows that. But you, the average user aren't installing a coding terminal on your computer. What you need to be able to do is go into Claude and unbeknownst to you, ask it to do something and it is coding. In the background, you never know what happens. So what they're saying is they're gonna take the power of this coding and they're gonna put it into every assistant and agent that you use, and you will now have the power to do all these things that coding enables without ever having to write a line of code.

[00:22:38] Mike Kaput: Wow. Yeah. That which is, which is kind of where we were about this time last year as Claude Code was exploding. A lot of that was not just Claude Code getting better at developers getting on board, that was also knowledge workers, non-engineers saying, wait a second, you can use these harnesses for way more than just coding.

[00:22:57] And that's been the story of the last [00:23:00] year.

[00:23:00] Paul Roetzer: Yep. So they continued. We are now building towards a world where Claude can be given open-ended tasks, again, across any knowledge, work lasting hours, days, or weeks, and then execute them autonomously with the judgment and initiative of a brilliant colleague. So now they're even saying judgment, which is largely the human domain still.

[00:23:17] They're basically saying you're gonna have these virtual collaborators who are as smart as your smartest colleagues, and it's gonna have judgment and initiative to go proactively do things. Ultimately, millions of virtual collaborators can run in parallel, each working independently or on rela, on re on, re, I'm sorry, each working independently on, on related problems or collaborating seamlessly to achieve a shared objective.

[00:23:43] The result is something with his without historical precedent, what we refer to as a country of geniuses in a data center available to every organization and sector of the economy. Then they give some context as to how early we are. So of all, the growth that they've been having more than about just are under [00:24:00] 12,000, customers have a hundred thousand or more in revenue that they've spent with them.

[00:24:04] So they don't, they don't have that many customers spending more than a a hundred thousand dollars annually. They have more than 1900, they say. So just under 2000 customers who spend a million or more annually, again. Enterprise software enterprise services, like a million's, not that much, but they only have about 2000 customers there.

[00:24:23] And then they have about 250 customers spending 10 million or more. they get into their internal usage, which I thought was fascinating. They said that the value of our internal Claude usage at list prices has grown by over 250 times this year from January to September of this year, and in April, 2026, the monthly value is first exceeded our monthly personnel expense, meaning they're spending more, if they were charging themselves what they charge the public.

[00:24:51] They're spending more on their use of Claude than they are on personnel. Which is crazy.

[00:24:57] Mike Kaput: Yeah.

[00:24:57] Paul Roetzer: They also have revenue per [00:25:00] employee exceeding 6.6 million. So every revenue is generating over $6 million in, every employee is generating over $6.6 million in annual revenue. just interesting facts. Dario makes about 700,000 a year payroll wise.

[00:25:16] They had that in there. they do get into, like AI is transforming intelligence, talks about like powerful AI and kind of what that means. and then they had a, a near term opportunity to transform cognitive work. I thought this is kinda like an interesting way and then kind of wrap up with these thoughts.

[00:25:31] So they said Frontier AI is transforming cognitive work, the application of knowledge, reasoning, and judgment to produce economic value. We believe it will make humans far more capable and expand the world's overall capacity for that work, including accelerating scientific discovery and helping businesses grow.

[00:25:46] Knowledge workers are the leading edge of this opportunity. As ai, as frontier AI advances more capable models can unlock entirely new use cases. While also efficiently serving a wide range of cognitive work at scale across the [00:26:00] capability, cost speed curve. We believe the work that knowledge workers perform represented more than 40 trillion in economic value globally in 2025.

[00:26:10] So there's their market potential. So knowledge workers generate 40 trillion in economic value and with Frontier AI's potential to make them meaningfully more productive, our near term opportunities to expand this value and accelerate economic growth. then related to this whole like lack of adoption in companies, they said in October, 2026, they launched the Claude Frontier Academy backed by a hundred million dollars commitment.

[00:26:35] To address a key constraint on enterprise AI adoption, the shortage of engineers who can take AI from an initial use case into production. So basically they're trying to train 10,000 engineers by the end of 2027, who can actually go into enterprises and help them figure out how to use Claude. And that is why we're so early is like most companies still have no idea what to do with this stuff, especially when we start [00:27:00] getting into the agents and like truly autonomous agents and agent swarms or groups of teams of agents.

[00:27:05] and so they recognize that. So it is a, it's a, it's a risk that adoption is so low, but it also means the market potential is massive and that's why they believe they could probably go from being a $1.5 trillion company to probably a $10 trillion company in the next five to 10 years.

[00:27:21] Mike Kaput: So just to kind of confirm here, it sounds like.

[00:27:25] They're betting on AI diffusing into every part of knowledge work. Not just coding, but also agentic ai, which is going to use dramatically more tokens usage in everything.

[00:27:38] Paul Roetzer: Yeah, it has to that that is the bet. Yes. Gotcha.

[00:27:42] The Rise of the Self-Driving Company?

[00:27:42] Mike Kaput: Alright, so next up, this past week, Replit CEO, Amjad Masad laid out a vision for the quote, self-driving company.

[00:27:52] This is something he explicitly predicts is how every Fortune 500 company will operate by the end of [00:28:00] 2030. So what is that? Well, in his view, a self-driving company is a business designed. So AI agents can carry out and coordinate much of the day-to-day work across departments. The agents use the company's data and software to pursue goals set by people who decide what matters, make judgment calls, and remain accountable for the results.

[00:28:19] So Replit builds. You know AI tools for building software, and they've actually been putting this idea into practice in their own business. They actually released a post in July where they described running elements of their company with internal agents that have access to systems such as Slack, GitHub, and the Replit knowledge base.

[00:28:39] Employees give agents goals and they can coordinate other agents to work through all the steps required to get work done as a result. Replit said that for a group of engineers at the company code output nearly tripled. In the first half of 2026, but at the same time, code review times, the rate of code changes being reverted [00:29:00] and software incidents stayed flat.

[00:29:02] So all good stuff in terms of more production, but not a rise in incidents. The company describes that this is a broader shift with employees moving from doers to directors of ai. So for example, Replit as a sales person, prepping for a customer meeting can use an agent to gather information on the customer's active projects and their use of Replit credits against their contract.

[00:29:25] The agent then packages that information into a customized presentation for a sales conversation. Marketing, for instance, can use agents to turn internal conversations and documents into product specs for launches. Support agents can investigate customer issues, prep responses, or send an issue to engineering with a summary of their findings.

[00:29:44] So Mossad says that the self-driving company needs what he calls a quote. Agent Habitat. This is an environment that connects agents to company knowledge and software while controlling what they can access, what they can do, and when they need human [00:30:00] approval. So Replit says it is building on its internal experience doing all of this to bring this approach to other companies through Replit for Business, which they recently announced this kind of self-driving aspect of it, on October 26th, I believe.

[00:30:14] So Paul, this is kind of a vision for exactly what it sounds like Anthropic is headed towards or expects companies to do. What did you make of Amjad Masad predictions here?

[00:30:25] Paul Roetzer: Yeah, I, I'm a, I'm a fan of Amjad Masad. I think. he's been innovating in this space for a long time. He's, he, he's pretty visionary. You know, I, I've always respected how he sees the market evolving and I think, you know, he got there early, obviously with Replit, fought the good fight for a long time.

[00:30:41] He founded the company in 2016. there, just for context, if you're not familiar with Replit, their latest round, I believe was in March of this year. They raised a $400 million series D round of funding that valued the company about $9 billion. so it's certainly a respected, you know, fast growing. I, I guess you could still call [00:31:00] 'em a startup, but it's 10, 10-year-old company.

[00:31:02] Hmm. yeah, so I think a lot of the ideas aren't necessarily new. They're, they're giving new names to some of this and they're living this themselves within Replit, which gives them the credibility to say, this is what it's actually looking like. And this is kind of how we're describing it internally and what we're thinking about.

[00:31:19] This would be conceptually new to a lot of organizations that probably aren't, you know, following along and doing this. Yeah. But just to revisit, again, you kind of summarized this, but for, so what they're calling a self-driving company, which, you know, we might call like an a, you know, an AI forward company on the frontier, or somebody might call it like an AGI company, whatever.

