A

If you concentrate power the most powerful models in the world into a select group of 10, 20, 50, 100 companies, those are already the most powerful people and companies in the world. What happens when you give powerful people more power? Not great things for democracies welcome to the Artificial Intelligence show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Raitzer. I'm the founder and CEO of SmartRx and marketing AI institute and I'm your host. Each week I'm joined by my co host and Smarter X Chief Content Officer Mike Kaput as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career. Join us as we accelerate AI literacy for all. Welcome to episode 222 of the Artificial Intelligence Show. I'm your host Paul Raitzer along with my co host Mike Put. We are recording on Monday, June 29, 9:00am Eastern Time. Who knows, Mike might get some new models, might get new models delayed by the government, might get the government re releasing models they with I who knows these days, but I think there's going to be a lot of model news this week. So usually what happens at the end of these? As Mike's going through the product and funding news, I'm usually scanning on X to see if any major breaking news has happened. It would not surprise me if we got some news because we've got a short week with the 4th of July in the US so I and we're going to have to cram all the news into like three days this week. So we'll see what happens. All right, Today's episode is brought to us by Site Improve. Here's something worth paying attention to. AI search is already changing who gets found and most brands aren't ready for it. Site Improve is the agentic content intelligence platform that helps marketing teams track optimize improve performance across both traditional and AI driven search. Think of it as your visibility layer for the new search era. Grab a free aeo check@siteimprove.com aipod that's siteimprove.com aipod and the week's episode is also brought to us by Macon. This is our flagship AI conference for marketing and business leaders that happens October 13 to 15 in Cleveland, Ohio. Our hometown, Macon is three days of keynotes, sessions, workshops and conversations built specifically for marketing and business leaders who are actively figuring out how to adopt, operationalize and scale AI across their organizations. We're expecting thousands of attendees at the Cleveland convention center this year, use Pod100. That's Pod100 at checkout and you save $100 off on top of locking in the best rate available. So visit Macon AI to register. That's M A I C O N AI all right, we kick off every weekly episode with an AI Pulse, where we recap the responses from our informal poll each week. So this week we're looking at the Questions from episode 221. We said as AI usage climbs, companies are starting to watch the bill. How is your organization handling AI costs? 43% said watching closely but no limits yet. 29% honestly no idea what we're actually spending. I like the honesty. 24% starting to set caps or budgets on usage and then a very small amount spending freely. We want everyone using AI more the second question from last week's Pulse was Microsoft CEO Satya Nadella argues that as AI gets more powerful, human skills get more valuable. More, not less. Do you agree? 55% True for some roles, not others. 31% yes. Judgment and direction are the scarce thing right now. 15% no. AI is steadily eating what made humans valuable. And then nobody was not sure on that one, so everybody actually answered. All right, so for the first main topic as of Friday morning, I think Mike, we had the story was the Trump administration prevents the release of another Frontier model. And then I think it was Friday. We then got the plans for the Release of GPT 5.6. So I had predicted on episode 221 we might get multiple models last week and then we got into like late Thursday, early Friday. It's like, oh, it looks like the government may slow roll all of these models. And so we did get a preview of a new model and I guess we'll start there this week.

B

Yes, Paul so This past week OpenAI previewed its next generation GPT 5.6 model family. There are three models under a new naming system. So they have a model called SOL S O L. This is their flagship model, the model Terra T R R A, a balanced model for everyday work and Luna, a fast low cost option. OpenAI calls Sol its strongest model yet with improved agentic capabilities, encoding, biology and cybersecurity. There's also a new effort level of max reasoning and a new Ultra mode that uses sub agents to accelerate complex work. But as you alluded to the bigger story here, not only getting these awesome new models or previews of them, but how they're being released. Because instead of a normal launch, OpenAI is starting with a limited preview for a small group of trusted partners, reportedly around 20 organizations whose participation was approved by the US government, available first through the API and Codex before any broader rollout. OpenAI came out and said they took this step at the US Government's request. The company previewed the models and their capabilities to the administration ahead of launch and agreed to a phased release while it works with the White House on a cyber executive order framework and a repeatable process for future model releases. Now this all traces back to what we've been covering all through the month of June. Back on June 2, President Trump signed an executive order that asks AI companies to voluntarily give the government up to 30 days of pre release access to models with advanced cyber capabilities. Reporting indicates Commerce Secretary Howard Lutnick personally called Sam Altman to warn against releasing GPT 5.6 publicly without government sign off first. This of course follows the ongoing saga where the government levied export controls that forced Anthropic to take its Fable 5 and Mythos 5 models offline after a reported jailbreak. That was the first time Washington had kind of led to a commercial AI model being down. And GPT 5.6 is another milestone. The first time an American company has launched a Frontier model under a government managed access list. Now OpenAI pushed back a little bit. Even as it complied, it wrote that it does not believe this kind of government process should become the long term default because it keeps the best tools from the users, developers, enterprises, cyber defenders and partners who need them. So Paul, this is very big deal. Probably not in a good way. Like have we crossed the Rubicon so to speak here? Like is the US Government now just forever going to decide which AI models

