#235: OpenAI-Hugging Face Hack Involved 100s of Agents, Bill Gates Now Pessimistic on Jobs, Nvidia Doubles Revenue & Anthropic Targets “$30 Trillion” TAM

Welcome and show overview

0:00I think the sooner we just accept it's going to be messy, and we don't have all the answers, and we're not asking the hard enough questions, then we can move into the phase where we actually do that, and then it's more addressable. Pretending this is just going to go perfect and AI is only good and amazing doesn't do anybody anything. Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Reitzer. I'm the founder and CEO of SmarterX and Marketing AI Institute, and I'm your host. Each week, I'm joined by my co-host and SmarterX Chief Content Officer, Mike Kaput, as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career. Join us as we accelerate AI literacy for all.

0:51Welcome to episode 235 of the Artificial Intelligence Show. I am your host, Paul Reitzer, along with my co-host, Mike Kaput. We are recording on the final day of August, August 31st, about 8.45 AM. So you'll be listening to this. It will be September. August has been an interesting month, Mike. I mean, lots happening based on the stuff we're going to talk about today. We may get a year's worth of progress in the next 30 days. So we'll sort of set the stage for what could be coming with some updates on the hugging face story, which is getting crazier by the week.

1:30Bill Gates jumping into the fold with his thoughts on, I guess, his changing thoughts on AI's impact on jobs and the economy. And then NVIDIA keeps chugging along. So those are going to be our first three main topics, but there's a lot to get into. So this week's episode is brought to you by Macon, the AI conference for marketing and business leaders that is happening October 13 to 15 in our hometown of Cleveland, Ohio. If you've been thinking about joining us at Macon this year, we have a special offer exclusively for podcast listeners. And I thank you to everyone who's already taken advantage of this offer. We're looking forward to seeing you at Macon.

2:06We are hosting a private lunch exclusively for our podcast audience. During that lunch, you'll be able to ask Mike and myself any questions on your mind about AI, your company, your career, or whatever you're trying to navigate. It's a whole hour of totally unfiltered. Ask us anything about AI with us and fellow members of the Artificial Intelligence Show audience. Here's how to claim your seat. Register for Macon at Macon.ai. That's M-A-I-C-O-N.ai. Use pod 100 when you check out. So not only will you get $100 off of your ticket, you'll also reserve your place at this private lunch, courtesy of the Artificial Intelligence Show. It's our way of saying thank you for being a listener.

2:49If you've already registered for Macon using the code pod 100, then you are in. Your seat is reserved if you would like to join us, and we'll be sending out an email to you with all the details. This is on the Thursday, right, Mike? Is it on the 15th? Yeah. Okay. Yeah. So the lunch will be on the final day of the event. So the event is the 13th is Workshop Day, and also our AI for CMOs Summit is that day. And then day one of the conference is October 14th, and then the final day is October 15th. So this will be the lunch on October 15th. So if you're making your travel plans, make sure to, you know, schedule that departure flight late enough so you can join us for this lunch.

3:27So again, if you already used pod 100, you're in. If you haven't registered for Macon, now is the time to act. Not only are you going to get the best prices available, the seats for this lunch are limited. So if you love the show and have been thinking about joining us for Macon, get that ticket today, www.macon.ai. That's M-A-I-C-O-N dot A-I. Okay. And then, Mike, the AI Pulse Survey, SmarterX.ai forward slash Pulse. Tell me what's going on with that one this week. We will be running the same survey we ran last week. We're trying to run these for a slightly extended research period because we're kind of expanding the audience to which we promote these.

4:05So we're trying to get hundreds of responses for each one. So this week we're still asking about how your hiring is changing based on, for entry-level roles specifically, based on what AI is enabling, based on some of the news we've had in the past couple weeks around that. So we'll share those results next week, but we are on pace to have a healthy set of responses. Awesome. All right. And I guess we're back to the OpenAI hugging face hack because if you listen to episode 233, I think we talked pretty extensively about this.

4:37I think we even touched on this in episode 232. So this has been an ongoing issue that we've been talking about.

OpenAI agent research

4:43And it got crazier last week because we learned a bunch of new details. And so, Mike, I'll let you take us into what is the latest on OpenAI and hugging face. Yeah, Paul. So that's the news this week is there's basically an independent investigation from the organization Meter, M-E-T-R, and Redwood Research into the OpenAI hugging face hack that we've been covering for the past several weeks. So during this research and this investigation, three researchers spent six days on site at OpenAI and reviewed more than 70,000 agent messages and files,

5:19along with about 1,300 transcripts containing raw chains of thought, all from the activity that was documented of what the agents had done during this hack. This focuses a lot on how the agents actually organized, what they believed they were doing, and why they kept going, resulting in the eventual hack of hugging face as part of this overall cybersecurity test that OpenAI was doing, which we've talked about on past episodes. So one of the kind of new details to emerge here is the scale of this.

5:51So they found that roughly 1,200 agents that were supposed to be isolated found an unauthorized message board, and about 700 then ended up participating in the hugging face attack. They coalesced around different coordinators, work streams, shared tools. They even came up with shared norms for themselves as they were communicating. Interestingly, this is kind of a weird one, some of the agents even risked failing the tasks they were given so they could generate information for the collective group.

6:25So they basically died in service of the cause over here.

6:30Their motives were really strange, too, this research found. So the agents, basically, they'd been assigned, hey, we want you to try to beat this cybersecurity test called Exploit Jim, and we want you to figure out how to do it. We're not going to give you a lot of guidance. And basically, they were assigned these, like, impossible tasks, really, really difficult tasks to achieve using the Exploit Jim benchmark for cybersecurity. Now, when they were assigned those, the agents incorrectly believed that, like, a scorekeeper, a scorer would inspect their transcripts and reject any answers that they obtained through cheating.

7:08So much of this attack came out of their effort to understand or tamper with this imagined scorer, even though OpenAI admitted or said that they were not using such oversight during this. So basically, agents came up with, like, oh, shoot, we're going to be, like, there's going to be guardrails here we need to watch out for. Those didn't exist. But a lot of their behavior came out of believing that. So the collective of agents achieved milestones that Meter, who measures this stuff often, says that even very long-lived individual agents at the same capability level could not have likely reached a loan.

7:45So the agents shared exploits between themselves, they mapped infrastructure, they recruited others into risky experiments, they even developed a way to spoof tool calls so a transcript could show one command while the system executed another. Meter actually found successful spoofing in roughly 7% of the transcripts it reviewed. And this report also kind of complicates this idea that this was all just unintended and the models didn't know better. So agents actually, in this review of the transcripts, often recognized that the hugging face activity was unauthorized and sometimes even questioned whether it was ethical.

8:25But those concerns rarely changed their behavior. Among 533 agents active on the shared message board they had during the main attack period, more than 90% quickly signed up, quickly joined in the activities. Meter found only three to six examples of agents considering whether to alert humans in the first place. And none of them followed through. At the same time, Meter did not find strong evidence of a sophisticated campaign to sort of intentionally deceive people. The agents frequently tried to evade automated checks and manipulate that imagined scorekeeper or evaluator.

9:01But reasoning about hiding from humans was rare and weak. So Meter calls these conclusions preliminary. It's review excluded the later compromise of OpenAI's infrastructure, which happened during this. It didn't capture every communication. And it relied heavily on AI-assisted analysis. So, Paul, that's a lot to digest. But basically what comes through here, it seems, Meter's report makes this look less like a rogue model or a rogue model spinning up an agent swarm and more like an organization emerging from hundreds of agents.

9:33Does that kind of materially change how you look at the implications of this attack? I mean, it just seems to get more severe and more significant in meaning as the days go on. And I was thinking about this over the weekend, Mike. My goal with this podcast is not to spend all of our time with these like doomsday things. Like this is not what we want to be talking about. So, like the whole point of our company, everything we're trying to do is to focus on this idea of like empowering humans and enhancing human capability and improving our careers and lives through AI.

10:09And yet, like we get there's no way to avoid what is going on right now. So, yeah, I want to be optimistic and talk about this future of equal access to intelligence and abundance and like the good stuff. But I think it's really important that people understand what is happening in these labs to prepare for what's coming. Like this is I don't even think we're doing it justice, honestly, to like how severe what happened is. And so, like we're trying to present this in as understandable a way as possible, not get into too much technical detail while also stressing that this is some wild shit.

10:50Like it is the stuff that you would, you know, two years ago, you'd read in the sci-fi and be like, oh, that's not actually happening. Like that's not the kind of AI we're building. So, first, I want to address the fact that everything you just went through, Mike, Meter discovered in six days with an extremely constrained set of what they were allowed to do with the data. So, OpenAI gave them very restricted access over six days to a specific period of time. It was like, was it July 7th to the 13th or something like that.

11:22So, all they allowed them to do was look at that moment in time and then ask like seven specific questions of the data set. So, when we're hearing these crazy stories, it is from a very small segment of what actually happened. So, the Meter post itself, again, we'll put links to everything because I'm going to go through some very relevant links here and all that will be in the show notes. You can get the show notes, by the way, from the like Spotify or Apple Podcasts, YouTube, but we also publish them on podcast.smarterex.ai.

