#221: Anthropic vs. the White House, Microsoft CEO on the Future of Firms & AI’s Token Crisis

A

I could actually see this administration banning the use of Chinese open source models by US companies within the next 30 days. Because what's happening now is to save on cost, companies are just going to start using deep SEQ models and things like that. I could totally see this government just going on saying you're not allowed to use foreign models. Welcome to the Artificial Intelligence show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Raetzer. 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 SmartRx 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 221 of the Artificial Intelligence Show. I'm your host Paul Raitzer along with my co host Mike Caput. We are recording Monday, June 22, 9am Eastern Time. We are expecting new models this week. We are expecting news out of cans. Is it can or cans? I was.

B

Yeah, I think it's can, but I'm not sure.

A

That's so funny. We've never gone obviously, but yeah, major conference going on. Lots of things already happening this morning. I was following news out of there this morning. So yeah, we're going to have a ton to talk about. But the model news this week is expected to be significant. We also may see movement on Anthropic versus the White House this week. I would, I would expect something. We're two weeks in now of very little movement, but I think that's about to change. So we are expecting a significant week. So tune in for sure next week as well to episode. I think that's going to be 222. I don't think we have any special episodes this week. So yeah, we're going to recap what's going on with Anthropica in the White House. To start off, we've got a letter from Satya Nadella from Microsoft that we're going to spend some time on this week and we are going to dive back into how in the world people are thinking about tokens and pricing and AI adoption within organizations because Mike and Taylor on our team have been doing some digging and it is probably more confusing than we all thought it was. So we're going to do our best to try and share some of that. All right, so this week's episode is brought to you by siteenprove. AI Search is changing what it means to be discoverable Site Improve is the agentic content intelligence platform that marketing teams use to track, optimize and prove performance across both traditional and AI driven search. From AEO visibility to content quality, siteimprove helps you stay ahead of the shift. Start with a free aeo check@siteimprove.com aipod I am actually going to be going and doing that as soon as we get off of this. I'm very curious to see how our site does on the AEO check. So again siteimprove.com aipod all right, our AI pulse if you're new to the weekly podcast format, we do a quick Pulse each week. It's an informal survey of our audience. It's SmartRx AI Pulse. You can participate in these we ask two to three questions each week following the episode based on something we talked about within that episode. So just sort of gets our audience a chance to provide some feedback. So the first question last week that we'll give you kind of the results of Anthropic just pulled access to Fable 5, its most capable model, after a US government directive. How do you feel about it? 42% said troubled. No one should be able to yank a model like that. 26% said it shakes my trust in building on any single AI provider. 19% I need more details before I can react. And 14% glad they complied. Safety has to come first. And then the second question was OpenAI just came kicked off its IPO process. If you could, would you buy in? This is almost split, like right down the middle. We have 35% yes, I'd buy at the IPO. 23% I'd wait and watch how it trades. 21% no, it's overvalued or too risky. And 21% I don't invest in individual stocks. That's interesting.

B

Yeah.

A

All right. So again, informal polls, just kind of a pulse of what our audience is feeling about different topics. At the end of today's episode, Mike will give us again reminder for that and a of what the questions are going to be. All right, Mike, we're going to turn it over to you to get us into Anthropic versus the US Government, the continuing saga.

B

Yes, it is continuing. So for some reminder and some context, Anthropic has spent kind of the past two weeks basically in a standoff with the Trump White House over its most advanced AI models. Things have not yet improved. So as we recall and we covered on the pod. On June 9, the company released Mythos 5 for trusted organizations, along with a public version called Fable 5. Three days later, the government basically passed export controls on it. That forced them to take it offline for all users. The trigger here was security scares. Fable 5 is essentially mythos with guardrails bolted on to block its most dangerous cyber, chemical, and biological capabilities. Amazon CEO Andy Jassy, whose company is one of Anthropic's biggest investors, actually called Treasury Secretary Scott Besant directly to flag an alleged way to jailbreak those guardrails. Apparently, the NSA reviewed this, concluded these protections they had put in place could be stripped, and the Commerce Department imposed export controls that eventually forced Anthropic to essentially just pull the models while they figure this out. Now, Anthropic had insisted the whole time. The concerns are overblown. The jailbreak was minor. A group of independent security researchers agreed with them, arguing in an open letter that the action took the best models away from defenders without any real risk to justify it. Since then, the administration's position has hardened a bit. It seems. Officials told why fired, that getting Fable 5 back is now Anthropic's problem to fix and that the company has to guarantee the guardrails cannot be jailbroken at all, which is something most experts say is just simply not possible. Now, we also just had the tone shifting again because Dario Amadei was at the G7 summit where Trump was, which we'll talk about in a bit. In another topic, a few days later, Trump told Axios he no longer sees Anthropic as a national security threat, saying they have behaved very responsibly. He separately told the Wall Street Journal that negotiat are, quote, going fine, but as of right now, nothing is truly resolved. Fable 5, Mythos 5, they're offline. You cannot use them. The formal commerce directive still stands, and Anthropic says only that it is confident the models will be back in coming days. Obviously, this is all unfolding as Anthropic moves towards what could be the largest IPO ever. So, Paul, it's been a couple weeks here since the whole debacle around Fable. Curious. Like, on the granular level, do you see any end in sight here? Do you see Fable coming back? But also, what are the bigger picture implications here?

A

I. I do think they'll resolve something. It seems like they're heading in the direction of some sort of agreement on, I guess, a more formal structure for how these things will be tested, not just Anthropics. Model, but other future models from other labs, because this is going to be an ongoing issue. We learned a few other things last week. So we have now the letter from US Commerce Secretary Howard Lutnick that warned Dario amadei on Friday, June 12, I think is when this letter was originally sent. It was. Bloomberg Got it on June 15th is when they published it. And so in that letter, the way that they did this is it says specifically the Bureau of Industry and Security, or BIS, U.S. department of Commerce is charged with administering and enforcing the Export Control Reform act of 2018. So they used that act as the way to force them to basically shut this down. Specifically, it says, authorizes BIS to establish interim controls on emerging and foundational technologies that are essential to the national security of the United States. So this is the key is they're saying this is a threat to the security of the United States because of its cyber capabilities. The letter goes on to say, I'm informing you that a license is required for the export re export or transfer in country, including deemed exports and deemed re exports of Anthropic's Claude Mythos 5 model and Claude Fable 5 to all destinations worldwide and to all foreign persons. Then on June 17, we got the Wired report that said that Trump administration told Wired that if Anthropic wants to re release Fable 5, it will need to ensure the model's guardrails can't be circumvented. Which you pointed out, Mike, is impossible. Not only for Fable 5, it's impossible for all models. It's not how these things work. It's not like patching a bug in software that you fix. These things inherently have the ability to be jailbroken and that doesn't go away. You can put guardrails in place, but there's nothing that that Anthropic or any other lab can do to fulfill 100% guarantee that they can't be jailbroken. So there's like, like anything where these, there's these back and forths, there's somewhere in the middle is the truth of all this. If the government is actually asking for 100% guarantees, then they don' understand how language models work. I don't believe that to be true. I assume someone in the government fully understands that that is an impossible ask. And so they're just whatever, they're just putting this information out there into the world for some reason. We also The Economist on June 18 had an article where they talked about maybe another trigger of this, maybe related to the Andy Jassy conversation, but something else that could possibly have been the cause for the export controls. So it said. On June 11, Mark Warner, the vice chair of the Senate Intelligence Committee, said that General Joshua Rudd, who leads National Security Agency, the nsa, and the Pentagon Cyber Command, had told him that, quote, mythos broke into almost all of our classified systems, not in weeks, but in hours. Now, that that quote was just sort of taken, like, someone tweeted it, and then it just got, like, run all over AI circles.

B

Yeah.

A

It was missing some context. So the journalist from the Economist who actually did the interview where that quote came from, Then posted on June 22, he said, this now widely circulated claim is based on a line I wrote last week. Now, I had read this article when it first came out, and I did think it was odd because that quote was just sort of in, like, the fifth paragraph.