[00:31:37] Like there's gonna be different names for it, but it's like how he's describing it. a business design. So AI agents can carry out and coordinate much of its day-to-day work across departments. The agents use the company's data and software to pursue goals set by people who decide what matters, make judgment calls, and remain accountable for the results.

[00:31:54] This is the virtual collaborators of, you know, of Anthropic's S one. So again, everybody's [00:32:00] kind of like heading in the same direction. So then you highlighted this, every company in the Fortune 500 will be a self-driving company by 2030. I, I do not believe that. I, I will, I will explain why in a moment, not.

[00:32:11] Not that the technology wouldn't allow it. Yeah, I just, I have a different perspective on how fast Fortune 500 companies can move. okay, so Replit for business you mentioned, so I think it's important context that they are moving in this direction and so they're trying to enable this kind of organization.

[00:32:28] And so what does that mean? So Replit for business, there's like four main areas they talk about on the website. One is empower subject matter experts. So cloud agents enable anyone to bring their ideas into the company code base. Makes sense. Bring your own connectors and cloud. so connect. And so this is what they're allowing within Replit for Business.

[00:32:46] Connect your internal systems with custom connectors and run agents in your own cloud. So they're trying to like empower you to do all of this in a, kind of a. A, a neutral way where you can bring everything to the table and they enable it in a safe [00:33:00] and secure way that maybe is like better than going right to openAI's or Anthropic, which I think is what they're competing with basically here.

[00:33:07] put agents to work securely. So built in isolation, identity and governance controls help your team stay in control. So this is trying to address the governance issue around agents and then drive collaboration and decisions. So every agent connects your, connects your teams. Every change makes your company smarter.

[00:33:23] Now it's like, why is this shift? They highlight the fact that companies have what they call coordination costs, management meetings, departments, reporting, bureaucracy, that for over a century we've accepted that this is just the nature of doing business. It's inefficient and it's okay. Like this is how we work.

[00:33:40] And he is saying like, that is not gonna be okay moving forward. Companies that become self-driving, he says, will force an existential choice on their competition. Change how you work or watch competitors do what you cannot. I, I agree. Like back in 2000. 21 or 22. I wrote The future of businesses, AI or obsolete.

[00:33:57] Yeah. Same concept that like, you [00:34:00] can not do this and that's cool, but like somebody else in your industry's gonna do it.

[00:34:04] Mike Kaput: Yeah.

[00:34:04] Paul Roetzer: So then what are the implications of becoming one of these self-driving companies? realizing potential require companies to rethink how they organize and make decisions. self-driving companies will make agents useful to everyone.

[00:34:16] The people closest to the problem should be able to put them to work without becoming engineers. This is an important concept. a company cannot become self-driving. It's only its technical teams can participate. This, to me, is a bit of a blind spot in Anthropic's view of the marketplace. So they're putting a hundred million dollars into training 10,000 engineers.

[00:34:34] I don't think engineers are the answer, like they're important, but increasingly you don't need to be technical to figure out how to do agents and assistance. Mike, like you and I have no. Real engineering background or tech, and we build agents all day. Like, yeah, I actually am a believer you should be empowering the business unit leaders and the director and above to build with no code, which is more aligned with like rep, Replit could be doing.

[00:34:59] So I, I don't know. I [00:35:00] think this is like something that a lot of the tech companies miss because they've always just worked with developers and engineers. They think that's the answer. And I actually think that that's where the shift has happened and the companies that realize that have a huge opportunity to get there first.

[00:35:13] governance of the habit will also, help the company become smarter when someone corrects an agent that corrections should be reviewed and made available to improve future work. This goes to that idea we talked about in recent episodes of like working in open with these agents so that other people also learn what's happening, right?

[00:35:29] self-driving companies must also remain free to change their models while retaining control of their knowledge systems and decisions. They said, as Satya Nadella puts it, quote, we need to separate the supply of intelligence from the authority over it. That, again, is a concept. What they're basically saying is, don't lock all of your intelligence into ChatGPT, because Claude may end up being the better model in a year or three months or whatever, Gemini might show back up and all of a sudden decide to be good again, which is what happened last week.

[00:35:55] And so like, you don't wanna be locked into a single solution. You want all of this to live [00:36:00] in an area where it can kind of grow. so how real is this? Is it like, are are we really gonna have self-driving companies by 2030? No. Like is, it's my very direct answer and I I would, I, I'm not a betting guy, but I would put some money on this one.

[00:36:13] So here's, I've said this multiple times, but I'll, I'll repeat this chat. GPT is now in three weeks, four years old. Most companies, including many of the Fortune 500 companies who I have met with, talk with their transformation leaders, run workshops for, they are still in the understanding and piloting phase.

[00:36:33] Of AI assistance. So the idea that they're gonna just like magically figure out agents in the next three years is absurd to me. the context, this is what we see. So gimme the take. I'll take any Fortune 500 company that we go into. We talk to What is happening is there are silos of teams that are racing ahead, that are building agents that are doing really fascinating [00:37:00] things.

[00:37:01] It's usually a team within a department. Imagine almost as like a center of excellence that maybe isn't even a formal center of excellence. Yet within that team, there's probably some technical people and there are some business leaders who are very AI savvy, and together they can look at the problems in the company and solve 'em more intelligently.

[00:37:21] They can look at the workflows, figure how to optimize what exists, or blow up what exists, and build something better like. That is not happening across 50,000 employees in any company I'm aware of other than maybe some technical, you know, like some high tech companies maybe. Well, no, I take that back. I have met with some of those big technical companies.

[00:37:39] It is not happening in those companies either. So what is happening when we talk about this widespread adoption of agents, I do believe it will happen within some teams and some departments and maybe even some broader business units over the next couple years. it will not happen across an entire 50 to a hundred thousand person organization.

[00:37:59] I, I don't [00:38:00] believe that that is even a viable option. we are entering 2027 in two months. My best estimate, and this is more to just to give like context about like some measurement of what I'm saying, I would estimate less than 1% of entire, like of you can give you the Fortune 500. Less than 1% of them are truly scaling AI agents across their organization.

[00:38:27] But less than 1% are scheme and scaling across like entire departments. So to think that by 2030 in three years we're gonna get to a hundred percent is crazy. my guess is if, like, if, if we took an entire organization and said, you're gonna be a self-driving organization and it's gonna look like what Amjad is explaining here, less than 5%, by 28, and then like play that out.

[00:38:52] It it just like, I, I don't know, Mike, you and I been doing research on the state of AI for seven years now. It never [00:39:00] takes leaps like this. When we talk about AI understanding AI adoption, AI literacy, do you have AI councils, do you have AI roadmaps? It moves by percentage points each year and never even double digit percentage points.

[00:39:12] Yeah. So something. Extremely significant would have to change in society, in the economy and business for me to even see this being 50% being self-driving by 2030. So again, great vision. I do think a lot of startups will function this way. AI native companies will be built this way.

[00:39:31] Mike Kaput: Yeah.

[00:39:31] Paul Roetzer: That will disrupt the legacy companies for sure.

[00:39:34] But legacy companies will not shift this fast. There's way too many unknowns. It's gonna take one to two years just to solve for all the governance and risks. To be able to start scaling the ai. So I would say by 28, 29, we may be at a point where they can feel confident to start scaling this more broadly.

[00:39:52] So, I don't know. And then just a couple chrono quick thoughts. We are thinking about this like as an organization. So we have our SmarterX Labs, [00:40:00] unit that like, it's basically, you know, me and Mike and Jeremy, you know, helps us with that too. But like, we're just trying to say, what does this look like?

[00:40:08] Like if we had unlimited tokens or compute, what would we be using, using agents to do? How can we centralize the process of building assistance and agents across the team to like operationalize the organization? It's all like stuff we're, we're actively thinking about. And it's part of what I'm thinking about with our AI Academy roadmap is like, how do we solve for this agentic frontier for the companies that do want to go this direction?

[00:40:30] And so I've, I've got a lot of thoughts leading into our November 19th, AI Academy Day where I'm probably gonna share some thoughts on how we can help organizations and individuals that want to go in this direction in a responsible way. But I just, I don't. I don't think it's like real, but I I, I appreciate the vision for it.

