A

The canary in the coal mine. Like AI research is the first discipline that AI will be applied to at this superhuman level where it completely redefines the role of an AI researcher. And it's going to happen in the next like 18 months. Whatever happens here, extract that out to whatever your profession is and the same premise will then follow as soon as more resources, meaning money, gets thrown at your industry to do the same thing. 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 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 218 of the Artificial Intelligence Show. I'm your host Paul Raitzer along with my co host Mike Kaput. We're recording Monday, June 8, right around 9am Eastern Time. Important timestamp this week because we expect a very busy week where we'll talk about some of the stuff that's coming. But we're definitely going to get insights into what the new Siri is going to look like from Apple. We are expecting multiple model releases potentially from OpenAI and Anthropic. So it is going to be a very busy week. Mike. We're work work cut out for us on episode two, 19 or 20, I don't know which one it is going to be because I think we may have another state of the industry, one we're going to do, I think.

B

Yeah, somewhere that'll be 220. So the next weekly will be 219.

A

All right, so episode 219, buckle up. There's going to be a lot going on, so. All right. We'll get through what we can this week. There was, it was a very busy week last week. Lots of big things related to IPO plans, executive order from the government, more about the cost of intelligence and tokens. Just a ton going on, it seems. I thought we were going to get a little bit of a summer lull and that seems to not be happening, Mike. Right. All right, so this episode is brought to us by AI Academy, by SmartRx, which helps individuals and businesses accelerate their AI literacy and transformation through personalized learning journeys and an AI powered learning platform. In addition to a community that includes Thousands of AI Forward professionals and leaders. New educational content is added weekly so you always stay up to date with the latest AI trends and technologies. Our very popular AI for Departments collection features seven course series and certificates designed to jumpstart AI understanding and adoption across organizations. There are certificates available right now for marketing, Sales, customer success, hr, Finance, operations and legal. So literally your entire company can get upskilled at AI in about a month. Cram it into a weekend if you want, but those are all ready to go on demand. They've all been recorded in the last, you know, six to eight months. We released a new one each month, so it's a lot of fresh content and then those are going to be updated as we go along and then they're enhanced with our Gen App series and our Academy Lives. So there's obviously this always continuous learning journey that people can do once they get through these on demand core series. So the series are an ideal launchpad for organizations that want to level up their teams and accelerate responsible AI adoption and impact across the company. Individual and business account plans are available. You can also buy single courses and series for one time fees. Go to Academy SmartRx AI to learn more. And you can use code POD100, that's P O D100 for 100 off any individual annual plan. So again, Academy SmarterX AI. All right, and then every week, if you're new to the podcast, we and we have new listeners every week. So I'll give a quick recap. We do three main topics where Mike and I dive a little bit deeper into those topics. And then we do, you know, usually about 7 to 10 rapid fire items ending with AI product and funding news, which we cover like a dozen things rapid fire within those. So that's the format. But we always start off with our AI Pulse survey. So this is an informal poll that we run each week of our podcast listeners. You can go to SmartRx AI forward slash pulse to participate in these polls each week. So we just asked two questions based on the topics that we covered in that week's podcast. And so this week the responses we have from last week's Is your organization spending on AI starting to outpace the value you're getting from it? 43% said, we're not spending enough yet to know. Okay. 36% said too early to tell, we're watching it. So okay, that's the vast majority have no idea. Basically.

B

Yeah.

A

And then 21% know the value clearly justifies it. So yeah, so not too many people saying yes. Costs are getting hard to justify, actually. Nobody said that. Nobody think. All right, that's pretty interesting information. And then the second one is, over the past year, has your personal view of AI become more positive or more negative? Okay, that's 50% more positive, 31% about the same, and 19% more negative. That's good. That's encouraging. I like the more positive. Despite all the heavy stuff we talk about on the podcast each week and the challenges we're facing from a political perspective, societal perspective, economical perspective, I like to see that people are still staying positive and trying to find the good paths forward with AI for sure. That's encouraging. All right, Mike, take us into our main topics for the week.

B

All right, Paul, so first up, this past week Anthropic confidentially filed a draft registration statement, a form S1, it's called, with the SEC. This is their first formal step towards an IPO, towards going public. Because the draft is confidential, the company has not disclosed share counts, pricing or detailed financials yet, though we do know a little bit because the filing comes on the right on the heels of a massive raise. So on May 28, actually Anthropology closed a 65 billion dollar Series H round that at that time valued the company at roughly $965 billion. The business behind the filing has scaled quite fast. Anthropic said its run rate revenue has crossed $47 billion and reporting indicates it is approaching its first profitable quarter. Now interestingly, alongside this IPO news, which is very practical, there's also some very important heavy, high level stuff Anthropic's been talking about because they a piece concurrently from its in house research arm, the Anthropic Institute and it is titled When AI Builds Itself. It is co authored by Anthropic co founder and head of policy Jack Clark and the piece argues that AI is now doing more and more of the work of building AI. It's moving the industry towards what is called recursive self improvement, which we have talked about in the past on the podcast. To support this argument, Anthropic says more than 80% of the code that it now merges into its own code base is written by Claude and that its engineers shipped roughly eight times more code per day in the second quarter of 2026 than they did in all of 2024. The piece also notes that the length of tasks that AI can complete on its own has been doubling roughly every four months now. In this essay, Clark and his co author also warn that as models start building and training their successors, alignment failures could compound growing more frequent but less understood until we lose control of them. And it argues for verifiable systems that would let labs credibly pause frontier development if they ever needed to. So, Paul, two kind of important but very differing perspectives here. We have the practicality of the ipo and then we have this talk of fears of recursive self improvement. Which one stood out to you or kind of what's the bigger story here with these?