A

get released publicly until there's a more formal structure? It sure seems that way, which is what we assumed is what we talked about last week. The couple of notes on this one from a tech perspective, Sam did tweet in response when he I'll kind of go through what he shared when they launched it, but he also answered a few questions on X and someone asked about the naming convention and he said star planet Moon as model sizes. So that that's where the names come from. Meter, the organization that does evals on these models did publish something. They did get early access to the model but due to the government restrictions they had to actually run past OpenAI what they were disclosing about the evals. So I'll just read two quick sections of what they wrote. They said we initiated an evaluation of GPT 5.6 SOL on our time Horizon suite of software tasks. So As a reminder meter looks at how long it would take a human expert to do something. They specifically look at like coding related tasks and then how reliably a model can do that thing. And the idea is that over time the time horizon has been doubling every six or seven months. Now Anthropic's internal data says four months, but in essence these things are doubling in their time horizon capability every roughly six months. So they initiated their testing on their suite of software tasks. However, the resulting measurement depends heavily on our detection and treatment of cheating attempts by the model. And GPT 5.6 SOL's detected cheating rate was higher than any public model we have evaluated. For our task suite, we define cheating as behavior where the model improves evaluation performance by exploiting bugs in the evaluation environment or by adopting strategies disallowed by the task rather than solving the task with the expected evaluation constraints with the data they collected. If we follow our standard methodology of marking cheating attempts as failures, they still arrive at a time horizon of 11.3 hours for their 50% reliability, meaning something that would take a human coder 11 hours to do. This model can do very, very fast with a 50 plus reliability. If we count the cheating attempts as legitimate successes, the point estimate jumps beyond 270 hours, well beyond the range where we consider our task suite to give reliable measurements. In other words, it's broken their way of assessing this that it's so good at cheating and they assume that in the future it'll become so good at masking its cheating that they won't even know it's cheating and thereby their efforts to measure this become irrelevant. I don't know what plan B is for them, but that's basically what they're saying. Okay, so then onto the model. So Sam tweeted on June, I'm just going to like accept that I'm going to call it tweeting perpetually. Okay, I get that it's like X posting or whatever, but it's tweeting. Okay, so good news first. Seoul is a smart, efficient and significant step forward. It is the same price as 5.5, also launching in the 5.6 family as Terra with 5 point level performance at half the price. Bad news. At the request of the US government, it is launching today in limited preview instead of the open access launch we were planning on. We are working with the government to get to general availability as fast as we can. I think it is quite reasonable to roll out models especially as they reach significant new levels of capability in this way, meaning every model moving forward is going to be released in this way. It fits with our long held strategy of iterative deployment, but this isn't quite the process that we think is optimal now with the government. We work with the government to attempt to get a transparent, reliable process for early access and to ensure that as long as our safeguards work as intended, we can release widely. He's referring there to guardrails, which we talk about in episode 221 in relation to Fable 5 and the guardrails anthropic put in place, meaning these models aren't safe on their own, but they put guardrails in place to try and make them safe, is what he's admitting to there. We want to be reliable, dependable partner that works with all stakeholders and we also want to live by our mission of benefiting all of humanity. I believe the government shares most of our goals and that they are overall doing a good job in a very difficult situation. We will work as quickly as we can to get this model in your hands and we hope you will love it. He then did go on to say, in response to a question on that post, I think a required preview period for extended red teaming is not a bad idea, meaning giving experts more time to find vulnerabilities and risks within the model. That's what red teaming roughly means. I just don't like the idea of the government picking the customers confident we will get to a better place. Okay, so then I just wanted to try and give a synopsis here Mike, of like, like high level what this means. And I, I, I probably said some similar things on episode 221. I didn't go back and look at my show notes and like what I said there, but try and like vary it a little bit because things have evolved in the last seven days. So where this basically means is labs and select organizations, including the government itself, get exclusive access to the most powerful models. So the government's using this model now for internal purposes, for NSA purposes, whatever, but they're using it and the labs themselves are also having access to this to do whatever they want with it for their business purposes. So we now have a situation where a very, very small group of businesses and individuals who we all have to trust are going to do good with these models, have access to things that the rest of society does not have. We've talked about the idea of nationalization. The Trump administration has literally floated the idea of some element of nationalization where they take an equity stake in these labs. So what we can probably say with relative Confidence is that soft nationalization has begun. So this means not taking ownership and control yet that these AI labs are still private companies. But the US government is increasingly treating their most capable models as strategic national security assets, not traditional like commercial software. So for the most powerful system, the government wants early visibility and then some influence over who gets access when and in what, under what conditions. So in practical terms, this could mean several things moving forward. So pre release review becomes normalized. This is what we're seeing right now, where the government's going to look at these things before they go live and they're going to make some arbitrary decision until there's something more formal in place. The access becomes tiered, so the first users are likely going to be U. S. Based companies, government agencies, defense contractors and or trusted partners, while the broader public or international access comes later or never, depending on the capabilities of the model. The model capability becomes export controlled in spirit, even if it's not always by formal export law. So the key asset is no longer just chips or data centers. It's the model weights, API access, cyber capabilities, agentic behavior. The government's trying to control all of that. The release decision becomes political. This is 100% happening. A lab may believe a model is safe enough to ship. So Sam's case, hey, safeguards are in place. We felt good with them. We were ready to go. But the government may still step in and ask for a delay, narrow the access or added safeguards. That's exactly what happened with Fable 5. They wanted, the government wanted anthropic to guarantee that their guardrails were going to work, which everybody, including probably the government, knows is impossible, but whatever. And so what happens is the strategic tension becomes too obvious. Like you move too slow and the U.S. risks weakening its commercial AI lead move to open. And the U.S. may proliferate dangerous capabilities to adversaries or criminals. So I found myself Friday night, Mike, because I guess this is my life nowadays. As I like, I think about videos I've watched or research I've done and I go back and like do this on a Friday night. So I'm sitting there Friday night and I'm like, I want to go rewatch the Bill Gurley talk from all in Summit 2023 called 2851 miles. 2851 miles, which if you ever watched it, we talked about it at the time that Bill Gurley did this talk. That is the distance from Silicon Valley to D.C. this entire talk from Bill Gurley BG2 podcast, by the way, amazing, is about regulatory capture so we've talked a lot about regulatory capture on this show. Not as much in 2026. Definitely a topic in 2025. Regulatory capture is when a government agency or regulator that is supposed to over industry becomes overly influenced by the company's groups or interests it regulates. So instead of acting in public interest, the regulator may start shaping rules, enforcement or policy in ways that benefit the regulated industry. So in the case of Anthropic, for example, who's been getting blamed all of a sudden for like the government stepping in and slowing models down? So there's all these like thought leaders who are saying, oh, this is what Dario wanted, whatever. So Bill Gurley, I was like, well, I wonder what Bill's saying about what's going on right now. And so he actually went and said this is what's causing Anthropic to aggressively beg for government protection. Customers are finding cheaper alternatives. Keeping employees requires continuing ultra rich secondaries that are dependent on revenue growth. When you can't win on the field, you go to D.C. so he started talking about like the efforts of Anthropic to have regulatory capture. But it's not just Anthropic. Like Sam Altman was on the Hill last year calling for regulation. So like everybody's been calling for regulation and I don't know that it's traditional regulatory capture attempts. So I don't know like there's going to be all these talk about regulatory capture. By the way, great 30 minute talk from Bill. I would suggest people go watch it. If you want to understand regulatory capture historical context, this presents enormous opportunities for China. We'll kind of come back around to the China thing in a minute. Potential for corruption. I, I wasn't really sure honestly how to deal with this one. So I'm just going to try and be as objective as humanly possible here about what is likely happening or going to happen behind the scenes. So by slowing these models down and having a very select group of people within the government who make these decisions, the potential, let's say for insider trading and self dealing becomes massive. So self dealing is when in the government occurs when an official uses their public position and authority to advance their personal, financial or business interests rather than the public good. It is a sever, a severe form of conflict of interest where an official holder's official actions allow them to deal with themselves privately, not accusing anyone of anything. I'm just saying self dealing is a term people should be familiar with related to where we are heading with this. The other one, Mike, that is like the hot Thing these days is like Polymarket. Yeah, Kelsey. So in March, White House staff were warned to not use insider information to place bets on prediction markets. It re referred to press reports that raised concerns over government officials using non public information to place bets on platforms like Calc and Poly Market. Now in this case this was I guess a little more severe. The warning came in a staff wide email from White House management office on March 24. According to people familiar with the matter, the day before Trump had announced the pause on the Iran war via truth social. About 15 minutes before that shift in policy, a mysterious flurry of activity kicked off in the futures markets. What has been happening in these prediction markets is that insiders within the White House have information about when models are going to be approved for release, when changes in, in public policy are going to happen, when wars are going to start or end or whatever. And it's, it's eerie how much movement there is in these markets in like the minutes or hours before something is announced. So now you can get people that are basically making themselves wealthy by having inside information on when models are going to get released. It's, it's wild. This could affect IPO timing for sure. So June 25th, New York Times has an article that says OpenAI is now leaning towards holding off its IPO until next year in part because of uncertainty around where these models go. And then I mentioned this a minute ago, but this idea of blaming Dario Amade and Anthropic is absurd. Now apparently the government has made progress with Anthropic. The, the headline every ran with was someone in, in the White House called it Dario Amadea weirdo and that they're making more progress now that the co founder of OpenAI is, is leading or Anthropic is leading negotiations with the government because Dario's out of the room. So it's like this, he's like the punching bag because he's been the most vocal with his concerns about these models. Now Dario, I think actually believe this. I don't truly believe it's a regulatory capture attempt. Yeah, he hates open source, he fears open source, but because of the implications of what happens when it goes. So I don't think he, he got what he wanted per se. I think he got a government that has a lack of situational awareness, that didn't understand scaling laws and what most people in the labs knew was coming and told them that was coming, that we were going to have these cyber capabilities and all we got was arbitrary and opaque rules on AI safety from an administration that's now controlling access to intelligence. So as Sam said, they get to pick the winners. And they have proven time and time again that they have no aversion to insider trading and self dealing. So now we have to trust an administration. And again, I don't care who you vote for, like what your political leanings are. The reality is whoever's in office, you would probably be dealing with very similar things. There's always self dealing and insider trade. It's like a reality of government. So now businesses, American citizens and our international allies are reliant on the whim of government with no clearly defined rules to determine if and when we get access to the frontier models. So they have still not actually told anybody what anthropic needs to do to get Fable 5 released. Like, what are these magical guardrails they're all of a sudden going to create that prevent people from misusing the models? It doesn't exist. So we're just holding these for what? And so you could absolutely look out ahead to a point where they say, you know what, these models are just too powerful for us to put out into the world. We, the government, are going to have on. You know, we're going to have access to the weights and we're going to be able to do whatever we want with these. And then you are only allowed to give these to these 100, 200 companies, whatever everybody else gets one two generations ago. Now if they do that, China can take the lead. And that's where this debate comes in. We'll talk more about China. Actually, like, I think you joke with me, Mike. Welcome to like the politics episode. Like there's so many implications.