[00:40:46] Mike Kaput: Yeah. It's super exciting. I'm super interested to see, I think you're right on the startups. The startups for sure are going to embrace this, but yeah. When he talks about your competitors at the enterprise level, figuring this out before you, it seems like [00:41:00] everyone's running a relatively slow race.

[00:41:03] Paul Roetzer: Yeah. I, again, most Fortune five hundreds we meet with are still assigning licenses to AI assistance with no personalized training and use cases being provided to their employees.

[00:41:12] Yeah. And that's so far from thinking that we're just gonna like, turn agents on and just automate all this work with no governance. No training. And certainly, I mean, again, you, you focus on the engineers and the technical people. That's fine. They're not the ones who are gonna be like governing these agents as they're going.

[00:41:28] It's gonna be managers and directors and VPs and like, who have to orchestrate all of this, who have no training how to do it.

[00:41:35] Google DeepMind and OpenAI Say AI Could Transform the Meaning of Work

[00:41:35] Mike Kaput: Our third big topic this week, researchers at Google's Deepmind Institute published two essays about how AGI, basically AI capable of performing a wide range of tasks at human levels, could transform the economy and change the meaning of work.

[00:41:51] So the first essay is titled The Economics of Intelligence. How and When Does AGI Lead to Economic Transformation? And this paper basically [00:42:00] asks why AI has not yet transformed the economy. So the authors focus on how AI agents or programs that carry out tasks work together. So they kind of play out these examples and say, look, a software team pair a strong coder and a strong writer together, they could outperform an AI team whose agents.

[00:42:18] Each beat the average person at both tasks, copies of the same AI ad capacity, but may share weaknesses. So if companies can rent AI cheaply to do the same work, the model that these researchers, suggest says that some workers' wages could fall even without job cuts. So, whereas on the flip side, specialists and managers whose skills AI can't match could earn more, but.

[00:42:43] If agents get much better at dividing work and sharing information, companies could build much larger AI teams with fewer of the coordination problems that slow human teams like we just discussed. So in this model, according to the papers, authors, a few such companies could outcompete rivals [00:43:00] in major industries rapidly changing jobs and wages.

[00:43:04] Today's agents, however, still coordinate poorly. The researchers say, and they don't have a timeline for when this scenario could play out. So they're kind of asking these questions and considering these paths of like, well, why have agents? Why is AI and why would a, why is AGI not transform the economy yet?

[00:43:21] Now, the second essay that the DeepMind Institute released at roughly the same time is titled What Is Work for Human Wellbeing and the Value of Work in the AGI Era? So it asks what people get from work beyond a paycheck, including things like purpose, social connection, status. A sense of agency, and while it acknowledges that many jobs are unpleasant or exploitative, it argues that replacing income alone should you be able to do that, may not replace everything a job provides people.

[00:43:52] So they did this survey of more than 8,000 working age adults in the us, uk, Denmark, and India, where most of the [00:44:00] workers said they would leave their job if their salary and benefits were guaranteed. One in five said they would stay even if offered up to 120% of their pay to leave. And the authors reviewed research on people who stopped working and found that the way in which they do that matters.

[00:44:16] So involuntary unemployment is associated with worse wellbeing while people who choose to retire often fair better. So the researchers suggest that caregiving, volunteering, other activities could provide some benefits People get from paid work. So even people who keep their jobs could see the meaning of their work change.

[00:44:34] If AI does more of it, the researchers. Suggest now separately with this, we'll probably get into this as well, openAI's release data on how individual ChatGPT users use the tool for work. They covered 165 types of work activities across a hundred countries, Mercor, which we've talked about in the past.

[00:44:51] And some researchers also released TaskNet, which uses AI interviews to map tasks people do in six occupations. So Paul kind of [00:45:00] interesting, Deepmind is releasing two pretty extensive essays at the same time. Basically trying to say, okay, is a GI going to lead to transformation of the economy and should that happen, what is the actual meaning of work?

[00:45:14] Paul Roetzer: Yeah, and I'm, I, I mean, I'm glad to see all this. It's, it is odd. Like I was trying to, usually when you see this trend where you see like four or five different companies all roughly dealing with the same issue. Yeah. Something has happened.

[00:45:26] Mike Kaput: Yep. Yep.

[00:45:27] Paul Roetzer: We may not like learn about it for a few months, but I think broadly speaking, everyone since fall of 2025 has seen agents becoming more autonomous, more reliable, able to complete longer horizon tasks.

[00:45:40] And they all probably started ramping up their efforts to understand this and where it goes because it was different than what's the impact of AI assistance on the economy. We are now talking about what's the impact of agents that can do long horizon tasks on the economy. And so my guess is just like, I don't know, everyone's been working on this all year and now we're starting to see the [00:46:00] fruits of their labors.

[00:46:01] Yeah. So just a quick retrospective. So DeepMind Institute, we talked about this at the time, but it was actually introduced September 16th of this year. So just less than a little less than two months ago. So the DeepMind Institute aims to advance bold thinking about the safe development of artificial general intelligence.

[00:46:19] Its beneficial use and its implication for society from the lab that pioneered the field. We focus on bringing creative foresight and grounded examination of AI to help tackle the most important questions of our time and navigate human's next era. So that's from the introducing the DeepMind Institute blog post on September 16th.

[00:46:38] so the first paper, the Economics of Intelligence, the, lead author was Alex Ammas, who's the director of AGI Economics at DeepMind. And he posted on x sharing this link and gave a little context. He said, we've seen exponential increases in what a, a AI agents can do, but the broader economy seems, almost remarkably unchanged, [00:47:00] which is probably why a lot of these.

[00:47:02] You know, when, when I'm sort of getting upset that the economists aren't trying to be more forward looking and they're just looking backwards and saying, oh, nothing's happening, or You're David Sachs, or all these people are just like, nothing's happening. It's all going, always gonna be good. I'm, I'm happy to see reports that are saying why the numbers aren't moving, but that doesn't mean it's not gonna, that it's gonna be all good.

[00:47:22] so he said part of the reason is that the solo AI agent is the wrong economic object. Economic transformation is an EQU equilibrium outcome where organizations compete. Looking at how intelligence functions within organizations. Transformation will depend on when two capability thresholds are crossed.

[00:47:43] Competence, which means gains on individual skills, like coding, math, things like that. Or you know, legal work, marketing, things like that. And coordination gains on ability of agents to cooperate and work together. And that seems to be like the real focal point here. So he says, we show that [00:48:00] what economic transformation looks like depends on whether AI becomes super competent.

[00:48:04] So it's like high levels, like a player stuff or beyond of doing the work. And super cooperative, meaning working across disciplines within an organization like you would work, like, let's say you're a. You know, an operation leader and you work really well with the marketing people or the salespeople, or you're an IT leader, but you totally understand business outcomes and goals and how the business units function.

[00:48:25] So that's what they mean by super cooperative. They're able to work across areas of expertise. once we have those or both the regime of artificial super intelligence, a SI, the traditional meaning of super intelligence tracking, this trans transition shows why the real visible transformation may come from improvements in coordination, a capability.

[00:48:45] Most benchmarks don't yet track. So we're not even like looking at this because it wasn't even real until like five months ago that they could really coordinate, at high levels. So the two main takeaways, Alex, I highlighted to understand the path of transformation, we should be [00:49:00] tracking AGI agentic coordination more closely.

[00:49:02] And two, the impact of AI will likely show up in wages before market structure or employment as you alluded to Mike. So a few other, things. that I'll highlight from the actual, research. Smarter AI will increase productivity, but real economic transformation hinges on the learning to coordinate organizations, assemble people with complimentary strengths and divide up tasks to create teams far more capable than any solar solo worker, because existing benchmarks only test whether AI can perform, outperform individuals.

[00:49:33] A model can exceed the performance of every person one-on-one. And still lose to a well-structured team. This is why organizations have sometimes been referred to as a form of super intelligence, but then they get into like, while the supply of human workers is fixed, AI agents and their associated capabilities can be scaled on demand at the cost of compute.

[00:49:54] Second, the constraints that limit the size of human teams, such as span of control, communication, frictions, [00:50:00] and misaligned incentives need not apply to AI agents when they can eventually work directly with one another rather than sitting inside human hierarchy. So rather than hiring like an AI. Agent, SDR or customer service rep, they're able to work across all these things that's not siloed into that specific role.

