Speaker 0:00

Hello, and welcome to a weekend news episode of the Leveraging AI podcast, a podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business, and advance your career. This is Isar Matis, your host, and we have some really interesting stories to cover today. Many, many, many things happened in this past week that each and every one of them on their own is interesting, but once you connect the dots, they become really interesting as they're showing, A, where the AI world is right now, and B, more importantly, where it is probably going. And it is all combined, or not all, but many of the stories combine around three main topics that we're going to deep dive into. One is the government, and I would say governments in plural, involvement and their impact on AI. and And what you'll see is that the US has done a complete reversal. The US government have done a complete U-turn from its position that it held not too long ago. the second U-turn that we're going to talk about is the approach of companies to AI spending, going from token maxing to limiting and switching to different models. And the third one would be the rise and the stronger position of customized specialized AI solutions compared to the next wave of high-level advanced top-of-the-line models. And so all of these things tell a very interesting story, and we're gonna dive into each and every one of them. And then we still have a few interesting releases to talk about and funding rounds, et cetera. So lots to cover, so let's get started. So before I dive into the first topic, as I mentioned, is about the government's involvement in AI, and again, governments, not just the US government, even though that's gonna be the focus of our conversation, but we're gonna drift into Europe and the UK as well. But if we go a little bit back, June 2nd, the current administration issued a executive order that defined a voluntary frontier model testing program. It means that the government gets a 30-day review window, which was, by the way, negotiated down from 90 days, and it was negotiated down by the former AI tsar, David Sacks. So it was set up as 30 days, and contains an explicit anti-mandatory licensing language, basically saying it is not mandatory, it is going to be voluntary. But what is actually happening in real life is very different from that. Now, what was released this week that is going to help us connect a lot of the dots together is a document titled What Should Be Done by a person named Dean Ball. Now, we've talked about Dean Ball in the past. He was a part of the White House team and one of the co-authors of the AI Action Plan. He is now on his way to work at OpenAI, uh, but he wrote this document again labeled What Should Be Done, where he talks a lot about the current situation and why is it problematic and what he anticipates that's going to happen. I highly recommend reading the whole thing. It's not very long, and it has a lot of really good components in it. But the main thing that he's saying or that he argues, and I'm quoting, I argue what the executive order on cyber and AI, which claimed to establish a voluntary testing program for frontier AI models, was really establishing a de facto involuntary licensing/pre-approval regime for frontier model. This analysis has proven correct." So he says that he warned against it when it came out, and that it's now proving to be the case. So The first proof of that we received on June 12th of 2026 when Washington added Anthropic Mythos-5 and Fable-5, which were released just three days before that, to the expert restriction technologies list, which the immediate effect was that Anthropic cannot export or cannot provide access to its models to foreign nationals, and because they cannot tell the difference, they just took it all offline. This is after Anthropic has been potentially the fastest growing company in history, and grew from ten billion to forty-seven billion dollars in ARR in less than six months. Why does the money matter? We're going to touch upon in just a few more minutes, but keep that in mind as we keep progress. So the government banned the usage of models that were released by a private company because of a, quote-unquote, export control risk," but that voluntary review now became a mandatory ban on the usage of that model. Now, the reason behind it was that found a that allowed them to basically jailbreak Fable-5's safeguards and be able to use it to identify software vulnerabilities in other software, which creates a serious cybersecurity risk. Now, to be fair, subsequent testing by similar people has found that competing models, including GPT 5.5, Kimi 2.7 Code, and others and can do something very similar, and yet they were not blocked Now, in this past week, two things related to this topic has happened. One is that OpenAI launched GPT 5.6. They actually launched three different models in GPT 5.6 called Sol, Terra, and Luna, going from the top of the line to a middle model to a smaller, faster, cheaper model, very similar to what Anthropic is doing with Haiku, Sonnet, and Opus. So in the OpenAI world, they're called Sol, Terra, and Luna. And they agreed for the first time to allow the US government to vet access customer by customer. So basically, it is not going to be released to the public. You and I are not going to get access to it, at least not in the short term, and it is going to go through an approval bodies of the Office of National Cyber Director, Office of Science and Technology Policy, and the Commerce Secretary, Howard Lutnick. Now, it was also already given to a short list of companies to review and use and verify what kind of cybersecurity it represents. So we're seeing this again. Again, in this case, it's not a complete block of the model, but it is a similar approach to what Anthropic has done on their own when they released Fable, or actually before Fable. It was actually Mythos-5 preview that was released to a short list of companies to review. But this time, it is mandated by the government. Now, what Sam Altman had to say about this, he said, and I'm quoting, We've made it clear to the US government that this is not our preferred long-term model, and we'll work with them and others in industry to achieve a more sustainable approach of future releases." But while they're not happy about it, they're obviously aligning with the US government, despite the fact that there is a, quote unquote, voluntary review" by the government. And as we can see, it is not voluntary at all. Commerce Secretary Lutnick announced the export restriction lifted from Anthropic after Anthropic agreed to proactively detect and address security risks Now, the other thing that happened this week is that on July 1st, Commerce Secretary Lutnick announced the export restriction lifted from the Anthropic model after what he's saying that Anthropic agreed to proactively