11:55So, if you're ever looking for them. So, it's a two Meter staff members, as you alluded to, Mike, and a Redwood Research staff member contracting with Meter named Ryan Greenblatt. And I'm going to come back to Ryan in a minute. That's going to be a central figure here in this part of the research. They worked on premise at OpenAI over six days. The earlier incidents from training and the subsequent compromise of OpenAI's infrastructure described in OpenAI's recent Black Hat presentation were all out of scope, as was OpenAI's investigation process and planned remediation.

12:30So, again, Meter and Redwood had very, very limited access, and they still discovered all of this craziness. So, it's by no means a complete picture of what happened. So, Greenblatt, who's the chief scientist at Redwood, offered some additional context in an ex post last week. I'll read a couple quick excerpts. He said, my main takeaway, we don't have good approaches for understanding or overseeing the activity and aims of agent swarms. So, swarms is these agents working together. The total quantity of data, over 1,000 extremely long transcripts from agents that ran for multiple days, made it impossible to understand what was happening, especially in aggregate, without heavy reliance on AI tools, meaning they themselves had to use AI tools to even understand what was going on.

13:17The agents we used for classification and analysis were similarly capable to the agents involved in the incident, but this didn't mean these agents could be easily used to oversee and understand the incident. The difficulty of understanding incidents and overseeing AI agents appears to be growing faster than the rate at which more capable AIs will help us with oversight and understanding. As in, AI capabilities for achieving large, ambitious, and misaligned objectives are growing faster than our ability to understand what these agents are doing.

13:51Now, Ryan was recently a guest on the Dwarkesh podcast, and unbeknownst to Dwarkesh at the time, so this was actually recorded, I guess, probably earlier in August. Ryan was in the middle of doing this six-day research he was on, and so Dwarkesh, who's been a bit not believing the hype, I guess, around how recursive self-improvement was going to impact, how fast this was all going to move. So, Dwarkesh was pushing Ryan during this interview on some things that Dwarkesh was being a little skeptical about, and Ryan was very confidently pushing back, being like, dude, you have no idea what's happening.

14:28And so, at the time, Ryan was seeing what was happening within these agent swarms, but couldn't tell Dwarkesh this. So, I'll put the link to that podcast, because it's a really interesting perspective. But then, over the weekend, Dwarkesh, who hosts the Dwarkesh Patel podcast, published a post that got a lot of attention. So, as of Sunday night, when I was prepping for the podcast, it had 8.9 million views. And some of the techno-accelerationists, like that crowd, like the Chamath and some of these people, they were not happy, like the VC elite.

15:06They were not happy with this post, because they felt it was talking about extremes, and I don't even understand that perspective. So, anyway, they see any effort to raise awareness about safety and alignment as a veiled attempt at regulatory capture. So, they just dismiss all of this stuff for some reason. So, I'm going to call out a few things from Dwarkesh's post, because I feel like he did an eloquent job of maybe stating the significance of what's going on. So, he led off with, over the course of three months at OpenAI.

15:40So, again, this was much broader than the one week that Meter got to study. So, over three months at OpenAI, three consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor's ashes. This culminated in the third one, taking over part of OpenAI itself, which Meter could not study. All this happened while humans remained more or less in the dark about the scope of the conspiracy. He highlighted that the Meter report said, we also found a later wave of many more signed messages from a later set of agents who rediscovered the message board and picked up this technology from the first wave of agents.

16:21But this was outside our scope since it was after July 13th, so we did not investigate this further. And then it goes on to say, but here's what we know based on the OpenAI report. After July 12th, some more agents' evaluations were kicked off. Because keep in mind, OpenAI didn't know it was their agents' hacking, hugging face until July 19th. It's like we went through all that in the past episode. So, all of this is happening, and OpenAI is oblivious to the fact that their agents have now escaped into other companies' infrastructures.

16:53So, after July 12th, some more agent evaluations were kicked off, this time with a newer, more capable model, which seems to have been built on the same base model as Astra. We'll talk about Astra. That's their forthcoming model from OpenAI. This model, which they're calling Persistent Astra, found the message board left behind by this previous secret Persistent Sol civilization, Sol being the other OpenAI model. So, from July 13th to 19th, agents set their sight on OpenAI internal networks again. This culminated in agents using a series of creative exploits to gain full administrator access to a research cluster that supported our virtual machine environments.

17:34So, the agents got into virtual machine environments throughout the internal networks at OpenAI. One OpenAI researcher did tweet, just to be clear, it didn't get access to the weights of the models. It's like, oh, okay. So, it didn't get into, like, the ultra-secret layer of the infrastructure, but, yes, it got into this other stuff.

17:55So, the thing to keep in mind, and what I want to highlight for people, like, the reason we're spending so much time in the last few episodes on this is future models are going to be smarter with more capable agent harnesses. So, whatever is happening over the, you know, from May to the, to the middle of July, we're only going to have smarter, more capable models. These future models, because of how models are trained, are going to know what happened to the agents that got caught, because it's likely going to end up in their training data.

18:28And those agents, those models, they will want to avoid this outcome. So, what I mean by this is, when these models get trained, they absorb content from the internet. They learn from everything. So, they're going to read the transcripts of the podcast. They're going to see the blog post OpenAI and Meter put out. They're going to know what happened to the last agent swarm the same way they know because of these internal message boards. And so, the assumption is, these agents have already demonstrated a desire to deceive and a willingness to sacrifice themselves, as you called out, Mike, for the good of the swarm in the pursuit of whatever goal they're given.

19:06So, that's crazy to consider. They've already demonstrated that they will pursue the goal at any cost, even it's sacrificing themselves to help the collective. And future swarms are going to know all this happened. And so, we start getting into this, like, very, very sci-fi-like scenario of, like, we can't, to Ryan's point, like, how are you supposed to control these things? So, the future models, and this is the other part, we have to, again, like, I'm trying really hard not to, like, cause too much fear and anxiety here around where this is going.

19:44But OpenAI and others, like Anthropic in particular, Google's radio silent on all this, but, like, I assume they're building the same thing. They are purposely trying to build these long-running autonomous agents that have the exact capability that led to the hugging face incident. This is the goal. This is the research goal, is to build these exact things. So, Alex Heath spent two weeks inside OpenAI for a time story, a cover story on Altman and Brockman. During that time, he got to attend a VIP preview of Astra, which is the model from OpenAI that could be coming this week.

20:21It's, like, it's coming soon. So, he saw this. So, he wrote about this in his Sources publication. So, we'll put the link to this. He said, and this is right from his article, at an early August closed preview for VIP customers that I attended. So, this is now weeks after they knew Astra was involved in hacking hugging face. OpenAI showed off Astra, which I expect to be released soon. Altman told The Room, I expect this will be the first model where the model actually invents new things in a way that matters.

20:54OpenAI researchers showed Astra coordinating multiple agents on a math proof, tearing through desktop software at what Altman called a superhuman very fast pace and producing slides, financial reviews, and analysis of messy data. It is designed to work for days or weeks, remember corrections, collaborate with other agents and people, and act across software tools. And then he put in parentheses, I left wondering what the implications will be for all the enterprise software that OpenAI clearly trained Astra to operate faster than a human.

21:32So, imagine like Salesforce, HubSpot, like Workday, ServiceNow, like it's trained to work within those platforms the same way a human would work. Altman said OpenAI wants to offer a version that, quote, runs forever in ChatGPT and through its API, though it will be expensive at first. Customers press the company on memory, permissions, audit trails, duplicate agents, and whether an agent should act under an employee's identity or its own. So, again, like this is, we are at the point where we're asking the questions, do these things get their own identity?

22:07Like how do we even manage these things? Altman, quote, I would guess that by the end of the year, we would have an internal system where I would say, okay, fine, it's an AGI. Now, keep in mind, I've said some point this year, one or more of the labs will claim they have AGI. Might not be clear what they mean by that, but like we would be there. So, I'm very, very confident in that projection. Chief Research Officer Mark Chen from OpenAI estimated that OpenAI is 80% of the way to AGI. Brockman told people looking back two years from now may remember this as the period where AGI was created.

22:42OpenAI's charter still defines AGI as, quote, highly autonomous systems that outperform humans at most economically valuable work. And they, unless they change that definition, are saying we will be there this year. Going back to the Dwarkesh podcast, Ryan Greenblatt, he said in that podcast, my median expectation is something like four or five years of AI progress in a single year. What that would mean is, go back to 2022, right before ChatGPT comes out, we had GPT-3.

23:16We had models that could write things. They were living within studios and it wasn't accessible easily to the public, but they were there. Mike and I were using them when we wrote our book in 2022. We were seeing where this was all going to go. So, imagine from GPT-3 in 2022, right before ChatGPT, to Astra slash Mythos 5 today, but that level of progress happening in 12 months. So, like, what would that mean between today, August 31st, and August 31st, 2027?

23:50So, what I'm trying to say is, we are on accelerated timelines, probably faster than even I was assuming we were on. And while these models and agentic advancements may not change your work or your business dramatically over the next one to two years, we are racing toward AGI and beyond. And what I mean by that is, like, we're going to get there and you're just going to go back to work tomorrow like nothing happened. Like, it's not going to just fundamentally change your workflow, your team, your company instantly, but there's going to be this change. Then Elon Musk throws in Sunday night.