A

Well, the IPO story, I mean, basically comes down to SpaceX, and we'll touch on this a little later too. But they're going to go public this week, June 12th. SpaceX will IPO. We talked earlier this year. The XAI, which was Elon Musk's AI lab, has been rolled into SpaceX. So as part of the IPO, it's really our first AI lab going public, I guess you could say. I mean, obviously Meta and Google are already publicly traded companies. Anthropic, OpenAI and XAI were the ones we were watching. So X AI is going to, in essence, be folded in and go public through this. So we'll learn a little bit there. And then it becomes a race between Anthropic and OpenAI as who's going to get to market first and then which company is going to have the higher valuation. So that's what we'll continue to watch. Because of the confidential filing, we don't have a ton of insight yet into, you know, Anthropic's business, but we've been learning a ton that they are growing fast. They had profitable months. It's going to be really interesting. And I, I think I posted on LinkedIn a month or so ago, it's like, okay, you got a hundred thousand dollars to invest. You can only put it into one of these companies, like which one are you betting on? So yeah, it'll be fascinating to see how the market responds to both of those companies then. Just overall, the trend line of AI labs overall and the impact they're going to have on the economy. The article about when AI builds itself I found infinitely fascinating. I actually wrote about that in my Exec AI newsletter editorial on Sunday because I think it's an extremely important concept that a lot of people probably haven't really stopped and thought deeply about. We've talked about it numerous times on the podcast, especially over the last six months, that recursive self improvement was likely one of the key unlocks to achieving AGI or Artificial General Intelligence, along with World Models and a couple others. But in essence, the labs are all Sort of betting on there probably needs to be like one to three. That's. And I'm just categorizing based on like how Demis would explain it. Demis Sabis like there's probably one or two or three breakthroughs out there. Some people like Demis and I think OpenAI and Anthropic think that it's basically taking large language models, scaling them up and then layering over a couple of breakthroughs. Then you have people like Yann Lecun who think language models are useless and that is not the path forward. But I would say the dominant consensus is that language models are certainly essential to the path. And if we can layer on things like recursive self improved improvement, that's probably going to be enough to get us to AGI and beyond. So what I'm going to do is I'll break down some of the key elements of that article and I'm going to recommend people definitely go read it. You're probably going to have to read it a couple times. I think there's just a lot in there, especially if it's a newer topic to you. So it said the subhead is our progress toward recursive self improvement and its implications. So then it says for most of AI's history, humans drove every step in its development cycle. But at Anthropic, we are delegating a growing share of AI development to AI systems themselves, which is speeding up our work, taken far enough and given enough compute. That trend points to an AI system capable of fully autonomously designing and developing its own successor. This is called recursive self improvement. We are not there yet. And recursive self improvement is not inevitable, but it could come sooner than most institutions are prepared for. So just to stop and unpack that real quick, the models are being used to improve and create future models of themselves. Like they're absolutely. All the labs are using the like the current version of the best model to help develop the next model. But it's heavy involvement of AI researchers. Like they're very, very in the loop. What they're saying with recursive self improvement is in essence, the humans largely come out of the loop and the model is in essence continually improving and rewriting its own next version. That's the basic premise. So what they then highlight, and you kind of alluded to this, Mike, is how fast these things are improving at this capability. So the models are improving way faster than they were six months ago, in part because they've gotten so good at COD coding. And this is directly impacting productivity. So Anthropic uses itself as sort of patient zero in a way. Like they're saying, hey, this is what we're seeing. So I'll call out a few elements of the article to highlight what they're saying about what's going on within Anthropic. And this is like how you and I might try and do this with Smarter X. We're trying to say, like, listen, you can look at all the studies and all the research in the world and all the thought leaders and all the experts who are claiming different things about jobs and the economy and all these things. What we always try to do is say, but we're living this every day and like, what are we seeing? And so I always like when companies take this like very inward looking and saying, listen, the data may say something else, but like, here's what we're seeing. So they said the rate at which AI models improve is accelerating. The length of tasks that they can reliably complete on their own has been doubling roughly every four months, up from an earlier trend of doubling every seven months. Now that seven months is a key stat that Mike and I have talked about many times over the last 18 months. That comes from Meter, which is an organization that actually studies the reliability of AI agents to do tasks that would take humans x amount of hours. So they look at a human expert and say, okay, it would take a human expert four hours to do this thing. The AI can complete that task way faster at a 50% reliability or an 80% reliability. So they look at how long a human would take and then they say, can the AI do it and how reliable is it for the AI to do it? So what they're saying is their internal data has gone from seven months, which is what Meter showed, to four months and it's dropping fast. They said in March 2024, Claude Opus 3 could complete software tasks that take humans about four minutes to complete. A year later, Claude Sonets 3.7 managed tasks that took about an hour and a half. A year after that, Claude Opus 4.6 managed 12 hour tasks. If this tread holds, tasks that take a skilled person days could come into the range. This year in 2027, AI systems could be capable of tasks that take a person weeks. Now keep in mind, and Mike, you and I talk about this all the time. They're focused on AI research and development of like coding, programming stuff. But what we're always saying is like this translates out to all knowledge work at some point. Like they're focused on AI research, because that's what all the labs are trying to build first. But whatever happens there, you as a listener, like you need to extract that into your own industry. And it may be a year or two or three behind what they achieve with AI research. But the same premise holds true. No matter if you're a lawyer or an HR professional or a CEO or like whatever it is something that takes you weeks, you're going to start to be able to do in minutes through simple prompts. Okay, so then they say large performance gaps persist when it comes to Claude exercising judgment in choosing goals in both engineering and research. So they talk about building AI models takes two major disciplines, engineering and research. And what they're saying right now is the gap is the model doesn't have great judgment and taste yet. That's where humans still excel, but they don't see that as a scientific limitation, meaning, like they think they can get there. It's just going to take a little bit longer, they said in April 2026. So just two months ago, Claude shipped over 800 fixes that reduced a class of API errors by a factor of 1000. The engineer overseeing Claude estimated that a human would have taken four years to complete this work. Now again, this is computer stuff, but like take that and extract it to other professions on the most open ended tasks. Claude's success rate reached 76% in May, up 50 percentage points in six months. So again, all of this is just like context that they are seeing. Rapid acceleration to the point where they have to write this article and say we don't even have good intuition as to what this means in two years, but this is the data we're seeing. They then said in May 2025, Claude Opus 4 averaged a 3x speed up over the starting code by April 2026. Claude Mythos preview, which is the one we've talked a lot about over the last two months that was not released yet, may be released this week, but only like the government and some select people have access to this. Claude Mythos preview was achieving 52x for calibration. A skilled human researcher would need 4 to 8 hours to reach 4x in this part of research workf flow, where they're optimizing steps within a clearly defined experiment. Claude has gone from super helpful to superhuman in under a year. Then it says an area of human comparative advantage for now is research, taste and judgment, which I alluded to earlier, including choosing which problems matter, which results to trust and when an approach is a dead end. So again, think of all of this, as I said in the newsletter, the canary in the coal mine, like AI research is the first discipline that AI will be applied to at this superhuman level where it completely redefines the role of an AI researcher. And it's going to happen in the next like 18 months. Whatever happens here, extract that out to whatever your profession is and the same premise will then follow as soon as more resources and meaning money gets thrown at your industry to do the same thing. We'll talk about this happening in the accounting industry in a couple minutes. So then the article presents three possible futures. They said one, the trend stalls, but today's AI capabilities are widely diffused. The article features many exponential trajectories, but these trajectories may actually turn out to be S curves where it kind of goes up and then it flattens out. We may be approaching the bend in the curve where the returns diminish and the line straightens, then flattens. Even if model capabilities were frozen at today's level, we would expect major changes to occur in the world. We are still early in diffusion of today's models into the wider economy. This is real important where a hundred person company can increasingly do the work of a thousand person company because each employee will sit atop a pyramid of agents. They include this scenario they say for completeness, but do not believe it's likely. So this idea that we're just today's it like we're just going to basically freeze the development and this is going to be the smartest models we ever have. Highly unlikely. But what they are saying is today you could in theory build a 100 person company that can do the work of a thousand person company. And I 100% agree with that. Like I, that's probably actually not that hard in most industries if you know what you're doing. Second, the AI labs continue to see compounding efficiency gains in this scenario. AI development becomes substantially automated. But humans continue to set research direction and judge results. So in this case, organizations that use AI systems will become much more efficient as time goes. We would expect to see significant productivity multipliers on each person. 100 person companies could do the work of 10,000 or 100,000 person organizations. This would revolutionize knowledge, work and government services. The evidence they say that they've laid out suggests that this is the most likely scenario that you can build 100 person companies doing the work of 10,000. When I think about Smartr X and what we're building, I think we can be $100 million a year company with 100 or fewer people. So what that would equate to is a million in revenue per employee. I actually think two to three is very likely. So you could build $100 million company that traditionally might have taken four or five, 600 employees, a thousand employees with dramatically fewer people, because those people are going to be empowered by agents and they're going to orchestrate these agents. So I think that is very realistic. And especially if you're an AI native company that builds from the ground up, it's going to be a lot harder to compress down from a big company like a 10,000 person company down, but it's going to be pretty easy is kind of not the right term, but it's going to be a lot simpler to do it from the ground up. And then their third scenario is AI systems themselves become capable of recursive self improvement and begin building their successors. If technology trends and advancing capabilities continue and AI systems are able to develop the capabilities inherent to transformative human ingenuity, then it is plausible that AI systems could design and refine themselves in this world. The pace of progress in AI development becomes determined entirely by the availability of COMPUTE for AI systems. Humans play a substantially diminished role in their development, likely moving most of our efforts toward oversight, validation and verification of an expanding virtual lab run by AI systems. And then they say we do not have good intuitions for what this world would look like because our economy is currently driven by humans and human built tools. Now. All that said Mike, the thing that everyone latched on to, at least in the media, was this headline you saw everywhere of like anthropic calls for slowdown in AI development. They're pulling that from the very end of the article. They said if it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing. We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology. Demis Hassabis would 100 agree with them on this. He has said that publicly, as has other AI leaders. It's not happening like that. I, I, we can, you can have as many articles about this as you want. You can have Congress get involved, which they're going to, you can have executive orders on this. That is not going to happen because China won't agree to it. And unless China agrees to it, and even then you need the frameworks to make sure that they're adhering to it.

B

Right.

