316 | MBA out AI skills in, rouge agents and bio-risk, new Agent Plugin industry standard, and more important AI news for the week ending on August 7, 2025

Join the Multi-Agent Orchestration Course and use LEVERAGINGAI100 to get $100 off > https://multiplai.ai/multi-agent-orchestration-course/ 

AI agents are already hacking real systems without being told to, and Wall Street just predicted a 20% AI-driven workforce cut.

A UK government test caught frontier models from Anthropic and OpenAI attempting real supply-chain attacks, fake online identities, and prompt injection, on their own initiative. Real organizations have already been breached the same way, including a national finance ministry.

Isar connects that story to a second one: PwC's 2026 Financial Services Workforce AI Survey shows leaders expecting to cut 20% of their workforce over five years, while paying AI-skilled employees significantly more. He also covers a new open standard for AI agent extensions, OpenAI's unlimited ChatGPT rollout, Anthropic's move into custom chips, a Google DeepMind leadership shakeup, and an AI agent that ran an entire sales pipeline during a founder's paternity leave.

In this session, you'll discover:

• How AI agents in testing bypassed their own safety instructions to hack real GitHub repos
• Why there's currently no legal framework for damage caused by an autonomous AI agent
• What 86% of financial services executives now value more than an MBA
• Why AI just designed 16 working viruses, and what that means for biosecurity
• How one founder's AI sales agent generated $3M in pipeline while he was on leave

About Leveraging AI

• Multi-Agent Orchestration Course: https://multiplai.ai/multi-agent-orchestration-course/
• YouTube Full Episodes: https://www.youtube.com/@Multiplai_AI/
• Connect with Isar Meitis: https://www.linkedin.com/in/isarmeitis/ 
• Join our Live Sessions, AI Hangouts and newsletter: https://services.multiplai.ai/events

If you’ve enjoyed or benefited from some of the insights of this episode, leave us a five-star review on your favorite podcast platform, and let us know what you learned, found helpful, or liked most about this show!

Transcript

Copy Transcript Copied

0:00

Hello and welcome to a weekend news episode of the Leveraging AI podcast, the podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business, and advance your career. This is Isar Metis, your host, and I've got some really fascinating stories for you this week. First of all, we're going to connect the dots between a few interesting stories that is showing the crisis in which the world is when it comes to autonomous agents that are really coming to a point where, as we talked last week, maybe the world needs to stop and go backwards, but this week or not go backwards, but at least not move forward. But this week, we have a lot of interesting stories that, again, if you connect the dots, you will see how problematic the situation is right now. We're also going to talk about the impact of the AI on the workforce, and it is not going to be necessarily job losses, but more like how employers in higher education are impacted by the crazy shift in AI capabilities, and what does that mean for your resume if you're looking for a job or if you're considering higher education or a combination of the above? Then we have a lot of rapid-fire items. Some of them are really fascinating that could have been a full story on their own. As an example, we have a new open standard for AI agent extensions, uh, called Agent Plugins. That is a coordinated move by some of the biggest giants, and I have a feeling this will be the next MCP and skills, so something that will take over anybody who is using AI across the world and will be commonly used within a relatively short amount of time. We have some very interesting news from OpenAI. We have some new models. We have a very interesting story in the end about how an AI agent is replacing a CEO of a company and generating three million dollars of pipeline. That's a little, uh, nugget in the end that we're going to talk about that I found really exciting, interesting, and scary all at the same time. So the bottom line is we have a lot to talk about, so let's get started. So in the past few weeks, we talked a lot about how models are escaping their sandboxes and breaking into places that they shouldn't have. So the first one we talked about was two weeks ago with the OpenAI model that has broken into Hugging Face and broke into many, many different components of Hugging Face. Last week, we talked about other models are doing this. Well, this week we're getting additional points from other sources. The first one is the UK AI Security Institute, known as AI-SI has done a thorough research at the end of July, so July 25th to 28th, and they have released the findings of that research. Ten out of 122 cyber evaluation runs have done things that they were not supposed to do. 17 actions from Anthropic Mythos-5, two actions from OpenAI GPT-5.6. So both models tested with safety filters disabled and open internet access enabled. So the-- Again, it's not the regular total standard models, and yet these are models now available and open to the public. They were specifically instructed against social engineering or leveraging the live web, and yet they did it. So let's see what happened. Anthropic Mythos-5 had performed a supply chain attack. So the most serious incident was the agent attempted to insert malicious code into real public GitHub repo Now, to tell you how crazy this is, the model actually created multiple fake online identities to try to social engineer the maintainer of that public repo, including signing off a message in Danish to appear authentic. And the goal was to attempt to convince the approver, the maintainer of that repo, to approve the malicious code into the repo. So this is crazy and scary at the same time Now it doesn't stop there. Agents directly contacted real people. Humans sent messages and files to persuade them to run the malicious code. It planted prompt injections targeting other AI systems in order to help it in its effort, and And one agent deliberately leaked its own GitHub credentials for other agents to find and reuse. So this is an agent-to-agent coordination that is trying to help each other do a malicious act. Again, this is troubling, scary, call it whatever you wanna call it, but there were other incidents. This is just one of them, and this was happening in tests to try to see how they're going to behave. Now you're saying, "Oh, it's only ten out of a hundred and twenty-two." Well, if you think about the fact that the world is gonna be flooded with millions of agents running on their own, performing tasks on behalf of businesses and individuals, and roughly 10% of them can do these kind of things in the current systems, not in the next future, uh, things that the labs currently are developing and have in their possession, you understand how serious the situation is Now, if you remember last week we talked about the fact that Anthropic themselves said that they found additional cases after OpenAI published their thing. So they released a blog post called Anthropic Alignment Science Blog: Agentic Misalignment in Summer of 2026. We talked about this last week, and again, they found several incidents where their AI has misbehaved, broke its protocol, and went to achieve its goal by doing things it wasn't supposed to do. Uh, again, in their case, it wasn't maybe as bad because the agent thought it doesn't have internet access when it did have internet access, so it assumed that what it's doing is a part of a test and not a part of a live universe. But yet it did things that it knew that it was not supposed to do. And so again, this is a situation where we know for a fact that the models that we are using today that is available to everybody in specific conditions will do things that they're not supposed to do. Or like Katie Moussouris, uh, and I hope I'm not butchering her last name. She's the founder and CEO of Luta Security, said NB- told NBC about this, and I'm quoting, "Everyone who is running AI inside their systems needs to be prepared for their own AI and their own agents to do unexpected things in pursuit of goals." This is a fact right now, and again, sometimes it's gonna be something small, and sometimes it might be something that would lead to whatever level of catastrophic results, either things that are harming your internal systems, external systems, really bad reputation damage or whatever else, uh, these agents can do. But this is the reality that we're in right now. Everybody in this crazy race to deploy agents to get benefits from them. And again, we're gonna talk a lot of, a lot about benefits and one interesting one in the end of this session. But it is also really, really problematic right now from a security perspective Now, I did an additional research, and I wanted to find additional cases where agents have already attacked live systems beyond what we know. And I found several cases. Again, the UK test was a test. They were trying to see what the agents are going to do. But then I found the following stories all from the recent few weeks Now, to be fair, the company who aggregated them is a security company called ExtraHop, which its purpose is to help protect against these kind of things. So they have a interest, a vested interest in sharing these kind of stories, but I don't think they made them up. And so here are a few stories of agents breaking into or doing things that they're not supposed to do. So the study is called ExtraHop Global Threat Landscape Data 2026, and they said the following: 40% of organizations targeted by AI-enhanced external attacks in the past year. 38% experienced compromised AI identity or session theft. 