B

Yeah.

A

With. No, then it was just. And then it just went. They went on to the next thing, and it was like. Like, whoa. Like, that seems like that's the lead to your story. Why? Why? And so what he says is, I accurately quoted Mark Warner, vice chair of the Senate Intelligence Committee, saying that the NSA chief had told him missiles broke into almost all of its classified systems. But it would be a mistake to read that literally. I think it surely depends on using Mythos alongside other tools under very particular conditions. I quoted it to give a sense of Mythos potent potency, but it was a mistake not to have added caveats. So now, all this being said, there is new research that shows that Mythos truly is a problem from cybersecurity perspective. So Epoch AI did some research where they looked at all the public evidence that they could get their hands on, most of it applying to Mythos Preview. But the conclusion should hold for Mythos 5, too. So they said, our current take on the public evidence is this. Mythos Preview was clearly a large improvement in exploit development, much better than GPT 5.5, and also seven months ahead of past trends. And Mythos 5 is modestly better still. So they go on to say, the Mythos family's cyber capabilities aren't just hype. If made widely available, these capabilities would likely move us into a new regime of cybersecurity, where vulnerabilities would need to be patched much faster to prevent a big increase in successful cyber attacks. So that's kind of a synopsis of why all of this is happening. So whether the government overreacted, whether they should or shouldn't have used expert controls, at the end of the day, all the evidence points to the fact that These models are more powerful and more capable from a cyber perspective than any previous model and anything that any of the other labs have, have put out. Now, Mike, if you'll recall, back in February 2025, this is like within a month of Trump coming into office, J.D. vance made an appearance at the Artificial Intelligence Action Summit in Paris, France. I was thinking about his comments as we were going through these last couple weeks and I thought it was worth revisiting what this administration said 15 months ago about artificial intelligence and regulation because it seems as though they either didn't understand scaling laws and where these models were going to go, or they just, I, I don't know, like it. I have to lean in the direction that they actually didn't understand what was going to happen because let me read you a few excerpts and then you ask yourself if this administration is adhering to its own very strong messaging about regulation and innovation that they got a lot of accolades for in the accelerationist and techno optimist crowd when they made this speech in February 2025. So here's J.D. vance. February 2025. I am not here this morning to talk about AI safety, which was the title of the conference a couple years ago. I'm here to talk about AI opportunity. When conferences like this convene to discuss a cutting edge technology, oftentimes I think our response is too self conscious, too risk averse. But never have I encountered a breakthrough in tech that so clearly calls us to do precisely the opposite. To restrict its development now would not only unfairly benefit incumbents in the space, it would mean paralyzing one of the most promising technologies we have seen in generations. This administration will ensure that American AI technology continues to be the gold standard worldwide. And we partner the partner of choice for others, foreign countries and certainly businesses as they expand their own use. Again, keep in mind they want to be the standard and yet they're, they're not allowing foreign actors to use the model like they're the ones that stepped in and did this. So when you're limiting access to your models, it becomes very hard to be that standard standard. He goes on to say, we believe that excessive regulation of the AI sector could kill a transformative industry just as it's taking off. And we'll make every effort to encourage pro growth AI policies. America wants to partner with all of you and he's speaking to an international audience of leaders and we want to embark on the AI revolution before us in a spirit of openness and collaboration. We need international regulatory regimes that foster the creation of AI technology rather than strangles it. By preserving an open regulatory environment, we've encouraged American innovators to experiment and to make unparalleled R and D investments. With the President's recent executive order on AI, we're developing an AI action plan that avoids an overly precautionary regulatory regime. The AI future is not going to be won by hand wringing about safety. It will be won by building. And then the final excerpt I'll read, and again, you can go read the whole transcript. But if you step back a moment and ask yourself who is most aggressively demanding that we, meaning political leaders gathered here today, do the most aggressive regulation? It is very often the people who already have an incumbent advantage in the market. And when a massive incumbent comes to us asking for safety regulations, we ought to ask whether that safety regulation is for the benefit of our people or whether it's for the benefit of the incumbent. So again, they either did not understand or believe the scaling laws that seem so obvious to everyone else, or why would he have given this speech? Because anyone who knew what they were looking at with the models and their obvious progression could have predicted, and it probably was predicted within the situation awareness document from Leon Bolt Aschenbrenner that this administration seemed to read. This was inevitable that these things would become superhuman at cyber. So something changed. And they haven't come out and said like what is it that they said 15 months ago that is so different now other than cyber, which they should have known? So I, I don't know. It's like it's really hard to look at what the government is doing right now and think they actually know what they're doing, otherwise they wouldn't have said what they said 15 months ago. So the other thing Mike, that I just noticed, every future model out of Google and OpenAI, you have to assume has similar capabilities like the models we're going to get this week potentially from, we're probably going to get one from Anthropic, we will likely get one from OpenAI. Google's got to show back up into the game at some point. They are very clearly a distant third at this point with their models. So either they, either the US government allows these labs to accelerate or China at some point they cede the lead to them, which they cannot do. Right. So I, I think what's going to happen, Mike, and I haven't seen this talked about much is I could actually see the, the, this administration banning the use of Chinese open source models by US companies within the next like 30 days like I. Because what's happening now is to, like, save on cost, companies are just going to start using like, Deep SEQ models and things like that. I could totally see this government just going on saying we're not, you're not allowed to use foreign models. Now what happens with the models that come out this week? Because the rumor is we will get a new Mythos model. There's no way they're going to release that to the public. But there's word of that and then there's also talk of like a Sonnet 5 type model coming from Anthropic. And then the, the word is that we're thinking it's GPT 5.6. I don't think they would go to 6.0. Has actually been in testing within ChatGPT, unbeknownst to users that sometimes when you're using 5.5, you're actually using 5.6 like their next model. So there's, there's reason to believe that the OpenAI model is coming soon. Now, again, if those models aren't on par with Ms. Oaths, then that seems like a Ms. By. By OpenAI. So you're in this, like, catch 22 where it's like, if you admit that your model isn't as good as the one that's being restricted by the government, you lose. If it is as good, then you're not going to be allowed to release it. And then we'll talk a little bit more about Google later on. But again, like, what are they doing? Like, where is Google in all of this? And maybe they have like 3.5. What are we. Which one are we on now?

B

There's 3.1 Pro, there's 3.5 Flash, I think. So I don't think we have a 3.5 Pro, but I might have to fact check that.

A

Yeah, so there's got to be like a 3.5 Pro. You would think that's sitting in approvals somewhere. For Gemini 4 has to be a leap ahead. Like, if all they do is catch up to Mythos, they are going to lose so much talent. We'll talk about all the talent that's already leaving DeepMind, but yeah, you are about to see the floodgates open if Gemini does not surpass the frontier models from Anthropic and Open Eye in the next, like three months. Like, they're. I don't know what they're doing. I have some more thoughts on what's going on at DeepMind, but yeah, so that I don't know. Mike. It's, there's just a lot going on, a lot to unpack with this topic.

B

Yeah. You know, with this, I just keep coming back to a question which I don't know if it has an answer. I haven't seen any reporting on it, but it's like, okay, let's assume the scaling laws are holding and everyone's roughly in the same position, which I think is roughly true. What is OpenAI doing differently about the next release? Because are we just in this world where they're going to release it two weeks later, it's going to get an export ban slapped on it, or they have to be, I assume plenty of lobbying and conversations behind the scenes, like at the G7 summit.

A

Probably.

B

But what are they doing differently? Or are we just in this place where it's going to be Band Aid solutions every two weeks or four weeks?

A

It's again, it's very hard to look at this and think that there isn't a vendetta against anthropic and preferential treatment for other labs, Xai OpenAI in particular. But if you just go look at the G7 summit, next to Trump was Demis and Sam. So those are the people flanking Trump at the table. And Dario was at the kids table across the room next to Benioff. And I forget who the other person was at the table. I, I mean, Greg Brockman is the largest donor to the Trump pack. That doesn't hurt when The President of OpenAI is one of your biggest donors. We, we know that helps with this administration and pretty much any administration. Sam plays the game really well. I, you know, I, I think that you have to understand how this administration functions. And he seems to be playing the game better right now. And so I do think that there's a chance that OpenAI may have an on par model and they're going to get preferential release approval over anthropic. And that's very slippery slope, but that is just how it works right now.