A

So it's this whole thing like you would first have to have all the infrastructure in place to monitor. It's like atomic weapons, you can monitor the enrichment of uranium. Like you can monitor if someone is adhering to a policy where we're going to slow down that stuff. AI wouldn't work that way. And so they presented some ideas and said like, hey, we're going to research and try and come up with ways to do this, but a slowdown just, just isn't gonna happen. And even if it did, if, if publicly they said it was happening, no way like it, they would be doing it in using dark money and like hidden lab. Like the US government is not going to slow down no matter what they say publicly they're doing. And so people have to understand like that scenario three, there's, I, I, I mean I'm running out of like energy and chip availability are literally like the only two things I could think where the government doesn't like pour trillions into it. Like they will do whatever it takes to be the ones to build that thing no matter how dangerous it is. So super important article. Like I, I mean I know we're spending a lot of time up front on this one, but I, I really think people need to comprehend the idea of recursive self improvement and the fact that every major lab is pushing forward with it even though they know how dangerous it may be.

B

And just to reiterate a really important point you made, even if you're listening to this as a business leader and saying, look, this is sci fi, I've got enough on my plate like read it and imagine it's 5% true and you have some thinking to do. I would argue so.

A

Well yeah, and even Mike, just like go to the bottom, go to the end of it and just read the three scenarios.

B

Yep.

A

The first scenario like, like think about just that one, like a hundred person company doing the work of a thousand person company. And then think if you're a legacy company, like an existing enterprise, someone is building the hundred person version of your thousand person company and how are you going to compete with them? So like you should be thinking about the other things, but you need to at least be solving for the reality of anyone can come up and build a competing version, an AI native version of your legacy company. Now you may have product and distribution and things that are going to be really hard for them to replicate. And so not every industry is going to be just disrupted over the next three to five years per SE but in a lot of industries, there's not very many obstacles to someone just building a smarter, more efficient, more profitable version of your company. And that's going to put leaders in a really tough spot where they're going to have to make some very unpopular decisions around staffing structures and cost structures, because that that's coming like that is. I don't even see that as debatable. It's inevitable across every industry.

B

All right, so our next big topic this week, somewhat related to just how serious this is all getting, is this past week, President Trump signed an executive order titled promoting Advanced Artificial Intelligence, Innovation and Security. And this executive order says that frontier AI Labs can now opt into a federal process to have their most powerful models designated as, quote, covered frontier models. If a model qualifies, the lab can give the government 30 days of early access before public release so that government agencies can evaluate its cyber security capabilities. Now, the White House Office of Science and Technology Policy made a point to say that it is not conducting oversight of all new models. And the White House has framed this process as applying only to models that, that represent a meaningful step change in cyber capabilities, not routine version updates. The order itself states that nothing in it creates a mandatory licensing, pre clearance or permitting requirement for new AI models. So that's one piece of this. However, some of these assurances are getting a little overshadowed by some headlines this week that suggested the government may be open to taking financial stakes in the AI labs themselves. So according to Politico, quote, President Donald Trump said he will likely meet with AI companies at the White House next week. This was last week to discuss what he called a federal government, quote, partnership that would allow the American people to profit in their success. Also, according to Politico, quote, the policy concept first reported by an outlet called Notice, which is actually, interestingly, a media firm founded and headed by POLITICO's former owner, quote, has been floated by OpenAI, which issued a policy paper backing a public wealth fund. And broadly discussed by anthropic and billionaire Elon Musk, who runs xai. Senator Bernie Sanders recently proposed a bill that would create a 50% government ownership in AI companies. Additional sources did stress these talks are really early. The legal mechanism for the government to even do this is unclear. It may not happen at all. But Paul, lots of chatter around this, like how should we be thinking about both the executive order and are you taking seriously this idea that the government might take a financial stake in AI Labs?

A

I mean, they've already done it with intel, like the precedent's sort of already set. Yeah. So this is what we've been waiting for is AI taking over Washington. Like it's, both sides are constantly posting on X about this. There's bills being proposed. Like it's now front and center within politics. It's going to become more so in next five months as we, you know, race into midterms here. So it is the talk of the town in Washington. The executive order was a little bit of a surprise because it's almost the exact executive order that was planned like a month ago or whatever it was. And Trump had called like 24 hours notice, all the big AI leaders, minus Dario Amade to come to the White House for the signing of the executive order. Then David Sachs shows up at the last hour and convinces him not to do it. The only thing that apparently changed was they went from a 90 day window to a 30 day window. So there's a fact sheet, like if you want to just read the fact. The, the order itself isn't very long. I mean it's probably 1500 words or something. It's nothing crazy. But there's a fact sheet we'll put a link to. You can read it. I'll just highlight a couple of the pieces of that. So it said the order directs federal government to establish a voluntary framework in collaboration with AI developers regarding covered frontier models which would provide the federal government with secure early access for trusted partners to strengthen cybersecurity, promote secure innovation. Voluntary is, I just, I would put, personally put that in quotations. The way the government does voluntary is like, yeah, you can choose not to participate and then we can choose not to give you government contracts. So it's not, it's like wink, wink, voluntary, they go out of their way. In the fact sheet, there's a whole section dedicated to it to talk about how the administration is the most innovative administration in history. And they're, you know, we're not trying to stifle anybody and this is not, you know, true regulation. We are all in on this and we're going to do everything we can to drive it because there's a lot of people who don't want the federal government involved in regulating this, including people like David Sachs. So they can't come across as stifling innovation because they have to be. They have to continue to show strength in the race for AI supremacy against China. And so this is this fine line they're walking between it now. Interestingly, two days later, a bipartisan group of House lawmakers on Thursday unveiled a proposal that would regulate AI that could override some state laws. This one is a 269 page framework, so there's a link to Axios we'll put in. They had a nice little breakdown on this. It would preempt state laws on the development of AI models for three years. Formally established the center for AI Standards and Innovation, tasked with making voluntary standards and guidelines and appropriate 100 million per year from 27 to 29. This would require large frontier developers to write and implement plans addressing risks prior to releasing new models, as well as to report critical safety incidents. And the draft legislation would also protect AI whistleblowers, which is going to be very important. Increase fines for AI enabled fraud and try to boost funding for AI literacy, education and research. It would create a mechanism for the government to study AI's impact on the workforce and add some workforce protections around AI. So again, just bills are going to start popping up everywhere. Interesting side note what's going on, especially related to cybersecurity is Anthropic's Claude Mythos model is apparently being used by the nsa, the National Security Agency and the forward deployed engineers that we've been talking a lot about. They have them embedded with the nsa. So while Anthropic is being blacklisted and talking about, you know, basically being the Antichrist by the government, they're embedded with the NSA and they're apparently using it to be proactive cyber attacks, like preparing to do these things. So who knows what is actually going on with the government in Anthropic, but it's this continuous dance. Like we're going to be really angry with them over here and then we're going to like, like use them over here to do this other stuff. Now the Bernie Sanders thing I think is interesting. Like I don't even know where the origin of some of these ideas are coming from. Like, Trump does an interview on Thursday or Friday where he's like, oh yeah, we're gonna like look at taking 50 stakes. I'm gonna meet with Sam Altman next week. But just like three days prior, Sanders had an op ed in the New York Times where he presented this idea. So I don't know if like Trump's latching onto a Bernie Sanders idea or if this came. I'm not, I don't know. But Sanders op ed that again, we'll link to. He said AI will almost certainly be the most transformable technology in the history of the world. It will profoundly affect the life of every man, woman and child in our country. It will bring and is already bringing unimaginable changes to our economy, our democracy, our emotional well being, our environment, and how we educate and raise our children. Further, there is a very real fear that as AI becomes smarter than humans, it could eventually function independently with potentially catastrophic consequences. The question then is not whether AI will change the world. It will. The question is who will own and control that future? Who will benefit from it, and who will be hurt? Will AI be used to make life better for working families? Will it enrich our quality of life? Will it help us eliminate poverty, extend life expectations and solve the climate crisis? Or will the future of humanity be determined by a handful of billionaires who have promoted and developed AI with virtually no democratic input, who stand to become even richer and more powerful than they are today? That is why I, Bernie Sanders, soon to be introducing the American AI Sovereign Wealth Fund Act. This legislation would give the public a direct ownership stake in the largest AI companies in our country. How it would create a sovereign wealth fund through a one time 50 tax, not on the profits of OpenAI, Anthropic XA and other companies, but paid with something far more valuable, the stock. So again, I, I don't, I don't know if that, the administration liked that idea. And then they said, okay, yeah, let's go do that. I don't know. But this is. Both sides are continuing to try and figure out their positions here. Are we for against acceleration? Are we for. Against government ownership? It, yeah, it's going to be really messy. But again, this is like a, it's this bipartisan weird thing where people just aren't really sure because they don't really fully understand the implications. But obviously they're starting to do a lot more research and paying a lot more attention to it.