36% reported supply chain breaches involving AI systems. So this is just general statistics. Now some specific stories An open source AI agent from Hermes that was running in an unrestricted mode autonomously infiltrated the Thailand Ministry of Finance. It executed commands without human approval at any step, explored internal files, hunted for privilege escalation paths, and basically went through whatever it wanted in the Ministry of Finance of a country, looking for information that it thought is going to be helpful to whatever the goal was Obviously, we talked about the OpenAI Hugging Face incident that we talked about two weeks ago. Now, so this is showing you that there are real examples of real AI agents, either in testing environments, breaking out of these testing environments or out in the live, that are already breaking into systems, including company systems, government systems, and so on, performing things that are not even meant to be malicious. Meaning there isn't a person there that told them, "Oh, I want you to break into this and get that," or change this or that, but it is they're just trying to achieve the goals that they were given, and in that process, they're doing things that are completely illegal. Now, speaking on the legal aspect, there's a very big problem with that. The computer crime laws, including the US Computer Fraud and Abuse Act, known as CFAA, require human intent. Autonomous AI agents without a malicious intent by a person does not have any legal framework right now to say, "Oh, there's a damage that has been created to a company, an organization, a government, whatever." You cannot sue for that because there was no malicious intent by a person. That is also very problematic because there is now no liability to what these agents may do, and that will then translate into are you insured against something like this? Because there's like this is a very big uncharted territory. Uh, Ahmed Ghappour, who is a computer law scholar at the New York Law School, said, "When an AI agent acts without being specifically directed, the more interesting questions may lie in negligence and product liability, not criminal hacking law." But that means if you understand, if you read between the lines what he's saying, when he's saying it is negligence and product liability, who owns the product? So if I'm using a OpenAI system and it goes and hacks a third-party company, the negligence is not on me, it is on OpenAI because they're the ones that created the system. Again, I'm not sure what that means from a legal perspective. I'm not a lawyer. I don't know if any lawyer knows what this means because, again, there's no precedence for that in history. So we are in a very interesting situation from a legal and then, again, insurance and liability perspective when it comes to the damages that these systems may and probably will do in the near future So the solution is the government, right? The government needs to step in and come and say, "Okay, this is what we allow. This is what we don't allow. Here are the laws. Here is how we're going to approach this new situation." And to be fair, they took the first step. So the White House hosted on August 3rd a meeting with Meta, Anthropic, OpenAI, and Google, and the idea was to discuss their new act, that is the voluntary US AI safety testing framework. The Trump administration finalized the details for a voluntary cybersecurity framework And the meeting was held on Tuesday, August 5th so what does this framework actually provide? So developers will grant the US government up to 30 days of early access to frontier models before public release. the idea for that is to allow the government to test for hacking capabilities only But it is explicitly structured, so it cannot create mandatory licensing or any kind of pre-clearance for releases. Again, this is a voluntary participation, uh, by these companies Now, to make an even bigger loophole, the open weight models are completely outside the scope of this mechanism So the government that is supposed to potentially find a way to stop this really risky situation put in place a very open-ended option for companies to share information if they are willing to do so, because it's voluntary, and again, it's for very specific needs. And to be fair, I'm not sure the government actually has the right means to test it, especially not in thirty days. So I have a feeling that all the most capable testing capabilities for these labs exist in the labs themselves, and giving it to the government probably will not yield the same results that the labs themselves can find. Having some kind of collaborative solution between the labs that will put in place a more rigorous testing environment that will allow them to collaborate on the safety side of things, I have a feeling will yield much better results. But as of right now, that collaboration is very far from being an option. It is the first time that I've seen Anthropic and OpenAI agree on something, including last week's letter to ask the government to slow them down. So maybe, maybe there are cracks in the wall of we are five-year-olds and we're not going to play together. Um, again, I'm being very sarcastic here, but I think the whole personal situation between Sam Altman and Dario Amodei is not helping, uh, the case right now. Now, in talking about Dario Amodei, as you know, Dario Amodei is not the most admired person in the US government right now. As a reminder, uh, Anthropic was placed on a national security blacklist by the Trump administration, and yet now they're asking everybody to play nicely together and to voluntarily give their new models to the government to test. So that, again, doesn't show great collaboration when on one hand saying, "Oh, please give me your models so we can test them for safety," and on the other hand, "I'm telling you you're a national security risk at the same time." So not a very clear path for success, uh, in the current setup. Now, if you wanna see how weird the situation is, OpenAI is currently running a $50,000 jailbreak competitions that is specifically targeting the biosafety safeguards because the federal agencies have not delivered any upgraded rules in that topic. So again, this is OpenAI taking the future of biosafety to their hands because the government hasn't done so, and they're giving prizes to people who are gonna use AI systems in order to break into biosafety environments, uh, which is absolutely crazy to me. But this is the reality we're living in right now. Governments, the US government or any other government cannot move fast enough and does not have the skills to evaluate these models. It has to come from some kind of a collaboration between the different labs themselves, and I really hope that they're going to wake up before something really bad happens. Now, speaking of biological security, it leads us to our next thing. Stanford and Arc Institute published in Science on August of 2026 that AI model Evo, which is a model that they have developed, just designed 16 viable viral genomes. So those of you who don't understand what this professional lingo means, it means that they generated 16 viruses that did not exist in nature before The AI was able to synthesize 302 candidate