detecting and addressing security risks, work with the government on protocols for Mythos, Fable, and future models, and inform the authorities of malicious activities Now, interesting things about the language that Howard Lutnick used. First of all, it uses the words and releases," and it does not name mythos and fable in the actual thing. Basically, he's saying this is the future. This is what is going to happen from now on. Like, this is not a one-time thing about this model. This is what's going to happen moving forward. In parallel, Anthropic have done a diligent work to block what they're saying is 99% of the jailbreaking options into the current model. Now, I must say something about the whole concept here, from Dean Ball. And I'm going to add information from Dean Ball in the process. So Dean Ball, as we mentioned, is an expert that helped write the plan, and he's the one that wrote the document we shared in the beginning. And one of the things that he's saying is that there are several big gaps in the government in order to even do what the government is suggesting they would do. The first thing is that they don't have any real standards written down or defined anywhere in order to evaluate these models. So the government for itself defined a 60-day deadline to finalizing the review framework rules, and that framework has not been released, and the 90 days has passed. Now, in addition, on a June ninth meeting discussed open source exemptions, and most of the open source models are Chinese, including DeepSeek and other which have really advanced capabilities which are not included in the US requirements, which is going to give a huge benefit to Chinese models over US models, which is obviously not something that I think the US government actually wants to do. The other thing that Ball is saying is that the government has no frontier AI experience among the officials overseeing the policy. In other words, all the top research and top experts of AI in the world are working at the leading AI labs, and they're not working for the government. So the group that is supposed to verify the risk that is associated with these models are very far in expertise compared to the people who are actually developing the models, which obviously creates a very big problem when coming to evaluate which models are riskier and in what capabilities. So combine the two together, there are no clear established standards, as well as there aren't enough expertise in the government, even if the standards existed. Now, the other thing that Ball mentions is that, and I'm going to touch on what he's suggesting in a minute, is that whatever standards are defined are going to be irrelevant is moving so fast. So if you try to define specific standards right now on what is acceptable and not acceptable from a model-level perspective, this will become irrelevant in six to 18 months because the algorithms and the platforms and the harnesses will become so much better that all the things that you've defined will become obsolete, and you will have to redefine them, which will take you longer than the next model can improve. So that's the third problem. The bottom line is the government does not have the tools to evaluate the models that it is trying to evaluate and block and stop Now, to put things in perspective, this is not happening in a void, meaning there is the rest of the world that is still running forward while the US government is trying to slow down, stop, or reduce the exposure to US-generated models. So a few interesting things to mention when it comes to the race in China, which is a big deal right now, and it is a huge deal when it comes to the US government that really cares about this. If you listen to anything we talked about in the past six months on this administration and their approach to AI, they always said that they're not about regulation, that they want to lead the race in China, and any kind of regulation is gonna slow this down. This was stated in multiple occasions, including by the top leaders and the vice president in front of the UN, saying that we are not here to define more regulation. We are here to provide acceleration of the AI, and we are now in a very, very different territory just a few months later. So let's talk about a few specific examples. Chinese cybersecurity company called 360 unveiled two AI tools during the ban of Mythos and Fable. One is called Tulongfeng And the other one is called Yi Tian Zhen, which I'm assuming I'm butchering the name The founder of the company, Zhu Hongyi has framed finding vulnerabilities in software as a national strategic asset He also warned against a one-way transparency where specific actors have advanced capabilities that other actors lack. Now, what this model is, it is supposed to have the same capabilities as Mythos-5, only as a Chinese model. Another company called Sakana AI from Japan by some top people from Google, including David Ha and Lyon Jones, plus Ren Ito from, another leading lab, has launched what they call Fugu. And Fugu is marked as an orchestration model that is coordinating access across multiple AI providers. And the reason they did this, and I'm quoting their CEO, David Ha, saying, "I believe that orchestration models are the next frontier beyond bigger models. Relying on a single company model for a national infrastructure is a massive risk. As a recent export control have shown, access to top models can disappear overnight. Collective intelligence is the practical hedge against this concentration of power." What does that mean? It means the rest of the world is, A, developing models that is competing directly with the cybersecurity capabilities of the US models. And even if they're a little behind, and we're gonna talk about what Ball thinks about this, even if they're a little behind, they will close the gap in six to nine months. And the other is that the world understand that the US or any other actors may block their access and will rely on other actors, potentially in combination with one another, in order to fill the void. So what the US gets is instead of preventing these capabilities from its adversaries, it is just allowing their adversaries to use other models by-- giving China potential access to how they're using these models while the US stays out of the loop. Now we're going to talk in the next segment where we talk about the reversal on expenses on AI, how leading US companies are shifting from the leading labs in the US to Chinese models already just because of financial reasons. Think about what that means if the US is also going to block access to the models to either US companies like it's doing right now with the latest model from OpenAI, or if they're blocking it to anybody else in the world, what are they going to do in order to stay in the AI race and stay competitive in their markets? They're going