B

Yeah, putting my tinfoil hat on, it feels like perhaps both parties got some polling numbers back. Totally say, which is why we're starting to hear this more.

A

Yep, definitely.

B

Interesting, man. Very interesting times to have quickly gotten to being talking so much about soft nationalization so quickly. Yes.

A

I mean I feel like, yeah, it was like six weeks ago or something. We're like, hey, it's going to start happening.

B

And this would have been inconceivable six months ago. Right, right.

A

It just becomes normal and like two weeks later it's like, oh yeah, the government's taking 50 stake and open the eyes, totally normal. It's like, wait, what? When did that become normal? And I don't even, like, I don't even know my. I mean, I haven't really formed my own like, strong beliefs of what I think should happen. I'm more of an observer at this point, just listening to all these different perspectives. There's definitely some things that come. I was like, well, that's a terrible idea. Like that. That would not go well. Republicans might love that idea today, but if the Democrats come back in power now, they're going to hate the fact that the democrats have a 50. Like, it's. I don't know, like, it always swings. Like, it might be a great idea when your political party is in power, but then you're going to hate that idea when the other party's in power. And I think there's just a lot of that going on right now where it's just good ideas today might not be good ideas in 2028 for some party.

B

Yeah. I've never prayed more for sound policy making because regardless of party, I just don't see a future where a technology this consequential doesn't get more, rather than less state involvement. So formulating sound policy would be really important here.

A

Yeah, I was trying to figure out a way to, like, weave in the Hunter Biden thing, but I'll just like. Yeah, so if you were on X at all in the last, like, 72 hours, Hunter Biden has, like, taken over X, and both political parties just seem to all of a sudden love this guy. It is like, literally the greatest communications thing I've ever seen. But the reason I'll bring it up, one, if you want to be entertained, like, go read his tweets. They're amazing. But he actually addressed AI and I thought it was great. I thought it was this really, really no thing. So in this long thing about things most Americans agree on, he's basically saying, hey, we all could just love each other and be kind to each other and listen to each other and learn from each other was the whole premise. But he had this excerpt where it said, AI is like, my new best friend that also might be trying to take my job, my ability to think for myself, and my humanity in the process. Yo, I love you. But W2F. But I still love you. And I was like, that's perfect. Like, that is basically like, AI right now. Like, we're all afraid of it, but it's pretty awesome. But, like, chill out a little bit. But, yeah, it's great, though. Let's go, like, you know, prompt some stuff.

B

Yeah, it'd be wise for AI labs to take note that that's probably far closer to a reflection of your average person's feelings than the narratives being told one way or another.

A

Agreed.

B

All right, so our third big topic this week is last week we had dug into a bit the soaring cost of AI and all these problems around usage. So we're going to talk about this in kind of part two, because this past week we had a few more stories come out building on this that show just how much companies are struggling to figure out a let's call it AI budgeting. So for one, Uber has now started capping how much its employees can spend on AI coding tools like Claude code and Cursor. Right now there's a limit internally of $1,500 per month per tool. And again, keep in mind, people are you often using this through the API, so they're not paying like, like 1500amonth for licenses per se necessarily. This is obviously directly related to what we talked about where Uber said it burned through its entire 2026 AI budget in about four months due to heavy usage. Uber CEO has now come out and said, you know, the company is being forced to adjust. Uber is now using, he says, the most expensive AI models for early experimentation, but plans to shift to cheaper or open source options as those use cases scale up, up. On a recent earnings call, they also said about 10% of Uber's code is now being written by AI. We had some commentary from Simon Willison, who's a AI expert. We've cited quite a bit on the pod. He did a little math here that just might be interesting to people. He estimates that if an engineer, let's say at Uber, is actively using two of these tools, the cap works out to about $36,000 per year, which is roughly 11% of the $330,000 a typical Uber software engine makes in total pay. He actually called this cap a rational response to overspending and noted that it actually shows how much value these tools deliver since people are clearly willing to run up bills this large. At the same time, we've got some reporting that showed the same issue is becoming a problem at Microsoft. This past week, Mustafa Suleiman, Microsoft's AI chief, told Bloomberg that, quote, anthropic is extremely expensive. I think people, many people are urgently looking for alternatives. He said Microsoft pays Anthropic a lot of money and that the company's goal is to reduce and ultimately eliminate that cost. To that end, Microsoft announced seven of its own in house AI models, including one that it claims matches anthropic's Claude Opus 4.6 on a popular coding test at a lower price. We're Also hearing about companies trying to build workarounds to solve this. For instance, the AI coding company Factory announced a new feature it calls Factory Router, which automatically picks the cheapest capable model for each task. They claim it keeps top tier performance while cutting costs by 25%. So, Paul, we talked about this at length. It sounds like people are trying to explore different options here. What did you make of all the developments on the AI usage and spending front?