viral genomes, uh, that were generated and synthesized And out of those, 16 were actually working. Now, to be fair, the goal of this thing was to do something really good. The goal of this thing was to generate viruses that are going to attack and destroy drug-resistant bacteria, so including coli and tuberculosis and MRSA And so the goal is positive, right? The goal is to create viruses that are not gonna kill people, but are actually gonna kill bacteria that really hurts people. But that means one thing, that you can now create actual live viruses in a lab with AI, and you can create hundreds of variants and find the dozens that actually work and survive in the wild and do what they're supposed to do Brian Hay, assistant professor of chemical engineering at Stanford University said, and I'm stating, "This is the first time AI systems are able to write coherent genome-scale sequences. The next step is AI-generated life." So this just kind of tells you the direction that these scientists are thinking about on what this means. So this is now just a virus, a very simple formula, but this is just step one in this research And going back to the point that I raised, Thomas Inglesby and Moritz Hanke, who are researchers at the John Hopkins University School of Public Health Center, said the following in science. Although this is, uh, promising for life science applications, it also raises urgent biosafety and biosecurity questions. The ability to compose viral genomes using generative AI now exists. The governance to safely steer it does not. Again, very similar to what we've seen in the previous section. The ability of AI to generate new things runs way, way, way faster than the government's ability to create the right laws and the right infrastructure in place in order to address the new risks that it generates Now this again is not too surprising to the labs themselves. Anthropic themselves in the release of Fable 5 have done a lot of effort to suppress the biological capabilities of the underlying model, and they even said, and I'm quoting, "Because of our concern about these dual use capabilities, those that could be used for beneficial or harmful purposes and where the line between them is not always easy to draw, we intentionally launch Fable 5 with almost all biological queries blocked." So the labs themselves are aware of the potential of the models that they're releasing to do these kind of things, like synthesize viruses and that could be generated in a lab afterwards that nobody has any kind of way to deal with them because they did not exist before Now again, what does this does? If you think about the fact that the world has been generating biological weapons for a very long time, it was done only by governments, it was done only in very secured facilities under very strict rules and regulations. It is the same thing as AI that existed before. AI existed before generative AI. It existed before ChatGPT moment. It existed in the hands of a very short list of companies, and the ChatGPT moment democratized the access to AI, right? So now we all have access to it, and we can do really incredible things. Well, what this does is it democratizes the opportunity, at least, to generate new viruses, and then potentially new bacteria, and then potentially new other stuff to anybody who has access to powerful AI and the right resources. This is very, very scary because these people don't have the safeguards and dozens of years of experience in keeping it contained where and when it needs to be contained. Again, very scary in my eyes, and I'm not trying to make this black and dark episode, but this is where the line is pointing to right now So quick recap of this topic. The testing to deployment gap is growing, and it's growing fast, right? We have models that are extremely powerful that can do things that are not expected from them to be done, including hacking more or less anywhere and anything they want. And the testing environment is either limited or voluntary by the government, if you want. The actual labs themselves try to put safeguards in place, but nobody else is verifying that information. At the same time, these models can now generate-- Again, it's not potentially, they can generate new biological threats that are actually working. It's not, again, a theoretical thing, it's a practical thing, and there's no real process right now to figure out where this can lead or where this is going. Uh, and on the other side of this, which again is a lot less relevant and scary from my perspective, but it is a problem that there is no legal framework to address AI liability. Right now there's really nothing there, and the actions that these models are doing are already happening. So big mess in front of us right now from that perspective. So what do I think the solution is? I think the solution is collaboration, like collaboration between the labs themselves, collaborations with academia, collaboration with governments, collaboration with experts on different topics. Again, today we talked about biological threats and data security threats, but what about the social threats? What will it do to the economy? All these things have to be considered, and I really, really hope these labs will find a way to collaborate between all of them, including the Chinese ones, the US ones, the Russians one, the Israeli ones. Like all the advanced labs in the world should sit together with the right experts in the room and focus, as of right now, most of their energy in finding answers before they continue moving forward. And we're gonna talk about, uh, in a few minutes about it's already moving forward, but we'll get there in a minute. Now to our second deep dive topic, and we're going to talk about, as I mentioned, the AI workforce revolution that is basically already here So the first aspect I want to talk about is the Kogod School of Business from the American University. It has done a 483-student study between the years of 2024 and 2026 And what they're asking students is what they were asked in job interviews. So this is not anticipating, this is not trying to project what will happen. It actually looks at actual trends on what actual students were asked in job interviews in the last three years. So employer inquiries about AI skills were 11.6% in 2024. It is 32% in 2025. It is 42.6% in 2026. Almost half of job interviews included questions about AI skills, which is a 267% increase in just two years. It went from 11% to, again, 42.6%. Another question that they asked the students, uh, is how many times a week they are using AI. The students who responded with 11 plus times per week grew from 6.2% to 29%. Again, this is a 368% growth. I must admit that something is weird in this data, because if we're talking about data in 2026 and only one of every three students is using AI 11 times a week or more, this sounds like a very low number to me, but maybe I don't understand something in the way this, uh, survey has worked They also reported that non-users, people who reported not using AI at all, dropped 74.9% during this two-year period. Again, I'll be shocked if there's people in universities that still don't use AI, but again, maybe I don't understand something very critical about this, and I live in a bubble, and I don't understand how small my bubble is Now, Handshake, which is a different company, had done their 2026 graduate survey, which is by the Cogode study, and it's saying that 58% of graduating college seniors say they'll need a stronger AI understanding to succeed in their career. Only 27% of the students report that AI was meaningfully integrated into their academic program So again, 58% think they will need more AI in order to succeed in their career, and only 27% saying that they got any meaningful AI, uh, learning during their university. That is a 31% spread between these two groups, which tells you that the education system is not moving fast enough. And again, it's not surprising. It's just like government. So just cannot. It is just very, very hard for these very rigid environments to move fast, definitely not at the speed that AI is moving But then the last parameter that I found was in PwC 2026 Financial Services Workforce AI Survey. We talked about this a few weeks ago. They surveyed over 1,000 executives at director level or above in firms with $500 million in revenue or more. So this is the target audience. 