[00:50:17] Then they gets in, gets into this whole idea of like how this is different. So they're not fixed supply like humans. AI removes the scarcity of, bottleneck at scale of like having expertise across these different areas. You can create one that's an expert in and then scale a hundred of them, and now they can work across the organization.

[00:50:33] they said today's AI systems coordinate poorly. And again, the quickest way to think about this is like the hugging face example. Mm. what we would call agent swarms seems to be the term that's, that's stuck. Yeah. Sam Altman was in an interview last week. He's like, God, I hate the swarms term. It's so negative.

[00:50:48] He likes to call it teams. And then he kept falling back himself and calling it swarms because God forget it. It's just, it is swarms, whatever. And so all that's gonna change and that, so at a high level, this one said, today's labor markets [00:51:00] reflects the current capabilities of AI systems, not mass unemployment or sudden AI monopolies, but leaner startups, flatter org charts, softer hiring and entry level roles closest to benchmark tasks and steady arising demand for experienced people who can direct complex work across teams.

[00:51:16] That is like. The main takeaway is like, we might not see massive unemployment, but you're gonna see leaner startups, flatter org charts, softer entry level hiring, and then the steady rise for experienced people, business people, not just engineers who can direct this work across systems. Like that to me, is like the most important thing to take away from that report.

[00:51:36] the human wellbeing one's interesting, you know, just looking at the impacts of AI on the labor market, but then going beyond that into the other areas. So that's a good read if you're interested in it. The Mercor one is basically taking on net, which is the database where you can go and look and say, okay, what does a marketer do?

[00:51:53] What does a lawyer do? And it gives you like the 20 standard tasks. We used this to build jobs GPT, yeah. three years ago. [00:52:00] So O*NETis a really useful database, but it doesn't capture like how the jobs change, and it doesn't like break the lawyer job into 10 different versions of it and say like, well, what does this person actually do?

[00:52:11] Hmm. So it's a, it's a good, like direction to look at, I guess, ironically. Then you also get the ChatGPT one for expanding the world of work, where they start to look at what is going on and how people are actually using chat. GPT, their data in the US shows the dominant use is planning and decisions, so making decisions and solving problems, and then thinking creatively.

[00:52:33] And then they also have work output as a category. Information input, developing and advising. Information processing and communications are sort of like buckets that they look at. So again, interesting data. And then one other one that just came out. Oddly enough, it was like they did this on like a Sunday night, if I'm not mistaken.

[00:52:51] Yeah. This came out at 9:00 PM Eastern time on Sunday, which all, I always like, my ears always perk up when like something gets released at a really odd [00:53:00] hour. Like it tells me like, why would you do that if you wanted anyone to see it? Yeah. So I'm not sure why they did this, but SemiAnalysis, announced GDPXA benchmark for measuring AGI agentic capabilities across the most valuable domains.

[00:53:14] so this is again, brand new, but it says, we believe the first trillion dollars of AI revenue will be concentrated in sectors that can deploy frontier intelligence at high ROI not broadly diffused throughout the economy. Building a general purpose benchmark based on a share of GDP or the own net occupational taxonomy is therefore a mistake.

[00:53:36] Interesting. Mm-hmm. if your goal is to track the progress of frontier models, you should focus on the domains. Where they'll be most impactful. This, is why we include chip design, cyber, bio over domains like government, real estate and in and insurance. Interestingly, this is like going back to my roots in AI 14, 15 years ago, I read Automate This by Christopher Steiner.

[00:53:58] Mm-hmm. And his equation [00:54:00] was like the potential to disrupt the reward for disruption. So if you're gonna look at industries and say, where should we focus our energy on applying agents? Where should we put our 10,000 engineers to focus on? They're gonna put it on the place that has the most value for disruption.

[00:54:14] And so that leads you to things like chip design and, you know, other areas. so yeah, just like a ton of ai. In economy and jobs and like the agent economy, and it's just fascinates me that it all happened in the last seven days. Like all this research sort of hit, so

[00:54:33] Mike Kaput: yeah. Yeah, the timing is interesting.

[00:54:35] Paul Roetzer: Yeah. Mm-hmm. Lots to dig into. If you're, if you're intrigued by, we put, and by the way, like if you go to podcast dot SmarterX dot ai and then click on show notes in the navigation, we put all of these links in the show notes. So if you've ever wondered like the things we're talking about in citing and mentioned like different sources, they're all listed in those show notes.

[00:54:51] You can go back and, you know, dig into 'em if you want.

[00:54:55] Mike Kaput: Before we dive into our rapid fire topics this week, Paul, this episode is brought [00:55:00] to you by the MAICON 2026 on demand bundle. So if you couldn't join us in Cleveland in mid-October for MA on, you can now watch recordings of main stage and featured sessions from the conference on your own schedule.

[00:55:13] So this bundle brings together big picture conversations about where AI is heading, and practical sessions on how to make it work in marketing. So you'll hear from incredible speakers like Karen Hao, Kevin Roose, Andrew Yang, alongside practitioners tackling the questions teams are working through right now.

[00:55:31] That includes a great session from Wil Reynolds on visibility and credibility and AI search. One from Andy Crestodina on turning a marketing process into an automated tool. One from Chris Penn on AI for analytics and from our director of research, Taylor Radey on AI research workflows. This also includes Paul's keynote on rethinking work in the age of AI and agents, and my session on building an AI native content teams.

[00:55:55] So you can go to MAICON.ai, M-A-I-C-O [00:56:00] n.ai. You'll see the full session list and pur you could purchase immediate access right there. There's a button right in the header to purchase MAICON2026 recordings. Okay, Paul.

[00:56:12] Google Unveils Gemini 4 Carbon and Genie 4 as OpenAI Releases GPT-6.1 Terra

[00:56:12] Mike Kaput: So first rapid fire. This past week, gem or Google announced Gemini 4 Carbon, which is its most capable AI model.

[00:56:19] Google says Carbon has is able to spend days reasoning through complex engineering tasks such as finding and fixing subtle bugs across large software systems. It's designed to find software vulnerabilities, test whether they can be exploited and right fixes. The benchmarking firm. Artificial analysis ranked carbon first on an index combining several AI tests.

[00:56:41] Google says it is twice as successful at the previous model Gemini 4 Argon at generating secure code in internal tests. So they are starting to roll out Gemini 4 Carbon to some cyber defenders. Then they plan to gradually add trusted partners. They are refining safeguards and participating. They said [00:57:00] involuntary US government testing wider access to Gemini 4 Carbon is planned, but there's so far no public date for that.

[00:57:07] Roughly around the same time, Google DeepMind also introduced Genie 4, A world model that generates environments that respond as you explore them. So you can describe a scene, move through it, and add characters, or change the weather in real time. It generates sound and characters can actually respond in conversation.

[00:57:25] And actually another Gemini model can actually help guide these worlds. So Google sees potential uses in filmmaking and robot robotics and AI agent training. Genie supports a few minutes of continuous interaction, though the visual details early testers report can drift. So Genie is a research preview for trusted testers with access plan for Google ai Ultra subscribers through Project Genie and select developer partners.

[00:57:52] And not to be left out, openAI's also released GPT 6.1 Terra, a fast model for understanding screens and operating [00:58:00] apps, apps at half. Its predecessor standard API price. This is available in ChatGPT work and Codex on eligible plans and through the API. So Paul Gemini, Fort Carbon is getting quite a lot of attention as Google being seemingly back in the race.

[00:58:15] Paul Roetzer: Yeah. In the world of X, people seem to be loving it. lots of positive response to it. Lots of strong early evals. yeah. So it seems like it's a, it's a good model. I, I don't know about you, Mike, but I've gotten to the point where like, all these labs, you know, we've, we've been on 'em for like four years about their inability to name models, and now they've all seemed to have landed on their own themes.

[00:58:38] Yeah. And now I can't keep track of what the hell they are. So like, I, I actually have to do, there's like a mental load to remember which version of the model is the strongest. And so like, Anthropic I think was sort of the, maybe the first to head in this direction where you had like Sonnet, there's one under Sonnet if I'm not

[00:58:57] Mike Kaput: Yeah.