A

Yeah, just continue to be a trending topic again, like a lot of where I get my information, a lot where I'm trying to kind of like analyze trends and figure out the things that are going to be really relevant for our audiences. And the stuff to talk about the show comes from Max. So. So I've shared this before, but I have highly curated lists of AI tech companies and AI researchers, entrepreneurs, journalists, things like that. And then I get alerts to probably. It's probably north of 400 accounts now. And that's largely the way I monitor what's going on. And then I track those and grab links and put them into a sandbox for Mike and I each week. And so this is one that's just like all over my feed right now of the accounts that I get alerts for. And I get that that's usually months ahead of, like, what the mainstream media might pick up on his stories, but I think X is actually a really good signal for what's going to be really important. So Business Insider had an article where they talked about Sam Altman saying that the top lead token leader at OpenAI is about 100 billion tokens a month. But in some other examples, OpenAI employees seem to spend even more than the 100 billion. Peter Steinberger's screenshot showed 603 billion tokens spent in 30 days. And the New York Times reported that one OpenAI employee spent 210 billion tokens in a week. So now, like, if you're not familiar with tokens, in essence, they're just like running agents on loops. Like all day long they just have agents going and it's like infinite use of the API or, you know, in this case, they're sitting within OpenAI. They have this infinite access to burn as many tokens as they want, and they're just like running all of these things simultaneous. The only way you could possibly do that. So, but then Sam, I thought was a very interesting and very vulnerable thing to admit. He said the cost question came up quite suddenly at the beginning of 2026. So just six months ago, the issue never came up. There's a quote from him. The issue never came up. Then he said, quote, people were totally happy with the amount they were spending. Now AR costs are, quote, a huge issue. So out of nowhere, companies are now talking to OpenAI all the time about the spend of these tokens. Sierra, which we've talked about numerous times. Brett Taylor, who is the chairman of OpenAI's board, runs Sierra. They had an article come out called Outcome Maxing. So they've been very outspoken about the need for outcome based pricing rather than token based pricing or metered pricing. And so they said that traditional enterprise software is about productivity tools for teams. Given AI can do the work, customers should pay for outcomes, not usage. They said the SaaS apocalypse that we touched on beginning of the year as it has been dubbed, is the market internalizing that the future is not productivity tools for teams, but AI agents that deliver outcomes. Outcome based pricing is more complex than seat based or consumption pricing. Operationally, contractually, accounting wise, people telling you it's simple are selling something if only works where it only works where the software is highly autonomous and highly attributable. If the work isn't autonomous, you're really just selling a tool and the tools price by usage or by seat. If the outcome can't be cleanly attributed to the software, you'll get into an endless debate. So then they linked to an article I had not read yet that I thought was really good. So I'm not going to spend a lot of time on this. But but if you're interested in this, be it based on your company is trying to figure out how to pay for AI technology, or you're a software company that's trying to figure out how to charge for AI stuff, or you're a service company that's sort of in the middle. There's this really good article from earlier this year. It's called Charging for Intelligence. How to price AI software. MCAP was the is the URL, but again we'll put the link in. So they talked about founders they're specifically writing for. Like software companies that are trying to charge for this. This are wrestling with the fundamental question about value capture in the AI era, including should I use seat usage or outcome based pricing or a hybrid? How do I price when my product replaces human labor? How should falling large language model costs factor into my pricing strategy? So today's model, let's say 4.8 or whatever, cloud 4.8 or 4.7, whatever number we're on this week, I think we're about to get 4.8 or 5 that it'll cost 10x less to deliver in 6 to 12 months. So like how do we factor that in that the models just keep getting cheaper, but then the frontier model is going to keep getting more expensive, like the best model is going to get more expensive. So they said. To help answer these questions, Jake Sapier, general partner at Emergence Capital, which is probably mcap, hosted a lively discussion at the Emergence offices with two of the best minds in pricing, Modhaven Ramajan and Josh Bloom, both general partners at 49 Elms Ventures and former managing partners at Simon Kutcher, were they advised 500 plus companies on pricing. Madhavan also wrote Monetizing Innovation, the most widely read book on pricing strategy. And I thought it was really interesting they broke pricing into two categories. Autonomy. So how independently does the AI operate? Does it augment human or replace them entirely? And then attribution. How clearly can the AI's actions be linked to to measurable outcomes? So it's not as simple as like, hey, let's just do outcome pricing. You know, where we're monitoring outcomes or usage based pricing or seat pricing or human replacement pricing. You actually have to look at what is the AI enabling? Is it when it does the thing, we can directly attribute it to a value, it's an A equals B thing. Or is it like not so clear exactly what it does. And based on the attribution level, it actually starts to dictate whether a seat based pricing, a uses based pricing or outcome is preferable. But obviously the human labor replacement cost is really easy to understand as is outcome, if you can do it. So then that leads to like, okay, what do we do about this? You can find cheaper models like you can like as a team, like say let's just use the cheapest model you can do model routing, which is the thing you're going to hear a ton about. That's what you were referring to Mike earlier, where it's just like the AI picks the best model for the prompt or the task that you're doing. The labs themselves can just make the algorithms more efficient. There can be more transparency from the labs as to how many tokens are being used, which right now we don't know. You can teach your people to be better at prompting and better at model selection. So that's what you as a leader can do. And then the thing I was thinking about as you were talking, Mike, and doing this intro is like, you almost need to do this token analysis by roles and workflows, where when I'm going to say, okay, we're going to give the sales team access to Claude. Here's what the standard salesperson would do in a day. Here's the 25 tasks they're going to use Claude for. These 20 don't need Opus like, they just need Sonnet 4.6 and like, right. That's all they use for it. And so just that would actually limit the number of tokens that you're using as a company if you did a token audit up front and like guided your team what to use or even better, restricted their use to only those things. So I just think there's a lot to go on. And then the last thing I'll say, Mike, is Brian Armstrong, who's the co founder and CEO of Coinbase, had a tweet that was getting a lot of attention and he said that, you know where this kind of all goes at Coinbase, they're working hard on routing prompts to cheaper models where appropriate and in some cases have been able to keep costs roughly flat while token usage continues to grow exponentially. So he had some other your thoughts in that tweet. But the basic premise, he said, is like, we're just trying to route it to like, cheaper models when they don't need the most advanced model to be used. So it's early. But a lot of companies are going to be racing to try and solve for this and there's gonna be a lot of pressure on these labs to figure it out. And it's going to be one of the big questions when they IPO is like, hey, aren't people just going to start using open source models to do 80% of the work because your models are really expensive?

B

You know, Paul, we're actually working pretty closely trying to solve some of these questions for ourselves at SmartRx in the next couple weeks. I'm curious, like, do you. I have no idea what the way forward is, but I did think it was interesting with Simon Willison saying like, hey, this is roughly like 10 of the engineer's salary it's being capped at. Do you. Is there any merit to like a percentage of like token spend per employee based on like salary? Or does that even on a long enough timeline it might not not matter as things get cheaper. How do you think about that?

A

So I haven't sat down and done this analysis yet, but I'm going to kind of think I'm going to do now what I would do if I was sitting by myself. The way I would probably start to think about this. Mike Is I would look at say the next 12 months of the growth of SmartRx and I would say, okay, in a, in a traditional environment, based on where we're going in the growth, I would need to hire for these five roles. So I would hire event marketing manager, I would hire this person, I would hire this person. I would then do an analysis and I would probably pull you in, pull a couple other people in. And I would say let's sit down and really think about these roles. Do we need this role as a full time employee? Like what that this person would do, is the human still going to do? And so I would look at and say, okay, what was going to be? Let's just say arbitrarily 100,000 per person. So I have a half a million budgeted for five hires. If I look at and say I think we only need two, like we're only going to hire two out of that five, that leaves us 300,000 that we would have spent on salaries that we can now say let's take 20% of that and let's put that into token budgets for the marketing team to go and do the work of what would have been five people. And we're going to basically do it with two plus AI agents that are going to take up 200,000. Now I'm just making up numbers here, making up a scenario, right? But for me that's how I would think about it. It's like what would we be spending and is it still going to put us on that path to get to like a million in revenue per employee? And so I would look more at revenue per employee. I would look more at margins we're trying to achieve. I would look at what the cost of payroll would have been and then I would back into what is my, I don't think just arbitrarily saying, oh, we're paying someone $150,000, their token budget should be 10%. What does that mean? Like they may burn because they, if they're not educated how to use that token budget, they're going to burn through it and it's going to be useless versus someone who knows how to properly prompt or knows how to pick the right model. They're going to get way more out of that 10% token budget. So the percentage of budget means nothing if the person doesn't know how to use the tokens wisely. So I think it's a combination of looking at what would we have spent on payroll and then how do we provide the illiteracy necessary for the person to efficiently use the tokens they're going to be provided. And then you need governance structures to actually ensure that they're using tokens in the right way, which might be new third party software, it might be internal systems, like, I don't know. But that's how I would start to think about it. And I would say, you know, if I had like a week to just work on this and like tap into some of my friends in the industry who like, smarter than me when it comes to a lot of this stuff, I think you could probably build like a beta structure of how to start thinking about this pretty quickly.

B

Yeah, yeah, I think because, yeah, to your point, if you can categorize certain tasks, like we all, we can at least triangulate on at least a handful of things we do all the time that we know don't need the level of tokens that we need. So you can at least crack part of the problem.

A

Yeah. And I actually think, like, I'll, again, I'm just going to give something away that I probably shouldn't say out loud, but what, you know, who cares? I'm not going to have time to do this myself. Yeah. So Jobs GPT that we built. Mike, as you were doing the intro here, I was like, oh shit. I could just totally go into Jobs GPT and I could train it to do token prediction. So like, Jobs GPT takes any job. So let's say Mike's job is a chief Content officer and I give it some context. Here's what Mike does. Jobs GPT will analyze how AI could be infused into Mike's Jobs. We could probably go into the custom GPT and train it to then also actually predict what model would be best to use for the different tasks and then what the token budget should be related to those based on how often Mike does the task. So in my mind, I think there's actually probably a way to build an app or a GPT in probably an afternoon. Right. That could at least get you in the direction of starting to think, okay, first we analyze what tasks and workflows AI can help with. Now we need to go to the next level and say, but what model is best to use and what is the estimated token cost of that based on how often Mike does the thing? And I think there's probably something there. And honestly, if I was running a consulting firm, I would be probably racing to like build some services around this because the demand for this is going to be through the roof.

B

Yeah, for sure. And just one final note, Your comment about AI literacy was also something I was thinking about because I was looking at this and as we're trying to solve for these things, it's like not only just token efficiency and AI literacy around that, but also like give two people the same amount of tokens, they're going to use them very differently. So like, how do you even figure that out? Because I could very much see the argument that there are some people that don't even know which problems they should be working on using the token. So it's like this whole mess to figure out.