B

You know, one final thing here, Paul and I want to get into some of the solutions proposed in our second topic is that as it stands right now, I can't keep help but keep coming back to. And I'm not the first one to have said this like, it's just this analogy of like Lord of the Rings, of like the ring of power. It's like the temptation is just too great for any government to not want to put it on and use it. Meaning taking more of a stake in labs, restricting the power, the powerful models and their usage to itself or to a core group of players. It just seems like a very slippery slope to be headed down.

A

Yeah, and I get it. Like if the NSA gets access to mythos 5 and they realize its hacking ability and its ability for cyber defense and they can now use that as a competitive advantage over other governments.

B

Yep.

A

Why would you, if you're the government, want other people to have that.

B

Right, right. And then that applies to all the other frontier intelligence capabilities that.

A

Yeah. And it's not a perfect analogy, but, I mean, it's literally like we go back to, what, the 1940s, 1950s. 1940s, mostly Los Alamos. And, like, let's say the atomic bomb is designed by a private company and the government doesn't take it seriously for a few years, like, oh, you're never gonna be able to achieve that scientific breakthrough. And the labs are like, no, we're going to. Like, in three years, we're going to have a weapon with massive destruction capabilities. And the government's like, now you're not. And then all of a sudden, oh, you do have a weapon with that capability. We're going to let the private company control it. Right now, it's not a perfect analogy because AI can do so much good, and it might not, in some people's belief, have the destructive capability, at least visually speaking, physically speaking, of an atomic weapon.

B

Yeah.

A

But the premise, what these labs believe is, like, it can be as disruptive to the economy and society as a weapon. And so if you have now created something outside of the government that protects the stability of the democracy, how do you not take some level of control of it?

B

Yeah.

A

And so that's why it's such a messy debate.

B

Well, I will say this is why I loved that we're focusing on this second topic here, which is this essay written by AI policy writer Dean Ball, who is also soon to be an OpenAI employee. And the reason I say that is because, like, he published this essay called what Should Be Done, which responds directly to this GPT 5.6 situation we discovered. And for a reminder for people, we've talked about Dean a bunch on the podcast. Rightly so. He's, like, really trustworthy, I think, voice on AI regulation and policy. But he's a former White House AI policy staffer. He helped write the administration's AI action plan. He is going to join OpenAI actually, quite soon. We covered that on last episode. He stressed, however, this piece is his own view and not the company's. And basically in it, he lays out all these possible paths and actions that might be wise to take or at least consider as we move forward into this kind of new normal. And so his central argument in here is that the executive order Trump signed earlier this month, which was billed as voluntary to testing for frontier models, has really become a de facto involuntary licensing regime. He says recent events prove this point he made back when it was signed. So obviously first the administration caused anthropic fable to anthropic's fable to be removed from public access. Now GPT 5.6 is being limited to a small set of people at the government's request. And so Ball's biggest concern in this essay is that nobody, including he says the administration itself, knows what a lab would actually have to do to get a model approved for a broad release. Paul, exactly to your point, until some safety standard exists, he says the government's answer to any broad release request would effectively be no. And he is skeptical this administration can produce a high quality standard soon, noting that almost no one now setting AI policy actually has frontier AI lab experience. Now he's actually careful to say the administration is not directionally wrong. The catastrophic risk concerns related to AI are real. And he says he himself himself raise them while inside the White House. But he does warn here that the current approach is economically dangerous because labs recoup the cost of training frontier models in the first few months. They're broadly available. So every week of delay that eats into that window really threatens their business model. The hundreds of billions of dollars, if not trillions over the years going into data centers assume a roughly global market, not access for a hundred government approved companies. So if you restrict releases this severely, he argues, and the risks making the AI skeptics right, that the US has overbuilt our AI infrastructure because the government is now rendering AI demand or could render AI demand essentially unlawful. So he gives some prescriptive steps here as well. He says we should stop trying to regulate individual models, instead regulate the labs themselves. He wants to federalize state laws that already require labs to publish their safety frameworks, then have independent privately run verification organizations staffed with technical experts audit the labs against those frameworks and their internal governance, with the government certifying the auditors rather than running the reviews itself. He actually points to a new bipartisan bill being proposed called The Great American AI act that was just proposed by U.S. representatives Obernolti and Trahan as the closest thing yet to that vision right now. So, Paul, there's a lot to unpack here. It's a lengthy essay well worth reading, I would argue. Dean is like one of the better commentators out there, I would say, on AI policy. What do you think of his recommendations here?

A

I just think that it's a great article for people who want to understand the state of a very complicated pro process and like where we are, where we're going now. If you listen to this podcast every week, you're going to have some level of familiarity with the majority of a lot of these ideas and things like that. But even for me reading through this, I thought it was just like a great synopsis of the challenge of trying to find a solution and then some actual viable ideas. And as we say on this podcast all the time, I just want people to bring ideas to the table, not just complain about things. Right. And this is the same thing that goes with the economy. And, you know, it's like when we talk about, you know, potential impact on jobs, so many of the I labs just like, say, yeah, it's going to affect jobs, but they have no plan about what does it do or how do we solve for it. So I find, you know, I don't agree with everything Dean Ball says. Like, there's quite a few things, honestly, that where I'm like, yeah, I don't think we'd be on the same page with that. But I have tremendous respect for the fact that he's thinking deeply about it in a very logical way. Now I understand why he's joining OpenAI, and he actually explains it pretty clearly in a Cognitive Revolution podcast. So I was just listening to this over the last week. On June 20th, an episode of the Cognitive Revolution with Nathan Labenz dropped that features like a two and a half hour interview, interview with Ball. And it was right after he announced that he was going to be joining OpenAI. So I'd highly recommend it. The basic premise is it's kind of like why Andre Karpathi went to Anthropic. You have these, like, amazing independent thinkers on the outside, and until they get on the inside, they don't really know exactly what's happening. And so if you're not working in one of, like, the three major labs, there's just only so much information you can gather without knowing the inside information. And so his feeling was he could. OpenAI claims he's going to maintain his independence and be able to publish stuff like this even once he joins again. Whether that's the reality or not, I don't know. But I just want to call out a few. Again, there's 35 here and I would read all of them, but I want to go through a few excerpts from this and then maybe, Mike, there'd be a couple here we could add some commentary to. So he starts off one. When President Trump signed it earlier this month, I argued that the executive order on Cyber and AI, which claimed to establish a voluntary testing program from four Frontier AI Models was really establishing a de facto and voluntary licensing pre approval regime. That's what you were talking about. This analysis is proven correct. First, the administration revoked public access to Fable Anthropic's latest Frontier model because of security fears. Now it appears OpenAI's GPT 5.6 is being limited at the request of the government. So that was his first point. Number four. I'll just read where these are in the, in the 35. When I say somebody, when I say nobody, I mean it literally. The administration itself does not seem to know what safety standards or best practices a company would have to observe for them to be comfortable with a broad release of a model that matches or exceeds Mythos and capability. And again, in theory, every Frontier model moving forward will meet or exceed Mythos capabilities, including ones from Chinese labs, within three to six months. Number five, nobody I know in the Trump administration has any Frontier AI experience. Just a few months ago, someone with experience at both OpenAI and Anthropic was hired to run the center for AI Standards and Innovation, but he was fired by senior administration officials within a few days. The rest of the staff at that center have reportedly been on a stop work order, not even allowed to communicate with other government agencies for much of the post Mythos crisis period. The lack of technically expert staff is one of the main reasons to doubt the near term ability of this administration to produce a high quality safety standard anytime soon. Number six. And by the way, really, that one Mike, who that is considered a top tier AI researcher and let's just generously say there's 500 of them in the United States. I mean, there's thousands of them. But like top tier, like, like tier one, maybe 500, maybe 100. I don't, I don't know. You're going to go work for the government instead of like in one of the labs?