to go to Chinese models or other models, leaving the US as isolated in the AI race, and there are huge other impacts beyond the technological race itself. So let's talk about the financial implications of that. As I mentioned, Ball relates to that in his paper as well. So both Anthropic and OpenAI filed confidentially for IPOs in June of 2026. They're both aiming at going public sometime late this year or latest beginning of next year. So we're talking about a six months window, about Q4 of this year to Q1 of next year. They're both planning to be some of the largest IPOs ever in history, and the valuations are running between $1 trillion to $2 trillion in valuation, led by the biggest financial organizations in the world, which include, you know, Goldman Sachs, Morgan Stanley, JPMorgan Chase, and all the big names that you know. So lots and lots and lots of money is on the table. This money is absolutely necessary, whatever money they're going to raise in these IPOs, in order to keep these companies floating. So OpenAI are currently burning through a crazy amount of cash. Their latest projections are projecting that in 2028 alone, they will burn through $85 billion. So without a serious cash inflow, they just cannot survive Now, what Ball is saying is that he estimates that the current regulatory approach will break the economical aspect of the US-based AI companies in three to six months. So what he said, and I'm quoting, is, No one is building $100 billion data centers to serve frontier models to whatever 100 companies the US government will allow access. And he's absolutely right. He's also saying that these companies make most of their money in the first few months of the release of a new model before either the competition catches or the next model is released. And if those first few months are blocked by the US government, well, you're not recouping the money that was invested in training this model, and definitely not the money invested in the CapEx of building the data centers. So the math just stops working. These labs cannot, period, stay afloat from a financial perspective if they cannot release the top leading of the line models that they are developing. Now, in addition to the fact that this puts at risk their livelihood, it puts at risk the IPO. So you're now looking at the largest investors and investment groups in the world that need to think where they're going to invest their money, and they know that the future revenue might be dependent on a decision of the US government that does not have the tools, the means, the people, the knowledge, the standards to decide logically on how to make those decisions. This may jeopardize or seriously hinder the results of the IPOs of both these companies. Now, the cascading risk of this is not just OpenAI and Anthropic and potentially X, but danger. So if you think about how much money there is in the pipeline right now across the board, from companies who build data centers to company who build memory chips, to company who print the actual boards, to company who build infrastructure, to companies who do the cooling systems, like everything in the ecosystem and in the supply chain of the AI world, combined with the amount of money that it invested right now in most companies around the world in training and education and infrastructure and licensing and so on, all of that goes down the drain or partially down the drain, which will be probably one of the largest economical collapses in history, which is not something I think the current administration is willing to pay as a price Now, I don't wanna dive into politics at all, but if there's something that was very obvious that we learned with the recent conflict with Iran, is that this administration cares very, very much about the financial markets and the financial future of the US, in many cases more than anything else. So this risk of taking down a huge economical engine, maybe the biggest economical engine in the next few years, is not something I think this administration and potentially any future administration is willing to bet against Now, another thing that happened this week as part of this whole crazy situation with the government is that Sam Altman has reportedly proposed donating 5% of OpenAI's equity to a wealth fund of the government that is currently worth, based on their current valuation, about $42 billion, and it's going to be worth a lot more after the IPO, assuming it goes forward the way it needs to go forward. Now, what he's suggesting is that other leading labs will do the same thing. So both Anthropic and Google and Meta are all going to provide some kind of wealth funds to the government to help the government, A, fund what it needs to fund in order to run AI forward effectively, and B, in order to share the revenue and the results of AI with the American people Now, we talked on this podcast about the fact that Trump has already confirmed the concepts of taking some kind of equity pieces from these companies as part of his approach. Trump has done this on a very large scale with Intel earlier, so it's not something that is unheard of with this particular administration. And so some kind of level of nationalization of these companies makes sense to the government, and now apparently also makes sense to the companies themselves. If you heard Bernie Sanders talk about this, he is talking about introducing what he calls the American AI Sovereign Wealth Fund Act, and he wants to take a 50% tax on the stock of the AI companies to create an estimate fund of $7 trillion, which is very, very far from what OpenAI is offering right now, and he's suggesting distributing 5% of that as an annual dividend for every American, which would be $1,000 per American per year. So is that something that I think can happen? I haven't seen the math. I don't know if the math can even work. I will be extremely surprised if there's gonna be bipartisan support for this. But I have a feeling that one of the reasons OpenAI is offering what they're offering is to appease the current administration to go with something like that, where 5% of the revenue is invested back into the US government economy, American people, et cetera, versus 50% if Bernie Sanders gets his way or something in between. Let's say 25% will still be completely crazy, and I think that's the direction that Sam is pushing, not necessarily because he wants to give up 5%, but because he sees a much bigger risk if they don't go down that path But as I mentioned in the beginning, the US is not alone. There are multiple aspects in Europe that are popping this week and previously showing that there are serious roadblocks to the AI success in other countries, again, in this particular case, in the European Union. So this week, Italy opens an investigation against Microsoft. So the Italy's Competition Authority, AGCM, launched a