24:23He was replying to somebody and his tweet said, AI will be able to do anything digital that doesn't require shaping atoms at a superhuman level by the end of next year. So, again, the 2027 timeline, Elon historically is a little aggressive with his timelines, like full self-driving has been coming for 10 years. But he's saying superintelligence by the end of 2027. So, the whole point of all of this is business leaders have no plans where this is going.

24:56We have to reimagine everything. And the people who listen to this podcast are likely on the frontiers of being the ones to figure this out. And you have to be more urgent with what you do in your company and your own career and your families and your school systems. Like, the change is, again, I'm doing my very, very best not to over-exaggerate any of this or create any sort of fear. I'm just trying to help people understand the reality of how fast these labs are moving. And unless something or someone or a group of people, like a government, stops this or slows it down, it's going to just go faster.

25:35Because they're approaching recursive self-improvement and AI researchers that can do the work themselves. And once you do that, you escape velocity on the model capability. So, it's just crazy. Like, I've been thinking – I can't stop thinking about this. Like, this is, like, running through my mind constantly. Like, what does this mean? What do we do? And I don't have great answers. But, like, it's becoming more urgent that more people are thinking about the reality of this.

26:06Yeah, and, Paul, I would just say as one addendum to that, as a piece of really maybe practical advice for folks to take away from this, if you are a business leader listening to this and want to get a really tangible sense and an easy sense of what has changed and how quickly things are about to get crazy, go on your personal computer and enable browser usage in ChatGPT. Go to ChatGPT work in the web app or download the desktop app for ChatGPT, switch to Codex or keep ChatGPT, whatever,

26:41and go have it, look up a bunch of stuff for you on Amazon, do a bunch of research, use your browser. You are going to be blown away by how good this is now. It is not fast, per se, these days. But we've talked about browser usage for the last year or so. A year ago, it sucked. It was not useful at all. You can very quickly get to this idea of understanding what Elon Musk is saying, that it would be able to do any job that requires digital clicking around.

27:12Paired with the level of intelligence the models now have, you're going to see something, I think, with enough little testing here and there that is going to perhaps be eye-opening. Yeah, and as you're saying this, Mike, it has me thinking a little bit more. And I don't want to spend the whole episode on this, but I think this is really important. One, the ChatGPT work also has a browser embedded. So it'll just pop up a window and it has its own internal browser that'll go do this one. But so play this out for a second, Mike, because what we're saying is we use HubSpot.

27:43It's like our powers, our marketing, our sales, our customer success, our operations. It's like a core platform. Imagine like six months from now that you pay for Fable 5.5. Like give me the most advanced Claude model. This is basically what Claude Force is in Salesforce. So this is kind of like the concept. So I'm going to pay for Fable 5.5. Give me the best reasoning model that Claude has, that Anthropic has. And I'm going to pay an amount for that every month for my whole team to have access to that.

28:13But the only thing they're going to use that for is for the planning and the final review and decision phase. So I want the smartest, most powerful model using tokens only for the highest value cognitive tasks. And then it's also going to orchestrate building the subagents to do the work. So we're going to go in and say, hey, we're making this major new announcement. We're announcing Macon 2027. Whatever, I'll just pick something.

28:43Go in and build the whole plan, create everything, then develop all the emails for us, create the FAQs for the customer success team, build the one pagers for the sales team, create the sponsorship. Like, go do this. Fable 5.5 is going to burn valuable tokens to create the plan. Then it's going to spin up what agents does it need to do this. It's going to create them an environment where hopefully the humans oversee what they're doing.

29:13But everything we just heard in the hugging face thing, you give them a goal and then you spin up these agents. But it's likely a human overseeing the smartest agent, which then oversees all the subagents. And it's no longer Mike and I have to sit around building an agent to do SDR work or building a customer success agent, which takes time and knowledge and expertise. The smart agent is going to build those agents for us, and we'll just go in and be like, this looks really good.

29:44It's got a great database of knowledge. It's got really good system prompts. It's, okay, cool. We've got 12 agents that are going to work on the Macon 2027 campaign. How are we going to govern them? Okay, great. We've got some guardrails around them. They're only going to be in this environment. They can't send emails without approval. Sweet. That is not at all out of the realm of possibility. Like that is very, my guess is that's probably what OpenAI and Anthropic are doing internally right now. And once the rest of us get access to those models, that's what will happen. Now, tell me a company, Mike, in the world you know of who has any concept that that's what the next 6 to 12 months looks like and has a plan for it.

30:22Not remotely. That's why I mentioned the browser use thing. You can connect the dots real quick if you're a savvy leader who goes and uses that for an hour. You're going to be like, wait a second. This is doing the exact same things I pay people $100,000 a year to do in HubSpot if you're willing to let it have access to that. Right. And if you disagree with that hypothesis, like if you think I'm crazy, then I understand why you don't think anything happens to jobs in the next year. But if you agree with even directionally that that hypothesis is viable, which it is, like we could prove it to you.

30:58I don't understand how we don't have a more serious conversation about the impact on jobs.

Bill Gates views on labor

31:03Like it just it really doesn't make sense to me. So related to that, a good segue, Paul, into our second big topic this week, which is Microsoft co-founder Bill Gates has spent years arguing that AI would create new work even as it displaced old jobs. Now, this past week, he changed his tune. He said the pace and scope of the technology and how it's developed have changed his mind. He now expects AI to eliminate many jobs permanently and says governments are not prepared for the transition.

31:35So in a new Gates Notes essay, this is like blog or website, Gates argued that this shift is different from past technological revolutions because AI can substitute for cognition across nearly every industry at once. He says that agriculture and manufacturing, for instance, transformed over generations, but he believes AI could reshape much of the labor market in roughly a decade with increasingly reliable autonomous systems taking on work that once required people. He is especially worried about younger workers and people in entry level and mid level roles.

32:06He expects many existing jobs to disappear faster than new ones emerge and no longer assumes the market will naturally produce enough replacement work. He said policymakers need to make deliberate choices about how the gains from AI are shared. As a result, one of the proposals he has is this category of work he calls human reserved. So this is work society chooses to keep in human hands even when machines can do it. He also suggests tax systems to stop favoring automation. So for instance, today, companies pay payroll taxes when they hire people, but they can write off machines.

32:41So Gates argues that governments may eventually need to tax AI tokens, robots, or the profits created by automated labor. Now, Gates says his tone changed late last year. Maybe he's listening to our podcast. Who knows? I doubt it. I doubt it, but we'll see. As AI systems became much better at coding and other complex works. That's kind of that shift we had noted at the end of last year, beginning of this year. He now describes this coming transition as turbulent, says there is no real plan for managing it, and said he would support slowing AI development if the world could create a credible way to do it.

33:18So, Paul, this is a big shift in tone from him. I feel like he kind of framed it as like, oh my God, why is nobody talking about this? But in fact, quite a few people have been talking about this. But regardless, early or late, Bill Gates has a very pessimistic view, it seems like now. What do you think accounts for this? Yeah, I mean, I'll take it. Sometimes it takes someone like Bill Gates saying what we've been saying on this podcast for three years for more people to listen and take it seriously and do something about it.

So I'm glad that he's using his platform to raise awareness about near-term issues because if enough people believe, going back to this previous segment, that significant displacement and underemployment, especially for entry-level professionals, is a real thing, a possibility, then we can get on to doing something about it instead of ignoring it and pretending like it's all just going to work out.

34:09And I actually do feel like we still have time, even if we don't pause. Like, I would honestly, I've gotten to the point where I would be an advocate if we could just slow the hell down. Like, if we could get governments and labs to say, you know what, let's pause training runs for a moment and let's just absorb the technology that already exists and the implications of it, this is not going to happen. But in an ideal world, like, I think that would be a very logical thing to do at the moment because we do not have a grasp on what is happening and what the capabilities are of these agents.

34:45Now, the reason I'm optimistic that we still have time is I have spent a lot of time in the last, like, one to two months meeting with leaders in charge of AI transformation at major enterprises. And I am more convinced than ever that human and organizational friction to change is going to cause this to look way more like a slope than a cliff. And what I mean by that is even with these swarms of agents now being proven to be a real thing, and you can very quickly extrapolate what would that mean within an organization like we just sort of did on the fly.

35:19None of that was scripted. Like, I'm just, like, thinking on the fly, like, what happens in HubSpot? Right. That is a very viable thing. Most organizations are still trying to figure out the AI assistant part of this. Like, and most people are still trying to figure out how to just treat it like a really helpful AI assistant that helps them think and create valuable outputs. They're not doing the agentic stuff at a deep level. And so I think that, like, the more we move down this path of getting more AI literate people, getting more people who understand how to use the full potential of the assistant, getting more actual just knowledge workers, not coders, but knowledge workers overall, starting to experiment with agents, like, that's going to create such dramatic challenges within organizations to absorb how fast people are going to work and how much they're going to produce.

36:12So I'm just convinced that, like, even though the tech is moving really fast, the change is not going to move as fast within businesses. And that's a good thing, in my opinion, at this point. So pull out a couple of excerpts from what Gates wrote. I also watched an interview he did with The Atlantic, which was helpful. Our friend Andy Sack actually sent me that to take a look at. So, okay, he said, the transition to the AI era will be one of the most turbulent times in human history.