[00:58:57] There's o or there's Sonnet Opus or [00:59:00] Haiku. Haiku is the

[00:59:01] Paul Roetzer: lowest one. That's the smallest.

[00:59:02] Mike Kaput: Yeah.

[00:59:02] Paul Roetzer: So Haiku, Sonnet, Opus. Mythos, which sort of break. Yeah. Broke from the naming sequence. But like Mythos is more powerful than Opus, but Opus 5.5 is actually a better model than Mythos 5.1. But like you have to stop and think like, wait, which is the most powerful model.

[00:59:19] And I'm definitely there with Gemini. I actually have no idea. Like argon and carbon. It's like what? I don't even. One, I have to go do research to figure out. Yeah. Which is more powerful. Right. And then I have to remember the sequence and then the, I actually liked the openAI's one initially, like the celestial naming.

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

[00:59:36] Paul Roetzer: But now I have to, like, again, I have to do mental gymnastics every time I want to pick. It's like, so Luna, Tara, soul Astra. Yeah. So Astras the biggest and Soul is the star. And then Tara is the land. And then l like, you have to like think, I don't understand why we can't just go like, small, medium, large.

[00:59:55] Yeah. Or like, I, I don't extra large. Like it would, it would've been way better if like the labs all [01:00:00] gotten cahoots like four years ago. And so let's just all make 'em small, you know, medium, large, extra large.

[01:00:05] Mike Kaput: It's deeply confusing.

[01:00:06] Paul Roetzer: It could've been way easier for everybody to, to understand and adopt these things, but yes, it sounds like Google, the genie stuff's wild.

[01:00:14] Like I, I I mentioned that I did the, what was it called? The playground, I think. Yes. Where my son and I created a video game together just with prompts. That's where all this goes. Like video game stocks, I would imagine are getting crushed. I think Roblox maybe got crushed, stock wise recently because you're gonna be able to create these like, like very real feeling games with just prompts and like anybody's gonna be able to build games.

[01:00:37] I think there's pretty broad implications to the Genie model. It's just like right now it's early and people aren't gonna really see it, but I think a year or two now you're gonna look back and realize the significance of the breakthroughs that they're making on that.

[01:00:51] Mike Kaput: And they mentioned in the announcement that it is critical they see it as critical for the development of essentially AGI because of the whole world model concept [01:01:00] that we've talked about in the past.

[01:01:01] Paul Roetzer: Yes. And I, I, I've always thought Google was way ahead on that game. Yeah. You know, there was a time when, what was the openAI's model that came out that they then like yanked back the video, the world model.

[01:01:14] Mike Kaput: shoot, what was

[01:01:15] Paul Roetzer: that? Begin with an SI think, didn't it?

[01:01:16] Mike Kaput: Sora.

[01:01:17] Paul Roetzer: Sora, yes. Where it was like, whoa, openAI's figured out some stuff with world models and then they couldn't, they couldn't afford the compute of it, so they just yanked it.

[01:01:26] yes, there's definitely a bet being made and that has been being made for a while by some leading a I AI labs that understanding the physical world was key to true super intelligence, and Google does seem to have an advantage there.

[01:01:44] Google, Microsoft, and Amazon Report Earnings

[01:01:44] Mike Kaput: All right. Next up, Google, Microsoft, and Amazon all reported earnings this past week for the quarter ending in September, giving us a look at AI demand, the company spending, and some interesting investor reactions.

[01:01:55] So first up, Google Parent Alphabet reported $127.6 [01:02:00] billion in revenue. Revenue. That's up 25%, but revenue and adjusted earnings per share actually missed Analyst estimates Google Cloud revenue more than doubled to $30.7 billion, driven partly by AI demand. Google says its Gemini app has 1 billion monthly active users.

[01:02:19] Alphabet expects to spend 195 to $205 billion in capital spending for 2026, but. Shares fell nearly 13% the next day as investors worried about slowing search advertising growth and heavy AI spending. So just a note on that. Google search and other advertising revenue reached 63.35 billion. That's about half of alphabet's total, but missed analyst expectations of roughly 65 billion.

[01:02:47] So search grew 12% year over year, but it was down from 17% the previous quarter. Google says year over year search query growth was the fastest in five years, but acknowledged that growing use of AI mode [01:03:00] contributed to the slowdown because it's currently less effective at generating ad revenue. Google said they claimed currency effects were the biggest factor in the slowdown.

[01:03:11] Microsoft reported $92.3 billion in revenue, up 19% with revenue and adjusted earnings per share, beating expectations. Azure Revenue rose48% to 33.2 billion and revenue from Foundry. Their platform for building AI applications more than tripled. Microsoft expects $170 billion roughly in calendar 2026.

[01:03:33] Capital spending, including some leases, shares initially rose after the report. As investors welcome strong cloud results and positive cash flow, but did give those gains back the next day. And then finally, Amazon reported $201.9 billion in revenue, up 12%, but sales and net income missed expectations.

[01:03:52] AWS Cloud revenue grew 39% to $46 billion. Amazon says it's bedrock service for using AI [01:04:00] models. Process nearly 17 times as many tokens as a year earlier. It expects roughly $220 billion in cash. Capital spending this year shares fell more than 6% after hours before closing, 1.4% lower Friday. Amid concerns about AI spending over 12 months through September net spending on property and equipment exceeded operating cash flow by $42.4 billion.

[01:04:26] So Paul, everyone's spending a ton of money on infrastructure. It sounds like Microsoft had pretty good results. Google appears to be getting crushed due to the implications of AI mode, perhaps on search.

[01:04:41] Paul Roetzer: Yeah, I mean, high level, the long game is still in play. Like there's nothing from any of these that says everybody overspent on, you know, data centers and CapEx, and so it's like, you know.

[01:04:54] All systems go still. Yeah. when you're looking out few years out, I, I, I don't, I mean, [01:05:00] I guess it doesn't get priced in, but like, I don't know how AI mode wasn't going to lead to a drop in ad revenue. maybe the, like the market just like expects it at a certain threshold and yeah, it was burning faster, but like, I don't know.

[01:05:13] I mean, eventually they gotta cannibalize it. It's been like one of the greatest business models in history for the last 26 years or whatever. And I, consumer behavior is changing the way people find information, sending their agents to do the research instead. Like it, they're gonna go through just a bumpy period where it's gotta transition, but the business is still strong.

[01:05:32] Google Cloud is obviously crushing it and maybe eventually that becomes the dominant business. I think maybe this was the first time where that the revenue from Google Cloud started crossing over. Yeah. Like that ad revenue. So, yeah, I don't know. I mean, again, if you're trading short term, you might get a little nervous on stuff like this.

[01:05:49] I don't ever look at any of this. I'm not a day trader. I don't ever look at anything in the perspective of what's gonna happen in the next 24 hours. nothing about any of this has changed my long-term views on any of [01:06:00] these companies. again, not, not investing advice, just advice. Strong, strong earnings.

[01:06:06] There's gonna be some things that are gonna cause some, some dips for a little while, but, you know, nothing earth shattering.

[01:06:13] Microsoft Offers Buyouts to 60,000 Employees

[01:06:13] Mike Kaput: So another story about Microsoft here this past week getting some headlines. So this past week, Microsoft said about 60,000 employees on selected teams in the us, Canada, the UK and India are eligible for what they call a voluntary exit package.

[01:06:28] So this is not an announcement that 60,000 jobs are being cut. In fact, Microsoft told Fortune that it expects fewer than 5% of eligible employees to accept this offer, which is an offer to leave the company voluntarily with financial support and career planning. They said they also do not intend a significant reduction in their overall workforce.

[01:06:50] However, this number of employees offered this option is more than a quarter. Of it's roughly 223,000 full-time employees that it reported worldwide in June. That [01:07:00] was just before it. It went through a round of layoffs in July. According to the Verge eligible workers include people in co-pilot windows, Xbox, finance, and human resource divisions, marketing, commercial business and security are not participating.

[01:07:14] And some AI and cloud teams are participating only in part eligible employees have until early December to decide to take the offer. And Microsoft says the package is comparable to its standard severance. Microsoft says layoffs remain an option depending on business conditions. So Paul, curious what you made of this.