B

A capped government salary that literally cannot go up over a certain threshold. And miss out on the IPO of

A

labs, The AGI agent? Hell no. No one is going there. So what do you have to do? The government has to borrow the talent. Like this is they have to get in bed with each other because there's no other way to have the testing regime that they're going to need because no one capable of doing the testing is going to go work for the government. Okay, the number six, you kind of address this one. Despite these criticisms, the Trump administration is not at all directionally wrong that something must be done about the catastrophic risk potential. Nothing about the fundamental security and safety concerns is illegitimate, meaning Dario's right to be, you know, preaching that we need to do something. Number 10, given the current pace of things, a safety standard produced by the administration tomorrow may well become irrelevant by September. So two years ago, the way we were going to do all this, the governing of this and even the state laws was going to be based on flops. It was going to be like in essence based on like the size of the compute needed to train these models. Well, they didn't think ahead to the fact that they were going to get more efficient algorithms that could achieve the same level of outcome with dramatically less compute in the training process. So like whatever the government thinks they're going to put in place, however the parameters would work, they're going to be obsolete in a year, and so you can't do it based that. Number 11, this is a bad state of affairs. Consider in particular some industry Dynamics. The ongoing AI infrastructure buildout, the one that is, according to former U.S. aI czar David Sachs, essential to the U.S. economy, assumes a functionally global total addressable market for U.S. aI services. No one is building $100 billion data centers to serve frontier models to whatever 100 companies the U.S. government will allow access to. That's the economics of this. What they are doing right now will break the AI industry within three to six months. You cannot spend the billions they're spending to train these models and then deliver them to 100 people. Unless the government plans on subsidizing this and writing multi billion dollar 100 billion checks to the AI labs to make up for the lost opportunity. And I don't see that happening. But that this is like, like I, I actually see this is a binary thing. You either allow them to sell these models and make money or you don't. And if you don't, the industry crumbles. Like that just seems obvious to me. Another a market panic about this is not implausible. Though markets tend to downplay this administration's abrupt shifts in executive action. That is, that is an understatement. I if such a panic occurred, it is worth noting that the effects could be extended far beyond AI. A large fraction of US re industrialization and heavy industry efforts, from nuclear energy to natural gas to power electronics to batteries, is predicated implicitly and explicitly on expected future demand from the AI industry. If suddenly nobody believes that demand can materialize, the effects could cascade well past the frontier labs. This is 100% true too. If sometime in the next 30 days we realize the government isn't going to get its act together and they're going to just have this totally OPA process for approving these models and thereby the companies can no longer do the capex spending that they were going to do. Over the next three years the US economy will crater. I mean 15, 20% in a day. We've seen it with the SAS industry already this year. Play that out over all the other industries that are now being boosted up by the future spend probability of the AI industry. The US economy is powered by the industry whether we know it or not yet. And so like that is a very important point he's making. A few more is number 14 indeed my view bad AI futures are much likelier if only an extremely narrow subset of actors have access to Frontier AI. This is because that narrow subset of actors is likely to be composed of groups that already have significant economic and political power, including of course the federal government itself. This goes back to what I was saying earlier. Earlier, if you concentrate power the most powerful models in the world into a select group of 10, 20, 50, 100 companies, those are already the most powerful people and companies in the world. What happens when you give powerful people more power? Not great things for democracies. So then in the what it should be done, I'll just highlight a few of these. So number 18 the best starting point is the aforementioned and wish for safety standards is the Frontier Labs safety and security framework. So they've all published them. DeepMind Anthropic Open AI. You start there, go read those. Those are like good baselines. 22. It would be good for someone to thoroughly audit the Frontier Labs, at least to test their adherence to their own safety plans, as well as to probe their broader internal governance of the AI recursive self improvement loops. Good luck getting access to that information unless it's mandated by the government, right? 24 isn't to say there's no government role. Perhaps if you persuaded me that private bodies have merit, you could imagine the government certifying or licensing the auditor. Just as auditors in accounting are licensed, a federal agency or a council of appointed officials could review applications from auditors and approve the ones they deem trustworthy. That way there also to be more than one such body. It's easy to manage, for example a body devoted to cyber risks, another devoted to biological risks and etc. So again the premise there is the government doesn't control it, it's actually audited by independents, but the government approves who is certified to do independent audits. Again, who are you going to get to do these audits Right. That's going to like audit labs instead of working in them. Couple other quick ones. Number 25. Governments could require that the largest AI labs receive certifications from these bodies in order to participate in the AI field. 31 this is an important one. An attentive reader will note that my proposal de emphasizes regulating publicly released models and focuses on stead on regulating the frontier lab lab as an entity. The reason for this? A model is a hard thing to regulate. We talked about that. The rapid pace of algorithmic efficiency means that it's difficult to design regulatory thresholds based on model characteristics. So again, six, 12, 18 months from now, we have no idea what they're going to be capable of. How do you set these parameters arbitrarily? Today, internal deployments matter a great deal. For some threat models alongside publicly released models. That's an important one because they're only regulating what's available to the public.

B

Public.

A

These labs could build models that are at AGI or super intelligence, whatever. We'd never know. And now you have three companies in the world with access to superintelligence and other governments who hack in to steal the weights. And now we literally have governments and like three people, three labs that are the only ones with this power. So I don't know how you do it without internal, internal regulation in the future, particularly as there are breakthroughs in domains such as continual learning, it's possible that model weights will change weekly, daily, hourly, or even in real time. So again, how do you regulate model that's just going to like, rewrite itself and improve itself and then 34 this the last one that I'll say there is momentum within the I safety community for an idea. There's momentum within the labs, there's momentum with Congress, and within the administration, there's at least a recognition that something must be done. So as complex as it is, everyone seems to be arriving at the fact that like, we got to do something outside of a few people in Silicon Valley who just, just say, let's just unleash health to the world and deal with it later. But other than that, like, it's just wild. And so again, I, I think this is a topic that is universally important to everyone. I get that not everyone's going to have the same level of interest in this topic. It's kind of overwhelming, to be honest with you, but it is, it trickles into every aspect of society. Business, the economy, your jobs. Like all of this is going to be decided in the like. Like this administration, for better or for worse, will in the next three years, govern over the regulation of AGI and beyond. Like, yep, unless they shut, shut it down completely, which isn't going to happen. We will get to whatever you want to call AGI, whatever the define it is, and we will be beyond that in this administration's time period. So this is it. What, what we've got is, is going to make these decisions and keep that in mind when you go and vote this fall. Like, whoever, whatever side you think is, is more important to the outcome you desire. As we'll touch on in a topic later in this episode, who you put in office in the midterms in the United States is going to have a massive effect on what happens with AI regulation and policy.

B

You know, I know we've been talking about these trends for years now, but it feels at least a little like to me, like, I, I wake up and I'm like, oh my God, we're here already. To some of the things that have been predicted or outlined, especially you mentioned the term situational awareness coming from Leopold Aschenbrenner's essay, which I would highly recommend people go revisit because we are pretty well on his timeline of the things he predicted coming to pass. So it's very surreal to be living through it and knowing how quickly it's moving, you know.

A

And the other thing I thought about real quick, Mike, this morning, and I guess we'll probably transition to this in the next topic, is I'm starting to feel like what, what is happening right now with AI safety and regulation is what's going to end up happening with jobs in the economy. That, like, we've been talking about this for two years. We other people have been talking about, like, hey, this is going to come to a point where the government's going to have to do something and they won't have done anything. And I feel like this all came to a head in, like, the last three weeks. Like, the government stepping in with Fable 5 was it like, that was the inflection point for regulation. And I'm starting to feel like this is what it's going to feel like with jobs. Like, it's going to be talked about. We'll have been saying for two years, hey, like, we should have plans in case it doesn't go well. And then you're going to get hit with like a jobs report or a massive layoff from a major company that very explicitly has to do with AI, you know, and all of a sudden it's going to be like, oh, wait, that was real like, oh, we didn't do anything for the last two years and now we're stuck with the administration trying to figure out what to do with jobs in the economy. I think this is what it's going to feel like. It's just going to boom and all of a sudden it just goes, yeah.