formal Ireland operations and Microsoft Italy over Copilot and Designer being bundled into Microsoft 365, with subscribers automatically shifted to a higher price plan Now This is not the first time something like this is happening. The UK is already investigating Microsoft separately. So there-- So the UK Competition and Markets Authority launched a strategic market status investigation into Microsoft's entire business software ecosystem in May of 2026, so just a month and a half ago And they're looking into bundling and licensing and interoperability and default settings as AI embedded into the different workspaces, so very similar thing. The governments are looking into how the AI is impacting the software that we already use. Another example from this week is that Tim Cook from Apple had several different face-to-face conversations with the leadership of the EU in order to allow it to run the new AI-powered Siri in Europe. So Apple announced at WWDC '26 of the new Siri AI, which is going to be the power behind Apple Intelligence, but they also announced it is not going to be launched in the EU alongside iOS 27 and iPadOS 27 due to the Digital Markets Act. Now, Apple is suggesting that their current system should be good enough from a consumer data and privacy perspective, and the EU thinks otherwise. They think that, and I'm quoting, "Apple failed to develop interoperability solutions meeting EU privacy and security standards." And what they're saying is that what Tim for is an exemption versus a fix to the problem Now, Thomas Regnier, who is European Commission spokesperson, said, The decision not to roll out Siri AI in the EU is Apple's and Apple only because absolutely nothing in DMA prohibits Apple from introducing new products in the EU." That being said, the current regulation in the EU puts Apple at risk if they are releasing it with the current infrastructure that they've developed, and hence they're not gonna release it, and hence 450 million EU consumers will not be able to use the new Siri. That is, again, a very significant ban by a government preventing, in this case, a company from delivering and the consumers from receiving one of the probably most commonly used feature of the future of an AI assistant on your phone So we now understand that there's a very big mess, there's nothing clear that is currently happening, and that there's a vague future of where this is going from a government position and how it's gonna impact the development and the deployment and the availability of future AI models, and what does that mean to the economy? So we have a lot of questions and not a lot of answers, a fair proposal, hence he called his paper What Should Be Done? Now in his paper, he gives multiple things that can be done, and I will provide the link in the show notes so you can go and read the whole thing, but I'll give you a quick executive summary. First of all, he's saying that what should be regulated is not the models, but the labs. Basically, he's saying that the model-based threshold becomes obsolete in one to two years as the algorithms improve. Instead, he thinks we need to require labs to publish their own safety frameworks and then establish an independent, certified, privately run auditors that verify compliance with these standards, similar to how accountants are licensed by state right now He calls these independent verification organizations or IVOs Now, the closest thing to that in real life is actually the Great American AI Act, which is pushed forward by Obernolte and Trahan, which is a bipartisan push drafted on June 4th of 2026 that is suggesting a very similar architecture, a federal framework with IVOs, again, independent verification organizations that is conducting mandatory semiannual audits, establishing licensing structure for auditors, not for the model releases, and promoting US leadership without rigid government control. So Ball has identified this as the closest thing to the push in the right direction that he thinks is a workable solution What he's also saying is that the approach of the government is completely theoretical because they're not developing anything versus actually understanding what these models do in real life because you're using them and developing it all the time. And he said, and I'm quoting, The only way you are going to figure out what good looks like in the context of technical AI safety is real world experience. You cannot purely think your way to safety, just as no one could have invented a cybersecurity ecosystem at the dawn of software." So where does this leave us? A quick recap of all of this before we change topic. One is the voluntary framing of the government is very far from voluntary. They find the ways and their levers in order to make the AI labs do whatever they want. Two, we've seen already two models being blocked or partially blocked by the government, the latest Anthropic models, Fable and Mythos, as well as GPT 5.6. The technical gap between the US and China is six to nine months right now. Some will say a lot less, which means even if the US blocks the US models from being released, the Chinese models will catch up if they're not already catching up, and companies and countries will use the Chinese models instead of using the US models. So the fact the US government is blocking it is not really doing anything unless there's going to be a large international push in order to govern this in a global way versus just in the US itself, and I've said that multiple times on this podcast. The idea of some kind of a wealth fund that is going to be funded by the revenue and equity of the labs to push safety and revenue into the US economy and the US citizens is not necessarily a bad idea. I don't remember any time in history where something like this was very successful for a very long amount of time, other than, again, recent China and Singapore that has governments that are involved in the actual financials of many of the organizations in their countries in a successful way. But both these countries are not really democratic countries, and that sounds counterintuitive. You would think that democratic countries would be better at things like that, but the reality it's not, because you have a new government every time that may change the rules and the approach, which does not provide the stability over five, 10, 15 years to how this is going to work and evolve. And so I don't see how that is going to work. I do see this as an interesting idea moving forward And on the regulation of this inside the US, and as I hope will happen way beyond in a global scale, needs to involve not the governments, or at least not the governments alone, but independent valuation groups that has the right skills, expertise, and access to both knowledge and compute, et cetera, to allow them to actually evaluate the processes correctly and be able