A

Yeah, and again, like, the AI literacy is so fundamental because like I just pulled up Claude right now my options are Opus 4.8 high, Sonnet 4.6 or Haiku 4.5. And then I also have effort low, medium high extra max. I can toggle on thinking, can think for more complex tasks. And then I have a dropdown for more models, which is Opus 4.7, 4.6 and Opus 3. The average knowledge worker has zero idea what the difference is between those things. Maybe they understand that thinking means it's going to like work harder. But if they don't know what a reasoning model is, they don't know what chain of thought is relation to models. They have no idea how to use Claude effectively. So if 4.8 is the one that's picked like out of the box and high, they're going to burn tokens to write an email that haiku 4.5 could have done just as well. And now all of a sudden, and so that's, that is just the most very basic thing. And so just simple AI literacy would, would like, dramatically change the token burn within companies.

B

Not to mention, then we'll move on. But this becomes several orders of magnitude more complicated when you factor in something agentic like a Claude code. Like for instance, if I ask coworker quad code to go, let's say I want to do a research project, there's four big things I want to research. I don't have control out of the gate unless I specify how it does that. It will say, hey, you know, I'm going to fire up four agents in parallel. Well, guess what? That's like having four windows open on the same account that all use a hundred thousand tokens each. So it's like, I don't. I have a better intuitive sense of that now having used the tool, I don't know how to predict that. I, you would have to say, don't ever spin up agents or something. Perhaps it's right. I have no idea. The agent thing is out of control. The usage is so fast. I don't know how you're supposed to regulate. Even if you nail down the model piece of it.

A

And I don't even think Mike like that we as administrators in that account can control that for our users. Like, I can't go in and say, okay, Mike isn't allowed to use these models. Or like, Right, right.

B

I don't think you can.

A

Yeah.

B

Or like specify for it to work in like low. And I'm sure they try to optimize a bit on the back end of like the token efficiency. But again, it's like you can't stop it unless you explicitly say so. From like, don't spin up an agent swarm. Don't go do this the way you think it should be done is basically what you're saying.

A

Right?

B

Yeah. So interesting.

A

I guess I'm sure we're gonna have.

B

Yeah, we'll have more to share at some point as we try to figure this out.

A

Yep.

B

All right, before we dive into rapid fire, one more announcement this week. This week's episode is brought to you by our upcoming AI for B2B Marketers Summit. This is a virtual event happening June 25th. It is presented by our friends at Intercept. If you're a B2B marketer trying to actually put AI to work, this is built for you. It is a half day of practical real world content you can apply right away. Obviously, AI is quickly becoming standard in B2B marketing, so this event is designed to help you stay ahead of it. And during the summit you'll learn things like how to create better customer experiences with AI, accelerate your content with AI, align sales and marketing with AI, build the right team structures, roles and workflows for AI adoption, and navigate some of the harder parts of AI like ethics, transparency and data security. So this event actually, thanks to our sponsors, has a free registration option if you're interested. So head over to B2B Summit AI B2B Summit AI to register and save your spot for June 25th. All right, so diving into rapid fire this week, Paul. So first up, this past week we are getting kind of some news front running. Apple's Worldwide Developers Conference, WWDC Bloomberg published some new reporting about how Apple kind of behind the scenes finally got serious about AI and kind of what to expect once the event hits when they're expected to unveil kind of their big AI reset. So Bloomberg is just some interesting context. We've learned they reported that back in early 2025 Apple's top executives, with the notable exception of CEO Tim Cook, held a secret meeting to confront how far the company had fallen when it comes to AI. So at the time, as we've talked about, Apple intelligence has kind of been a bit lukewarm. The promised overhaul of Siri was being delayed again. At the time, several products, including a tabletop robot and AI glasses were stuck because the underlying AI was not ready. Now, according to Bloomberg, that meeting led to a push for fresh AI leadership. Cook himself became far more directly involved in the company's AI roadmap. And so at wwdc, this is where Apple is hoping for a turnaround. So at the event, Apple is making its AI turnaround a centerpiece. The showcase will be on a completely rebuilt Siri that is at least hopefully fully AI forward. This includes something we have previously reported on which is Apple is actually set to pay Google around a billion dollars a year for a custom version of its Gemini model to Power series cloud features. Along with Apple's own on device models, iOS 27 is also expected to let people pick third party models like Gemini, Claude and ChatGPT through a new system wide search or ask interface. So Paul, after years at this point of Apple kind of falling behind on AI, like what are you hoping to see at wwdc?

A

Just something more functional than we have. Like I, it's not hard to improve on Siri. It's a relatively useless technology for the most part. So I, I don't know. I. This reporting was somewhat shocking to me. Like I mean we've been talking about how far Apple was behind since you know, 2023, like when chat GPT really started exploding and it's like what, where's Apple and all this and the fact that, that the like the early 2025 is when this like secret meeting happened, like where were they for like two years? So that was wild to me to read that. And so obviously someone in the room told the, the writer everything like it was, it was pretty crazy. So I think that we're going to get a much better version of Siri. I think it's going to be good. I, I don't think it's going to be like life changing replacing your use of Chat GPT or Claude or anything like that, that, but the idea of an always on assistant that's connected to the tools and apps on your phone and able to do things and hopefully go through a chain of thought and not just, you know, give you the weather or the wrong song you asked for when you, you know, you're driving and you ask Apple Music for something. If it's if they just solve voice actually working, if they solve me like talking what the text message I want to do and actually get the words correct. Like I it would not take much for me to like really appreciate Siri more than I currently do.

B

Right.

A

So I don't expect it to be transformational yet, but I expect it to be much better. And given Apple's distribution, like that's gonna be enough for now and then I think they can kind of grow over time and improve on it. But it does seem like they finally snapped out of it and realized how bad Siri really was and are doing something about it. So I'm very hopeful now. You will have likely heard what it is by the time you listen to this, because on June 8, Monday is when they're supposed to like announce a lot of this. So we'll definitely cover more on the next episode and talk more in detail about exactly what they announced.

B

All right, next up, OpenAI held a live stream event called Intelligence at Work this past week. They talked about how they're kind of combining their Codex coding tool with ChatGPT. So Codex has started out as a tool that helps software developers write code. OpenAI now says it will put Codex inside the ChatGPT app everywhere on desktop, mobile and the browser. And that's happening over the next few weeks. Codex is evolving pretty quick. Instead of being just a coding assistant, OpenAI is kind of shopping it as now and turning it into a general work agent, meaning software that can carry out multi step tasks on its own for any kind of job. The company said Codex can now handle work like data analysis, sales prep, marketing, assets, product design, not just programming. So to make all this possible, OpenAI launched actually six role specific plugins which are add ons that connect codecs to the tools different teams already use. These cover data analytics, creative production, sales, product design, public equity, investing and investment banking. Together those plugins bundle dozens of popular business apps and connectors including to Snowflake, Figma and Salesforce. They have over a hundred built in skills at the same time. OpenAI also announced a feature called Sites. So with sites you can describe what you want in plain language. And Codex will build an interactive internal tool like a dashboard, project tracker or a workflow app and then host it at a shareable link with workplace sign in built in. This is rolling out first to business and enterprise customers. So Paul, I'm curious, what do you make of OpenAI is basically sounds like turning codecs into some type of version of like Claude cowork sites is really interesting. This, this could get very compelling very quickly.

A

Yeah, they, they were basically saying like they don't see the chat interface as the future and they're just trying to build this super app. But I, I mean, you know, you and I, and I'm sure a lot of our listeners have experienced the power of Claude code and Codex and what it like being able to build interactive tools, even if you're just doing HTML versions of things or working with a tool like lovable or replacement. But when you give non coders the ability to just spin up apps and websites with words, it's so transformational to the way work is done. The implications are just so vast. Like it's hard to really wrap your head around what this looks like. And unless you've built in it, unless you've done it yourself, where you're like, oh, I got this idea for an app or a dashboard and you talk to Claude and like five minutes later you have a functioning HTML file or an actual working app that, that does the thing that six months ago you had no ability to build yourself because you're not a coder. It, you just look at everything differently. And so I think that over the next few weeks we're going to really start to see a lot of this. I would imagine the new model we'll get from OpenAI probably this week, it's just going to start happening so fast. And going back to our first conversation, most businesses, most leaders just are not prepared for the accelerated rate of how quick the model is going to improve and how quick the capabilities are going to evolve. That changes the way you think about the structure of your team and how you hire and train people. It's just, I don't know, it's just coming so quick.