B

Yeah. All right, next up, we have an interesting post this week from Microsoft CEO Satya Nadella, who published a longer post on X that laid out his view of what he calls the, of the firm in an AI driven economy. This has racked up, as of, you know, this morning, more than 65 million views. And basically he opens with this line that a frontier without an ecosystem is not stable. So he goes to this essay arguing that the AI transition is different from previous platform shifts. So for the first time he says companies can create a real, what he calls cognitive loop between people and digital systems. And the real question here is how organizations keep learning and differentiating in a world where AI models can continuously absorb human and organizational expertise and then commoditize it. So the core kind of sound bite here is he frames this whole thing around two kinds of capital. Every company will need to build human capital, which is the knowledge, judgment, relationships and pattern recognition of its people, and token capital, which is the AI capability a firm builds and owns. So Nadella claims that the human capital actually does not get less valuable as AI grows. It gets more valuable, in his words, quote, without human direction, you have compute running in circles. The opportunity here, he says, is not about picking the best model or tool, but building a learning loop on top of models where human and token capital compound. So a company should be able to swap out a generalist model without losing that kind of company veteran expertise that's built into its own systems. And some ways to do this are through private evaluations, internal reinforcement, learning, and a queryable knowledge base. He calls this all the new IP of the firm and a machine that compounds over time and is hard for competitors to replicate. He also included some warnings saying that no one should want a world where every company seeds value to a handful of models that eat everything they see, comparing it to how globalization hollowed out industrial economies through outsourcing. So his prescription is to build a frontier ecosystem, not just a handful of frontier models, so value flows across every company, industry and country possible. So, Paul, there's a lot to unpack in this, but I really like how he framed the relationship and the value between and the value of human capital. And token capital is a really good kind of term to use here, I think.

A

Yeah, it definitely resonated on X at least. Like I have no idea again, if this was covered anywhere outside of, you know, social channels, but 65 million plus people and lots of comments from people thinking like this. This made a lot of sense to them.

B

Yeah.

A

So, you know, certainly worth a conversation. It's interesting to me the continued break from their own reliance on OpenAI's models and just the overall concept of not centralizing the power and knowledge into like, you know, the three core model companies. This builds on Mustafa Sullivan's June 2nd article that we talked about about building a hill climbing machine and even used that same language in some ways. So just a quick recap if you didn't listen to the episode where we talked that a few weeks ago in this article, Mustafa talked about this idea when they were announcing the launch of seven new models that Microsoft was building themselves from the ground up. Big focus on reinforcement learning environments that allow the Microsoft AI models to learn directly from workflows. He talked about thinking of them as training gyms for AI that are accessible only to you. So again, it's the same concept that these things continuously learn in this loop based on what you're doing, he said in that article. You're building your own model, trained on your data, within your environment, controlled by you. Your institutional knowledge becomes part of the model and it stays yours. The goal, as Mustafa said, is to build what they think of as a hill climbing machine. People really like this term. I haven't seen it used until recently, but it's become kind of the way at least Microsoft's described it. Actually, I heard Demis talk about a hill climbing machine. So an organization that continuously improves cycle after cycle as they apply more compute, better data and sharper evaluations. And then the key, I think is what Mustafa has said and they talked about this at first in 2025 and then they kind of doubled down on it this year with the article on June 2 is the idea of humanist superintelligence. So this is Microsoft's kind of flag in the ground. They define that to mean advanced AI systems that are designed to serve serve people and organizations, not replace them. These systems must remain tools shaped by human intent, accountable to human oversight, and ultimately subordinate to human goals. People, you must always remain in control. Now, I've mentioned this book numerous times on the podcast, but if you go read the Infinity Machine, there is a lot of background on Mustafa Solomon and how he became a co founder of DeepMind. And then his focus, the humanist superintelligence thing, is very, very aligned with how Mustafa has seen AI for the last 15 years. I actually thought it was going to be a bit of a break from how Microsoft would think about it because I don't know that it's the most capitalistic centric approach to this. And I think that I still do think at some point this messaging, while it sounds amazing, doesn't hold up necessarily to the economic models that are going to be required for Microsoft to compete here. But I like it. I agree with it. I feel like I want this idea to win. I just don't know how viable it's going to end up being. So Satya says a lot of really positive things that I like to hear, but I just don't know that it holds up. So a couple of things I'll call Mike says AI model. So this is now Satya's post. AI Models can continuously absorb the expertise of humans and organizations and commoditize it. He is seemingly taking a shot at like Meta and Zuckerberg there where everybody can monitor what your humans are doing. Some of you will monitor what we're doing in order to replace the need to have those people. So we're just going to have the software developers train the AI how to be software developers so we don't need software developers anymore. It's kind of like the contrast where he's at least positioning AI as augmentation of humans, not not learn from them so we can get rid of them. He then says human agency will be the driver of token capital growth. Humans will set ambitious goals, connect dots across domains, build relationships and recognize patterns that matter most. That is 100 true. What isn't addressed in his article and Microsoft hasn't really addressed at all that I've seen is, but you're going to need fewer humans to do that. Like, yes, that is 100% true. That is the role of humans, humans. But one human can probably do the work of 10, eventually 100 humans in that role as the agents get better. Now, what this hopefully means is Microsoft's aware of that and they're going to continue to think about what that means, the implications to the number of humans that are needed within jobs. Then it's the future of the firm is the ability to compound the learning across people in AI. 100% agree companies need to turn their workflows, domain knowledge and accumulated judgment into AI systems that approve with each user juice. Totally. Like, it's like right now we all have this very limited version of this where you turn memory on in ChatGPT or Claude and it starts to remember things and eventually like remembers it across your whole team and they get smarter based on the memory and the context they hold. We're in like the first at bat of the first inning of how context and memory applies to these things. So totally. Like we can all look ahead and say how amazing will that be when it learns with every prompt, every interaction, every piece of knowledge that's given to the system. And it doesn't matter if I move across models, like all the models we use have that same knowledge base that that would be very, very powerful. And then the final thing, I'll note that Satya said that I 100% agree with and we've been talking about this for like two years on this show. Private Eval should capture whether a model is actually improving against outcomes that matter to the business, not external benchmarks that is is like 100% true. It keeps saying it's like how an AI lab compares their models to other models and IQ tests and things like that is 100% irrelevant to your business. Like, they're all super smart, they're all super capable. We can stop looking at those standard evals that come out when these models are released. All that matters is the evals of the work you actually do in your company. And is this model better at that eval than it was at the previous Model 1? That's all that's going to matter moving forward. So I think there's going to be a lot more systems put in place, technology created to help establish and, and monitor those evals.

B

Yeah, I kept coming back reading this to that point about context and human capital, just because it's especially, you know, in 2026, something I've been personally interested in and working on, you know, at an individual level is like, how do you give these tools all the context and judgment and background they need to actually do hyper personalized work really well or outcomes really well? And I think that that is where, you know, when we teach learners about this through our AI academy, when we do talks, workshops, whatever, I think it's a part people miss quite a bit because it's not always intuitive to say like, wow, I should be collecting all these documents or information or institutional knowledge to give to this kind of machine brain. But it can be such a differentiator. If you do it right. You don't need Fable 5. That would be amazing. But you don't need Fable 5 to move ahead very, very fast if you have that context layer.