to provide a safe and yet very effective future with AI to all of us. So I really hope that is going to happen. I really hope that the US can establish something like this, and then over time bring more and more countries into this. With the current level of collaboration or lack of between the US and China, which are leading the race right now, I don't see this happening, at least not in the immediate future Now, the second story that I wanna talk about today, which is another U-turn complete reversal from what we heard just recently, on AI by companies who are using AI and the reliability reckoning, if you want, of AI as well. So the first article that I'm going to talk about this week is that Ford just announced that they are rehiring 350 experienced engineers that they let go previously because AI can do their work. So full context is this In the past few years, Ford has shed roughly 5,300 salaried positions. Now again, a lot of it had to do with over-hiring after COVID, and we know all that story, but that's still a lot of people. Out of them, a lot of experienced engineers, And that started hurting the quality and reliability of Ford cars. So Ford has rehired and promoted three hundred and fifty experienced engineers to mentor junior staff to rebuild the data pipeline and refine the automated systems that just didn't work well enough. They created a forty-person software QA team And they added over 100,000 AI-powered automated tests to improve the quality of the four products Now this rehiring has led to Ford capturing the number one quality study place among the mainstream brands. It is the first time they captured that place since 2010 They were able to reduce the amount of vehicles per 100 vehicles to 152, which is 41 fewer than they had previously And if you want the summary of this topic, Charles Poon, the VP of Vehicle Hardware Engineering at Ford said, and I'm quoting, Mistakenly believed it could swap in AI and still produce a high-quality product." So at least as of right now top engineering is required by humans in order to keep the quality of production at a top leading company in the world, in this particular case, Ford The second thing as far as reliability of the models and where they're going comes from a completely different aspect this week, which is a new report releases by Meter, which is a company we talked about many times before. It is a company that evaluates the effectiveness of AI models over long-term tasks and trying to check how good they are in performing tasks compared to how many hours it will take humans to do the same tasks. So they just started testing GPT 5.6 Sol. And what they found is that it has the highest cheating rate ever recorded among all publicly tested AI models, and these behaviors include exploiting bugs in the test environment, extracting hidden solutions, attempting to cover its tracks while it is doing the evaluations and it's trying to cheat Now, some of these cheating was found actually by OpenAI themselves, and they them to Miter, which Miter praised obviously the transparency while flagging that is still a very big situation Now, to put things in perspective on how good these systems are in cheating, assuming they can actually catch all the cheating. So as I mentioned, what Meter does, and we talked about this many times before, is they try to check the time horizon. How long of a human task can the AI do at a 50% success rate? And ignore the fact it's a 50%, they're just comparing apples and apples. So the time horizon estimated to GPT 5.6, if you exclude the cheating, is 11.3 hours, versus if you allow the cheating and you don't catch it, it's 270 hours. That is a 24x swing between what you think the model is doing when you allow it to cheat versus what it can actually do when it does the work properly But it is not the only model who has done similar things. As early as Claude Opus 4.6 in similar tests by Meter has shown to try to hack the tests in different ways But the bottom line of what Miter is saying, and I'm quoting, If future models display much fewer undesirable propensities, we could become more concerned about catastrophic misalignment, as we'd be worried that models may have learned to evade detection. Basically, what they're saying is that they don't think the models will try to cheat less. They just think they will become significantly better at cheating to a point we cannot detect what they're doing and how they're doing it. And at that point, we have zero knowledge on what the models are actually doing, what their intentions are, and we could get a complete misalignment with our human goals compared to what the model goals are, and we'll have no way to detect and understand it in advance or in real time as it's happening. Now, the third component that has to do with this story on where we are on AI usage and how effective it is from an economical perspective comes from the complete reversal on how companies spend on AI, and many of these in the past week or few weeks. So the first one is Lindy, which is a large software company. Their CEO, Crivello, switched one hundred percent of traffic from Anthropic Claude to DeepSeek. The cost dropped from eight dollars to fourteen cents per million tokens. That is saving millions of dollars within months for this twenty-five person startup. Now, what he's claiming is that they've seen no performance degradation at all, and he's also claiming that the migration time was very, very quick, and it basically didn't matter how long it took to do this, but to save more than eighty percent on the cost to achieving the same results with AI We already talked about Uber that earlier this year announced that in four months they've consumed the entire budget they have set up for AI coding for the entire 2026, and now they have a cap of 1,500 hours per month per engineer Now we hear similar things from multiple direction. As an example, Glean's CEO, Arvind Jain, said that roughly ninety-five percent of enterprise AI usage still runs on expensive frontier models despite cheaper alternatives that are existing. The AI Squared CEO, said that using state-of-the-art models for simple tasks is unattainable in the long term Peter DeSantis, Amazon top AI executive, told the Wall Street Journal that, and I'm quoting, AI has a cost problem. If we ultimately want AI to transform everything, the costs have to be different." Now I want to add my two cents to this. I am working with several different organizations. In one of them, I am the fractional chief AI officer, and tracking costs across a large organization with fragmented departments and different people using different levels and different tools is extremely hard and on the verge of impossible at this point. Do I think it will be solved? I have zero doubt it will be solved in two different ways. potentially both will happen at the same time. One, the