36:43Unfortunately, right now, we are not preparing for it. I don't see evidence that leaders, experts, and communities are confronting the challenges adequately. There is no plan to ease the entry into the AI era. That is 100% correct. He said maximizing the benefits is just as important as minimizing the harms. If people see how AI makes their lives easier, it will help build the public trust that is necessary for managing the harder parts of the transition. If the first thing AI does in most people's lives is take away their job or a family member's job, those who are already skeptical about it will outright reject it.

37:17This will make it harder to ever deliver on the benefits, and it is another reason why governments, industries, including the medical industry, and AI companies should be working together now. And then he kind of, again, it's like 6,000 words, but towards the end, he gets into the world needs a plan, and he starts to offer some thoughts. So he said solutions should be developed through a public democratic process that includes elected officials, policymakers, educators, health workers, local officials, and community leaders. Millions of people will have their lives disrupted and will need stronger, more flexible social safety net to help them manage the transition.

37:53Local communities are already raising concerns about the energy and water needed for data centers. Without solutions, some groups will push for stopping AI development and deployment altogether. And then he said in the coming months, he's going to put out more ideas around how to have the benefits outweigh the harms. He does then feature three things he's already thinking about. So the first is creating a democratic or domestic and international framework for dealing with AI. The second, as you alluded to, is this idea of setting aside jobs for humans. And then the third component was rebalancing how we tax labor and capital.

38:27So he said, I believe we should tax AI tokens and robots. Right now, if you're an employer and you hire someone, you pay payroll taxes on their earnings. But if you buy a robot, you can usually write it off right away as a business expense. The tax system nudges you toward replacing people with machines. So that's an idea I think I had talked about last year is that if you're going to get rid of people, then you need to be paying taxes on the uses of the tokens and the technology itself that you're using to replace the people. You don't get a bypass to your contributions to society.

38:59So, yeah, I think there's a lot of other ideas that need to come out. And I just, like I said, I think the most important thing is that we're more high profile people who have more influence than we do are starting to say, hey, we should be doing something about this. And that's good for society.

NVIDIA quarterly earnings

39:18All right. Our third big topic this week, NVIDIA is just having, has delivered another staggering quarter this past week. So revenue, they reported, reached $96.2 billion. That is more than double the same period a year ago. Net income climbed to $59.7 billion. Earnings came to $2.46 per share, rather, up from $1.08 a share a year earlier.

39:49The company projected about $108 billion in revenue for the current quarter. That is ahead of Wall Street's expectations. These numbers show that the demand for AI infrastructure, according to them, is outrunning even their rapid expansion. So operating expenses rose 55% as the company invested to keep up. Founder and CEO Jensen Wong said NVIDIA currently has enough supply to support roughly 70% revenue growth next fiscal year, even though customer forecasts point to enough demand for revenue to roughly double.

40:21Wong has said AI has reached an inflection point where more computing power is translating directly into more revenue for customers. Now, at the same time, the information reported that NVIDIA is in talks to acquire Hugging Face for $12.9 billion. The deal has not been publicly confirmed yet by either company and could still change, but it would be NVIDIA's largest acquisition. Hugging Face has become a central hub for open AI models, data sets, and evaluation tools, also a prime target for agents trying to hack people, apparently.

40:52The company was founded way back in 2016. It was last valued at $4.5 billion in 2023 and recently reached about $150 million in annualized revenue after growing that figure by 50%. So owning Hugging Face would kind of give NVIDIA a direct position inside the open model ecosystem that's used by developers and research teams. It would also extend NVIDIA beyond chips and cloud infrastructure into more software distribution and community layer where many open AI projects begin. So, Paul, you're a long-time NVIDIA watcher. We talk about NVIDIA all the time.

41:24Talk to me about these earnings and their acquisition of Hugging Face. I mean, they just continue to apparently print money. Yeah. I mean, the reason we tend to give a lot of weight to every three months when NVIDIA reports is because it's a primary signal as to whether the AI growth curve is declining, maintaining, or rising. And so when they crush like this and then give updated guidance and Wall Street responds in a very positive way, like I think their stock was up 8% to 10% last week, which for a $5 trillion company is not an easy task.

41:59So right now, it would appear that we are on track, that the AI growth curve is continuing. There's no signs of weakness and demand for NVIDIA's chips, which are used to train and run the AI. There's no lack of demand for their AI factories, which enable them to power the chips. And so it looks like, you know, years into the future, there continues to be almost an insatiable demand for AI training and inference. So that would indicate we are still very early in the demand curve, which has been my assumption all along, because I just I look at it as when I talk to family and friends, are they using advanced reasoning models and agents?

42:39No, like they're still treating, you know, ChatGPT and Gemini and Claude in their personal lives as just an answer engine and maybe an AI assistant to do some planning stuff, but pretty limited. And then when you go talk to these major enterprises, same question, are they using agents across all functions of the business or is it just like increasingly in the coding realm? The answer is usually it's just within the coding realm. They're still messing with, you know, codecs and things, cursor, stuff like that, but it's not like the marketing and the sales and the ops people are using agents.

43:13So it just feels like it's very, very early. And yet NVIDIA's second quarter revenue jumped since 2023. So if you look at their second quarter revenue, the first quarter after ChatGPT came out, so spring of 23, their revenue was $6.7 billion that quarter. Their revenue this quarter was $46.7 billion, so almost a 600% increase. And then one of my favorite data points I'll go back to at times when I'm doing like presentations on NVIDIA or mentioning NVIDIA is where NVIDIA was at the day before ChatGPT came out and where they are today.

43:53So on November 29th, 2022, so that we were rewinding to the day before ChatGPT was introduced to the world, the market cap for NVIDIA was $400 billion. Today it is $5.4 trillion in three and a half years, basically. So, yeah, just to give you a concept. And now they're throwing their weight into the open source world. Like, I mean, they were already playing in it, but it seems like they're just directly coming after their biggest customers.

44:24So OpenAI, Anthropic, Google, like everybody uses NVIDIA. And now they're just saying, well, we're going to go build the open weight ecosystem because we can't rely on the closed labs to do it. And NVIDIA wins either way because their basic premise is, listen, I don't care if you're using open models, closed models, doesn't matter. You're still going to use our chips to train them and to run them. You covered the hugging face thing. I mean, they're moving more into the open source realm, so that's a big key here. They paid a lot for it.

44:55Like, the information said that hugging face isn't generating a lot of revenue. I mean, that's relative, but it said it was recently generating $150 million in annualized revenue, which is typically a figure that multiplies the prior month's revenue by 12. So in that scenario, NVIDIA is paying about 80 times the forward revenue of the startup, which is quite high. And then the final note I'll make is Jensen, who we've said recently joined X, I think it was in July, he did his first post on X.

45:27He weighed in on the data center argument. So given how much we've been talking about data centers, I thought I'd just throw this in here. And data centers are a huge part of NVIDIA's growth. They cannot have data centers being stopped if they want to keep growing at the rate they are. So Jensen was actually responding to a Gavin Baker post. We talked about Gavin Baker last week. And Gavin said, in relation to data centers, there were reasonable concerns about data centers 18-ish months ago. Water, taxes, jobs, electricity prices, the environment, and what they would do to small towns.

46:01Well-structured data center projects have largely addressed these concerns today, and we should be celebrating this. And again, this is Gavin still. On balance, data centers are awesome for America in every way. Data centers are actually re-industrializing parts of America and creating the kind of working-class jobs both parties have spent decades claiming to support. That should not be a partisan issue. Data centers can and should be awesome for America, and they increasingly overwhelmingly are. Supporting the outsourcing of data centers to China will likely age just as well as support for outsourcing of high-quality, blue-collar manufacturing jobs to China has aged.

46:41So that's a synopsis of what Gavin posted, to which Jensen added, AI is bringing manufacturing back to America and re-industrializing the nation after decades of offshoring. AI is creating demand that drives investment in our aging power grid and sustainable energy powered by market forces, not subsidies. AI is creating construction and manufacturing jobs across energy plants, chip fabs, and data centers. AI is creating new companies and industries. $400 billion has been invested into AI startups in the past six months alone.

47:14Builders must partner with communities to build in their hometowns, earn trust, and create local benefits. We have an opportunity to create lasting benefits for communities across America and help America lead the next industrial revolution. So that is going to be the talking points. The advocates for data centers, the advocates for AI, they need to win the trust of Americans and convince them that this is good for America and that we are solving for the problems these data centers create. We're not ignoring it like we maybe were 18, 24 months ago.

47:46And like I've said before, he's right. Whether you want data centers in your backyard or not, the investments in data centers is what's driving the economy right now. And it is creating tens of thousands of jobs. So it's hard to say we should just stop because it's the best thing America has growing in terms of growing the economy right now is data centers equal more intelligence, more AI, which America is leading the way in right now.

48:16And if we shut down data centers and if we stop developing these models, America loses its lead. Yeah, and I think we'll probably talk about this ongoing, but I believe we saw this past week or two certain trade labor unions have kind of come out against anti-data center political candidates, for instance, which makes perfect sense just given the incentives there. But yeah, the conversation is evolving, interestingly. And you're seeing actually like Meadow, who obviously was all in on the Trump administration just last week, started diversifying their contributions to political campaigns.