A

Yeah. A real practical example of this I'm thinking about is like, let's say you have a team of 100 people who all have access to Claude, and let's say there's five to 10 power users in there who are like every day are making breakthroughs that are like 10xing their productivity, their innovation, their creativity. And then everybody else is sort of at differing levels of their capability. But let's say like Mike is one of those 10 people who are just like crushing it and figuring new things out every day, new ways to prompt, new ways to work with the system, new outputs to ask for, giving it access to new knowledge bases that like, you know, levels up its capabilities. And let's say that the model learns from Mike's use of the model. And now rather than isolating those breakthroughs that Mike is having in terms of using the model every day, the model can say to me, When I go in today to work with it. Listen, one of your co workers has actually found a way, created a project that does this. Should I apply that to what you're doing right now? It's like, holy, I had no idea. Mike did. Like, right now you get into all kinds of, like, governance and like, what access does Mike have to data That I may or may not, but. But the premise is when one person makes a breakthrough or gives the model something, it knows that context that now helps it. Shouldn't that then spread to the other hundred people? Shouldn't that learning transfer to the 100 users? Right, but right now it won't. Unless Mike happens to post it somewhere or runs a lunch and learn or like, whatever, Mike's just going to go crush it and him and the nine other people are going to like 10x themselves and everybody else is going to, you know, be questioning whether or not they're getting an ROI from it. So that's like, I think where this goes is this idea of this compounding learning with every single interaction across an entire team. And nobody is close to solving how to do that.

B

Not at all. No. But that is truly going to be an incredible unlock, if we can.

A

Yeah.

B

All right, so for our third big topic this week, we're talking again about AI pricing and usage. Because this has become one of the bigger issues in AI people are trying to solve for. It's something we keep coming back to on the show. There's an interesting piece in Wired this past week that kind of went inside some, like, examples of this emerging scramble inside companies to manage the soaring cost of AI usage. So for instance, you know, AI vendors charged by the tokens, just the unit of text going into and coming out of a model, everything you send, everything it generates gets counted and build. So the more a company leans on AI, the faster that meter, so to speak, can run. So, like, they shared some examples of, like Royal bank of Canada's CEO said its token usage jumped 500% in six months. Cisco CEO called its own usage quote pretty, pretty crazy. Roughly 300 companies brought up AI token costs on earning calls in April and May, up from just 93 a year before. Some, like Meta, Uber, Salesforce, we've talked about a couple of these, have started capping how much employees can use. So, like, overall, Paul here, like, the honest truth is like almost nobody, us included, has kind of fully figured out how to think about this yet. So. So what we wanted to kind of do in this segment is I was just going to kind of walk through Some of what we've actually learned through some of the research, myself, Taylor Rady, our director of research, have been doing into figuring out like, okay, help me wrap my head around what's going on with like usage and billing, because it should seem you would expect it to be pretty straightforward. You would be incorrect. But I think a lot of people assume this is more, more clear cut than it actually is. So I'm going to dive into a few quick findings here and if you have anything we want to talk about, we should, we should talk about it.

A

Yeah. And real quick, just context. So the way this came about is like we were running into these Claude usage limits internally and it was becoming a major problem. And you know, give you a really specific example. Like let's say Mike or Taylor has a course that they're creating on a Friday and we run into a limit by 10am Friday morning. And like they can't do their course because now they have a five hour wait time before they can get back in and use Claude fraud. So we're like, what the hell? Like how, how are we supposed to solve for this? So they go off for a couple weeks, they do a bunch of research. Mike explains to me like last Wednesday or Thursday, like, hey man, here's what we found. It's crazier than you thought. And I was like, dude, just share this on the podcast. Like we. So this is literally real time. I've seen some of what Mike is about to share with you all, but like, we're just trying to figure this out like everybody else, as Mike said.

B

So before, I share kind of some of the complexities we've run into. I just want to give a quick refresher to people. So I actually found this really help. Some of this might be a little basic for some people, but I think it's really worth quickly understanding how this works. So we're all in one way or another charged per token, right? So tokens are these small individual units of language that are basically output and input by AI model. So for something like Claude, a token is approximately 3.5 English characters. So what you're running into is about 750 words, roughly equals about a thousand tokens or so. Everything you send and everything the model says back to you gets counted in these units. Now, as if this needed to be more complicated, there are two different types of ways that labs charge for these. So input tokens are everything you put into the model. Output tokens are everything the model generates back. Output tokens cost significantly more than input tokens it's typically two to five times more than input tokens because the model can pretty efficiently read all your inputs in one pass. But then when it generates the output one token at a time, it's predicting the next token, the next one, the next one. More work per token means higher prices per output token. So basically you can have two workloads that use the same amount of tokens. But depending what type of tokens they are and what they use them for, they can be priced totally differently. So providers often list a price per million tokens with separate input and output numbers. You'll see this on the pricing page of anyone's, you know, API section on their pricing page. So as an example, you could give Cloudsonnet 4.6 a 1500 word brief, which is let's say roughly 2000 input tokens. Ask for a 600 word summary, that's roughly 800 output tokens. This would cost overall, once you really do those input and output calculations, about 1.8 cents. Nothing wrong with that. That's pretty cheap, right? But that's one request, right? Organizations can run millions of these. So this is how you really with that kind of volum over time as AI adoption increases, this is why usage and bills go up. But really what we've talked about that's exploded costs is, you know, late last year, as things like Claude code and Codex become even more popular, we now have conversations going into agent loops or it being more agentic. So in those things entire prior context is resent as the input on every single turn. So you know, if you're on phase 10 of your conversation, it has to go reread phases one through nine. AI coding agents, for instance that like read a project, open files, run tests, feed errors back. That's stacking more and more input onto every later step. So this is why you're seeing these companies suddenly all have this exploding usage. It's often just due to agentic AI. So like for one other example that I found that was really helpful is like imagine a customer support assistant. So to answer well, let's say it needs a knowledge bas like we just talked about. Let's say it needs a 20,000 token knowledge base. Let's say it handles a thousand questions a day. In many cases that knowledge base is resent as input on every request. So like you have like if you do 20,000 tokens times a thousand requests. I won't get too much more into the math here, but it's like 20 million tokens a day, which is like 60 bucks a day just to reread the same knowledge base. So let's talk about like what do you do about all this? There's no perfect answer, but I do want to get into some of the new nuances because this is where a lot of our research focus. So one big thing at the API level is prompt caching. So if a big chunk of inputs repeat, like system prompts, reference documents, providers will actually cache it and charge a fraction of what they normally charge. So the fun part of this, every vendor offers this, but do it totally differently. So anthropic, you have to actually mark apparently what's cache. I have not done this. We have not had cause to do this yet. If you're on a paid like flat fee plan, this all happens on the model provider side. They do all this on their end to keep things cheaper for themselves. OpenAI's is automatic. Google has like two modes, apparently. So without getting in the weeds. The point is there's this method that you can use that is the predominant method to dramatically lower your costs, which people take advantage of. But the other ways to do this are ways we've talked about to save money in the past. So like for instance, right, sizing the model, using cheaper, smaller models. If at the API level you can actually batch or do asynchronous runs of stuff, which is less money output is the expensive half. So if you're capping your response lengths and things that can help and then also trimming context, getting strategic about what you're actually putting in. Again, these are all band aids though I would say like, except for maybe model routing is potentially a longer term solution here. But here's the thing, Paul, like all of these levers for the most part live at the API charges you go per token level. But most businesses just buy per seat licenses like us, right? And you still have all these usage limits to figure out. But the thing is once you start saying okay, well if you're on a license, you have a set amount of usage. Here's what our research found. Good luck figuring any of that out because everyone treats it differently. For instance, we are looking into, to begin with a Gemini Enterprise which we are using Gemini Enterprise Business Edition and the To Google's credit, their reps have been super helpful answering a bunch of questions. But something like Claude, right? You just have like a Claude team account. Everyone has their own usage limits. You don't really know exactly what those are. You just see the meter running and you can know when you're approaching it. Gemini Enterprise has And I'm just going to read some things that they sent me. In Gemini Enterprise Business Edition, quotas are pooled across your entire organization. Each purchase license contributes a set amount of capacity, for instance, 120 queries per day to the organization's total shared pool. So in some cases you may have individual usage limits, like we do under a CLAUDE team plan. If we move to a Claude Enterprise plan in, we would now be on like pooled usage, meaning the organization is taking from everyone else. Gemini Enterprise works like this, out of the box. So we say every user seat gives us more usage, but we're still all pulling from the same pool for the most part. Now also Gemini Enterprise, when you hit your quota limit, it just stops working. So that, that's interesting. So there's all these nuances again. This stuff could change tomorrow. This is not how OpenAI works, it's not how Claude work. They're all different. So you've quickly run into this problem of not only untangling how your license and usage works, and it differs depending on the license within a single model provider, not just between anthropic and OpenAI and Google. There's like eight different ways this can work if you have a specific tool. So figuring that out alone, pooled usage versus individual is critical. But even within that, unless it is quotas like Google is offering, we say, okay, it's a certain amount of messages, but like, is anyone like sitting here tallying and counting? I mean, it's like I, I come back to like, I appreciate there's some clarity around usage and around how the rules work. And I get that it has to have some limits, but I don't really see a sustainable way to be sitting here. Whether it's tokens, a pooled usage, individual usage quota. Am I just gonna have to sit here with like a calculator tallying things while I go? Seems insane to me, so I'll stop there. It's a lot to process. For the moment, we're making progress, understanding better how this is working. But in terms of solutions, I'm really struggling sometimes to untangle this. And this doesn't even get into when we are, I assume as we deploy more agents outside of using agentic tools on a paid set plan. Getting into the API is a whole new interesting animal. And most organizations are doing a combo of both, it looks like. So it's not one or the other. Neither one is clear and super easy to solve.