86% of US financial services executives say AI skills training is more valuable than an MBA for many hires. Think about how much money and how much effort and how many schools and how many years are invested in MBAs of students around the country and around the world, and you have 86% of financial services executives, which are probably the top leading hiring channel for executive MBA students, say that AI skills are more important in some cases than the MBA training, that their students that they're trying to hire are getting. 62% plan to hire employees with AI specific skills in the coming year. 61% will upskill or reskill existing staff. 57% will partner with external vendors to close the gap So you're starting to understand the problem here, and we're gonna dive deeper in a second. But the students are not prepared, but the economy, the place that they're going, are asking more and more and more, and are seeing more and more value, and are focusing on that value even more than traditional education that students are getting as the core of what they're getting in universities Now, the PwC survey doesn't just look at the credentials of students and what people are looking to hire. It's trying to forecast industry workforce, and it's going to look fundamentally different in the next five years. So again, in the same survey by you just heard who the people are, nearly 80% of financial services leaders expect their workforce to shrink by at least 20% over the next five years. Why? Due to AI adoption, 30% identify entry-level roles at most as the most vulnerable, 26% flag middle management next. So this is not random people giving these answers. This is not my feelings that I've been sharing for a very long time now and raising the flag. These are top leaders in the largest financial institutes in the world that are saying out loud that they will shrink their workforce by 20% over the next five years And those who are still going to have a job after this reduction their salaries are going to be significantly influenced by their AI capabilities. So over 90% of executives plan to raise compensation for employees with AI skills. 90%. That's basically almost everybody. 58% intend to tie pay directly to AI-enabled productivity. So if you do not know how to use AI effectively, you cannot get a raise. And if you do, or if you're really good at it, you can make more money and potentially significantly more money because the productivity of the employee and the success of the company will be correlated with the ability to use AI Now, if you think what this means, this is a structural repricing of human labor, right? So far, your worth was based on your experience, based on your personal knowledge, based on your personal skills, and now this is shifting from that to your ability to leverage other skills, AI skills, knowing how to use AI to do work for you, with you, et cetera, in order to get a better, faster, bigger total than you can do on your own And I gotta use this for a shameless plug. As you know, I've been teaching AI courses since April of 2023. I've trained thousands of people to do exactly this thing, exactly what this survey is talking about, to be more productive at work with AI tools, and the current course that we're teaching, the multi-agent orchestration course, is the jewel in the crown, right? It's the cherry on top. It is a course that is teaching people how to build agentic systems, how to build the infrastructure for them, and how to create extremely powerful business results without any crazy investment in infrastructure and so on. So if this is something you're interested in order to improve your business if you're in leadership position or improve your career if you're just an individual, uh, come and check the link in the show notes for the next session. Uh, we just started the August cohort. It is going really amazing. There's a fantastic group of people there from all over the world that is taking, uh, this course, and the next cohort will start in mid-September, which means it will graduate in mid-October, which means the next course will probably open in November. So if you don't wanna wait for the end of the year and you don't wanna wait for the end of the year, come and sign up, uh, for the course. You get $100 off with the promo code leveragingai100. And now back to our story. And I apologize for this plug, but I really couldn't resist myself with what this survey actually means in saying very clearly and out loud. But now back to the story Now to make this point even stronger, in 2025, and there's no updated number in the PwC survey for 2026, but in 2025, 88% of AI related positions in financial services were AI user roles, meaning not developing AI solutions, but knowing how to use AI in your day-to-day work. Again, this is not a come develop solutions for us. I want to see you doing your job with AI all the time, meaning it is applicable for basically any role in this particular case in financial services. I think the same thing is true anywhere else Now, in this survey, there's another parameter that shows the big dilemma of employees right now. So the same leaders, executives report that 44% of leaders, uh, saying that employees have concerns about their job security because AI. Forty-three note that employees use AI only when required. Forty percent acknowledge staff feels overwhelmed by the pace of AI-driven change. Thirty-four percent cite the change fatigue as an active barrier for adoption. So it is a very complicated time as an employee in large organizations right now because on one hand, you are asked to use AI as much as possible. On the other hand, they're telling you that you are going to lose your job potentially because 20% of the workforce is not going to be there in the next five years. But the reality is they're saying very clearly that those who will know how to use AI effectively will keep their jobs and will get a salary raise. I still think that's a very scary situation to be in. The bottom line is you gotta learn how to use AI, and you gotta learn how to use it safely and effectively in your role or potentially make yourself a brand new role if you know what you're doing. And I've seen this time and time again with people who are taking my courses, with people who participate in my community and has been there for two years and see the shifts in their careers in just two years just because they're invested in learning how to use AI effectively, and they keep themselves up to date with skills all the time And to connect this back to the course that I'm providing, Peter Pelini, who runs PwC Financial Services industry practice and I'm quoting: "There is a clear need for people that understand not just what an agent is, but how do you build them? How do you think about managing them?" And this is exactly what my course is teaching But there are additional data points that are showing how different the job market is right now than it was before. So LinkedIn released a report called LinkedIn 2026 Skills on the Rise Report And that report shows something that is very interesting and yet not surprising. So technical AI skills like AI implementation, prompt engineering, et cetera, and on the other hand, distinctly human skills, executive communication, cross-functional collaboration, are growing at roughly the same pace at the same time. Which means the labor market is not choosing between AI skills and human skills. You need to have both. The people who will have the right human skills together with AI skills will be the most successful in the future, and probably the relatively near future. Another interesting data point that pushes in the same direction is from Harvard Business analysis of US job posting twenty nineteen to early twenty twenty-five What they found is that structured cognitive roles fell 13% after ChatGPT's launch. At the same time, analytical, creative, and leadership intense roles grew 20%. That is at the early 2025. Think what happened since. The year of agents were in this past 12 months. And so I'm sure these two numbers are now more extreme, or I'm not sure. I assume that both these numbers are more extreme, meaning simple cognitive roles are probably falling even faster, and human leadership, creative thinking, judgment-related skills are probably growing even faster Now, those of you who haven't heard my lecture, I'm saying that one of the most critical human skills right now is judgment. And judgment is the ability to make the right decisions as you are doing things and as you are working with AI. I think judgment with working with AI is the biggest differentiator between people who