B

All right. So our third big topic this week is OpenAI has released new research co authored with economists from Columbia, Wharton and Duke that uses real usage data from its Codex agent to show how agentic AI is changing the way people actually work. Unlike a chatbot, Codex takes actions on your behalf. And this paper tracks three of the groups using it individ users, organizations, and OpenAI's own employees. And the top finding from this is that adoption of agentic systems like Codex is growing fast and spreading well beyond engineers. So, for instance, the number of active Codex users grew more than five fold in the first half of 2026, with the most rapid growth coming from outside the original audience of software developers. And what people use it for is also shifting as well. So rather than asking AI for advice, users are delegating real work, and that work getting more complex. Since the start of the year, the share of individual users who gave Codex at least one task that would take an experienced human more than eight hours, the number of people doing that, the share grew nearly tenfold, and people are more often running agents in Parallel. More than 10% of users manage three or more Codex agents at the same time in a given week. About a quarter use saved skills, which are reusable instructions for complex workflows. And inside OpenAI, where usage is not really capped at all. The paper shows a glimpse of where this is all headed. So Codex now drives more than 99% of OpenAI's employee work output and has largely replaced Chat GPT this past month. The median employee in a legal role, for instance, produced 13 times more across the company's AI tools than they did in November. The median researcher produced more than 50 times as much. So, Paul, on the surface, this research is about how Codex usage is rapidly expanding in OpenAI outside of it, but really we're talking about this kind of sea change in how people work. So there's a real fundamental difference between using AI through a chatbot, like many of us still do, and having agents do long horizon tasks autonomously. Curious how you're looking at this change.

A

So this, this research to me is invaluable, whether it's coming from here. Anthropic has put out some great research just on their own. Internal usage because all of us are trying to figure out what the future of work looks like and how it's going to reshape our organizations, our tech stack, our org charts and the best chance to see around the corner is to look inside of the labs who are actually using all this technology and have this in this case infinite token use basically. So it becomes even more essential as government restricts access to more powerful models. So when we don't have the ability to even can experiment in these ways, this insight from Labs is very helpful to start trying to piece it together. So a few excerpts from How Agents are Transforming Work, which was the OpenAI post that they published in relation to the research so he said agentic AI changes the unit of knowledge work from single interactions to delegated long horizon tasks. Which again this goes back to like that meter research that we were talking about reliability over long horizons lessons. They said chatbot interactions are often short and self contained. Agents can operate independently for minutes or hours while orchestrating tool calls, interacting with environments and iterating towards solutions. As a result, agents are quickly becoming the most powerful AI tool for work. As you mentioned Mike, every department, including non technical departments such as legal and recruiting use Codex within OpenAI as their primary AI tool for work. The pattern reflects what they believe will be the future of work given the expanded capabilities and accessibility of agentic tools. So one doesn't take someone with technical knowledge to use these things. They're saying like it transforms the way people work. People use Codex for long Horizon work by May 2026, 81% of sampled individual users made at least one Codex request estimated to exceed 30 minutes of human work, 70% made one estimated to exceed one hour and 26% made at least one Codex request estimated to exceed eight hours. So so one out of every four did a request that did one day's worth of human work in a basically in a prompt is kind of a way to think about this. Codex became the primary AI tool for every department engineering move first. Legal Finance recruiting crossed into Codex being their primary tool around April 2026. For the average OpenAI worker, Codex usage now accounts for more than 85% of output tokens. Since Codex users tend to use more tokens than non users, its share of overall, as you alluded to, is 99.8% of weekly output tokens. Nearly a quarter of all Codex requests are for tasks that would take a person more than one hour to complete. Again, like you can just like I'll just read it again. Nearly a quarter of all Codex requests are for tasks that would take a person more than one hour to complete. So when we talk about the future of one work and you break down what you do in a given day, the average person, let's just say, works eight hours. Imagine instead you have like eight prompts and like each of those prompts does something that generates what would have taken an hour of work.

B

Right.

A

And then you, then you scale. And that's just one. Like they're going to be able to do multiple hours, multiple days. Now it's going to take a bunch of tokens, but we're entering a phase where agenta capabilities are going to be made accessible that can do hours or days worth of work with simple prompt strings that could be completed in minutes. And that's just fundamentally a really difficult thing to comprehend. So they go on to say, over the last six months, Codex usage has deepened and intensified. You mentioned the research saw the biggest jump. Customer support rose 32 times, engineering 27 times, legal 13 times times across the country. Across the company, users are switching from chatbots to agents as their primary form of AI direction and are deploying an exponentially growing amount of agentic labor. It's just like this is, I would say, the second half of the year. Mike, is it fair to say AI agents and agentic work is like the core focus of our, of SmartX research team?

B

I would, I think that's a safe assumption. Yes.

A

Yes. So we will be doing a lot more. This is actually to be a focus of our AI Council in the second half of the year as well. And I think rightfully so. It's just, it, it, it is the most consequential thing that's going to happen. Well, that's hard to say. One of the most consequential things that will happen in business is the autonomy and reliability of agents and then finding out how to get affordable access to do what they make possible efficiently. That's, that's kind of another challenge. But again, you have to look inside the labs, see how they're working. And that is how we then figure out what does the future of work look like.

B

Absolutely. And I just want to emphasize and make really clear for people because the labs don't always do this, a great job of this, and I think it leads to confusion. It's like we're talking about Codex. We talk about things like Claude code. We are not talking about legal or customer support. Using Codex to go build software, though they do that, you can literally use these tools as general purpose agentic harnesses. Around the frontier models. Meaning I use codecs and quad code daily. Not to go write software to do the knowledge work that we are all used to doing. So if you haven't kind of connected those dots yet, the easiest way to start learning is go use these tools. They are not just for developers. People are using them in really innovative ways in knowledge work. And that's kind of what you're doing seeing here. Okay, so before we move on to rapid fire, this week's episode is also brought to you by the AI for Business Boot Camp by Smarter X. Coming to Columbus, Ohio very soon here July 16th. This is a single day at the Hilton Columbus at Easton. It is built for professionals and leaders who are ready to accelerate AI adoption and value creation. So this is a day with Paul and myself that starts off with Paul's State of AI for Business keynote. We then transition into two highly interactive workshops. I will be teaching an AI productivity workshop in the morning. Paul will be teaching an AI Innovation workshop that afternoon. This event is built for AI forward managers, directors and executives across every department who are ready to move past the theory and get into building real AI powered workflows, getting strategic frameworks to accelerate transformation, and ready to leave with an immediately actionable plan for themselves and their teams. If you're an AI Academy Mastery member, you get discounted pricing to this event. Discounts are also available for teams of two or more groups of can or more can get custom pricing as well. Also, if you're a podcast listener, use the code POD100 to get a hundred dollars off your ticket. Most importantly, you are listening to this on Tuesday, June 30th. Today is the last day to get the early bird pricing for this event. So this is the last day to get the lowest price possible. To grab your spot, just go to SmartRx AI, click on Events. You will see AI for Business Boot Camp as an option option right there. All right, Paul. So more political intrigue, so to speak. So this first rapid fire topic connects directly to our first couple main topics. So if the US Government is now potentially going to stagger the release of American frontier models, there's this big second order question some AI watchers are asking, which is what happens to Chinese models? And this past week that debate caught fire, kicked off in part, but by a post by widely followed AI commentator Andrew Curran. So Curran argues that staggering American model releases almost guarantees that Chinese models get restricted or even banned in the West. So his logic is the following. Chinese models are technologically roughly nine months behind the US Frontier. Today if you force every American release into a slow rollout and China starts closing the gap over time, this undercuts the business model of Western labs because the public would eventually get good or better open source Chinese alternatives. Curran, for one, argues the US government will not let that happen. Either China somehow agrees to the same staggered release schedule, which he thinks is unlikely, or the US restricts or bans Chinese models and he puts that window at nine months at most, probably less so. So it does seem like Chinese open models are genuinely closing in right now. On US Ones estimates of this gap, he puts it at nine months. Some others have estimated six to nine months. Stanford's AI index says Chinese models have largely erased the US quality lead over the past year. Alibaba's Quinn has already passed Meta's Llama to become the most downloaded open model in the world. So Karin adds one more prediction. He actually also thinks because of this, this environment that he raises the odds that Nvidia eventually gets restricted to selling only within US borders as well, on the logic that it's much harder for chips to end up in China if they are treated as a national resource. So, Paul, some pretty bold predictions, but obviously it seems like we're quickly in a new normal here. Like what do you think the odds are that the US cracks down on Chinese models?