A

Well, and you're one. That was a great breakdown. Super confusing. As I expected, you're focused on like the three major labs who are selling direct to consumers. Imagine now like we use HubSpot. Well, they're not selling direct. They're passing the cost of the tokens through on anything we do within HubSpot with their margin on top of it. And it's like, good luck figuring that out like it is. So I guess what we're saying is we'll know we've reached AGI when this isn't a problem anymore, when we have a model smart enough to solve how to price models for business usage. I mean, as you explain this, I'm thinking like, okay, maybe someone in IT and procurement is like capable of solving this. But do you think like ahead of sales, ahead of marketing, ahead of customer success, who now has to like manage these agents, has any clue, like any context whatsoever to have ever dealt with a pricing model like this? So yeah, just, I'm mean again, like we've talked about this maybe three or four straight episodes now you understand why this is not some simple like just do it this way. And I, I lean in the direction still that, you know, the simplest solution is probably the answer. And I think the simplest solution comes back to a per seat license where you're just subsidizing the most powered users. Yeah. But then you have a bunch of people who aren't using their limits and it's like Everybody just pays 200amonth or 500amonth or like whatever that number is, you just charge by more. But the labs don't have enough adoption and there isn't enough AI literacy within enterprises for them to know how to get that value out of it. So if you just raise your rate all of a sudden, say 500 bucks per month per license. Someone like smarter X, like, yeah, cool, great. Like we're getting that ROI and it's not even debatable. But if you do that to a major enterprise that has yet to do AI literacy training, specialized like personal personalized adoption training, they're gonna be like, no way in hell we're getting 500. Like we can't justify that. It's like, okay, well the labs have more work to do before they have enough of an educated buyer to know the roi. So they can't just raise the per seat license. You could do human replacement cost of like, hey, you got 10 marketers, you're only going to need two now. So you, you can use some of the difference, but no one can market that because they would lose customers because they're replacing humans and they can't Say that.

B

Right.

A

So it's like the simplest answers to how to solve this aren't really super viable right now for a variety of reasons.

B

And you know, there's, I'm sure some more technical people that would say like, oh, okay, well, you can use a really cheap, fast model and through the API you can get a set understanding of how much it'll cost to build something maybe like that works for one off services and agents you're building to accomplish a task. But like, but like there's so much amorphous strategic knowledge work being done. Like, it's just like how, how much does it cost to just have a straight deep dive strategy conversation for three hours with Claude? I don't know, who knows? Like, I'm not going to sit there and be like, oh, I'm going to make sure this one stays at 50 chat. That's not how thinking works.

A

Especially if you're in voice and you're just talking. Right.

B

It's like crazy. For all this knowledge work stuff that's going to be transformative, that's not building repeatable AI powered services. I get that that part's a little more quantifiable and manageable, but for this stuff. And also we haven't even gotten into like the transformation part of it. This is just to like do stuff we're already doing more efficiently and better. But the innovation piece, who knows?

A

Yeah. And we'll talk about Sakana, the Japanese lab at the end and the product updates. But they just came out with a model. There was getting a lot of rave Reviews first like 12 hours. It came out like last night, I think. But in essence, like my initial research, it's basically just a routing system. So then you have this opaque routing thing where it's like, okay, so let's say I go into Claude and all my model choice is gone. I can't pick of my seven or whatever models that you currently can pick when you're in Claude. And so now I have employees who maybe aren't as advanced in their usage, who can't pick like Opus 4.8 and like Burn through tokens unnecessarily.

B

Yeah.

A

But now it's like totally opaque to everybody. And now when I want to do my deep thinking, I don't even know which model model I'm getting. I'm assuming they're giving me the model because it costs them less per tokens or we're running up against our budget. So it arbitrarily moves us to like the more efficient model because it Knows otherwise I'm gonna get pissed and like run into my lint. Like, oh my God. Like, this stuff is so complicated and it just feels like all the labs are literally just like making it up. Like, all right, let's, let's try this model right now, right? See what happens, happens. And all of this pricing is built for developers. That's the other thing that, like, annoys the hell out of me right now is none of these models make sense to the average knowledge worker. They are built for researchers and developers who understand how to work with APIs and the metered access to all this stuff. It is not built for marketers and executives and stuff like that to, to comprehend at all.

B

Yeah. And just one final thing I would say about it is like, again, people might argue, well, okay, like this is going to get solved at some point as model prices drop and compute, more compute comes online. Well, A, sure, but that's going to take some time. B, like in the next three years, business A, we're going to be hitting almost super intelligence in the next, like three years. So like, businesses are trying to plan and position. You cannot do that without some predictability. And the, the lack of predictability here, I think is really what gets me and what is very difficult because, like, how are you supposed to adopt all this great technology? How are you supposed to diffuse it everywhere if I have no idea what we're getting charged for certain things.

A

Yeah. And then the final note again, we could probably talk for another hour about this because every time like now my brain's just like going on this. What they all want to have happen, what the labs are positioning again. I said this last week, like the it word in Silicon Valley is loop. And we just heard it a bunch more times from Satya this week. And hill climbing. Apparently what they want to have happen. And again, they can't say this yet. They want to take an enterprise that's 10,000 people and let's say your marketing team is 300 people. Their belief is likely that you can run that 300 person marketing team with like 30 people or less.

B

Yeah.

A

But to do that, you have an email marketing agent, an SEO or AEO agent, you have a strategy agent, and they are running on 24, seven loops. So your marketing team never sleeps. So you only need like 15 to 30 people to orchestrate all of this. But your AI is running all the time. You are literally just burning tokens all day long. And it's learning from itself and it's running a campaign and it's tweaking the campaign and it's updating the creative and it's updating the media by and like. That is the inevitable outcome of these models becoming generally capable and autonomous and reliable. In the next, like two to three years, the models should be at the point where they can largely do what I just explained across every function, across every department. But imagine the amount of tokens you're going to burn to do that. Now, could you get, like, let's say we get Sonnet 5 this week and three years from now, Sonnet 5 is 1000 times more efficient than it's going to be when it comes out. Then in theory, the cost to run that model 247 across hundreds of agents, maybe it becomes right. I don't know.

B

Like, how do you even plan? Yeah, People would say, like, okay, well, isn't it worth whatever money you're saving on those people? Maybe. But like, how are you plan out any of this? Like, expecting an average enterprise or company, like an entrepreneur's at the bleeding edge. Great, go do this and say, I'm going to spend a million bucks to get 5 million of value. Amazing. But like your average company, it's insane to think they're just going to go all in on something like this without any type of predictability or measurability, right?