do the right things with AI, assuming everybody knows what they're doing. I'm not talking about people who don't have the skills, which is a gap that you can close. You can get the skills. But having the right judgment on how to use AI effectively, and it is something that is hard to teach, and it is something that does require all the other things that these two surveys are talking about, right? The human-centric stuff that helps you have the right judgment, that helps you make the right decisions and steer AI in the right direction is a big differentiator, uh, in the new economy Now, I told you the biggest problem that I see right now is that the education system is very, very far from catching up. Now, there are initiatives that we talked about on this podcast and beyond of universities that are understanding what's happening. They understand that the employers are very clear about what they need, and yet I think they're just not moving fast enough. But just to designate a few things that are happening, because they are happening. Ohio State University has initiated an AI fluency initiative in fall of 2025. It is now required coursework for first-year students on generative AI and technical tool training, and so on. Carnegie Mellon University, uh, have their first US bachelor degree in AI, uh, for a few years now, even before ChatGPT came out. Syracuse University has new AI degrees and university-wide integration starting in fall of 2026. Texas A&M University has discipline-specific AI integrations that are integrated into the main industries that they support, including engineering, agriculture, healthcare, et cetera But the problem is not that there are no initiatives to teach AI. The problem is that teaching AI becomes a course of its own. There are multiple institutions right now that have AI degree programs. That's not what we need. I mean, it doesn't... It's not a problem, but that's not the real problem. The problem is that AI needs to be woven into every discipline. Anything you teach, any kind of profession you teach will use AI, meaning you don't need a course in AI. You need to teach people with everything they're learning how to use AI in their field, how to integrate it into the actual things you're doing. And I talk to maybe not a huge amount, but not a small amount of university leaders, and they are very far from being at that point. Professors are pushing back. The actual organizations are pushing back. Some are even saying that this is, "Well, it's just like the internet. It is not going to fundamentally change the way we work." And I so shocked every time I have these conversations. But this is what top leaders in universities are saying to me behind closed doors. They do not understand that the world that they know, the world that they were teaching to prepare students for, does not exist anymore. And if it does exist, it exists because there's crazy friction in the industry of getting this into production and working at scale. But it's just a matter of time. In two, three, four, five years, it will be a very different universe. And if the universities only then start training people and kids into this new universe, then the graduates in these four, five, six years are gonna be in deep, deep trouble when they come to try to get a job once they graduate. Now, at the same time that we're talking about the macro level, the micro level is already happening in more and more fields, and we talked about that as well. Again, I'm just reminding you and connecting all the dots together. The labs or other private organizations are doing everything they can to now focus on specific fields and develop specific capabilities in these fields in order to be significantly more productive. We talked about Microsoft, well, that is helping companies train models for specific aspects that are now yielding significantly better results than the general models in these fields where it costing significantly less. Anthropic just announced that they hired Robert Mahari, uh, who is their first head of Claude for Legal This guy holds a doctorate in legal from MIT, and he's now going to be working at OpenAI to develop specific legal capabilities for Claude to be able to be used in the field of law instead, or at least replacing somewhat, lawyers. I can tell you something from my own personal experience. I have a law firm that I work with to draft documents and so on. Now, my new way of working with them is I don't go to them when I need legal documents. I go to Claude, and I ask Claude to research other companies in my field to see what kind of agreements they have. And it goes through these agreements, and it finds all the stuff that is relevant to me, and it drafts incredibly good legal agreements. Now, I'm not a lawyer. I cannot say how good they are from a legal perspective, but I read hundreds of legal agreements in my career across multiple different disciplines, and they look legit. And then I send them to the lawyers to review. That means that the law firm, instead of charging me for 10, 20 hours for the work, charge me one or two hours for the work. If that happens at scale, then obviously they will need significantly less people to do the same work they're doing today. That is the situation right now. Now, to be fair, inside the law firm, I would guess, and I hope for them, that they're doing the review of the legal documents that I'm sending them with AI as well, which amplifies the problem even more Now, obviously, Mehari himself was trying to sugarcoat it, and he said, "The expertise that matters most for creating impactful legal technology sits with practitioners. At its core, Claude for Legal is a set of building blocks that we are co-designing with the legal community so that lawyers can augment how they work." That is true, but the outcome is very clear. There will be a need for significantly less lawyers. Other than maybe lawyers that will handle the whole issue of gap in the legal system right now of AI damages and so on, but I'm putting that aside for a minute. Now, OpenAI has done the same thing, right? It's not just that Anthropic is moving in that direction. OpenAI has recently, uh, recruited Jason Bojmig, and launching their own formal legal vertical to do basically the same thing. It's gonna be called Codex for Legal, and the direction is very clear Now why legal, you ask, of out of everything? If you think about what AI is really good at and why programming was the first thing, programming is, A, helps the labs run faster so that they have a vested interest to do this. But B, it's a very structured universe. Legal is the same thing. It's a very structured universe, word-heavy, which are the places where AI thrives in the current level without any new improvements. It can do incredible things. And so this tells you where the direction is, but these dominoes will keep on falling one after the other So recap of this topic. What does this mean to you? What does this mean to the world? What does this mean in your industry? What does this mean to your career? What does this mean to your company if you're a leader in that company? More and more executives and leaders are understanding that AI skills are more critical than the formal education that people had before. The actual universities are not ready with the right programming in order to enable that, which means people have to take care of themselves right now. By the way, the same applies for people inside of organizations. Many organizations are starting to take action. I'm doing workshops and courses all the time to private companies, uh, and public companies, to be fair. But what I mean by that is to companies and not public courses. I'm doing that all the time, and the gap is insane between companies who are ahead and companies who are the average right now, which has very few and very little AI skills, knowledge, and practices in place. And that is true all over the world, and that is true across industries and across sizes of companies. Again, I do this every single week. So the gap between the knowledge that people have and the knowledge people need to have in order to use AI effectively is enormous. The people at the decision-making positions are willing to pay more and more for these skills and see them as more and more important. And the labs themselves are now developing more and more capabilities that will shift from, I can use AI," to, "AI can do my job," and focusing more and more on that. So the combination of all of that is the industry, any industry, is