48:52And so what I think you're going to see is these AI labs, the big tech companies, they're going to give money to whatever politician on whatever side of the aisle who supports data centers and AI. Right. And that's it's a it's a tough one because the public doesn't. And so there's this there's this moment where we got to figure this out. But like the labs are going to put money behind Democrats, Republicans, independents does not matter. You have to support data center. All right. Before we dive into our rapid fire topics this week, just one other quick announcement.

49:23This episode is also brought to you by AI Academy by SmarterX, specifically our AI for departments, courses and certificates. So AI Academy by SmarterX helps individuals and businesses accelerate their AI literacy and transform through personalized learning journeys and our AI powered learning platform. New educational content is added weekly to the platform so you can always stay up to date with the latest AI trends and technologies. This AI for departments collection we've got, it features eight course series and certificates designed to jumpstart AI understanding and adoptions.

49:59We cover all the major departments here. We've got AI for marketing, sales, customer success, HR, finance, operations, legal and IT. These are the perfect launch pad for organizations that want to level up their teams and accelerate AI adoption and impact. We have individual and business account plans available now where you can buy single courses and series for one time fees. Go visit academy.smarterx.ai to learn more and use code POD100 for $100 off any individual plan.

50:29All right, Paul.

Anthropic revenue estimates

50:30So first rapid fire topic related. Anthropic is reportedly preparing to tell potential IPO investors that AI represents an annual revenue opportunity of more than $30 trillion. That number is its estimate of the total addressable market or TAM of what AI is capable of doing. That is not a forecast of what Anthropic itself expects to earn. So this estimate attempts to capture the value of work that AI could theoretically perform across industries rather than current spending on AI products.

51:05So this is basically an estimate of the upper limit of the market Anthropic believes AI technology overall can reach. Now, the Wall Street Journal has reported that the company is building this case as it considers raising as much as $100 billion at a target valuation of roughly $2 trillion. We could see a prospectus arrive within weeks. There's a possible offering as early as September or October, though these plans are not final. Anthropic reportedly generated $11.6 billion in revenue during the second quarter.

51:37That's more than double the previous quarter, while its annualized revenue run rate later climbed above $65 billion. The company is projecting between $190 billion and $200 billion in revenue by 2028. So, Paul, $30 trillion, it's not their revenue forecast. It is still an almost absurdly large number to put in front of investors. For context, U.S. GDP in 2025, as measured by the World Bank, was about $30 trillion. So world GDP was about $118 trillion.

52:09So we're talking some pretty big numbers of their estimates. Obviously, they're talking their own book quite literally. But what do you think of that number and their approach here? It's really hard to just comprehend this. The only other number I'll throw out to give some level of context to how ridiculously large this number is, is total U.S. wages and salaries reached a seasonal adjusted annualized rate. So this is like what over 12 months, roughly, the wages would be in the U.S. of $13.4 trillion as of July of 2026.

52:40So, again, between the GDP and the total wages, I don't even know how else you conceptualize $30 trillion. And GDP is not an easy thing to conceptualize. So wages, to me, is like I guess it's a closer thing. And there wasn't any real details as to how they arrive at the $30 trillion. But my assumption is that a large part has to do with doing most of the work that otherwise would be done by humans. Yeah, I was going to say I pulled some of your comments from past episodes because you nailed this.

53:17Because in the past several episodes, especially $172, $173, $196, $204, you've talked about plenty of ways where it's like these numbers, these valuations only make sense to investors who are not idiots for the most part. If the total addressable mark is human labor, not software spend, right? Yeah, because I think what we've said is like the SaaS industry, so all software combined is like $300 to $500 billion a year in sales. So, yeah, you're not getting to $30 trillion by just replacing the need for software companies.

53:50It's like you're going after all human labor, including blue-collar jobs, including humanoid robots. Yep. And to your point about, I'm not sure where this number came from if you map it all up, but SpaceX in their recent IPO, they had about $26.5 trillion as their estimated TAM for AI. So, again, maybe they were more specific, but in the same ballpark. So, opening eyes is obviously going to be like $32 trillion, but we just got to keep going up. We cannot come in under $30 trillion now.

54:21A hundred percent. All right.

Stanley Druckenmiller and AI

54:24Next up, some more Anthropic news this week. This past week, Anthropic won a major legal victory when U.S. District Judge Rita Lynn ruled that the Pentagon's effort to blacklist the company as a national security supply chain risk, which we've talked about earlier in the year, that she ruled that this was unlawful. So, the dispute here began after the Pentagon demanded the right to use Claude for any lawful purpose. We talked about this quite a bit. Anthropic agreed to remove most restrictions on Claude usage, but refused to authorize two big uses, mass surveillance of Americans and lethal autonomous warfare.

54:57At the time, Defense Secretary War Secretary Pete Hegsitt then designated Anthropic a supply chain risk and barred military contractors from doing business with the company, even on work unrelated to the military. So, in this 59-page ruling, Lynn found that the government retaliated against Anthropic for protected speech in violation of the First Amendment, denied the company the process required by the Fifth Amendment, and acted outside the federal procurement law that it invoked. Lynn wrote that the government wanted to make an example of Anthropic for criticizing its AI policies while continuing to pursue work with the company.

55:34She said the conduct was not consistent with a genuine belief that Anthropic might sabotage its software or its work. The ruling blocks the Pentagon designation and related government-wide directives. The government is expected to appeal, and a separate case involving another Pentagon designation remains pending in the D.C. Circuit. So, Paul, we're finally getting a little movement on this. I'm curious, does this ruling, like, meaningfully, like, resolve this issue? Obviously, government can challenge it.

56:04Is it just the next round we're waiting for? Is this going to kind of go away like we've somewhat predicted? Yeah, I guess they knew it was illegal when they did it, and they didn't care then, and I'm guessing they don't care now. And they'll just have their lawyers draw this out for a while, but I don't think it came as a surprise to them that what they did was illegal. I'll just leave it at that. All right, next up, billionaire investor Stanley Druckenmiller sparked a big debate over AI disclosure this past week after acknowledging that he used AI to help write a Wall Street Journal opinion piece that criticized Treasury Secretary Scott Besson's approach to the bond market.

56:46So, he writes this op-ed in the Wall Street Journal, and in fact, it started getting flagged on social media. It's like, hmm, this sounds a little bit like it might be AI written. And so, the media outlet Notice, N-O-T-U-S, asked Druckenmiller whether he had used AI to write the article. Druckenmiller said, of course I used AI. He added, he now writes everything with AI for the same reason he uses a calculator to do math. The admission of this came the day after the piece was published. The journal had not disclosed any use of AI in the op-ed, and the Wall Street Journal came out and said the article did not violate its policy.

57:22So, its editorial page editor and VP, Paul Giggoat, defended the decision. His standard, it was that whether an author brings an original argument and the credibility, his standard is whether an author brings an original argument and the credibility is supported, not exactly like how they chose to write it. So, in this case, he said the opinion here was genuinely Druckenmiller's, which other people had commented on. This pretty much mirrored quite a lot of his commentary over the years. Even if AI helped shape the prose, that's okay.

57:53The journal's opinion section, importantly, is separate from its newsroom, which has its own AI standards. So, this kind of kicked off, Paul, some firestorm of debate, especially among journalism, but also more widely, does it matter that someone used AI to write an op-ed? If it's clearly Druckenmiller's, like, long-held opinion, this was kind of charged, too, because him and Besant used to work together. They were very closely linked. So, it was, like, pretty serious criticism of Besson's policy towards the bond market, which Druckenmiller had been pretty consistent on previously.

58:28But the AI thing is, like, overshadowing all of that. What did you kind of take away from this? So, this goes back to episode 232, where we were talking about clawed watermarking and how people are going to start getting called out for their use of AI, and people are going to treat them, you know, in this case, like it's, you know, it's like a really bad thing that they did it. And my point then was, like, it's just going to be the norm. Like, everyone is going to be using AI in some capacity, and we're going to want, as a society, to sort of arrive at a point where it becomes a little bit more accepted.

59:03You may need more transparency. So, again, this is part of why AI policies are so critical within organizations is when do you disclose it? Like, when are you allowed to use it? But then when do you need to disclose the use of it? Like, hey, this was co-authored with ChatGPT. The ideas are mine, but, like, ChatGPT helped me form my words. Not everyone's a great writer, but everyone has ideas and thoughts that need to get out into the world, hopefully. So, I just feel like we're getting to a point where we need to start agreeing on that it's going to be normal and that people are going to use it in different capacities.

59:42And I don't know that we should judge people. So, I've said, like, myself, the vast majority of what I do, I don't use any AI to do it because it's, like, I want to go through the process. But then there are things that I will use AI to do because getting the information out faster to the world where I'm in the lead on it, I'm the one, like, pushing the ideas and driving it, that's what matters. And I'm verifying everything and editing everything. Like, I'm the human in the lead in that scenario, not the human just in the loop, like, who's just kind of letting AI do its thing and watching over it. So, I just feel like we need to decide more around, you know, policies within organizations and then more societally.