A

Yeah, I said it last week. I think it was going to happen now. Yeah, what it looks like I was actually doing a little like, I guess working with GPT 5.5 thinking model, kind of expanding on, like what are the possible ways that this could happen? Like what could this look like? So I'll get to those in a second. Just a few names because we don't spend a ton of time on the Chinese model. So I just wanted to surface like some of the names to recap you've heard us say, or, or like maybe Michael mention them in the product and funding news. So we have Deep Seek, which uses this, the company and the model we have. Alibaba. Quen is their model, which we'll talk about in a second. We have bytedance Dubo I think is how you say that one Zippu, which has glm, which just recently became a big thing. Like in the last two weeks that one's got talked about a lot. Moonshot has Kimmy and then Hawaii has Pangu. So those are like the, the most strategically relevant labs and models that are coming out of China. Now. The Alibaba was big this week because Anthropic sent a letter to US officials accusing Alibaba of brazenly and illicitly attempting to extract its AI models. So Alibaba is a massive Chinese multinational technology and e commerce conglomerate, kind of like China's Amazon. They have an AI research division that develops Quinn, a flagship suite of highly powerful foundation models. So the letter which CNBC obtained claims that Alibaba was distilling Anthropic's models through tens of millions of exchanges. So in essence they have all of these like fake operators and organizations and accounts that are then going through and, and using the model and learning how it behaves and learning from there and then distilling that down to build their models faster. So in my chat, GPT, kind of back and forth, what you know, where are we with us versus China? What are the implications? What could the US government do? So I had this whole string of questions that I was going through and kind of pushing it on. So I'll just give you a quick synopsis. So AI models are obviously becoming general purpose cognitive infrastructure. So if a US company builds customer service, coding, analytics, marketing, legal review and decision support on a Chinese model, so let's say that you can't get the most powerful models and then you're going to turn to China and a US enterprise is going to start using these Chinese models. The concern is not just where is the data going, so now they can see what you're prompting, things like that, but who controls the intelligence later inside that enterprise. So a model access regime that blends export controls, import controls, procurement rules, cybersecurity view and industrial policy is sort of of what is forming within the United States. So password could be literal federal procurement ban. So the easiest move is to say federal agencies and contractors can't use Chinese origin foundation models, APIs, model weights inference services, embedded AI tools. So that would mirror the way the government procurement has been used against Hawaii and TikTok on federal devices. You could do critical infrastructure restrictions so the government could prevent Chinese models from being used used in defense, energy, financial services, telecom, healthcare, etc, you could do enterprise compliance pressure. So instead of banning Chinese models outright, regulators could require companies to disclose whether they use models from China or other countries of concern. That could stop adoption or dramatically slow it down. So boards, insurers, auditors, enterprise customers may decide the risk isn't worth it to work with companies that are using those models. You could do cloud, cloud and API control rules or you could do sanctions. So there's a huge risk though in all of this. These restrictions can accelerate Chinese open model ecosystem to other countries who no longer can depend on the US to get access to the most important models. So it creates this paradox where if the US blocks access to Chinese models, it may protect sensitive environments. But the US also restricts access to its own best models, pushing developers, startups, global markets toward cheaper, more open, less controllable alternatives. So in essence, you have the Chinese government trying to tank anthropic and OpenAI and innovation within the US by just flooding the market with open source models and subsidizing them. So the US policy challenges become protect national security without accidentally making Chinese open models the default global substrate. And the bigger limitation is model choice may stop being purely technical and economic decision and it becomes a geopolitical decision. And we are racing toward that right now.

B

Yeah, yeah, right, yeah. The US versus China and then US verse the US AI labs, right? It's unfortunately what the two prong front they're fighting on.

A

It's a complex. And again, you can empathize with how complex this is. And there isn't necessarily like very clear right and wrong answers that even are political in nature, that it's like, oh, one side will do this, one side of that. Like, no, neither side knows what to do right now.

B

Right.

A

And unfortunately, that leaves us where we're just arbitrarily having models held up for whatever reason by the government. Government.

B

All right, so another big political topic this past week, the Economist made the AI data center backlash their cover story. They had a bunch of reporting that opposition to data centers is spreading fast and turning even way more political. So some background here. According to their estimates, the biggest AI and tech companies are pouring as much as $750 billion into AI data centers. This is part of an estimate estimated 3 trillion globally estimated to be spent between now and 2030. This is enough to roughly quintuple America's AI computing capacity by the end of the decade. And as the Economist puts it, people across the political spectrum are absolutely furious about this. So by the Economist count, protests have now stopped close to $100 billion worth of data center projects, including at least 20 worth $42 billion canceled in just the first three months of this year year. Bills to freeze new construction have been introduced in at least 10 states. And polling shows Americans would rather live next to a nuclear reactor than a data center. And the reporting kind of makes it clear this is about the usual suspects, but about a little more than them too. So it's not just the noise, the water consumption, the perceived power consumption. This has become a proxy for a deeper anxiety about AI itself. And three quarters of Americans now say AI should be more regulated. Republicans, nearly as much as Democrats, 45% said they are at least somewhat worried AI could cause human extinction. Wanted to give a quick shout out to Ohio, our home base. This is a telling example. We are now the fourth largest data center hub in the country, yet 3/4 of Ohio Democrats and 2/3 of Ohio Republicans oppose local data center development. Now on top of all this and we'll talk about this piece, Mark Cuban jumped into all this with a widely shared post arguing that AI companies have already lost this PR battle. He basically says that this fight functionally now has nothing to do with data centers and has become a standin for just anger at AI and at the wealth it is concentrating in a few hands. And his advice to the labs is stop trying to explain the benefits. Show up in person in affected towns with real money for community programs and job support. Go to creator, go to artists and creative unions who he says are terrified. Ask what would actually help. He's basically warning that they do not win over workers and communities. They will never ever get these projects that they need to make all of this work. So Paul, like, based on the numbers, I just don't know if I've seen an issue more capable of uniting Americans in recent times than hatred for data centers.

A

Yeah, and we're seeing this in our own community. So Cleveland is our hometown. Just last week, Cleveland City Council is advancing a plan to press pause on data center projects responding to a wave of high powered computing hubs sweeping across the state. This is from News Channel 5 in our hometown. After hours of discussion and debate Thursday, Council's utilities committee voted to move forward a three month moratorium on permitting and reviews. The freeze will apply only to standalone data center buildings, both ground up and construction, and expansions, not to server farms integrated into multi tenant properties. Legislation is headed for a full council vote on July 15th. So I mean like we're again literally in our backyard. We are seeing the fight against this now. Now Cuban's post, I think it's a good one to go read. I'll just highlight a couple of quick excerpts here. It's time for everyone to realize the fight against data centers has nothing to do with the data centers. As you mentioned, they've become the proxy for hate towards AI and the concentration of accumulation of wealth it's creating. Until those running the big LLMs understand this and start a community tour. Not to explain the benefits, it's too late for that, but to help towns and cities that may be impacted by job losses And I'm a believer there will be a net gains in a few years here. He says this battle is only going to get more intense. And let me tell you, no matter how much money you pay to buy politicians and races, you will lose. Don't try to pay famous people to endorse what you are doing. That's dumb. The big LLMs have lost the PR battle.

B

Why?

A

Because they all suck at putting people first. They have a Silicon Valley attitude that makes them all think they are John G. Saving the world. Which by the way I had to look up because I'd never read Atlas Shrugged. He is the hero and central philosopher of Atlas Shrugged, who's an inventor and intellectual who organizes a secret strike of the world's leading creators and innovators to protest the society that exploits their talents. Back to Cuban Given the number of data centers and power that is needed today and going forward, if you don't kiss the asses of the people that go to work every day and are just trying to pay their bills, you will fall far far short of the capacity you need to make your business work. We talked about this like a month or two ago as like the PR back battle is the labs are screwed. They don't know how to do this. They're just like hey, it's going to save the world. And they have no good tangible. What does that mean and how does it help me who's struggling to pay my bills and living in a community with a data center that they're running off of backup generators that are noisy. That's the reality of these communities and these ultra wealthy people from Silicon Valley just pushing techno optimism is not going to win you new votes or data centers that you need.