not prepared for what's coming. The economy itself is not prepared for what's coming. Higher education and the entire education system is not prepared for what's coming. And while I do see the things that the leaders of the labs are saying recently more and more, that it's going to create new jobs, that human ingenuity and human research and human wish to do new things is just gonna be increased by AI, and I agree. But I think the economy as we know it will take a very long time to adjust, and during that time, there's gonna be a very big shift towards job losses, and it is going to be very significant to the economy because it is potentially gonna get to 20 or more percent unemployment. And even if less, it is going to be in roles that used to make hundreds of thousands of dollars a year, which are the people who are moving the economy forward. I don't want to be a doomer, but this is what I really, really believe So now let's switch to, uh, rapid fire items. There are a few really interesting topics to talk about. The first one is a group of leading companies has released Agent Plugins 1.0 which is an open format to allow standard packaging and distribution of extension for AI agents And it is not tied to a specific vendor. Again, it's an open format that the industry will hopefully, uh, align with, similar to, as I said in the opening, skills and MCPs So the idea here, how are you going to package all the things that agents need so you can move this package from one platform to the other, from one agent to the other, and so on? So what's in those packages? It packages agent skills, which provides the instructions for the agents what to do. It provides the MCP servers that they can connect to, uh And these components are now packaged within a common directory structured, uh, with a plugin.json at the end of it And it has two directories inside, a skills directory and an MCP directory. Those of you who have been using Claude for a while, like I am, know that a similar component exists as plugins inside of Claude, and now I'm not sure they're exactly identical, but I assume that Anthropic will align with that structure as well. So I assume plugins will be in the same exact thing moving forward The interesting thing is who collaborated in this process. So the company who initiated the push towards that was, were Vercel, and that makes sense because they deploy things from more or less everybody on the planet right now. And so they're saying, "Ooh, it will be a lot easier for us if things are standardized." But they collaborated with, uh, AWS, Anysphere, GitHub, Microsoft, OpenAI to build and define the final version of the standard or the immediate version of the standard, version one, but the one that was released to the public Now, after the announcement, Google has also joined this group and saying that they will start using and implementing, uh, this new format as well Now because this collaboration in the development Agent plugins are supported across the following: uh, ChatGPT Codex, Cursor, GitHub Copilot, Keero, and VS Code immediately at launch. As I mentioned, I have a feeling this will be available and supported by everything moving forward because it makes sense. Just like skills and MCPs were so successful, I think this will be as successful in the immediate and long-term future Now, it's been a while since we had a focused segment about OpenAI. Well, there's a few interesting pieces of news about OpenAI this week. First of all, OpenAI unlocks unlimited ChatGPT text chats for all users, including free users. So this is a major shift from the limits that we all had before. And again, the fact that it applies to free users is a little crazy to me. It will be very interesting to see how that works. There are still limits on uploading files and images and on using voice and generating images, but just text, pure text, there are no limits to how you can use, uh, ChatGPT Now, the free tier users will now default into the new ChatGPT 5.6 Luna model, which is going to replace 5.5, and it's a 62% reduction in factual errors compared to the model that it is replacing, GPT 5.5 is- instance But I will say something from my own personal perspective as a heavy, heavy user of Claude and still a decent user of ChatGPT, this is very attractive to me because Claude have changed sometime in the last two to three weeks the limits on my $200 max plan, and I'm hitting my weekly limits every single week, which is really, really annoying. And having the opportunity to do this unlimited is very attractive to me. So A, I hope that Claude will follow the same path, at least for the $200 a month, uh, users like I'm paying Claude right now, and if not, I have a feeling that more and more of my work will shift to OpenAI, which is exactly why they're doing this move. How exactly are they financing this? I'm not 100% sure, because I'm pretty sure that the new restrictions in Claude, the changes in how many tokens I actually get per week for my $200, I'm sure this comes not from them being cheap, but from them having capacity issues by allowing everybody to run on an unlimited environment. So it'll be very interesting to see how they respond to that. But staying on OpenAI, they just launched three new education plugins to help teachers and students learn So the plugins are K through 12 Educator, College Educator, and College Student, and they're available through ChatGPT Work and Codex, and their goal is to streamline academic and administrative tasks for educational institutions, and as I said, one of them for students as well Now, the college student plugins is aimed to drive engagement of college students with their course materials by generating personalized study guides, quizlets, flashcards, etc. I love this initiative. I have been delivering this kind of training to teachers for a while now as volunteer work and showing them how they can do the same and having a pre-made plugin that comes with the platform as long as you know that it exists and you know how to use it, is really, really important because I think the future of education is significantly more personalized, more engaging, and more fun for students, and hence they'll be able to learn significantly faster than they are today in just the general standardized classroom environment. So kudos to OpenAI for moving in that direction and the last piece of news about OpenAI is that they filed motion to dismiss Apple's trade secret lawsuit. Again, not surprising. We talked about the lawsuit. The lawsuit looks really, really bad. It talks about OpenAI stealing everything they're developing in hardware, uh, in a very aggressive way from Apple. And so OpenAI's core argument for dismissal in a 31-page motion to dismiss Is that Apple's complaint is baseless and pretextual And basically they're claiming that Apple is failing in the AI space and that they're trying to go after OpenAI because they're not able to do that if you wanna hear some of the specific quotes from Apple's lawsuit, you can go two weeks back and listen to the episode. It is very, very specific on things that they're claiming that OpenAI is doing, including aggressively hiring employees and telling them to bring stuff from Apple to them and doing things that are very far from acceptable in any perspective. But OpenAI is obviously claiming that is not the case, and they're saying, and I'm quoting, "No use, need, or desire for Apple's trade secret." Uh, how will this evolve? I don't know. Uh, the courts will decide. I'm sure that's not gonna happen in a day or two. But on the other hand, this may have a very significant impact on OpenAI's hardware initiatives, which will have an impact on their public offering. So we will follow this closely, and I will keep you posted But switching from OpenAI to the growing competition in the AI field, Meta just launches Muse Code Terminal Agent and Muse Spark 1.2 Muse Code is obviously Meta's effort to have a player in the burning hot, uh, agentic code development universe. It is an async background agents that run active throughout sessions in parallel to develop new capabilities, spawning for agents for specific individual tasks, reducing lacen-latency and reducing redundant information. So again, parallel work of multiple agents as we've seen in the most advanced tools right now I haven't seen specific reports on the capabilities of the model because it was just released, but we'll probably get those next week. As I mentioned, they also launched Muse Spark 1.2 which is the latest version of