[01:07:33] Obviously nothing has happened yet with 60,000 people taking offers or getting laid off, but that is a pretty significant number to extend the option to.

[01:07:42] Paul Roetzer: Yeah, so high level, obviously it's not directly related to ai. They're not looking and saying AI is now doing the work of all these people and we don't need them.

[01:07:52] Go back to Amjad Masad's vision though. If, if we have self-driving companies by [01:08:00] 2030 or even make significant progress, the companies that are gonna do it first are going to be Microsoft, Amazon, Google. Like, it's gonna be the tech companies, the cloud companies. Yeah, the labs. I, if, if you start having 10, 20,000 high paying jobs at a time that do take the options.

[01:08:24] 'cause Microsoft's doing this, Google's gonna do this, Meta's gonna do this. Like, everybody's gonna do this. Yeah. Like, it'll, it's a prelude too. It is like they're looking out and saying, okay, let's, let's just stay on the self-driving concept. Let's say that by 28, 29 we're there, like we, we can build self-driving companies largely where humans really are more orchestrating the work of agents and, you know, mix of humans and agents if they're, if they believe that to be true.

[01:08:50] You don't have to make wholesale changes right now, but you start to do things like, well, let's do some strategic buyouts. Yeah, let's see if we can reduce the workforce by [01:09:00] 5% today and maybe, maybe we like another 10% next year and maybe 15% the year after that because we just don't need as many humans.

[01:09:07] Like that is very clear. And they definitely all believe that, that they don't need as many humans to do the work that they're doing and likely even as they grow, because if they're seeing Anthropic doing 6.6 million in revenue per employee and they're doing 750,000 revenue per employee, like, I don't know what it is, but like that would be a under a million would be a good number traditionally in the technology world.

[01:09:30] But if the new measuring stick is 6 million per employee, then all of a sudden all these other tech companies start to look really human capital inefficient and the pressure to like. Raise that starts to become significant. And so you need to raise revenue by not raising the equivalent amount of people with that revenue.

[01:09:52] And so you start to wonder like, wow. Like what if we do look out one to two years and all of a sudden there's tens of [01:10:00] thousands, hundreds of thousands of buyouts of people, they're not going and getting other jobs. Like you're not

[01:10:05] Mike Kaput: Yeah.

[01:10:05] Paul Roetzer: Making 350,000 a year at Microsoft and there aren't 10 jobs waiting for you when all the other tech companies are trying to do the same thing.

[01:10:14] Yeah. So this is the kind of stuff that worries me on the surface. They can PR this and say, not a big deal. It's just like, we're doing this. It's not gonna affect a ton of people. But when you zoom out and realize everybody else is gonna do the same thing. It adds up. and then they wouldn't even ever say, Hey, AI cost all these job displacement.

[01:10:33] Again, the data might not tell you that's what's happening. Yeah. But it's, it's absolutely what's happening.

[01:10:38] Mike Kaput: Crazy.

[01:10:38] Paul Roetzer: Like, I don't know,

[01:10:39] Mike Kaput: even some of these earlier in the year, some of these, like CEO memos and layoffs that we covered were basically, they said they were positioning, you know, like the Atlassians of the world.

[01:10:48] Were positioning the company for a new era. Yeah. So they're getting ahead of it.

[01:10:52] Paul Roetzer: Yeah. Yeah. And it might not always look like we cut 20% of our staff today because of ai, because Right. It's gonna Right. It's gonna look at this like [01:11:00] two to three year strategic initiative where they never say that's what's happening and they never have to admit that they cut 20,000 people because of ai.

[01:11:10] Because you just do it over this time period. Yeah. And it's like optional buyouts and. That I hope there's economists studying that playing out, like

[01:11:19] Mike Kaput: Yeah.

[01:11:20] Paul Roetzer: You know, kind of gaming out, like what if this is what happens, what happens to those communities like Seattle?

[01:11:27] Mike Kaput: Sure.

[01:11:27] Paul Roetzer: Dependant upon Microsoft, like Yeah.

[01:11:30] Yeah. I don't, these are, these are the kinds of things I think about a lot. Like I don't, I mean we talk about some on the podcast, but like all other risks being equal and like all the things that could go on, right. This to me feels like the most real thing that is very easy to sort of project out being a viable thing that could happen.

[01:11:53] where you have this like three year period where there's just this like, doesn't feel like a rapid [01:12:00] transition, but three years is fast in historical context where you get to 2030 and you're like, oh my gosh. Like you go back and look at 26 to 28 and you're like, wow, we should have prepared better for that period and human history.

[01:12:13] Wow. Yeah.

[01:12:15] Amazon Refocuses Its AI Strategy on Frontier Models and Business Results

[01:12:15] Mike Kaput: Okay, some news this week about Amazon. So some of this is some news that kind of flew under the radar this past summer, with some business insider reporting in July that Amazon is scaling back many in-house AI models and focusing its strategy on fewer projects, including a new flagship model.

[01:12:33] Now, this comes also with some more recent news. We'll talk about both these things where Amazon also just released an engineering memo on October 27th, warning against token maxing or maximizing AI token usage. So together these announcements and developments may signal a bigger shift in Amazon's strategy.

[01:12:52] So first. Let me go through this reporting from July. Then we'll jump into this token maxing memo. So from July, Amazon has this [01:13:00] family of AI models called nova. So businesses can use these through AWS to build applications that work with text, images and video. So Business Insider reported that Amazon was deprioritizing models, including Nova Premier and Omni Plus its real video and canvas image models.

[01:13:18] At the, at the time some employees said the models were being kept running for existing customers, but were no longer major development priorities. Instead under executive Peter DeSantis, business Insider said Amazon was shifting engineers and compute power towards fewer projects, including a new frontier model.

[01:13:36] So it said researcher Pieter Abbeel was leading the effort with this model expected at Amazon's annual reinvent conference. Later this year, I looked this up. The dates appear to be November 30th, I think, to December 4th. the company also cut some AI jobs and closed a research lab in July, so second, so kind of keep that in mind.

[01:13:56] Second, they released this like memo about [01:14:00] token maxing, where they circulated to software engineers and managers. A memo in September, which was reported on this past week, where Amazon says it is updating expectations for engineers around ai. One challenge they're facing is that long-term outcomes can be slow and difficult to measure.

[01:14:16] Amazon says it has used proxies like token usage, AI tool adoption. Code changes and reviews in the past. To do that, though, these are flawed, the memo warns treating those metrics as targets and says they can create the wrong incentives. So with tokens you can drive up AI use and costs without necessarily improving the work.

[01:14:35] We've talked about kind of token maxing before, and the big thing here is CNBC reports that Amazon shut down an internal AI use leaderboard in May after employees ran up tokens to climb their ranks. So Paul, interesting, I guess no surprise, token maxing was not a long-term strategy.

[01:14:53] Paul Roetzer: Token maxing is what happens when the technical people are the ones leading the AI strategy.

[01:14:59] Mm-hmm. [01:15:00] I, I don't know any business leader who would've ever thought token maxing was actually a viable way to do this. Right. And we talked about it at the time, like Right. What an absurd KPI it was. Yeah. I just, in retrospect, like how, how did that become a thing people actually cared about? yeah. I don't know.

[01:15:19] This is interesting. You know, and again, it goes back to Amjad Masad's points about like, you kind of need to be model agnostic because you just never know. Like, you know, maybe one of the labs you're dependent upon models just go through a phase where for three to six months they're just nowhere near, like, I mean, Google would be a good example.

[01:15:36] Mike Kaput: Yeah.

[01:15:36] Paul Roetzer: For the last like six to nine months, Gemini was not even in the conversation. Yep. As being a top model wasn't even a top five. Like there was two generation old Claude models that were better than Gemini's best model. Yep. And then it changes. And so you have to be flexible with this. And I think Amazon is realizing this.

[01:15:54] So if you bet two years ago on Amazon's AI efforts and you built around what they were selling as their vision for [01:16:00] this, and then they rug pull that, it's like, oh, okay. Like that's can't build around that anymore. Yeah. But Bedrock allows you to move between models and Yeah. I don't know. It's just like, it's this constant shift.