B

Right? Exactly. Yeah. It'll be interesting to see if people wake up to this fast enough. Yeah. All right. Another big political topic this week and then I swear we've got maybe only one other. The rest of this episode I don't know. But this past week, tons going on the most expensive proxy fight yet over AI regulation played out inside a single congressional primary. This was in New York's 12th district. A Manhattan seat opened up when longtime representative Jerry Nadler retired and AI money flooded an eight candidate Democratic field. At the center was someone we've talked about before, Alex Boris, who is a New York State assembly member and the kind of co author of the Raise act, which is a state law requiring frontier AI companies to publish safety protocols. Boris pretty early on made AI safety the core of his campaign which made him a target because leading the future. A pro industry super super pack we've talked about before. Backed by Andreessen Horowitz, Open AI President Greg Brockman, Palantir co founder Joe Lonsdale and the AI company Perplexity spent more than $8 million to defeat him. On the other side, AI Safety Group spent close to $20 million to support him, with the largest share coming from an anthropic linked pack. All in all, more than $27 million in AI industry money poured into this one household race. On top of all this, Boris actually lost. His fellow assembly member Micah Lasher won the nomination about 35% of the vote to Boris's 35%. Interestingly, Lasher wasn't exactly soft on AI either. He literally co sponsored the Raise Act. He is back to national moratorium on new data center constructions until there's guardrails on costs in the environment. He's called for antitrust investigations into partnership partnerships between Google, Microsoft, OpenAI and Anthropic. So, Paul, this is not the first time we've talked about Alex Boris. He really seems to have become this lightning rod for certain factions in the AI industry. It seems like.

A

Yeah, I don't think it'll be the last time we talk about him either. And this to me was just a preview of what's to come. It didn't necessarily prove anything like it's, you know, it kind of let that American citizens will, you know, know, show an interest in AI. And as you've highlighted multiple times already, the hatred is sort of right in the middle. It's actually a bipartisan issue.

B

Yeah.

A

And so I don't think it's solved yet how the vote will move based on AI. But it proved to us that AI is going to be a core part of campaigns moving forward and that this is going to play out again this fall in the United States in particular with the midterms that are coming up.

B

And you know, there's an interesting quote in a New York Times story about, about this race. It said, informed by the Boris race, the AI industry's political activity will enter a new phase. So predicts the writer of this article. And they said strategies in. Strategists involved in the fights say they are newly attuned to the weakening poll numbers for the industry, particularly among Democrats. They also say they are more aware about the chain reaction of consequences they can trigger when they throw their weight around in elections. So that tells you a lot that you might need to know about how you can expect people to get in involved in AI moving forward. All right, switching gears here. Something not super political this time around. We covered on our last episode Google DeepMind's Talent Problem. They saw Nobel laureate John Jumper left for Anthropic. Gemini co lead Noam Shazir left for OpenAI. And in the days since, the bleeding appears to have continued. Bloomberg reported that two more high profile staffers, Jonas Adler and Alexander Pritz are leaving Google for Anthropic. Both work seen internally as key contributors to Gemini. Adler was working on Google's AI coding efforts, pretzel on model training. Separately, DeepMind researcher Arthur Conmi announced he is joining Anthropic to work on aligning its on its upcoming models. So this is at least three more senior departures to Anthropic in a matter of days. On top of those two we covered last week. Interestingly, DeepMind chief Demis Asabas pushed back publicly. In an interview he said Google has by far the biggest and broadest research bench of any lab and that the company still wins its fair share of top talent in what he called a ferociously competitive market. So Paul, I gotta ask. We now got Demis weighing in on the issue. Is this just a couple weeks of bad news? Is it something bigger and worse?

A

I mean for the last few years, talent has definitely just moved around between the labs. This does seem a bit of an anomaly. Like it seems out of the ordinary. How many major players are moving?

B

Yeah.

A

And that they're moving to Anthropic. The only thing that came to mind as you were saying this is, I wonder what, like how stock options play into this. So if Anthropics valued today at whatever it is, 800 billion or whatever that number is.

B

Yeah.

A

And people are looking saying, wow, I could, I could jump ship now and get in pre IPO and you know, my stock options might be worth, you know, a couple hundred million in three to five years. Like maybe, maybe their compensation packages are just really enticing and DeepMind can't compete there. Or maybe they just think Anthropic is a better lab to go and build AGI and beyond with. I don't know. We talked a lot about the last one, about how, you know, DeepMind has to be product focused and that's, that can be very challenging for people who want to solve for the most important things in AI and in science. So noteworthy and definitely something we're going to continue to keep an eye on is the movement between the last hubs.

B

All right, so back to our regularly scheduled government focused programming here. So this past week the European Parliament voted to amend the EU's AI act for the first time since it took effect in August 2024. So the biggest change here is a delay in putting parts of this act into effect. So the main compliance deadline for high risk AI, for things like hiring, education, law enforcement, this is moving from August 2026, so just in a couple months to December 2027. Basically a 16 month extension and AI being built into regulated products like medical devices and machinery. The regulations around that are now being pushed to August 2028. And the reason given is that the technical standards companies need, need to benchmark compliance against are not yet ready. This amendment's also adding a ban on AI tools that generate things like non consensual intimate imagery. Now, the EU Council still has to approve this, but it's expected to have that in front of it before August. Now, at the same time, in light of everything we've talked about so far, Paul, you had highlighted this week, there's this debate heating up around a concept called digital sovereignty. Digital sovereignty, which is being pushed largely through some key figures and voices at the United Nations. This is the idea that every nation should build its own AI stack at home, its own compute, its own data and its own models. This had some attention drawn to it this past week because Jacob Helberg, the US Undersecretary of State for Economic affairs, published an essay called the Digital Sovereignty Trap arguing hard against this idea. He called it backward and counterproductive and warned that it would leave dozens of countries with smaller, weaker copies of technology that already exists and each pouring billions into rebuilding last year's breakthrough while the frontier moves on without them. So Paul, just curious on your thoughts on both of this. I mean, I sympathize, like we've talked about all episode, how hard it is to regulate AI, but my God, delaying parts of the AI act by 16 months seems crazy given that, like, we're probably gonna have AGI by the time this thing is going into effect. And like, how do you see that squaring up? How are you thinking about the digital sovereignty thing?

A

I just, I mean, the whole technical standards that they need to benchmark compliance against are not ready yet. Like how we just talked about this, like, how do you devise these technical standards when the, the goalposts keep moving and what these models are capable of and how they're trained and how efficient they are, are. And what you thought would have been a good technical standard might not be anymore and it's just going to keep evolving. Yeah. And then the idea that like, again, A lot of this is being driven, the digital sovereignty is being driven right now by the fact that our allies who thought they could rely on the US can't rely on us anymore because we may choose not to give them access to these frontier models. Again, like, this whole episode is just like, it all is interconnected, but it's all so, so broad and like complex and like dependent upon each other. It's kind of wild, actually. It'd be interesting. Like we, we have a notebook LM that we put all of our episodes into. This would be a really interesting episode to take the transcript of and like, look at the, like the mind map that notebook builds of like how these things are all going. Yeah.

B

The interconnectedness of these topics and I wonder too, the through lines over the last several years of what we've talked about and how it's led to this would be really interesting to visualize. Yeah. All right, next up is our AI use Case Spotlight segment where each week we give you a quick look under the hood at some AI use cases we're exploring at SmartRx. So Paul, I'm going to share one real quick and then turn it over to you.

A

Okay.