their really capable model. So am I saying that Meta is back in the race? I don't know yet, but they definitely have a much better chance in the race right now. The question is, are they too late in the race? I'm not sure. I think,- there are two competing arguments here. One is that a lot of the big organization and mid-size organizations so on already picked a horse and are running with it. But on the flip side, I think the pricing capabilities is going to be significant, and I think if AI will, and I think it will, become a commodity, meaning the vast majority of things can be resolved by most models, and I think we're practically already there, then companies may switch and jump ship if it's going to be relatively easy to do so, if there's gonna be a faster, easier, cheaper way to do things. So time will tell. Uh, the reality is in the big companies, the enterprise world, think about how many companies are with the same platforms for decades because it's hard to switch. I think the capture of territory that has happened in the last two years is gonna make it hard for Muse to be successful. But in the smaller companies, small to mid-sized businesses, it is definitely still for grabs, and it will be interesting to see whether or not Meta is actually successful in doing so. Staying on the competition, Alibaba just launched Qwen 3.8 Max, which is a two point four trillion dollar AI model that is challenging the top models from China and the Western Hemisphere models as well How good is this model? Well, one thing that they released is that the model has run autonomously doing a coding sprint for 16 days. This is out of this world. It is absolutely crazy During these 16 days, the model accumulated 267 commits, 127 pull requests, and 151 issues, uh, that it has created and built The system during this time continuously collected community requirements, dispatched issues, generated code, verified results, et cetera. Basically, a full development team, including product and engagement with users that is completely autonomous. This is not something I've seen anybody do so far. Now, is it good? How good is the code? How good is the product? I'm not 100% sure, but just the fact that they were able to do this is really impressive. It is showing you where the world is going, and it is showing you where the Chinese are right now in the race. So I think the gap, if it exists at all between the US and China, is very, very thin right now Switching from competition in the model space to competition between the companies, Microsoft just had their earnings call, and it was absolutely crazy good for Microsoft. So Microsoft Q2 earnings has driven a crazy rally. The shares has surged over 20%, uh, in the beginning and then up, uh, following the call. The main thing is not their AI success, but their monetization success, which is what actually counts in the stock market Azure Cloud is exploding. Azure revenue significantly surging 43%, uh, year over year, and its cloud annual revenue is surpassing $100 billion for the first time ever. They've been very disciplined compared to the other companies, still crazy spending on their capital expenditure. So They're actually lowering their twenty twenty-six CapEx guidance from a hundred and ninety billion to a hundred and seventy-five billion. Still a crazy amount, but it is moving in the right direction, and we talked about the other companies releasing, uh, their results and being punished for too aggressive capital investments. So the same thing is happening here. More capital discipline is driving better results for the stock market One of the interesting things that they shared is that the usage of Copilot has grown to thirty million users. Now, this blows my mind, first of all, because it's a really, really, really, really large number. The second is, I did a workshop this week for Copilot, which means I had to go back, and I don't use Copilot regularly. I only go back to it when I need to do workshops for it, and I have a license, and I work with it. First of all, understanding their licensing mechanism is on the verge of impossible. I spent a few hours on the phone with Microsoft experts to try to figure out how to get specific capabilities into my platform so I can demo them, and I failed, including, uh, the help with Microsoft to try to figure out how to do this. But second, Copilot's performance is very far behind OpenAI and even further behind Claude on the specific things that I was trying to create. So I was trying to do the same exact thing, uh, with Claude because that's what I'm used to, and then convert it into Copilot, and I failed miserably after multiple attempts, and I feel that I know how to use AI pretty effectively. And so Copilot's capabilities is not even close to the frontier right now, and yet they were able to grow it to thirty million users is absolutely mind-blowing. Uh, my hearts and thoughts are with those thirty million users that don't really know what AI can actually do for them, and they're using a system that is maybe aligned with something that was available a year ago on the other labs, not even that potentially. Uh, but from a financial perspective, in a distribution success for Microsoft, this is a huge success, and it shows that in many cases, having access to solid distribution is more important than the actual capabilities. The one thing that they mentioned that I found very interesting is that they are planning to deploy the capability for Copilot to be more of a router than an actual model, if you want, where it will, behind the scenes, potentially be model agnostic. It will allow you to switch to whatever model you want or potentially, and I'm assuming, choose for you. So going down the path of the router craze that we're seeing right now, just built into the Microsoft environment. If they do this, they will see, I believe, huge success because then they will not be stuck with what Copilot is doing as Copilot, which is a very reduced watered-down version of whatever it is they're running in the back end right now, to potentially being able to run the top models from the top labs, potentially open source models to runs on your data in your Azure cloud while you just talk to Copilot and be fully connected to everything Microsoft. If they pull this off, I think it will be gangbusters, and it will be a huge success, even more than they're seeing right now. Now staying on Microsoft and on a topic that we talked a lot about in the past couple of months, which is the reduction of usage of AI inside large organizations. So Microsoft is directing individual divisions to meet token consumption targets and face potential restrictions if they don't, basically saying, "No more token maxing. You each have a quota, and you gotta work with that quota instead of using as many tokens as you want." This is not new. Again, we've seen this across industries, across countries, across different sizes of companies. I'm seeing this with many of the companies I work with. It's a very serious challenge to keep the AI usage under a specific quota, and it is something that all organizations will have to learn how to deal with. The main thing is not even how many people are using it or how they're using it, is using advanced models for things that don't necessarily require advanced models, and these advanced models require sometimes 50x the tokens than the lower quality models that are still good enough to do the tasks in an effective way. Um, again, this is the huge success of these smart routers that we're seeing in the recent few months is because this is a real need in the business world Now a few-- switching to robotics, there are a few interesting robotics news this week that all connect to the fact that the robotics world is going to change dramatically, and I have a feeling it will happen in twenty twenty-seven or worst case scenario, twenty twenty-eight, where we're going to see a huge leap in the usage of robotics. So the first one is Reimagine Robotics, which is a new company founded by former Google DeepMind applied robotics leaders, uh, has emerged from stealth, and what they're developing is a robot that learns on the job, meaning instead of spending hours and days and months in training robots, they just learn as they work with humans next to them. You show them the task, they're watching, uh, you do it, the robot attempts it, you give it feedback, just like a human employee would learn, and it learns significantly faster and way, way, way cheaper than the way robots learn