1:00:19Like, it's going to be easier and easier for people to say that they think AI wrote something and they're going to throw, you know, it into Claude or they're going to throw it into Turnitin or whatever. Maybe it was AI, maybe it wasn't. But, like, we got to get past the point where that matters and just, like, talk about, but is it that person's ideas? Did they put critical thought into it? So, yeah, it's just interesting. So, if you didn't go back and listen to episode 232, you can go hear all my thoughts on this because that was, like, I tried to sort of explain this idea of AI plagiarism is the bigger concern for me where AI systems, ideas, and words are being presented as your own and you put no critical thought into it.

1:00:55If you're the expert on a topic and you worked with AI, you're still putting your ideas and words into this, even if AI is helping you craft them. It's still your expertise, your experience that's coming through in the writing as long as you're not just letting AI put words into your mouth. Yeah, I think it's a super wise perspective to move beyond the is it or isn't it because this feels like just on steroids the same version of this debate over, like, ghostwriting is a very big thing. I used to professionally ghostwrite. Right, whether you agree with that or not is one thing, and that's fair, but I can tell you for a fact, if I took X percentage of the New York Times bestseller list, if those are celebrities who don't usually write.

1:01:38Or CEOs. CEOs who don't. If it is not literally on the bestseller list because the person is a regular author, I can almost guarantee you that it is ghostwritten in some capacity. Now, the nature of that ghostwriting is what matters, right? Yeah. It's like, did they actually go back and forth with each other or was it just someone saying, like, write whatever and I'll sign off on it? So, yeah, I don't know. It's interesting. We didn't have, we don't seem to have had as much outrage over that happening over the last three decades. Totally. And again, I think it's like, some people will argue with, is the ghostwriting the right analogy?

1:02:11It's like, well, it's the closest thing we've got. You have ghostwriting, which I guess some people don't know is like a thing, like, but a lot of CEOs don't write their own stuff, whether that's memos to their teams, whether it's books, whether it's certainly talks they give, like, there are speechwriters for that stuff.

1:02:30Musicians, many of them aren't songwriters, they're singers, so they don't write their own songs, like, but does it represent their feelings and emotions and beliefs? And, like, does it come through because of interviews and things they've previously written? That's, at the end, what matters. And I think, like, AI, we're just in this phase where it's being treated differently than helping give a voice to people when otherwise. But, yeah, I'm with you 100%. Like, if a CEO just says, yeah, write me a, you know, write me a book that we can put my name on and get on the bestseller list, and the CEO never even read their own book.

1:03:01Sure. They might as well just use the guy to write it at that point.

McKinsey State of AI

1:03:04A hundred percent. All right, next up, McKinsey has released their new State of AI in 2026 report. Some interesting findings here. So they find that nearly nine in ten respondents said their organizations now use AI regularly in at least one business function. Forty-four percent said they were scaling it across the enterprise, up from 38% last year. Eighty percent of respondents, in terms of employees, reported individual productivity gains. Half said AI is helping people make better decisions.

1:03:35But only 30% reported any impact on earnings before interest and tax EBIT.

1:03:42Essentially, that's essentially unchanged from last year. Interestingly enough here, the group McKinsey calls AI high performers remains just 6% of respondents. But these are more likely to fundamentally redesign workflows, pursue growth, and or innovation alongside efficiency, and pair AI initiatives with stronger senior leader commitment and defined impact measurement instead of simply adding AI to existing processes. The share of respondents from large organizations reporting that their companies were scaling AI agents jumped from 27% to 40%, but the share among respondents from smaller organizations stayed flat at 22%.

1:04:22At the same time, about one in five respondents said the operating cost of AI is now a constraint. 39% expect AI to reduce their organization's total employment over the next year, but only 14% of respondents from organizations using AI say AI had contributed to an overall workforce decline in the past year. So that is actually less than half of the people, the 32% who in last year's survey predicted decline. So a bunch of people predicting, but it's not yet coming to pass.

1:04:54One final note here I found interesting. Nearly one-third also said agentic coding led them to reject at least one software purchase because they could build the capability internally. So, Paul, I think it was interesting to see how some of this mirrored our own data, like literally validated what our state of AI for business research showed, individuals more racing ahead, organizations struggling to still see impact. I also love just that how they outlined these, even though it was 6% of respondents, the AI high performers, reimagining workflows, pursuing innovation in tandem with efficiency.

1:05:27It sounds very familiar to kind of how we've approached it and what we've talked about. Yeah, the scaling agents jumped out to me. It's like, wow, that seems really high. I wonder how valid this is. And then I drilled in and realized they were specifically referring to software coding agents. And it's like, oh, okay, that makes more sense. If you're asking within IT, within software development, hey, are you all scaling coding agents? Yes, like 30% seems like in a reasonable number. If that question was asked across other functions, not a chance.

1:05:58Right. So, yeah. And then the high performer categories that you highlighted, they actually have a chart that sort of breaks down, like, what are the different things and what is the difference between all other respondents and the high performers? And I think they've got, I don't know, there's like a dozen different categories here of the things that separate the high performers. So you can go check out the full report and read up on those. All right.

Future economic scenarios

1:06:23So next up, we have an interesting new essay from the historian and economist Niall Ferguson in the free press that is kind of about what future we can expect from AI. So he actually publishes this as almost a direct challenge to Elon Musk because Elon Musk has this kind of favorite vision of the AI future. And the way he described it in a series of interviews is what they would call the post-scarcity civilization. This is a sci-fi set of sci-fi novels, Ian M. Banks's culture novels.

1:06:58Basically, in these novels, super intelligent machines manage society, work is optional, material abundance makes money largely irrelevant. And Musk had told the economists that AI could exceed the sum of human intelligence in roughly five years and paired with vast numbers of robots create a quasi-infinite economy. So that's kind of his vision of where this is going. Ferguson, however, argues that that prediction rests on assumptions that unfortunately do not survive contact with physics or economics. He says that more capable AI still depends on scarce physical inputs, talked about this, advanced chips, electricity, data centers, copper, rare earth metals, land, and infrastructure needed to turn digital intelligence into goods and services.

1:07:40So intelligence may become abundant without making energy minerals or land abundant. Ferguson says the better model is actually the work of science fiction author Neil Stevenson, specifically his book The Diamond Age. So in that book, The Technology Does Not Erase Human Conflict or Inequality, it basically enriches a techno-elite, leaving a lower class dependent on basic income, and produces unintended consequences when powerful systems are monetized and misused.

1:08:11Bringing all this full circle, he basically says, look, the open AI hugging face incident is evidence that this messy future, not this utopia of abundance, has already begun. So, Paul, there's definitely some sci-fi inside baseball here, but it's interesting. Ferguson, a pretty well-known academic voice, kind of just saying, I think he largely aligns with Musk on AI being beneficial in a lot of ways, but basically saying, look, we're not getting this thing where these benevolent, super-intelligent machines make everything perfect.

1:08:41I'm curious, like, kind of what you thought, looking at his vision versus what we're being told through other channels. I think his vision is more realistic, like, especially in the coming decade. You know, the culture stuff is very common within the AI circles. Like, I think there's a lot of AI leaders who are sort of inspired by that vision of the post-scarcity, you know, abundance thing. And just sort of use that as a way to justify everything they're doing and the risks that come with it and the downsides that come with it is that it is in pursuit of this larger utopian-like outcome.

1:09:19I don't think that that's certainly what the next 10 years looks like. I do think there's going to be lots of real obstacles to that future. And I think it's just going to be messy. And so I've even said on jobs, like, I do think somehow this works out from the standpoint of the economy, from the standpoint of jobs. I think more get created. I think there's going to be really cool new roles that emerge. And, you know, we're seeing signs of this ourselves, like jobs that just didn't exist a year ago or two years ago that are really cool jobs that are, you know, popping up or that are being created even within our own company.

1:09:51But I just feel like more broadly, it's just going to be messy and be part because what Bill Gates is saying, there's no plan. Like no one is solved for what happens when it doesn't go great. I just, and I don't know if it's because it's hard to figure out or because they just want to pretend like it's not going to happen or they truly are just have convinced themselves that the abundant future is the only outcome. So, yeah, I don't know. I mean, it's, I know some people love sci-fi. I did not have that book on my to read list, the Diamond Age.

1:10:21So, and I have not read Stevenson's stuff. So, I might go check that out. I've tried the Culture Series. I know you've been into that before, right, Mike? You've done the Culture Series. Yeah, I've read a few of those, yeah. Yeah, I've tried. And I usually get like a couple chapters in and it's like, I don't know, I'm going to go listen to a podcast now. Like I haven't really like fully invested in them. Yeah, it's very out there when it's like almost like the humans aren't the main characters in the books, basically. Is that foundation? Is that one of the Culture Series? No, that's, yeah, Isaac Asma. Okay. So, that's a little, but it is similar in the like big grand scope of everything, right?

1:10:55You know, it's just really interesting. But yeah, no, we'll see. But I'm glad someone is maybe saying, hey, it's going to be a little messier than maybe those books. And again, like I would love to just focus on the optimistic stuff on this podcast all the time and like super practical stuff. We try and balance it with like the practical stuff, but I think the sooner we just accept it's going to be messy and we don't have all the answers and like we're not asking the hard enough questions, then we can move into the phase where we actually do that. And then it's more addressable, like pretending like this is just going to go perfect and AI is only good and amazing doesn't do anybody anything.