[01:16:09] And we've seen this a lot, even with like Meta for like the last year, Meta looked like maybe they were just gonna become a cloud company and not compete at the frontier. But then you realize that, you know, Alexander and and Mark were just like taking their time and building the next thing. Yeah. Elon Musk seems to waver every other week.

[01:16:27] Is, is he gonna build frontier models or is he gonna just like, use everybody else's and become a cloud company? So it's really hard right now to figure out like who actually ends up staying, but yeah, when you look at a company like Amazon, there's all the resources in the world. Yeah. and they're just like complete shift in strategies

[01:16:45] Mike Kaput: on Unreal.

[01:16:45] Yeah. Well, we'll find out more at reinvent it sounds like.

[01:16:48] Paul Roetzer: Yep.

[01:16:49] AI Is Forcing a New Security Agenda

[01:16:49] Mike Kaput: All right, so another topic this week around kinda cybersecurity venture firm, a16z Andreessen Horowitz. They have a new essay called the New Security Agenda, where they [01:17:00] say that companies face two big security jobs, one, secure the AI tools they're deploying, and two, use AI to defend themselves.

[01:17:09] So they actually surveyed a hundred security chiefs at companies with over a thousand employees. 87% expect cybersecurity spending to rise in 2027. Nearly half expect their teams to grow. Most already have generative AI tools in use. Nearly half worry employees could leak sensitive data into them. 45% plan to buy or upgrade tools within a year to secure their internal AI agents.

[01:17:37] And interestingly, we had on the heels of this some recent cybersecurity incidents that seemed to confirm the importance of doing all this. So a handful that have come to light. In the last week we had cursor and AI coding platform said that tampered versions of Anthropic's Claude code extension in its marketplace attempted to give attackers remote access to users' computers.

[01:17:59] The [01:18:00] model company cohere reported that it asked Alibaba's Open Weight Qwen model to write a security rule. But the model. Apparently tried to send test text to what appeared to be an Alibaba Cloud storage address. The Washington Post reported that researchers at the Cybersecurity startup Cathedral used AI agents to help find ways into Anthropic's business systems through third party software.

[01:18:25] And the vulnerabilities in it, Anthropic confirmed the researchers accessed a limited set of its sales, HR, and IT systems. separately we had a huge story with the cryptocurrency exchange Binance, saying attackers stole $6.3 billion in digital assets from its inter, internet connected wallets. the company says it is replenishing 90% of the affected amount using its own funds and reserves while insti investigating how the attack happened.

[01:18:57] Paul, there's a lot of stuff going on here. [01:19:00]

[01:19:00] Paul Roetzer: And just to be clear, the cohere one that you kind of like went pretty quick over what cohere is an AI lab out of Canada is saying, is that an open weight model? An open weight Chinese model was secretly sending data to a Chinese lab.

[01:19:18] Mike Kaput: Yes. Yes.

[01:19:20] Paul Roetzer: Just so we're clear.

[01:19:21] So this goes back to what I've said all along, which is the US government at some point I believe is going to outlaw the use of Chinese models in corporations. This is the kind of thing that could accelerate that sort of stuff. And it's the thing I always say, it's like if again, I'm all for like open weights.

[01:19:43] You have to know what you're doing. Like don't, don't just assume, and this may end up being nothing. it is possible. It was just like some flaw within the model and it wasn't actually trying to send private data to servers in China. but maybe it was, and we [01:20:00] just have to be aware that we are in a very new era and we're all trying to like understand how these things actually work, what the real risks are.

[01:20:09] The cybersecurity stuff is by the day getting crazier. I did see a funny, like a little big humor to this. So Chris Bakke, I'm, I'm not sure how you say his name, but he posted on X, and this made me laugh, he said, your friend who pivoted into crypto in 2021 and then to AI in 2024 is about one week away from announcing their new cybersecurity startup.

[01:20:34] It's just, it is very true. Like cyber is gonna be the hot space.

[01:20:38] Mike Kaput: Yeah.

[01:20:38] Paul Roetzer: So yeah. Again, Have, have the experts on call, have plans for everything. Work with the people who know what they're doing in this space. Don't, don't try and fake it. if you don't understand deeply how all this stuff is working, get advisors who do, lean on the staff if you have IT professionals in [01:21:00] your organization.

[01:21:01] Security professionals who want you to be a little bit more cautious, listen to them. Like, we, we have to work together on this and, and do this as safely, as possible.

[01:21:11] Mike Kaput: Yeah. And the cohere post about the incident is quite dense. I would say, though I think it's worth the read, but it is worth mentioning.

[01:21:19] Like you said, Paul, they said, we cannot answer whether this is deliberate or a leftover glitch from training. But defenders and policymakers face the same reality. A wait file, meaning the open weights file or the weights of the model, a wait file you can't trust is still a black box and running it yourself just moves the black box inside your firewall.

[01:21:41] So keep that in mind. All right.

[01:21:47] AI Use Case Spotlight

[01:21:47] Mike Kaput: So Paul, next step, we have our AI use case spotlight, where every week we look under the hood at some real AI use cases we are exploring at SmarterX. So I was gonna share one real quick and then turn it over to you. So this past week, I am in the process of working to scale up our AI transformation series where I interview leaders about how they're putting AI to work in their companies.

[01:22:09] So we're lucky enough to have tons of people in our network who are potential fits for this series, but I also wanted to know what I was missing and build a database of great stories across different industries that I was unaware of. So I actually used Codex to build an extensive research database with potential guests, their stories and links to public sources to learn more about their work with ai.

[01:22:33] So interestingly, this was not just about finding some people with like cool AI case studies. I wanted to find really, really compelling stories that were flying under the radar and that met my strict standards for what I find to be a great story that translates across industries. So I actually worked pretty extensively upfront with Codex before doing a single task to kind of describe exactly.

[01:22:56] What I was looking for, we essentially collaborated to build kind of like quote [01:23:00] unquote like algorithm that defined the qualities of a great transformation story. This like took me a couple hours, honestly, and that also included which sources to prioritize. So I actually focused first on people who had.

[01:23:11] Done recent podcast interviews since that tells me they're open to podcast appearances. it gives me a sense of how well they interview. I followed that with, have they appeared in talks or panels if they haven't been on a podcast. And I wanted to scale this up across industries. So I started with healthcare and did a test of like a handful of having it go research candidate stories in healthcare.

[01:23:34] I refined that, gave feedback to Codex and turned it loose across healthcare. So in short order, I got 40. Incredible stories for healthcare, many of which honestly felt like diamonds in the rough, like going through their podcast interviews is I don't think I would've found many of these otherwise. They were kind of buried in like niche industry podcasts that didn't even know existed.

[01:23:56] And then from there it kind of gave Codex another round of feedback and then turned it [01:24:00] loose on five other industries. So in an afternoon while I went and did other things, after I'd spent all this time upfront training it to do what I needed it to do, it built me a database of hundreds of stories across six industries that I then went through manually to select the best candidates.

[01:24:16] So that was really cool. And I had it checked things like, Hey, are they still at their company? Obviously if on LinkedIn it shows they moved, we can't probably interview them about their past work. So. A bunch of other, you know, post research checks to make a really great list.

[01:24:31] Paul Roetzer: That's really cool. Now, when you're saying you're using Codex, are you using the terminal on your desktop?

[01:24:35] Mike Kaput: I use the Codex app. I just find it much easier. The desktop app, it's just structured chats.

[01:24:42] Paul Roetzer: So like why don't you just use work in ChatGPT? 'cause Codex Powers work. Like I couldn't. Is there a reason why you. Would you stay it? and this is, this is just me like outta curiosity. Yeah. 'cause I don't use Codex like intentionally, but I use it as like under powering underneath work.

[01:24:56] Mike Kaput: You could use ChatGPT work to do everything I just described. Okay. I think [01:25:00] habitually I got used to Codex, the desktop app I have gotten used to using, because in certain scenarios if you want to generate local files and things, that is how you would do that with Codex on your desktop. But you could also probably do that through the app using ChatGPT as well.

[01:25:17] Okay. So just force of habit. But you, you can't do something like this as extensively as I did it with just ChatGPT chat. But you could very well do it with ChatGPT work.