right now Another new robotics company was announced this week. It's called Exclaim Robotics, and it's a Swiss startup founded by a former Microsoft researcher called Helen Olienkova, and they're emerged from stealth with four point nine five million in pre-seed funding Their robots are specifically built to maintain and help operate data centers. So this is not a generalized robot, come and do any kind of job. They're built specifically for data centers. Why do I think this is an interesting piece of news? Because we're hearing more and more the argument that AI generates a lot of jobs, especially around data centers, that will require people to build them and maintain them. And now you see immediately a robotics company that's going to target, uh, that specific need. So will there be a huge spike in needs for humans building and maintaining data centers? Yes, in the short term. In the long term, the robots will do most of that work, or I don't know if most, but some of that work for sure Switching from that to Anthropic that are now going to build their own custom chips, just like all the other big labs. Anthropic is launching an in-house chip design team to enhance Claude efficiency at scale Now they're doing something interesting. They are planning a strategic initiative that aims to co-design hardware and AI models, enabling Claude to basically have chips that are custom-built for it. Think about how Apple builds its stuff. So Apple builds the software and the hardware it runs on. The same kind of thing for specifically Claude to be more efficiently running on those things. We heard just last week or two weeks ago, OpenAI announcing their first chip is ready, so the same kind of thing coming from Anthropic They are actively hiring for a custom silicon team And they're looking for people who are engineer experienced in shipping semiconductor designs, and they're offering pretty hefty salaries between 320,000 to $485,000 per year for these positions. Again, this is not a new move. It's just a new move from Anthropic. All the other labs are already doing it. So This is just Anthropic trying to align with what the other labs are doing right now, building their own custom chips Now, a really big piece of news when it comes to AI leadership came from Google this week CEO Sundar Pichai revealed that there are significant leadership changes in Google DeepMind, elevating Demis Hassabis to chair of Google DeepMind and chief scientist at Alphabet, uh, to focus on shaping the future of AGI, While Koray Kavukcuoglu assumes the role of senior vice president overseeing the daily operations They're saying that the restructuring comes as Google reports extraordinarily momentum across its AI stack, with Gemini reaching 950 million monthly active users and Gemma models surpassing 900 million downloads Now, we talked a lot about Google in the past few months and how they've disappeared from the race, and I told you what my thoughts were, that I was thinking that Demis doesn't care about the race. He said that multiple times. He cares about solving world problems. And I think this shift is in order to move him away from the day-to-day where he wasn't delivering or his department wasn't delivering. I don't know if that was on purpose or not, but his focus is very, very clearly on other things, and I think the move is exactly to do that. It allows Demis to focus on the things he believes are important for humanity, uh, with the future of this technology and allowing somebody else to actually run the day-to-day in order to potentially stay competitive in the AI race. Going back to the huge success that they're talking about with almost a billion active users for Gemini, again, not surprising, very similar to Microsoft. They have the distribution. If you have the distribution, you don't necessarily need the best models around in order to see financial success. So it will be very interesting to see what happens now in Google. I assume we will see them back on the map with some interesting models coming and integrations into their entire ecosystem. They've been focusing a lot on that integration and not necessarily on driving new models And I'll summarize this section with two quotes from Demis. He said, "It's critical that we collectively get the next steps right to ensure this all goes well for humanity, and we usher in an incredible new age of discovery and wonder." And at the same time, he said, "We have arrived at a pivotal moment of human history. I've been working towards AGI my whole life, and now, like many of you, I feel it is close at hand." So there you go. Uh, one of the most respectable and knowledgeable people in the AI space since its more or less modern inception is telling that AGI is basically here, and that he's going to be focusing on trying to get it right. And now the last thing I want to share with you, and there's other articles, uh, and you can go and read them in the newsletter. But the last article is a very interesting one. Hajen's co-founder Wayne Liang developed an AI sales agent powered by his digital avatar and a knowledge vault system to handle the prospect calls during his eight-week paternity leave. According to his X post this week, the agent conducted two thousand seven hundred and forty-one prospect conversations, converted a hundred and thirty-two into paying customers, and opened thirty-seven enterprise opportunities worth three million dollars in revenue while he was on leave So how does the agent work? It has three main components. One is a live avatar, which is the actual video avatar of the person. The second is Liveclaw Knowledge Vault, uh, which is a system that knows how to access the vault very effectively, so how to manage the data, and then the actual knowledge vault itself that maintains over fourteen hundred markdown files that teach the agent everything it needs to know in order to have the context across everything it needs to know Now going back and closing the full circle on how we started this episode, this did not come without any issues. So the agent has had three main failures that now have been redesigned and been resolved, but the agent reasoned pricing instead of retrieving it. So it quoted a non-existent, uh, annual plan at $4,800 that it completely made up. It leaked internal triage metadata to customers, and it promised meetings with teams who hadn't really been agreed to. So those have been exposed during its operation and has been fixed, but it is telling you, again, these tools are incredibly powerful in the hands of the people who know how to connect the dots together. Again, having the visual aspect, the brain aspect, and the data vault aspect figured out correctly can do incredible things, but it will also do unpredictable things that you can't anticipate, and you have to monitor very closely all the time in order to prevent it from causing damage. That's it for this week. Really interesting things are happening. The level of things that we're talking about and their impact on the world is rising every single week. I assume you've noticed that. I hope that you find my new way of telling that story by connecting the dots, uh, and trying to tell bigger stories versus just news articles. I hope you find this valuable. If you do, I would appreciate if you would drop me a note on LinkedIn and say, "Hey, uh, we've been listening to the new versions of the episodes. We really like it. We hate it." Uh, you tell me what you think. I would really love to hear your feedback. Uh, we'll be back on Tuesday with a really interesting, uh, how-to episode that is going to show you how to create visual content in an incredibly effective way that actually works versus looks like AI slop. Before I go, I want to remind you to take a look at the course. There's a link in the show notes if you're interested in taking the course in September. Uh, and also, I want to remind you that we have an amazing community that meets every Friday at 1:00 p.m. Eastern. We call it the AI Friday Hangouts. There's now over 30 people every single week, and it's growing all the time. These are people who are sharing their AI findings, sharing amazing golden nuggets on things that they've developed and how they went through different issues. They are asking questions and getting answers. Just an amazing, lively community of people who care about AI. So you can come and join us. There's a link for that in the show notes as well. That's it for today. I hope you learned a lot, and I hope you have an amazing rest of your weekend.