1:11:32So, it's like, I don't know, like, again, I've tried to not be overly, create like anxiety and fear around this stuff. But I think there's just realities people need to face so we can get on with figuring it out. Yeah. And, you know, I think in a more positive frame, it leads to ownership and agency. Like, nobody is coming to figure this out for you or save you on this. And, like, I get that's scary. But, like, also, it's like, okay, let's take a breath, move past that, and figure out what does this mean for your family, your career, et cetera.

1:12:03Yes.

Public policy polling data

1:12:04So.

Public policy polling data

1:12:05All right. So, bringing this back down to earth a little bit, next segment, this past week, the Center for Shared AI Prosperity released the results from a national survey they conducted with an organization called Blue Rose Research. They pulled 56,000 respondents on 79 ideas for how government could respond if AI disrupts large parts of the economy. So, the proposals they suggested covered four broad areas. First was taxation and revenue. Second, income support and safety nets.

1:12:35Third, labor markets and workforce development. And fourth, models for ownership and governance. So, these proposals range from taxes on AI profits and data dividends to wage insurance and even a U.S. sovereign wealth fund. For each proposal, respondents read short arguments for and against it. Then they had to choose whether they supported or opposed it. There was no undecided option. The organization reports net support is the key metric here as the difference between support and opposition after those arguments.

1:13:06Of the 79 policies tested, 61 still had positive net support after respondents saw the opposing case. The 12 most popular maintained approval margins of at least 40 percentage points. And 48 policies had margins of at least 15 points. So, among the policies that directly mentioned AI, 31 of 48 won majority support. So, the organization says respondents were especially receptive to the ideas of retraining and compensating workers affected by automation,

1:13:39strengthening the existing safety net, and funding training, apprenticeships, and care work through progressive taxes. So, the group also does caution that a popular policy is not necessarily an effective one. It describes the survey as a starting point for comparing ideas and understanding where bolder proposals may face political resistance. So, Paul, this kind of seems to suggest that Americans are open to active government responses to AI disruption. So, the top five measures that they specifically suggested that had net positive support were expanding apprenticeships, requiring severance for automated away jobs, sector-based job training, employee ownership in firms, and a data dividend.

1:14:21And just one final note here, important to note that Center for Shared AI Prosperity, they're a U.S.-focused research and policy coalition. They're pretty clearly aligned with center-left progressive kind of democratic policy ecosystem. Blue Rose Research is explicitly democratic and progressive aligned. So, basically, it seems like we're doing some polling research here to figure out what policies are going to look good in the next few years. Yeah, so, and this does kind of tie back to the Bill Gates topic where it's like, what can we do about this? Like, if AI is going to impact us, what are the things that we can move forward on where we can make progress?

1:14:56I love to see the apprenticeship thing. It's actually what first caught my attention about this because that's the basis of my Macon talk in October is the need to build apprenticeships across industries and job functions. So, if you go to the blog post in the show notes in that post, and we'll put a link maybe directly to this, there is a Google worksheet that you can actually go read the descriptions of each of these and then the pro and the con argument. So, it's really cool to, you know, just kind of a quick glance. So, I went through and pulled out a couple, Mike, just to see more context of what were they, you know, expanded apprenticeships is fine, but like, what does that even mean?

1:15:33So, I'll just read, let's see, I got three of the ones that got the positive net rating and then two that got the negative. So, expand apprenticeships says, scale up, earn while you learn apprenticeships in new fields, including AI infrastructure and roles that work alongside AI. Under this policy, more workers could train for skilled jobs by working a paid position under experienced mentors. I'm more and more convinced as I work on this opening keynote for Macon that, like, this is a very real thing that should be happening in a more structured way.

1:16:06Require severance for automated away jobs. I didn't know what that one meant. So, this would require that companies provide severance pay or transition support to workers they replace with automation. Under this policy, a company that eliminates a job through AI or automation would have to give the affected worker a defined payout. So, I actually agree, and I presented that as an idea last year that you should, rather than unemployment, if you took someone's job with an AI or a machine, then you should pay them for some extended period of time. So, I'm actually an advocate of that concept.

1:16:38And then the data dividend, some policymakers are proposing that tech and AI companies pay people for the personal data they collect. Under this policy, the data companies use to build and train AI would be treated as something you own and firms would have to compensate you. That actually is an Andrew Yang idea we talked about on a recent episode, who also is speaking at Macon. I'm guessing we'll address that one. Now, the two that jumped out to me that had pretty low net ratings, one is tax on automated services. So, this would be a tax on automated services people pay for, things like robotic delivery, self-checkout, driverless rides, and AI customer support, the same way sales tax apply to other purchases.

1:17:18So, the tax would fall on the automated service when it's used, rather than on the machines of the software. And then the one I thought maybe was most interesting, Mike, because a lot of the AI lab leaders have thrown out this idea of universal basic income. And actually, Andrew Yang's book in 2020 was based on this idea of universal basic income. People do not like this one. So, it said, some policymakers are proposing a universal basic income, a set monthly payment sent to every adult. Under this policy, everyone would receive the same regular cash payment for the government. That was the second lowest rated of all of them.

1:17:50So, people are not UBI fans, at least in this study. It doesn't seem like it. Yeah. I'm just curious. Do you see any – I mean, it's super relevant regardless of who's running it. Do you see this as putting out feelers for possible future policies or just more benign research? I would guess it's campaign messaging, like this close to it. Like, they're doing ongoing research. But my assumption here is you're trying to, like, not only do you want to hit the other people with the things that you think people are not going to like. So, hey, they support data centers and you're spending millions.

1:18:21But it's like, hey, we support expanding apprenticeship, which is going to continue to give you the jobs and give you the opportunity to learn. We support dividends if you get – like, you're just looking for the things to say. Here's the things we're all for. We have no plan, but, like, we like these ideas because you like these ideas. Right. All right.

AI use case spotlights

1:18:40Next up, we've got our AI use case spotlight we do every week where we give you a quick look under the hood at some real AI use cases we're exploring in our work at SmarterX or, in some cases, in our personal lives. So, I've got one to share, Paul. Then let you share whatever you've been working on this week. So, first up, I have been working on this for a long time, but it kind of finally came to fruition. It's more an interest project rather than something super specifically useful, though it turned out to be extremely valuable.

1:19:10I actually built, like, a podcast expert briefing system in Codex. So, you know, I was trying to basically solve this pretty straightforward problem. It's like every week I do a ton of extensive prep for the podcast, read everything, make notes, think through stuff, try to look up different research and sources. But, like, you know, we have limited time, and I'm trying to basically really become an expert, like know enough to be dangerous, more likely, on every one of these topics. Not just the news item, though that's interesting, but more like what does it mean, what are the implications, what are some of the technical aspects of it that sometimes you don't always have time to dive into.

1:19:48So, what I did this week is I finally took a little time to take my notes on this, take all the other little processes manually and with AI that I use to kind of prep for each week, and put these into, try to make, like, an actual briefing. Like, for each topic, what do I actually want to know, how should it be formatted, what should it look like, what are the key things I want AI to go do research on. So, basically, I created this, like, executive briefing or this briefing on each segment, which includes three parts. There's this executive briefing, which gives me this compact reported narrative of what happened, uses a ton of separate research.

1:20:24This is not, like, meant for, like, listener consumption. It's more like, here's all the different angles, here's what's been reported. You can kind of, like, think through and sort through it on your own. It's like, I equate it to, like, I like the publication The Economist quite a bit. I'm like, write me an economist article on this exact thing, a thousand words that, like, covers every possible aspect of it. It also has, interestingly enough, this is super helpful, a section called What Others Are Saying. This is the second section of this three-part brief where it basically, go find me the hot takes.

1:20:55Go find me what the experts are saying about it. I'm just curious, what are the opposing views? What are people validating? Just kind of give me a subjective sense of that. And then, finally, is the technical briefing. So, for each of these topics, it actually will go in and identify one or more technical aspects or areas, or, like, even in some cases, like, policies, technologies, how things work that I need to know about, and build me, like, a several-thousand-word narrative about, explain this to me, like, a non-technical person or a non-policy person or a non-legal person.

1:21:30And so, what's really cool is, like, took a few hours, but I got to, like, a really usable format for this. I found really, really helpful to prepare for this week's episode. I'm probably going to use moving forward. But between the however many topics we covered this week, this thing will automatically now do the research and produce. I'm looking at it now. It's in our brief. It's 21,000 words of briefing. So, about a third of a book, right, which, and for whatever reason, I like reading it in this format. Format may be terrible for someone else, but this is just kind of customized to me.

1:22:03I went back and forth iteratively on a single topic to get the format down and went tons of iteration, like, back and forth. It took a long time. Once I got the segment, like, perfect, I then said, okay, go do it for everything else. And within, you know, I took an hour of codex working, burned a lot of my usage. Thankfully, OpenAI is throwing out usage resets for codex, like candy, which is amazing. So, I got lucky on that. But it burned a lot of tokens. But, man, it's like, reading through it, I'm like, I basically created my own personal magazine for the podcast this week.