A

Well, yeah, because it requires an understanding of what these models are going to be capable of in 18 to 24 months. It's very rare I have a conversation with a business executive who can even comprehend deeply where we already are today, today and, like, what they're already capable of. So it's going to be a very, very select few organizations that can truly, like, live in the future, I would say. And it's going to look very weird to companies on the outside and the comp, the competitors who can't keep up with what they're doing.

B

All right, Paul, before we dive into our rapid fire this week, another announcement here. This week's episode is also brought to you by the AI for Business Boot Camp by SmartRx. We are coming to Columbus, Ohio on July 16th. This is a single day, 8:30am 30pm at the Hilton Columbus at Easton. This is a full day set of workshops and content built for professionals and leaders who are ready to accelerate AI adoption and value creation. So the day is going to start with a keynote from Paul. State of AI for Business keynote that transitions into two highly interactive workshops with Paul and myself. So I'll lead an AI productivity workshop in the morning. Paul will lead an AI innovation workshop in the afternoon. And this is built for any AI forward. Managers, directors, professionals, executives across every department who are ready to move past AI theory. We're going to build real AI powered workflows, get strategic frameworks to accelerate transformation. You also leave with an immediately actionable plan for yourself and your team all in one single day. If you're one of our AI Academy Mastery members, you get discounted pricing. Discounts are also available for teams of two or more. Groups of 10 or more can get custom pricing. So be sure to join us if you're available. You can use the code POD100 to take 100 bucks off your ticket to grab your spot. And spots are limited. Just go to Smarter X AI, click on events. You'll see the AI for business boot camp as an option right there.

A

My guess is we will also have a lot of time at that event to sit down and just talk about this. Things like the token, but like just get ideas from other people who are going through this because that's the best part of these live in person events is like the collaboration, the ideation, commiseration, like whatever you want to call it.

B

Yeah. All right, let's dive into some rapid fire this week. So first up we've got some big talent shakeups at some of the big labs, but the biggest one so far is Noam Shazir who's a vice president, was a vice president of engineering at Google and co led co lead of its Gemini models. Announced this past week he's leaving to join Open AI. So he has a bit of a history here. There's a bit of a big reversal based on this because less than two years ago Google reportedly paid $2.7 billion to bring shazier back along with a team of his researchers after he had left in 2021 one to co found the chatbot startup character AI. After that he went on to co lead Gemini. He's widely credited as a key figure in Gemini closing the gap with Chat GPT. He is also basically AI royalty. He first joined Google back in 2000. He was a co author of the seminal attention is all you need paper that introduced the transformer. That's the architecture underneath basically every modern chatbot. At Open AI he's going to lead architecture research and OpenAI's Mark Chen welcomed him as quote extremely AGI pilled. He's not the only big name on the move. John Jumper, the DeepMind scientist who shared a Nobel Prize for AlphaFold announced he is leaving Google DeepMind after nearly nine years to join Anthropic OpenAI also hired Dean Ball, who we've quoted a number of times on the podcast. He is a former top White House AI advisor. He is the lead author of the administration's AI action plan plan. They have hired Ball as the head of Strategic Futures. Basically. This is kind of interesting too, because Ball has recently become a vocal critic of Trump's feud with Anthropic. And on the churn side, Barrett zaf, a senior OpenAI research leader who helped build Chat GPT. He had come back to the company about five months ago after leaving Mira Morati, former CTO's thinking machines lab. He is out after just five months back and we'll see what happens to Barrett. So Paul, not everyone may know these names nearly as much as Sam Sam or Dario or Demis, but these are, these people are a big deal. This is a big deal. Some of these shakeups are happening huge.

A

So I mean Google DeepMind has a very deep bench. That's not debatable. But when you just look at the recent months, like David Silver, again, probably not a name a ton of people know, but if you watch the AlphaGo documentary, you know who David Silver is. College friend of of Demis. Major, major player in the breakthroughs the last decade. He left. No, I'm leaving is a massive deal. John Jumper. That one shocked me because again, if you read Infinity Machine, you'll, you'll get a little bit more of the backstory on, on Jumper. But while it's obvious that Gemini currently has fallen pretty far behind chat GPT and, and Claude, I assumed DeepMind is where you were if you were serious about science, that, that the one thing it seemed like Demis and Google were going to own was scientific breakthroughs. And for jumper to leave DeepMind seems problematic. And again like we're all just sort of looking from the outside in and trying to kind of connect the dots here. But the thing I can't shake, Mike and I, I haven't really said this on the podcast yet, but like when you read the Infinity Machine, the vibes you get from Demis are, are like there was a number of times where I would like listen to his, his quotes from the author and I would be like, I think he might leave. Like, like you, you just, you really. There's a lot of times where you listen to his interviews and you think he, he never wanted to run a product driven lab like he, he had to. He is a researcher through and through who wants to solve the biggest challenges in the universe. And, and when I see these major players, including like Silver that he's so close with Jumper. It's like something is off. Like, there's just. I don't know what it is, but like, and I don't. I'm not saying I think Demis is actually going to leave, but if like tomorrow somebody tweeted that Demis has left to start his. Restart his own lab again, focused on pure research, I would look at that as like, yeah, that, that makes sense. Like, I would not be shocked at all all if that was something that actually occurred. Which is not great, like, for, for that. But again, if you read that book, I, I don't know how you come away with that thinking that Demis is like, locked in and totally happy doing what he's doing right now with DeepMind and, and happy with the progress that they're making. So again, I come back to what I said earlier. Like, DeepMind and Google, they need to show, show up big in the next few months. Like just some transformative things because I could see a lot of top talent there if they continue to be third place or worse that are like just jump ship and go to anthropic or open AI and then you get into a situation where I'm not, I'm not saying this is what's going to happen with DeepMind, but like, where Elon Musk at XAI was like, oh, screw it. We're just, we're not going to be one of the frontier models that anthropic and OpenAI have that right? And I wonder if at some point Google's like, hey, we own 14% of anthropic. Like, let's just pour that in. Let's do you know what we need to do in our products? But like, Gemini is just not going to be like a. Again, I don't think that's what's going to happen. But I also am getting to the point now where it's like, I wouldn't even be surprised if things like that started transpiring because of everything that's going on. But seeing this level of talent leave is very weird in a very short time period.

B

You know, it's total speculation, but I'm just curious about your take here.

A

So.

B

So let's say Demis, I think, has been on the record, probably pretty safe to say he believes we'll have some form of AGI by 2030, if not sooner. So you've got a few years window here. Like, he obviously could raise any amount of money he wants. Like, is it easier for him to just do his own thing? Or would he be looking to just say, okay, I've got a three year window to usher this technology into whatever form it's in. Do I just join one of the other two at this point? Big labs.

A

I can't see him just joining a lab. I mean, I know him and Dario are close. Yeah. Again, like when you, when you listen to his interviews, because in the Infinity Machine, he did like 40 hours of interviews with the author and there were a lot of like, very vulnerable conversations where it was just like sitting in a cafe and like sharing his thoughts. And again, I'm not reading into anything. He, he literally says, like, once we get the AGI, I think I'm just going to go away for a while. Like it's, you know, he, I think he is on this sor of treadmill where he's just trying to get to the end game. And up until now, but he tried to negotiate his way out of Google multiple times prior. You know, prior to this, he saw Google as the, the most likely path forward to achieving his, his mission.

B

Yeah.

A

And if at some point that changes or the need to be so product driven continues to get in the way of him solving, you know, the biggest challenges and questions in the universe and he feels like he just needs to go somewhere else to do it. He's a very miss mission driven person. And, and actually the analogy, Mike, I thought of this morning when I was kind of getting ready for this. There's the story they tell in the book about, you know, his pursuit of AI, where he loses this chess match. I think he was like 12 years old and he basically was in this like extended match and he, he made a mistake in the end and he lost when he shouldn't have lost. He just kind of got fatigued. And he said like that night in the hotel room, he's realizing that like all of this brain power is sitting in this room playing chess when they should be out solving like the most complex things in the world. Right. And that's when he started looking at like, AI is like the path to do this and like more people should be working on these things. And so it's not unprecedented that at some point, and this is where I guess where I started kind of drawing the parallel where if he feels like his brain power is being wasted on productizing this technology instead of applying it to solve these problems, he has shown throughout his life, he will just leave and go do the thing that needs to be done.