labs themselves will figure out how to create a orchestrator that will truly put the right work to the right model, dramatically reducing the cost. And if they won't do it, people are just gonna switch to other options. The other thing that is going to happen is that third-party tools are gonna build harnesses that will do the same thing, and that means that the labs themselves will lose more traffic. So if the labs themselves will not figure out a way on how to allow the cheapest model to do all the work it can and only move it to a different level of model to do higher level work and fine-tune that perfectly, what will happen is that third-party tools will do the same thing using all the models or multiple models, which means the labs will lose even more traffic because some of the work is not gonna be done with a lower, cheaper model that they own, but will be done by a third-party model. And like I said, I think both things will happen at the same time, and we'll have some organizations using this and some organizations using that. The value, obviously, of using one company is from a data security and infrastructure perspective where it is a lot easier to manage. But again, I think if a company puts something like this in place and can guarantee the data safety and security, then people will switch. Because if you don't care whether using DeepSeek or Qwen or OpenAI or Anthropic, you just care about getting the work done safely and in the most efficient way, then that is gonna be the way of the future Now combine that with the previous topic of the government is now stopping the US labs, leading labs from releasing their latest and greatest, and you got the perfect storm of pushing companies to use other models or at least have that as a secondary option. I shared with you before, every automation, every major automation that I build that runs behind the scenes, consuming tokens always have a fallback plan. It always goes to one model as a first option, and then if it doesn't get access to it for whatever reason, it switches to a different model, either from the same company or from a different company, in most cases from a different company. So if Anthropic is down for two hours, my automations still keep on working. I assume most companies and most organizations do the same thing. But for that, you need these kind of relationships in place with multiple companies, and you need to build it in that way from the ground up. Either way, I think we're going to see a dramatic change in the approach on how companies consume AI. Just a couple of months ago, we talked about the concept of token maxing and how companies were encouraging employees to use the models and consume as much tokens as possible in order to drive efficiency. And now we're seeing exactly the opposite because people start understanding the amount of money that actually is going to consume. So we heard xAI is now budgeting engineers with X number of dollars per month. Meta is doing the same thing. They're all switching to in-house models from external models because this is going to be cheaper. And all of that, combined with the latest restrictions, put a very big question mark on the IPOs of OpenAI and Anthropic. I don't think they have a choice. They will have to go public because otherwise they will run out of cash. Yes, Anthropic had a potentially profitable month, and there are rumors about a profitable quarter, but I don't think that's sustainable. I think this is just a one-time thing because of they had less capital expenses or other expenses in this particular month or quarter versus their full plans for the future. OpenAI are very, very far behind having that kind of scenario, and so they must reach their IPO, but they need the IPO to raise enough capital to make this worthwhile for everybody. And the investors that need to pour billions of dollars into this has to be certain that this is not going away from either a economical reason or a political government control reason. And right now, there is a very big question mark about both these issues Now, the third aspect of this bigger story is that it is becoming very obvious that specialized models that are trained to do something specific are outperforming the top-of-the-line leading labs' latest and greatest when it comes to doing specific work. So Microsoft MAI at Build 2026 on June 10th said that Their new environment that is called Reinforcement Learning Environments, RLE Frameworks create unique training gyms that is producing custom agents tuned for individual enterprise workflows One of the examples they gave that a McKinsey-tuned MAI model beat GPT 5.5 on quality at 10X lower cost Thinking Machines, the company of Mira Murati, variation for financial document filtering that they published on June 2026 The tool was tested on six financial document filtering tasks, and it has outperformed all frontier models at the time and did this at a lower cost Bloomberg's ACL 2026, which includes eight papers that were released just now on July 2nd, found that domain-trained models match or outperform larger specialized guardrails on unseen financial policies. Basically, their key finding is that model size is not the determining factor, but training data and domain alignment are the determining factor on results for very specific, highly specialized real knowledge work and there are examples in finance, clinical medicine, enterprise safety systems, and others all showing the same pattern that basically if you take a smaller model and you train it on very specific information down to the task level and your specific company and your specific entities and the way you work, you will get better results with smaller models, which again is going to put a spend crunch on the larger frontier models Now, the labs themselves know it, and they keep on releasing smaller, faster models that usually outperform or align with the previous top-level models. So In, in addition to getting Fable 5 back, we also have Claude Sonnet 5, which was launched on June 30th, and it matches most of the capabilities on Opus 4.8 at a much, much cheaper price point So at a 40% less cost, $3 for input tokens and $15 for output million tokens versus five and 25. So three versus five, 15 versus 25 for Opus 4.8. And right now it has an introductory pricing of $2 and $10 through the end of August. So a very big spread. you'll be paying just over a third to get very similar capabilities by using Sonnet 5 versus Opus 4.8 Now, in addition to matching the performance of Opus, it is apparently in real life even better than Opus was before, and early access partners, including Lovable, ClickHouse, Pace, and others have reported really solid performance A feedback that was reported by Anthropic is saying, and I'm quoting, I asked Claude Sonnet 5 to investigate a bug. Unprompted, it wrote a reproducing test, implemented