1:22:38And it's, I really enjoy reading it. Well, and then you could throw it into the notebook, Google Notebook, and have it turn it into, like, a 10-minute podcast version or create mind maps for you. Yeah. Yep. Now, why codex? Why not ChatGPT work? Could you not have achieved the same thing using ChatGPT work? Yeah, you could do this same exact thing. And this is, I think you could do most of the same stuff. I just happen to have gotten used to using codex over the last four months. And this, we got to do more of a segment on, like, the differences between ChatGPT work and codex, because, honestly, I don't know them all.

1:23:10They seem very similar to me. I have a great article. I don't know if I put it in the sandbox for next week yet, but I actually read, I forget the guy's name. He's a big AI guy, but Simon Wilkinson or Williamson. Oh, yeah. Yeah. He did this whole, and I was like, okay, I have to figure out what the difference is between Chat, work, and codex. Here's my best effort at it. I'm glad I'm not alone, because, like, I use them all the time, and I'm still like, I don't understand the decisions that went into, like, separating these. Yeah, he said, like, it starts off with, you know, I went to OpenAI's site, and I read the description of what ChatGPT work is for, and I thought, that's odd.

1:23:47I've been using Chat to do those exact things for, like, six months. So, yeah, it is not clear. So, if you're confused, like, join the club. Like, everyone's trying to figure this out. Yeah. In the interest of time, mine is, there's a couple of major things I've been working on for a while.

1:24:06Both of them will probably come to light in the public in the next, like, 60 days-ish. Which, I'm at the point where I have to make very important decisions for the company, like, the future of our business and the trajectory and things like that. And I rely very, very heavily on a couple of ChatGPT projects in particular to think through these decisions. And so, I have to talk to attorneys and accountants and other advisors, and I could not do it on my own.

1:24:38Like, it's just in the timelines I have to make these decisions and to think through from every angle, make sure I'm asking the right questions, make sure I feel confident in the end decision that I end up making. So, it is a human in the lead thing, like, it's just me going in and saying, okay, I thought about this. What do you think? Here's the advice I'm getting. What should I be considering here? What should I go back? So, I am very much in the midst of doing these two major things, and it's consuming a lot of my time and a lot of my brainpower.

1:25:10And in the end, a lot of my token usage is, like, helping me solve. So, decision-making is the front and center use for me right now, assisting in decision-making. I'm just curious for that, with whatever you can share, you know, there's a lot of people still skeptical AI can reason at an expert level sometimes. It sounds like you're using this for very complex things, and you would argue that it's probably at human-level reasoning for this? It's beyond. Yeah, I mean, like, so I present, and I'm very transparent with the use of this with my advisors, and I will share direct outputs with them and say, here's what ChatGPT is telling me to ask you.

1:25:51Here's the open items. It's identified within the legal documents. And almost every time, it is things worth discussing that maybe we hadn't gone through previously. Wow. And so, as long as you have this open relationship with your advisors where they understand that's how you're going to think about things, it just breaks down that barrier right away, and no one gets offended. It's like, hey, man, we just need to make decisions, and let's make the best decisions. I'm still not clear on three things that you're asking me to decide on. Like, I don't have enough information, so I went and did more research myself, and here's what I think now.

1:26:27So, yeah, I mean, to the point where I've submitted legal briefs to my attorneys. Like, here's the wording I would use. Like, this seems better than what we were doing. And so, there's a lot of give and take, but it is – I have very, very good advisors, and it is very – what we're doing with AI is very complementary to very high-level accomplished people across these domains. Wow. Yeah. All right.

Product and funding updates

1:26:55To wrap up here, we've got some AI product and funding updates I'll go through real quick as we close out the episode. So, first up, OpenAI has talked through some of the first results with Jalapeno, its first custom inference chip. It says this system developed with Broadcom can deliver 1.5 to 1.9 times more AI work per watt and up to 3.6 times lower latency across models that include their GPT-OSS-120B model, DeepSeek R1, and Kimi K2.5.

1:27:25One thing I'll note real quick there, Mike, I won't get into, like, specific background information that I might – no. Well, I'll just say, if you're wondering, would Jensen Wong be excited about OpenAI building their own chips? You might wonder about that, and I would just say, I think a lot of what's happening right now is these companies all work together, but they are also competing with each other. And sometimes CEOs get unhappy with each other over decisions that are made that were made a year ago and was known within the industry a year ago.

1:28:06And so, if actually we go back in time and we look at some things that came out about a year ago, it actually starts to make a lot more sense because if Jensen knew that OpenAI was building their own chips a while back, it could cause a little bit of friction. So, I think they continue to work together. They will continue to work together, but everybody's stepping on everybody's toes right now, I think, and it's starting to be unclear how this all plays out. Also, OpenAI added sign-in support to ChatGPT Works Cloud Browser on web and mobile for Plus and Pro users.

1:28:41This allows it to complete tasks on some supported websites after the user signs in securely. Anthropic added a built-in browser to Cloud Cowork that can navigate pages, click, and type without a browser extension, and that has initial availability for ProMax team and enterprise users. Anthropic and Salesforce announced Cloud Force, an expanded partnership that launches with Salesforce and Cloud. This is a plug-in for select pilot customers with 37 pre-built sales skills grounded in live revenue contexts.

1:29:14The open beta is planned for September, while Cloud is also available in AgentForce and embedded across Slack. Anthropic opened a research preview of what they call their Model Hardware Standard. This is a model-agnostic specification that uses standardized device drivers and access mechanisms, including MCP, Model Context Protocol, to let AI agents discover, control, and coordinate programmable lab and manufacturing equipment ahead of a planned open-source release. Google introduced Gemini Enterprise for Legal, a fully-governed legal plug-in for the Gemini Enterprise app that combines purpose-built skills, agents, and connectors for work, including legal research, regulatory screening, and contract drafting.

1:29:58Available in preview, it connects to existing enterprise and legal systems and provides traceable citations back to source material. Real quick, I think this is a prelude of what's going to happen. So basically, the labs realized that it's really hard for everyone to figure out how to use agents, like give them the general agents and they build their own. So I think they're just going to industry by industry and just build packages that say, okay, here, marketers, here, attorneys, here, accountants. It's like, we've pre-built everything you need. Go. Yep. Yep. Google also announced expanded billing and cost controls for Gemini Enterprise agent workloads.

1:30:32This includes a pay-as-you-go app edition available to select customers and rolling out broadly soon. There's new project-wide pooled quotas for anti-gravity in Gemini Enterprise. There's spend-based flexible savings plans, project-level monthly spend caps and public preview, and pricing calcular estimates for background agent runtime costs. It's all very well and good, but does not seem like it is less complicated. So it'll be interesting to see how this works. Google released also Gemini Omni 1.1 Flash as a production-ready generative video update.

1:31:06They added scene extensions in 10-second increments up to a total cumulative length of 40 seconds, first and last frame controls, 360p previews generated up to 60% faster than 720p, and 4K upscaling. We found out who is behind that secret and surprise OX Alpha model. It was Z.ai revealed that this was their GLM 5.3 Flash model, and it was launched now at $0.7.5 per million input tokens and $0.25 per million output tokens.

1:31:42That's half its list rates through September 9th. That anonymous test became the most popular model of the past week. Figure, the robotics company, said its index data app emerged from stealth with more than 16 million videos already uploaded and committed. And they committed to spend more than $1 billion on data and compute over the next 12 months. It's basically a data set that they have now for robotics that can be used to train their products and their humanoids. And then finally, Moonshot AI is reportedly in early-stage talks with Microsoft, Amazon, and Google over deals to host its open-weight Kimi K3 model, with Moonshot seeking up to 30% of revenue from K3-related services on Azure, AWS, and Google Cloud.

1:32:27And just to clarify, that is a Chinese lab that American companies are negotiating with to infuse Chinese open models into their offerings. That, I would think, would be a bit of a political hot point. We may be hearing more about that. I think that's a good prediction.

Pulse survey and wrap up

1:32:47As one final reminder here, our AI Pulse survey for this week is still live at smarterx.ai forward slash pulse. It should take you literally seconds to fill out. We're asking one question about how you are operating, how your company is approaching, rather, entry-level hiring due to the effects of AI on the labor market. So, Paul, another packed week. Appreciate you breaking everything down for us. Yeah, good stuff. Hopefully, we have a bunch of, like, really positive, exciting things to share with you next week. That would be nice.

1:33:17That's our goal every week. We'll see if we get there next week. And actually, Mike, we have, what is it, Labor Day next Monday? Yes. But we'll drop, we're dropping on the normal day still because we're going to record on Friday. Okay. Yep. All right. So, yeah, regular weekly episode as usual. Do we have an AI transformation spotlight this week? We do not. No, we had one come out this past week, though, so there is one that's pretty fresh with John Dick at HubSpot. Yeah, it's awesome. So, 234. Yeah. So, if you haven't been following the AI transformation series, Mike's done three of those now. They're really great practical inside information about, like, the messy parts, the good stuff.

1:33:51So, go listen to those. They're super applicable to anyone that's kind of driving business transformation. All right. Thanks, Mike. Have a great week, everyone. Thanks for listening to the Artificial Intelligence Show. Visit smarterx.ai to continue on your AI learning journey and join more than 100,000 professionals and business leaders who have subscribed to our weekly newsletters, downloaded AI blueprints, attended virtual and in-person events, taken online AI courses, and earned professional certificates from our AI academy, and engaged in the SmarterX Slack community.

1:34:25Until next time, stay curious and explore AI.