B

Yeah.

A

And so I think that's why I feel like there's Some something has to give here. I don't think Google is moving as quickly on the frontier advancements as maybe, maybe he's going to want it to be. And, and like something's just got to change internally or external. I don't know. Like it's. Something isn't right. Like it seems unsettled right now.

B

Gotcha. So next up, the G7 summit this past week happened in France and it turned into a bit of a major AI event. There was a closed door working lunch on Wednesday that brought about a dozen tech executives together with heads of state, including President Trump. So we saw at this meeting Anthropic CEO Amade, Google DeepMind CEO Demis Savas. Sam was there. They, they. But Dario and Demis used this meeting to actually call for a US led international coalition to set the rules and standards around AI. And Sam, also in the room, backed them up, pushing for a global forum to test models and establish shared standards. Amade also told the leaders to, quote, resist the temptation to splinter over AI. He argued that cooperation should include structured access to frontier tier models and a chip and component supply chain that excludes China puts joint work on the risks of AI in joint work on the risks of AI in cyber bioterrorism and intelligence. Obviously this is all playing out against the backdrop of the White House fighting with anthropic. Separately, G7 leaders discussed a quote, trusted partners plan that would give select allied countries or companies a path around those restrictions to access advanced US models. So Paul, just as a refresher, you know, the G7 of countries is the U.S. france, UK, Germany, Italy, Canada, Japan. We had a bunch of people in the room, including Trump Secretary of the Treasury Scott Besant, Lutnick and Rubio as well. Along with these AI leaders, there are a bunch of AI leaders from other companies, basically other AI champions in G7 countries. Countries. But kind of interesting to see CEOs standing right in front of Trump wanting a US led global coalition when the Trump administration is also banning US models.

A

Yeah, I really feel like if Dario was willing to play the game like we talked about earlier, he would have a very, very important seat at the table right now. Like, I think Anthropic has demonstrated the need for greater collaboration with other democracies, democratic countries and allies. And I feel like they just don't want to give in to him because of how he has approached the administration so far.

B

Yeah.

A

But I do think that Demis has played the game well. Sam has. And as weird as it is is I could see Sam and Dario actually finding a way to work together. And maybe Demis is the bridge to bring them together to do this. But they all see the same thing. They all know what these models are going to be capable of. They all talk about them differently and work with the administration in different ways right now. But there's no denying the power of these models and the importance of them to national security and democracy and all these things. And I just feel like they've got to find a way to, to do something together. I just don't know what it's going to be and I, you know, maybe just getting them all in the room like this is a good start, but something's got to happen sooner than later or else we're just going to be in a very weird regulatory regime that's going to slow down innovation, which is not what the government wants.

B

Yeah. One, I'll just say one other note, as I was doing some research on this, like, obviously we talk about constantly and we know, know totally that like the US is an AI leader, but when you look at like the other companies that were at the table that are all basically one company representing each country, essentially you have Sakana AI, which is interesting for Japan. There was a company called Domen, which I'd never heard of, which is Italian. They're an AI, sovereign AI company for regulated industries. But they were Italy's first unicorn, actually, and one billion dollar valuation in 2024. London based, Britain's most valuable generative AI startup, Synthesia, was present and represent. But that's a four billion dollar company. Germany had something called Black Forest Labs represented. They make the Flux image model. They were people who created stable diffusion. They left stability AI Again, these are all really interesting companies and like, no knock on them, but you're like, wow, these are just like your average startup back in the us, not like a national AI champion. And it was right. Really stark to see that.

A

Let's just say, yeah, the US dominates.

B

Unreal this conversation. Yeah, sure. All right, so next up, Mid Journey, the popular AI image generator used this past week to announce something pretty out of left field a little bit. They announced Mid Journey Medical, their first hardware product, and they move into healthcare. The centerpiece of their announcement was something called the Mid Journey scanner, which is a full body ultrasound system scanner. The way they describe this, basically you step onto a platform that will lower you into a shallow pool of water. You'll pass through a ring of 350,000 or more tiny sensors that basically act as both a speaker and microphone, sending Ultrasonic waves through your body from every angle. This whole scan takes 60 seconds about and produces a 3D map of your insides down to a fraction of a millimeter. CEO David Holtz says it aims for image quality comparable to an MRI in many ways at nearly a 100 times the speed with no radiation or magnets. This was built with the ultrasound company Butterfly Network. About a dozen people have been scanned so far. The idea here is that thanks to this type of technology, scanning would become casual and routine. You could do it as often as every day. Holtz claims that early imaging and doing enough of it could eventually avoid 30% of all deaths and half of all healthcare costs. To facilitate all this, Mid Journey plans to open what they call a Mid Journey Spa in San Francisco in 2027. This is going to have their scanning technology along with saunas, cold plunges, hot tub tiles style scanning rooms. And they have a goal of more than 50,000 scanners worldwide doing a billion scans a month by 2031. For now, the company is starting with body composition maps that sidestep the FDA clearance required for a true diagnostic imaging. Right now this is kind of a launch event with not that many details. The Verge did note, and I would agree, it's still not entirely clear what Mid Journey's image generation technology actually has to do with any of this. If anything, they speculate it might just be a new business to use unused AI compute. So Paul, I'm curious what you thought of this. Is this like legit? Is this a big swing at preventative health? Is it a gimmick? How did you land here?

A

Yeah, it seems like it's a viable path. I mean I've read a lot of reviews online of actual like doctors and you know, certainly from the AI entrepreneur side, it was getting a lot of positive feedback as like, oh, this could actually work. This is an interesting thing. So overall I'm extremely excited and bullish on the next like 10 years of medical advancements powered by AI. I think it's going to be, be, you know, maybe the most important thing that AI brings us in the next five to 10 years. And so I think it's noteworthy when these sorts of things happen. I think you're going to see a lot of these out of left field things where it's like, wait, what? Like Mid Journeys doing MRI machine? But yeah, I don't know, I, I think it's positive. I, I like any advancements that are, you know, going to make diagnostics more affordable and more accessible to, you know, the general public and hopefully, you know, solve a lot of long term illnesses that otherwise just go unnoticed. So yeah, I don't know, positive stuff. I, I like it.

B

Yeah. And you know, like we've tried to do sometimes highlighting more AI for good, we'll see if it works. But it's, hey, it's not all doom and gloom out there. People are doing some really interesting things.

A

Yep.

B

All right, next up, a new study from the University of Oxford and the UK AI Security Institute out this past week found that frontier AI systems can now rely on viably out persuade expert humans in conversation. This research was led by Kobe Hackenberg and it ran four pre registered experiments spanning nearly 19000 conversations with almost 7000 people. So the test went like this. Participants rated how much they agreed with a contested policy stance, in this case specifically in the uk, something like allowing more immigration or raising the pension age. They basically, basically rated how much they agreed with this on a 0 to 100 scale. They had a short text chat around 7 messages over 14 minutes with either an AI or a human trying to change their mind. They rated their agreement again afterwards and that shift in the score was the measure of who persuaded better. On the human side, researchers recruited the most capable persuaders they could find. They got professional political canvas canvassers, elite competitive debaters that included four world champions. The AI still still won. Even when the humans picked their own topics, prepped in advance, practice for hours, and even when they were paid thousand pound bonuses to win. AI beat the elite debaters by 4.6 points of attitude change. And coaching those debaters against the very AI that beat them did not close the gap. In the final study, the AI was nearly three times more effective in the real world than professional fundraisers at getting people to donate money to a charity called Save the Children. They tested models including Claude open chat, GPT, GPT, 5.4 Grock and versions of Gemini. Interestingly though, an important caveat, it seems like the AI was able to do so well because it could flood the conversation with more info than your average human. The AI averaged 294 words per reply with sub second latency. So it was able to pack in about 37 fact checkable claims per conversation. When they throttled it down to how fast humans could type and what length they were able to do in an average or normal conversation, that dropped to about 12 fact checkable claims per conversation and the advantage of the AI essentially collapsed to zero. So Paul, like long story short here based on this research at least it does sound like AI does not have to be superhumanly intelligent to actually be better than humans at persuadance. It's the scale and speed thing.