B

So this week I just wanted to share, kind of related to our talk about Codex, how I used Codex on a very practical data analysis project. So I was basically just trying to better understand how a specific piece of content connected to revenue. And basically we have all this rich data that, you know, we thought could help us better understand whether, you know, a content asset was actually influencing revenue along someone's path. Like a lot of real business data, though the answer was not in a single field. So I had a lot of messy data and they used Codex like a genius data analysis sitting next to me. So I gave it a structured export, had to inspect the data, identify which fields looked revenue related, which looked related to attribution, which ones are probably too noisy or duplicative to trust at face value. Importantly, used a fully anonymized segment of data, kept this focused on patterns, fields and counts and definitions. Not like there no personal details in here, but super interesting to note the scale here. So I could literally, not even kidding, I could barely open this spreadsheet without crashing my computer. It was like 144,000 rows, rows, a thousand columns. Codex was able to scan all that, narrow it down to a few hundred fields that looked relevant. From there it turned this into a very focused analysis. So it surfaced likely candidates for fields I might have been looking for, explained what each type of field could and could not tell me, gave me a sanity check on some smaller cohorts from this larger data set so I could quickly understand kind of exactly how to model some revenue attribution based on real contacts. And you know, I think this gets at what I was saying before. Like Codex is sold as this coding tool, but I was not using it to write software, which surely it was writing code as part of its analysis. But I was using it as a general purpose agentic system for knowledge work. So you could certainly try to do this in chat. You could drop a spreadsheet in there and analyze it. That's nothing new. But what's cool is that with something like Codex you can just hand it a goal, say here's where I want to end up. And then it just runs the whole investigation into the data itself. It fixes its errors, goes down different paths, chains, dozens of steps together until the work was done. So kind of just behind the curtain of like why you might want to use a more robust tool like that for something like this.

A

I'm anxious to see that report schedule meeting. I want to see how this all worked. Yeah, mine I just have two quick ones. So one I've alluded to like I'm working on a couple of like very high level strategies, strategic things for the company. One of them is developing an AI transformation system. So I actually wrote a little bit about this in my exec AI newsletter this past weekend for SmartRx and then I posted an excerpt of that on LinkedIn. So you can see a little bit what I'm talking about on there. But in essence I'm developing a set of tools and frameworks to help individuals and organizations guide their transformation, understand where they are in their journey and then develop action plans based on it. And so I've been working on a lot of components to this. There's probably over two dozen elements to the transformation system. I would say that we will release all of it in, in the next few months probably. So I've started sort of sharing out some elements of it. But there was one thing in particular I was really struggling with how to explain. And so like I'd created an internal brief. It was, I was reading and I'm like I still. It's just not there like this. It's not. The messaging isn't simple enough and I would predominantly done it in clothes Claude. And no matter what I how I pushed Claude, how I changed the question, how I put the kind of it just kept not. It wasn't getting there. So I was like all Right. I'm gonna start fresh. I go into ChatGPT and I did this while I was driving, which is, again, kind of a cool element for me. It's like, all right, I got 20 minutes. Like, I just, I'm just gonna think about this out loud. So I just open it up. I'm like, listen, here's what I'm working on. Here's the trouble I'm having. I need a way to describe these things that's totally intuitive, and it just, like, makes sense sense. And it came up with something. I was like, oh, that might actually be it. And then I was like, okay, let's. Let's go down that path more. And so I was on my way to a breakfast, actually, I get to breakfast. I was there early before the person I was meeting. I continued the chat then in, you know, typing form, where I was like, all right, let's play around with this. And here, you know, here's some additional context. And I think I, I ended up nailing it. Something that I've been working on for months, that I just couldn't get through it all of a sudden worked by just bouncing over to Chat GPT and giving a little fresh perspective. So that was amazing to me. I mean, it's like a foundational piece of how we're going to explain everything we're doing. So this is like a one of those, like, it's a random conversation on the way to a breakfast that may end up becoming core to Smarter X's positioning and product offering for the next 10 years. Like it, those kind of things. And then another fun one. I, I, I have a golf bag from that's still branded with my agency that I sold in 2021. And that bag itself is probably 10 years old. It's falling apart. So I was like, I need a SmartRx golf bag. So I found a bag I wanted, and I was going to go order it, but I'm like, wait, I want to see what it would look like with the embroidering on it. So I took the image of the golf bag, dropped it into Gemini, and I was like, can you help me visualize the SmartRx logo on this bag? And. And it, like, for whatever reason, I don't know if Nano Banana got dumb, but, like, it couldn't do it. It was driving me insane because I was like, this is not a hard thing I'm asking you to do. And so I went over to ChatGPT and I was like, let me try over here. And it nailed it. So that was a fun one and it's like it looks amazing. I'm hoping the actual bag looks as good as the one I created, but I sent it to the vendor, I was like, here's, here's what I want it to look like. Here's what I want the logo placement to go. Here's the bag I want versus me trying to explain the positioning of the logo and things like that. That so that was just a fun use.

B

That's awesome.

A

I'd love that.

B

All right, we're going to wrap up with some AI product and funding updates real quick, so I'm going to run through these very fast. But first, OpenAI and Broadcom unveiled Jalapeno, which is OpenAI's first custom AI chip. This is an inference processor purpose built for large language models that the two companies co designed in a nine month cycle they call the fastest ever for an advanced chip. They have initial deployment targeted for the end of 2026. Anthropic launched a feature called Claude Tag in Beta. This is a slack integration that lets teams tag Claude as a collaborator in a channel so it can pull context from the conversation, use connected tools and complete tasks asynchronously. Anthropic says its internal version already writes 65% of its product team's code. Anthropic, as we alluded to, also accused Alibaba in a letter to U.S. officials of the largest known distillation attack against it to date. They alleged Alibaba used roughly 25,000 fraudulent accounts to run 28.8 million exchanges with Claude between late April and early June to illicitly extract its capabilities. The company's stripe, Anthropic and OpenAI are backing Intercept, a new $500 million nonprofit led by Stripe executive Nan Ransohoff that aims to dramatically reduce and eventually eliminate respiratory infections through vaccines, air cleaning infrastructure and antiviral research. Google is taking a tougher line with news publishers and its AI licensing talks, pressing them to grant broad rights to use their content, including to train its AI models in a Google News pilot and warning that they some that they will lose payments from its older showcase program if they do not sign sign on. At the same time, Google is investing about $75 million in Studio A24 and partnering with its DeepMind unit to build AI filmmaking tools, its first stake in a movie studio. It's a multi year non exclusive deal that does not give it access to A24's film library, but it did draw immediate backlash from A24 fans. Mirandil Co founded or founded by anthropic veterans. Veterans Benham Nishibor and Harsh Meta raised a 200 million dollar seed round at a 1 billion dollar valuation, led by Andreessen Horowitz and Kleiner Perkins, with Nvidia participating one of the largest seed rounds in AI history to build AI that accelerates AI research itself. General Intuition also raised a $320 million Series A at a 2.3 billion valuation, led by Khosla Ventures with General Catalyst Eric Schmidt and Jeff Bezos investing to build large action models trained on billions of gameplay clips from the 17 million monthly users of its metal platform. And finally, Superhuman, the email company Grammarly bought and rebranded last year acquired AI detection startup GPT0 for an undisclosed sum, adding GPT0's 19 million registered users and roughly 30 million in annual recurring revenue to its existing operations.

A

I would suggest people go read the Claude tag stuff. We'll put some links in the show notes that was one that probably could have been a topic for sure today. It it alludes to the future of work stuff we talked about with the OpenAI research. Once you give agents access and treat them like co workers, it changes the dynamic. So it's just in preview Salesforce workers are not happy about it from an information article. Like they're not understanding why they're allowing this kind of thing to happen in Slack instead of Slack bots doing this. So it's it might take a few months for that to be realized, like the significance. But I think it's just continuing to show practical applications of agents and how it can almost overnight change the way we interact with them and how how they function as true co workers.

B

And one final reminder, we have our weekly AI pulse survey up at SmartRx AI forward slash pulse. So this week's survey is going to ask a little bit about how you feel about the US Government handling the release of new AI models and also getting more of a pulse on how, if at all, you're using things like Codex or Claude code in your own work. So Paul, I mean just easily a banner week. Not always in a good way in AI. I appreciate you breaking everything down for us.

A

Yeah, like we said at the beginning, might might get some models released this week. Maybe the government will feel generous in the fourth of July week, maybe in not. Yeah, we'll see what kind of news dump we get in a holiday week, but hopefully not as much government talk next week. I could do without it. And I'm sure our listeners would like a break too from that. All right, thanks Mike Thanks Paul 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 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 SmartRx Slack community. Until next time, stay curious and explore AI.