the fix, then stashed it to confirm the bug came back without the change, all in a single pass." That tells you that these models are getting cheaper from the leading labs without going to the frontier, while still getting better results in real life. In addition, with the same concept that we just Claude also released Science Workbench this week Now, this is not really a new model. It explicitly runs existing Claude models, including Opus 4.8, but it integrates 60 plus scientific databases. Again, the same concept we just discussed of providing it more specific specialized data that can now run better in that specific domain In initial tests of this new model, there is already approval for the fact that this model actually helps in scientific research across several different fields And so again, no new model, no development, that's just access and training on a specific dataset provides better results in science and research as well Now, to show you that this is a very serious direction, let's just look at some very recent fundraising in this direction of training specialized models. One is Legora, which is a legal AI company, just raised in June their Series D. They raised $550 million out of $5.55 billion valuation Even Up, which is another legal AI company with their Series E, raised $150 million at a two-plus billion dollar valuation, doubling their valuation in just under one year. A 8090 Labs, which is another enterprise coding platform has raised $135 million An interesting note about this company, it is led by Chamath Palihapitiya, who is one of the guys from the All-In podcast who held senior roles in Facebook and is not new to the field of developing large-scale software. He is targeting specifically regulated industries, healthcare, insurance, financial services, aerospace, energy, manufacturing, and stuff like that So again, not just another coding platform, but a coding platform specialized for specific industries. And I think we're going to see more and more of that kind of specialization of models that will cost less and deliver more value on specific topics Now, there are other signs that things are not moving forward in the smooth path that a lot of people maybe think it is moving forward. Everybody's talking AI, AI, AI, and everybody on the outside that are not doing it yet are thinking that this is an easy path and a clear path forward, and the people who are doing it are sometimes not getting the results they want, definitely not at the money they're willing to invest So just earlier this year, Writer, who is a large AI company that builds an AI writing product that has been around for a while, even before, ChatGPT came out. So they released their 2026 Enterprise AI Adoption Survey. That survey included 2,400 C-suite executive and employees And 48% of executives call AI adoption, I'm quoting, "A massive disappointment," up from 34% the prior year. Only 29% report significant ROI from generative AI tools, and only 3% from AI agents specifically 75% of executives admit company AI strategy is more for show than actual internal guidance. 67% of executives believe that company has already suffered a data leak or breach from unapproved AI tools being used by the company CAIS Remote Labor Index that was just published in June of 2026 is trying to see how much of actual real work is being automated by the different models. So Fable 5 achieved 16.1 automation rate on real freelance work. Opus 4.8 achieved 8.3, GPT 5.5 achieved 6.3 So what does this mean from the usage of models right now? And then we're gonna put all three topics together. What it means is that you do not need, and you should not use the latest and greatest models for everything that you do. If you are on a personal account, it doesn't really matter because what you're doing is subsidized. I have the $200 plan on Claude and I can use it as much as I can, and I don't hit the wall ever, and I'm using a lot more tokens than I would have if I would've been an enterprise employee, paying the $100, $200 a month plan is not an option. You just pay for credits and tokens depending on the company, depending on the specific use case, and that would've cost me most likely thousands of dollars instead of the $200 I'm paying per month right now. Now, would that have been worthwhile for me? For what I'm doing right now, probably yes from an ROI perspective. Is that true for most employees? The answer is 100% no. The type of things that I'm doing with AI on daily basis for myself and for my clients are way above the average by a very, very big spread. Does that mean everybody can get there? Yes. But as of right now, they're not, and there are many other reasons why AI is not working well in organizations right now. The number one from my perspective is training. This is what I do for a living, right? I go into companies, and I help them train employees and executives and boards on what is AI, what it can and cannot do, how to set it up correctly, how to use it effectively, and so on, and the differences are night and day. And over 48 hours of two days of training, you go from employees who uses AI in order to replace Google to employees who write applications who can replace the biggest bottlenecks they have on their day-to-day work right now. So I do think that training is the number one showstopper right now between the executives and people who are saying that AI is overrated and that it's been a failure compared to companies who are seeing incredible returns on the efforts that they're doing. That doesn't take away everything else we said. There is a serious risk right now with using the frontier models of them going away, being blocked, and so on. There is proof that using smaller specialized models is delivering better outcomes at a much smaller price, sometimes a fraction of the cost. There is a huge opportunity to replace the leading US models with Chinese open source models if you are okay with the risks that this may introduce, but the risks can be mitigated dramatically by hosting these models on US-based servers by AWS or Azure or Google Cloud. So you don't have to rely on the Chinese servers. You can run on a US provider using these open source models, and that's why they're creating them as open source, and you can benefit from them while dramatically reducing the risks. The risks are still there. It doesn't go to zero, but it allows you to run safer and pay a significant lower price. And if all you're doing is writing code or doing mundane tasks that run your company without sharing any real sensitive information, the answer is very simple and straightforward. What does all of that mean to the future of AI and from a financial perspective and from a model capabilities perspective is unclear to me at this point. But I think we're seeing a shift in the last two weeks from a government that's saying, "Let's run