A

Sounds like there's a Sam Altman tweet that I've gone back to numerous times and mentioned on the podcast a few times. So back in October 2023, he tweeted, I expect AI to be capable of superhuman persuasion well before it is superhuman at general intelligence, which may lead to some very strange outcomes. Now, he tweeted that 11 months before the release of the first reasoning model 01, but right in the midst of what was called Project Strawberry, where they were working on reasoning capability abilities. And so I've always felt that the models were already superhuman at persuasion. I've felt like earlier versions, but there was guard rails put on them to prevent them from doing it, because superhuman persuasion can be used in very nefarious ways.

B

Yeah.

A

And so it's a very dangerous thing to have the models be able to do, but it's inevitability that they're going to be. And so it can be used for really positive things. But that is, it's a, it's a dangerous thing.

B

Yeah. And I mean, we've already talked about so many times, people already, you know, having AI psychosis or weird relationships with AI, and that's not even a fraction of probably the persuasion capabilities that totally it actually has.

A

Yeah, we're not. When you're working with ChatGPT and Cloud, you're not seeing anywhere near the full capability from persuasion perspective.

B

All right, so Paul, next up, we've got our AI Use Case Spotlight. Every week we give you a quick look under the hood at real AI use cases. We're exploring, building or deploying in our own work at SmartRx. So, Paul, I'm just going to share a real quick one then. If you have anything to share, we can talk through that too. So for me, this past week I was actually preparing for a couple of upcoming talks and I was interested to see if I could actually get a better way of critiquing my speaking style. So I built a way to record myself, practice, practicing and have Claude actually critique it. Now, you can of course use a natively multimodal model, like Gemini, for instance. Just drop like a recording in. It'll help you do some of this. I was curious just doing this with a different model because Claude can't do this natively. So Claude can't natively watch video or really listen to audio. So. But what's really cool is that you can just Decompose the recording into things that can actually analyze and do this locally on your machine. So basically, it's worked in three quick layers. First, it transcribed the entire talk locally with Whisper, like an open source transcription model, so you could get the exact words, my pace, every filler word. Second, this was the part Claude Code figured out how to do, and I had no idea what's even possible. It then runs the audio through a speech analysis tool called Prot. That's like an open source tool that pulls out your pitch and volume as actual numbers and a chart, so you can literally see if you're going monotone, trailing off, et cetera. And third, it suggested sampling a grid of still frames from the webcam video to check eye contact, energy, and gesture. So, not perfect. But it did read all that and hand me back a prioritized critique that I found genuinely useful. It caught some things I was wrong about. I. I thought I had actually gone too slow. The data said I was actually running about 170 words a minute on the fast side, so I actually just needed to, like, trim content. It flagged that I leaned on certain words a little too much, which was helpful. Definitely told me what was working as well, which was really cool. And so the whole point here is like, you can kind of take it or leave it for how much you want to take the advice of AI when doing a talk or something. But the point was I didn't have to, like, sit down and vibe code anything. And that's awesome. I love vibe coding, but I think people assume, oh, I got to go vibe code. So something. It's like, you don't always have to. You can just describe what you want, strategize back and forth with AI to refine your idea, and then something agentic like Claude Code or Codex can just figure out, like, what pieces of software can we put together to make this work? What strategies can we use to actually do this thing? So that was the bigger point for me there.

A

Impressive. Yeah, I. Mine was, again, strategic. Most of my uses are strategic related. I was working on something that required some pretty advanced legal and accounting details that I'm not an expert in. And so basically the way I do it is like I just come up with all the questions I would have if I was talking to my attorney or my accountant, and all the questions I have, you know, as an entrepreneur and someone trying to make some decisions around, you know, the growth of the company. And so I keep that kind of scratch pad where I'm just constantly adding more questions, more Thoughts and then I'll like pop them into the model and ask these questions. Questions. But when I got to a point on Friday where I, I kind of arrived at what I thought the decisions should be, at least what I thought the path forward was. But then I needed to share it with my attorney, I actually went and said, okay, let's now build a brief that I can send to my attorney. Here is what I want. Like these are the decisions we've made. This is what I think the path forward should be. After each thing, I would like you to give me the sample legal language that would end up in the legal document so that my attorney can re react to actual legal language that I would never be able to write on my own. And So I created two documents. Each one was probably between 15 and 20 pages that I then went through, did deep edits of myself, but it was all just regurgitating what we had decided together, me and Claude or ChatGPT and, and then put it into a format I could then send off to my advisors. And so in the end one, it gives us like a way for me to give clarity. And honestly, when I sat down on Friday, I was like, I have to finish this like as soon as possible. Did not think it was going to be even a possibility that I would have done it on Friday. And I actually was able to do all of the work and create both the documents by lunchtime on Friday. So like three hours, I did what, no joke, was 30 to 50 hours of work. Like it. If I had, if I didn't have those models three years ago, like I easily 30 to 50 hours, which means it would have just taken me another couple months because for me to find 30 hours to work on anything is almost impossible. So yeah, just again, one of those things where like if, if you would have just said like, you get to do this one thing with chat GPT all year and it's going to cost you $500, whatever. I'd be like done like, yeah. So that project alone is, pays for my license 100 times over in terms of the value it created for me.

B

That's incredible. I love that. All right, so we're going to end up with some product and funding updates as usual. So I'm going to run through these real quick. So first up, SpaceX struck a $60 billion all stock deal to acquire AI coding startup Cursor with a $10 billion breakup fee if it falls through. This was just days after SpaceX's own blockbuster IPO. This was a move to bolster the AI efforts of its Xai division. Chinese AI lab deep seat closed its first ever external funding round raising more than $7 billion at a valuation north of 50 billion. There's an unusual structure here where founder Liang Wenfang put in about 3 billion of his own money and outside investors received no voting rights. In a five year lockup, OpenAI launched the OpenAI Partner Network $150 million program with launch partners including Accenture, BCG, McKinsey, Bain and PwC aimed at training 300,000 certified consultants by the end of 2026. To help enterprises actually deploy its models, Microsoft made Copilot cowork generally available. This is an agentic system that runs complex long running multi step tasks and end across Microsoft 365 apps. And this is for co pilot 365 copilot subscribers. There is usage based billing and admin set spending limits.

A

Good luck with those tokens.

B

I was going to say good luck. Let us report back with your results please. Chinese startup Z AI released something called GLM 5.2 that's getting quite a bit of attention online. This is an open weights model under an MIT license that tops tops the open model leaderboards and beats GPT 5.5 on several long horizon coding benchmarks at roughly a sixth of the cost. And then finally Sakana AI released something called Fugu, which is a model that works by orchestrating a pool of other AI models behind the scenes, including copies of itself, and can automatically route around any provider that gets cut off, which the company is pitching as resilience against export controls. So everything comes full circle here. Paul, thank you for breaking everything down. Just a quick reminder here before I'll turn it over to you to wrap up, but our AI Pulse survey will be live when you're listening to this for this week. So SmartRx AI forward slash pulse. And this week we are going to ask about AI usage and also ask about your take on Microsoft CEO Satya Nadella's essay that we discussed this week. So Paul, thanks again for breaking everything down for us.

A

Yeah and again reminder this episode is brought to us by Siteimprove. Go check out that a e o check@siteimprove.com aipod that is something that Mike and I are definitely going to be taking a look at and seeing the ways that site improve can help with agentic content intelligence platform for our marketing team. So again siteimprove.com AI iPod we expect as I said earlier, multiple new models this week that is on if the government allows the release of multi new models, I guess we've arrived at the point where we have to give the caveat. If the government lets it happen, we might have a couple new models to talk about next week. So thanks as always, Mike. It should be an interesting week ahead.

B

Can't wait.

A

All right, talk to everyone later. Have a great week. 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 SmartRx Slack community. Until next time, stay curious and explore AI.