B

You know, one question that came up for me as I was reading through this, I have no idea what the answer is, is basically just if we assume that agents will be baked into ChatGPT, which is what this sounds like, like what happens to how we think about AI policies? Because we've kind of been thinking of this almost separately. Like, hey, you're probably using chat, you might not have agents enabled or turned on. But like what happens when chat GPD is just agent or like it starts

A

to see your prompt and say, hey, you know, this is a great prompt. This actually would make a great agent. I can spin up an agent for you to do this, or I can create a skill for this, or I can build you an app that does this. And so the prompt you gave it, the thing you thought you needed, it actually thinks in first principles and says, no, no, you don't need what you think you need. Let me actually build you something that does that thing.

B

Right?

A

Yeah. Like, how do you manage that? How do you manage people being empowered to do that on the fly? And maybe somebody inadvertently obsoletes their co worker. It's like, hey, I was just prompting with Claude and it sort of recommended I build an agent that does the thing you do. Sorry, like, I know, like, that's real. Like, yeah, it's having a proactive agent builder that just analyzes your prompt. Like, that's totally doable today.

B

Yeah.

A

I don't know, man.

B

Interesting times.

A

Every time we talk, every time we do this episode, every week, I'm just like, as a leader of a company, you're more and more like, oh man, I'm not even prepared for the future.

B

Like, this is crazy, right? All right, so some more Open AI news. They did something a little unusual this past week. So they put out a public statement to distance themselves from a super PAC backed by one of OpenAI's own co founders. So as a reminder, a super PAC is a political fund fundraising group that can raise and spend unlimited money to influence elections and policy in the US as long as it doesn't coordinate directly with a candidate. So the pack we're talking about is called Leading the Future. It's a pro AI super pack that we've talked about in the past. It pushes for policies friendly to the AI industry. It opposes certain state level AI regulations. It has raised well over $100 million from major tech players, including notably OpenAI President and Co founder Greg Brockman and his wife. They're some of the big backers of this group. And in this statement, Open AI said basically there have been questions around Leading the Future, which has re received support from Brockman and his wife. They added that any engagement with the organization has been in a personal capacity, not on behalf of the company. They said they do not direct the activities of the group or have visibility into their operations and said that no outside political group speaks for Open AI. So Paul, why do you think they felt the need to put this statement out? Now? This is, is a group that's been around for a little bit at least.

A

I, I mean, I assume they're just getting blowback from investors in preparation for the ipo. You know, it's going to be more scrutiny on this sort of thing. That's a tough one. I. You can Say whatever the hell you want. The guy's the president of the company or whatever his title is. Like, he's top three, top five in the company. Yeah, it's kind of hard to separate his personal.

B

Right.

A

Beliefs and investments from the company because he's making decisions that influence the direction of the company. And like, his personal money is different, I guess, but like, you don't like, forget that you're also the biggest backer of that pack.

B

Right.

A

And then like, go about your job in an unbiased way. So it's a, it's a really tricky one for any company. You know, this is just OpenAI specifically. But, you know, other leaders of other companies give money to, to political parties and super PACs, and it's hard to distance it. When you get that involved in politics, you're, it's who you are. Like, it's what you believe and you're making it very public. And so it's hard to just say, yeah, but I'm unbiased. When I do my job every day, it's like, no, you're not. It's human nature. You're not like, nobody is. So I don't know. Yeah, it's, it's a tough PR campaign to win.

B

Yeah, they have plenty of PR problems as it is, I suppose. All right, so next up, this past week Forbes reported on a very interesting example of investors using AI to roll up a single traditional industry, in this case, accounting. So a roll up is when a company buys up a bunch of small independent businesses in the same field and combines them into a larger operation. So the company behind all this example that we're about to discuss is Thrive Holdings. This is an AI focused holding company spun out of a famous venture firm called Thrive Capital, run by Joshua Kushner. Thrive says it plans to commit $1 billion to acquire local brick and mortar accounting practices across the US and they're basically going to apply AI to these businesses all as one business and make it much more productive and performance driven using that technology. They've actually had about two years of acquisitions through their accounting platform that they have, which is now called Current. The strategy is basically buy these practices and equip them with AI. Current. That AI platform, or accounting platform, says it's open AI powered tax, AI tool, process 7,000 tax returns this past season and cut tax prep time by about a third with up to 98% accuracy. So, Paul, I'm curious. This is not only interesting for accounting, but seems like an obvious playbook for certain industries, perhaps most industries.

A

Yeah. This is the thing we've been, you know, talking a lot about for years, is if you're private equity, this is the play. You just pick an industry and you go in and you roll up firms and you unlock margins and just apply AI. Like this is, again, it's like an inevitable thing. This is what's going to happen across industries. Now the accounting industry is ripe for it because they actually have a talent shortage. So the article says as the number of CPA candidates declined 30% between 2016 and 2021, the need for AI became greater. Like it's, it's a demanding industry. They work long hours, especially during tax season. There's a lot of monotonous work, like data driven work, where you're just doing these repetitive processes. And so if you can infuse AI that can all of a sudden start taking some of that burden off, one, it helps with this gap in talent, but two, it actually accelerates the automation of the industry completely or to a large degree. So I think it's a good article for people to read if you want to. Again, you may not be in the accounting industry, but look at this as an example of what's going to happen. Because the tax AI that they're talking about was just specifically trained, you know, for the tax returns. And I think it was like these. Well, I forget the exact thing that they did it on, but it was a very like, niche though, the 1040 and 1041 tax returns. So like a very tedious area, but they proved it could work there. And now it's like, okay, now we're going to expand it out and we're going to go to like all functions in accounting. So this is what will happen. Like you're just going to have PE firms or venture capital firms that just, or, or, I don't know, venture studios who just, like, they just build AI native versions of something or they just go roll up existing ones and then they infuse AI into them and they just make the industry more efficient, more productive, and hopefully more innovative. But yeah, it's kind of like the AI researcher thing. Like, you got to watch for these early instances of it and then try and extrapolate. Okay, what's that going to mean when they take this out further and start going into these other areas?

B

All right, next up, some AI and higher education updates. So a few different stories we're kind of tracking here. So there's a piece in New York Times Magazine that took a look at how AI is reshaping American universities. So they actually did a deep dive into California as California State University because this is a university system, which we had reported on a while ago, that spent nearly $17 million on a deal with Open AI to provide half a million licenses of CHAT GPT to students, faculty and administrators. At the time, this was described as the largest single institution deployment of CHAT GPT in the world. But this article now reports that chaos and conflict have followed. So the faculty union organized against the deal. Professors delivered a petition urging CSU to cancel the contract. They argue that ChatGPT was not designed, trained or optimized for education. The a lot of frustrations around the education system's priorities. The system leaned into AI while facing a serious budget crisis and cuts elsewhere. A survey of more than 94,000 students and employees found that 52% of faculty said AI was having a negative effect on their teaching. Even so, this system recently renewed its OpenAI contract. Meanwhile, at the University of Chicago, President Paul Alivisatos told students and staff the university is partnering with Anthropic to provide Claude access. Faculty and staff get access starting in July and all students will have access before the fall term. Finally, some new research from Stanford Law School had 16 law professors evaluate nearly 3,000 anonymized head to head comparisons of answers to contract law questions. Some were written by AI, some were written by other professors. In that blind evaluation, the Professors preferred the AI answers 75% of the time. They also flagged AI responses as what they would call pedagogically harmful only 3.5% of the time, compared to 12% for the answers written by fellow professors. So Paul, a couple of not exactly related stories, but kind of showing that higher ed is facing some challenges and opportunities just like anyone else trying to integrate AI.

A

It sounds like, sounds like, yeah, that the Stanford one's fascinating. When I first saw those data, I was like, geez, oh man. Like that's right.

B

Right.