[01:25:28] Paul Roetzer: Okay. yeah, mine was actually, I've been building a lot of very narrow assistance for things. So I have my co CEO but then I also like build assistants at things I, I regularly do.

[01:25:40] And then also like trying to figure out what I turn into agents and what, you know, remains just an assistant. And when I build an assistant, I usually build it as a project within ChatGPT That's like my dominant use for it. I will do the same in Claude sometimes. Every once in a while I'll experiment with Gemini, but like mostly ChatGPT is my home base.

[01:25:56] So, I actually had to, book travel for a couple of trips I have coming [01:26:00] up in November, December. And I, I, I will use at times, like, especially for personal stuff, I'll use Gemini for personal travel, but I don't think I ever built a travel assistant. So I was like, I'm gonna take 20 minutes. I'm actually gonna build this so I can use it every time.

[01:26:12] Yeah. So just quick, like how I did it, I actually went into a normal chat and ChatGPT, so it has all the context of like who I am and my company and all this stuff. I said I travel a lot for business. I wanna build a travel assistant that helps me plan my trips, including the best flights. And then I actually asked that I curiosity, should I build this as an agent or a project?

[01:26:29] It said project, which I assumed I said, okay, great. draft the travel profile for me. Cleveland Hopkins is my home airport, but I also can travel a out of Akron Canton, which is a regional 45 minutes south that has like some direct flights. I prefer United Business class like dah, dah, dah, dah. Here's my status on these different airlines.

[01:26:46] Like if I have to take Spirit Southwestern Allegiant like I will, but like, please don't ever make me do that unless I have to. and so like, here's my back. I prefer nonstop flights when possible. So I just gave it like some basics, like, all right, lemme see what it does. So it came back with that and I was like, oh, [01:27:00] that's not, not bad.

[01:27:00] And I was like, okay, so for hotels, here's my status with the different hotel chains. So like, know that information and then, I also said for context, like my flights are usually for business or speaking engagements where it's being paid for. So it's not, I'm not as price sensitive because it's being paid for by the client.

[01:27:17] So like, don't make that the primary variable. Focus more on like convenience and the, you know, the benefit of the flight. And then it gave me that. I was like, okay, great. Turn this all into like one instruction for a new travel assistant project. It wrote it. I opened up a new project, I gave it the instructions, and I planned my travel for the next trip I'm doing in two weeks.

[01:27:34] Yes.

[01:27:34] Mike Kaput: That's awesome.

[01:27:35] Paul Roetzer: So like just a, again, like whenever you have something like that, I think it's a good example of like, I did this in like, I don't know, 10 minutes and that those are literally all the pro, I just told you the entire chat. I had to create this project. And then what I do is I create the, now I have the project, I go into travel assistant and then for each trip I take, I just start a new chat.

[01:27:54] Okay, here's where we're going. Mm-hmm. Here's the details, here's where my talk is. You know, the thing now what I could [01:28:00] also add in there is like, hey, like give me some recommendations if I've, if it's an area I haven't traveled before, like I've got 24 hours in that city, what are some things to see? I can kind keep expanding this.

[01:28:09] but yeah, I mean, usually it's like me and my wife plan my travel mostly and then some people on the team. And so this is just like an assistant for us to use to like, you know, coordinate it. And then it also coordinates my personal schedule around it so my wife knows what's going on and where I'm gonna be and stuff like that.

[01:28:24] Mike Kaput: That's super helpful.

[01:28:26] AI Product and Funding Updates

[01:28:26] Mike Kaput: All right, we're gonna wrap up with some quick product and funding updates. We've only got a handful this week. first up, OpenAI updated what it calls it's Model Spec. This is its public guide to how its AI models should behave. They organized it around five values and added guidance on human intent and agency while retiring.

[01:28:46] Much of its detailed style guidance. Anthropic introduced a gallery of ready to use Claude artifacts, including interactive tools. Legal teams can customize and share to help review [01:29:00] contracts, track regulatory changes, and manage incoming legal requests. That's interesting. In the light of what we discussed in the IPO.

[01:29:08] Mercor says it has paid businesses at least $50 million for data and offers to de-identify and license workplace records, including emails and Slack messages to AI labs, building models for real world work. And Nvidia says it plans to invest $3 billion in META'S stock offering as the companies expand their work on Meta's personal AI Agent Muse and its family of AI models.

[01:29:35] So Paul, super packed week. Appreciate you breaking it all down.

[01:29:39] Paul Roetzer: Yeah, those two middle ones are really interesting to me. The Anthropic for legal, it's sort of a, you know, an example of what could happen. I don't know what that means to like Harvey's valuation, but they're just coming after like specific verticals.

[01:29:51] They'll keep doing that. And then the Mercor one band paying for people's data. That's such a, such an interesting business model. [01:30:00]

[01:30:00] Mike Kaput: Yeah.

[01:30:01] Paul Roetzer: Alright. Good stuff Mike. Thanks as always. Yeah, the pulse survey, smarter.x.ai/pulse, check that out. we'll be back with episode 250. Yeah. Next week Mike, believe it or not, huge, we started doing the weekly, the month before chat, GPT.

[01:30:15] So October of 2021 we started doing this weekly podcast. We've been doing this for five years now, so

[01:30:20] Mike Kaput: that's awesome.

[01:30:20] Paul Roetzer: Episode 250, we may do some cool stuff. I don't know, I actually messaged with team this morning. I was like, what should we do for the 250th episode we got? So, I don't know. We got some ideas brewing.

[01:30:28] We'll see what happens. We might do something cool next week.

[01:30:31] Mike Kaput: I can’t wait.

[01:30:31] Paul Roetzer: Yeah. Alright everyone have a great week. We will talk to you then and, stay safe during the midterms, you know. Life moves on the next day, we'll all figure it out and let's work together. That's all I wanted, say. I just want to like be good to each other and figure out how to move this forward and, in a lot of ways.

[01:30:48] But we've obviously focused on ai. We got a lot of work to do with AI and its impact on the economy and jobs and society. And so like, let's just, Wednesday morning, let's figure out ways to work together. All right. [01:31:00] All right. Thanks everyone. Thanks for listening to the Artificial Intelligence Show.

[01:31:05] 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.

[01:31:27] 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.

[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

This Week's AI Pulse

Listen Now

Watch the Video

Timestamps

Read the Transcription

[00:06:57] Anthropic’s IPO Filing

[00:27:42] The Rise of the Self-Driving Company?

[00:41:35] Google DeepMind and OpenAI Say AI Could Transform the Meaning of Work

[00:56:12] Google Unveils Gemini 4 Carbon and Genie 4 as OpenAI Releases GPT-6.1 Terra

[01:01:44] Google, Microsoft, and Amazon Report Earnings

[01:06:13] Microsoft Offers Buyouts to 60,000 Employees

[01:12:15] Amazon Refocuses Its AI Strategy on Frontier Models and Business Results

[01:16:49] AI Is Forcing a New Security Agenda

[01:21:47] AI Use Case Spotlight

[01:28:26] AI Product and Funding Updates

Claire Prudhomme

Explore Our Brands

Education

About

[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

This Week's AI Pulse

Listen Now

Watch the Video

Timestamps

Read the Transcription

[00:06:57] Anthropic’s IPO Filing

[00:27:42] The Rise of the Self-Driving Company?

[00:41:35] Google DeepMind and OpenAI Say AI Could Transform the Meaning of Work

[00:56:12] Google Unveils Gemini 4 Carbon and Genie 4 as OpenAI Releases GPT-6.1 Terra

[01:01:44] Google, Microsoft, and Amazon Report Earnings

[01:06:13] Microsoft Offers Buyouts to 60,000 Employees

[01:12:15] Amazon Refocuses Its AI Strategy on Frontier Models and Business Results

[01:16:49] AI Is Forcing a New Security Agenda

[01:21:47] AI Use Case Spotlight

[01:28:26] AI Product and Funding Updates

Claire Prudhomme

Recent Posts

[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

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

[The AI Show Episode 247]: OpenAI’s Mounting Safety Concerns, Anthropic’s AI Expectations Gap, MAICON 2026 Recap, AI Score & Gemini 4 Argon