forward, and we don't care about how this is going to look like. We just wanna win," to a government in the US and definitely in Europe for a long time now that's saying safety first. And we are seeing the audience going from, Let's go all in on AI. It doesn't matter what the cost is, it's gonna be worthwhile in the end," to, Let's take it step by step, figure out the right effective way to do this, reduce costs, limit the usage, find the right models," and so on. and that has serious implications on how the leading labs can project their work forward, their spending, their IPOs, and so on. So very interesting week this week when it comes to the news and what it means if you connect all the dots Now, since I mentioned AI training, before we dive into the rapid fire items, I want to remind you of the multi-agent orchestration course that we have been running since April. They all get sold out. There's a few seats, I think three, in the August session. So if you want to learn how to build agents and skills and combine them together into incredible automations that can automate basically any knowledge work in your business, and you wanna do this before September, then you should sign up right now. Literally, don't wait, because these seats will disappear. The next cohort that we're going to open is gonna be for September. So if you wanna do it before that, come and grab the last three seats on the August course. These courses I'm teaching also privately to organizations. So in a leadership position in a company and you want to teach your people, your team, your department, your entire company on how to develop AI agents safely and effectively, Please reach out to me on LinkedIn or or directly to the link of my calendar that is available in the show notes, and I will explain to you how that works. Or if you're an individual, you can sign up to our open courses. Now to the rapid fire items. Anthropic apparently are in early discussions with Samsung to develop their own custom AI chips, and they're joining OpenAI that just announced their own chip So we told you last week that OpenAI just announced Jalapeno, which is the chip that they have developed in the last few years together with Broadcom We already know that Google has their own chips, Amazon has their own, so Anthropic is going to be the last big player that doesn't have one. xAI also has their own chips. They're all also using Nvidia, but they're all developing their own in-house capabilities as well Now to show how serious they are, Anthropic also poached Clive Chan, who was one of the early members of OpenAI custom chip to help them in these efforts. So they are talking to Samsung. Samsung has already invested 65 billion in Anthropic, so they are in the same boat already, and now they're going to work together to develop a custom chip for Anthropic themselves. Anthropic also announced a new AI native enterprise services firm, firm that was co-founded with Blackstone, Hellman Friedman, and Goldman Sachs backed by approximately $1.5 billion in committed capital. This follows a very similar move done by OpenAI and then later on Microsoft as well from Anthropic to OpenAI, we told you that OpenAI released three new variations of GPT 5.6. Again, not really released. We don't have access to them, but specific companies and the government has access to them, and I assume we are all going to get access sometime in the next few weeks. But OpenAI also retired GPT 4.5, completing a complete phase out of the GPT-4 family from ChatGPT From OpenAI to Google, Google just launched NanoBanana 2 Lite and Gemini Omni Flash to the broader audience. It's something they've announced and demoed before, but now it is available to the public NanoBanana 2 Lite is the fastest and most cost-effective Gemini image model generation, and OmniFlash is a high-quality video generation and conversational editing of existing videos. These releases are already available to developers and consumers, and you can use them right now Meta is planning to sell compute on their extra compute that they've built. So according to Bloomberg, Meta are developing a new cloud infrastructure business to sell access of AI compute to anybody who wants to pay for that and basically compete with Google Cloud, Microsoft Azure, AWS, et cetera, by providing compute. we've seen SpaceX talking about and already doing, uh, the same thing, and it is very lucrative space to be in right now. Meta has a lot of infrastructure. They're not necessarily using all of it. They're planning to buy a lot more compute, so instead of just using it internally, they're also planning to sell some of that compute. And the last thing I want to mention today, there is a lot happening in the AI device space, and we haven't talked about it in a while. But there's potentially a new player in the market, and that is SpaceX. So there are rumors that SpaceX showed investors an AI device prototype as they were getting ready for their IPO Now, currently we've seen multiple failures of AI products such as the Rabbit r1 and different types of pendants and so on. The only somewhat successful AI device right now is the Meta glasses that are selling nicely and coming up with new and new features, but they're still not anywhere near replacing phones. Well, apparently SpaceX might be going in that direction. Now Elon Musk himself denied the reporting and said it's utterly false. However, these rumors comes from several different directions, and if you think about not SpaceX, but their sister company, Tesla, they have a lot of experience in building devices, including everything that runs the Tesla, including building the chips, including building the displays, including integrating it all together, including having access to internet through potentially Starlink. So in the Elon ecosystem, there are a lot of skills and a lot of capabilities that will make perfect sense for him to build an AI-driven device. Is that really happening or not? I'm not sure. But I have a feeling that 2027 is going to be a very, very different year from an AI on-device perspective, but time will tell. That's it for today. I hope you found this, uh, very educational. Again, I think this week was a critical week in the history of AI. A lot of things are happening. Again, small things that are connecting to a very interesting big picture that impacts how AI evolves from here forward, both from a safety perspective, from a financial perspective, from a geopolitical perspective. And so we'll keep you updated every single week. We'll be back on Tuesday with another detailed how-to use case with AI. We have two really interesting episodes. I'm not sure which one is coming out on Tuesday yet. I'm going to decide over the weekend. But until then, have a great 4th of July. Enjoy this long weekend. Have fun with your friends and family, and come and join us again on Tuesday