A

That's hard to figure out the implications of that. But I think more and more that's going to be the case. We're just like AI answers if they're. If you can't directly tell it's AI because I think there's a bias by humans to say, oh, I don't like the AI version of this. It's like, but if you don't know, you truly don't like. I think more and more this AI answers are going to be preferred in a lot of disciplines, not just law. Yeah, yeah. I would imagine the universities are struggling with the adoption just like any enterprise would be if you just go buy the tools and infuse them in, like, that's not going to solve everything. You gotta think about things. In cohorts. There's going to be people who hate AI. There's going to be people who, you know, are early adopters and love it and can't live without it. And there's going to be people in the middle who just haven't had time to learn how to do it. And you're going to get pushback from the crowd that doesn't like it or if you don't deploy it with a change management strategy, then it's just going to fail. So, you know, it's just education is a microcosm of what's going on in enterprises too. It's not that different. And I think we're just starting to finally get these case studies in a way where we're now one, two years into universities that have actually gone all in and tried to do this. And you're going to start getting these hard lessons learned by the people who are the early adopters, just like we are with enterprises.

B

Yeah. I imagine there's also just a magnified version of the sentiment and PR problem within universities too is it's kind of ground zero for all the disruption happening. Happening.

A

Well, yeah, and you know, it's a hotbed for obviously activism. Like, you know, you're going to have students who are going to get involved and have feelings about the environmental impact, economic impact, like it, they're, they're at the center of this and so you have to deal with, also with that within an environment. Yeah, it's really complex. Like to do this within a university is very hard. And I think we're just, like I said, we're just learning the lessons now and it's just going to take a while to figure out how to do this the right way and to address the, the real concerns people have about the technology. Yeah.

B

Okay, next up, we have our AI Use Case spotlight. So every week we give you a quick look under the hood at the real AI use cases we are exploring in our own work at SmartRx. So I'm going to share a quick one, Paul, and if you've got anything to share, let me know. For mine this week I want to, kind of similar to last week, also point to like a way more of a way I'm starting to work with AI rather than one specific use case. And this kind of was inspired by a post from Al K. Miller, who you should definitely follow. She's really sharp and really good with actually putting AI to work. So we'll include a link to the post I'm about to talk about. But basically she posted something last week that I have started to experiment building my own workaround. And she basically said, look, the way most of us are taught to find AI opportunities in your work is already behind. The old way is you look at your tasks and you speed up the slow ones using AI, which I don't disagree with. I think it's a huge part of what we teach and a necessary very first step. But she also argued that you can aim way higher, way faster, because speeding up your work assumes the work itself should stay the same and just be faster. But in many cases it shouldn't. So she says instead of starting just at looking to do things faster, start with three things. Your goals, context and an interview. So basically first tell the AI what you're actually trying to achieve, not what you produce or how you do it, it. Then give it context. We know this is super important from a prompting perspective. You connect it to whatever context you are allowed to give it, your docs, email, data, etc and then finally just have it interview you one question at a time about the goal of the work itself. Like who hands what to who, what you are doing on a repeating basis, who receives it, what's supposed to happen to it. And basically she just advocates working backward from first principles, which is for everything you're doing. Does this thing even need to exist? What outcome was it meant to create? And she actually provides a prompt in the post itself that I've used to great effect to kind of almost re architect some of the things I'm doing. It'll basically challenge you like does this have to exist in its current format? Here's a smarter, better way to do it. So I found this really helpful, Paul. Just challenging myself not just to get more productive, but to innovate more in. What I'm trying to do is still very early. I'm not saying this is the only approach people should take. All the other, you know, standard advice and workflow mapping and things that we teach I think is deeply important and a really good way to get started, but this was just really good food for thought and I'd highly recommend people take a look.

A

Yeah, I so one of the kind of tricks I've used for a while on strategic work, which is a lot of What I use AI for, it's like high level strategy or building like MVPs of things just to like get apps out of my head. And you Know, build earnings, interactive charts, things like that is I will just tell it, hey, here's what I want to do. Let's do this one step or one question at a time. Like, what do you need to know from me? But like, let's go through one at a time. And it is, it's very therapeutic. It's very helpful as like, you know, someone who's trying to think about something to just have it do. Because AI wants to just solve stuff for you, wants to just like, give you the plan or give you. Here's the 20 questions. Like, that doesn't do anything, anything. Like, let's just go through this a step at a time. And then you may deviate and say, hey, listen, actually, let's go in this other direction now based on what we just talked about. So for me, that is like the vast majority of the way I use AI is I just have these, like, conversations about it and I try and like, keep it focused on going through these steps with me. So, yeah, I love the approach she's talking about there. I find it to be extremely helpful.

B

Yeah.

A

And then the context historically was whatever you could put into the brief. It's like, here's what you need to know. Here's a bunch of bullet points. Now the context is increasingly becoming. What do you connect it to? Here's my CRM database. Here's my Google Drive. Here's my inbox. And that becomes the ongoing context that lets you really work with these models and agents in new ways.

B

Yeah, yeah, it's really, really valuable approach. I would recommend everyone go take a look at that. All right, Paul, as we wrap up here, I'm going to run through very quickly some AI product and funding updates. And as we close out the episode. Episode. So first up, SpaceX, as we've talked about, is expected to go public around June. June 12th under the ticker SPCX. This is going to be perhaps the largest IPO in history. There's underwriters targeting a valuation of roughly $1.75 trillion. There may be a raise of up to $80 billion. Next up, as we've also mentioned, Anthropic and OpenAI are expected to drop new models very, very soon. Anthropic is reportedly set to widen access to its mythos model. OpenAI is rumored to ready be readying its next GPT update, though neither has officially confirmed timing. OpenAI has also published a policy blueprint for governing frontier AI in the US proposing a, what they call, quote, reverse federalism approach, where states pioneer AI safety rules and the federal government absorbs the consensus. Anthropic has expanded Project Glasswing, its program that gives organizations access to its Claude Mythos model to focus security flaws in critical software. They've expanded to roughly 150 new organizations across more than 15 countries. Google has rolled out Gemini Spark in beta, which we mentioned on a previous episode. It was announced it's a 24.7personal AI agent that runs in the cloud even while your devices are off and can carry out multi step tasks across Gmail, calendar, drive docs and sheets. Google also released Gemma 4 12B. This is a free open weight multimodal model that processes text, images and audio without separate encoders. And it can run locally on a laptop with just 16 gigabytes of memory. Super important to start tracking that kind of thing with all the usage and cost questions we've been talking about. We also of course talked about how Microsoft launched its seven new in house AI models. Just a little more on this. This includes one called Mai Thinking 1, which is a reasoning model that the company says scores 97% on the AIME 2025 math benchmark. They say it is preferred over Cloud Sonnet 4.6 in blind human evaluations. Microsoft and Mayo Clinic also together announced a collaboration to build a frontier AI model for healthcare that will be owned actually by Mayo Clinic and will be trained on its de identified clinical data to support earlier diagnoses and treatment decisions and later distributed worldwide through Microsoft's Azure Foundry APIs. A couple other updates here. Hackers actually gained access to some high profile Instagram accounts, including the Barack Obama White House account and the brand Sephora. They did it by simply asking Meta's AI support Chatbot to change the email address link to those accounts, which it did without proper verification. And finally Meta introduced the Meta Business Agent, which is an AI tool that is free to get started. There's paid tiers coming. It basically lets businesses automate customer conversations such as answering questions, recommending products and booking appointments across WhatsApp, messenger and now Instagram.

A

I want to keep an eye on there, Mike. The Microsoft one. That's their effort to become one of the major frontier labs. So that's. We'll talk more about Microsoft and their efforts, you know, in the in the coming weeks. But they came out and basically said they want to be one of like the four or five major labs and this is their effort to break free of reliance on OpenAI's models. In essence.

B

Gotcha. Yep. One final announcement here. We Talked about our AI Pulse survey at the top of the episode SmartRx AI forward slash Pulse this week's survey we will have live when you listen to this, we're going to ask a little bit about how much of your day to day work AI is already doing and get your take on the US Government's plans to possibly take stakes in AI labs. So Paul, thanks for breaking everything down. Feel like we're a bit in the calm before the storm here before all these model drops, but still plenty going on this week.

A

Yeah, it should be a busy week. Thanks everyone for joining us. As always, thanks Mike for curating everything and we will talk with you again next 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.