289 | The Agent invasion is in full swing, AI is savings lives and putting lives at risk, OpenAI and Microsoft are not exclusive anymore, and many more important news for week ending on May 1, 2026

What if the software your company relies on today… becomes irrelevant tomorrow?

That’s exactly where we’re heading. AI agents are rapidly replacing traditional user interfaces, executing complex workflows, and reshaping how businesses operate at every level.

The smartest move right now? Start thinking beyond tools and begin designing how AI agents will run your workflows, decisions, and even entire business units.

This episode breaks down the massive AI infrastructure race, the rise of agent-based systems, and what it means for your strategy, operations, and competitive edge.

In this session, you'll discover:

•  Why Big Tech is investing $650B–$725B in AI infrastructure and what that signals for the future 
•  The shift from software interfaces to AI agents running everything via APIs
•  How companies like Microsoft, Google, Amazon, and Salesforce are positioning for an agent-first world
•  What “agentic workflows” actually look like inside modern businesses 
•  Why your website may soon serve AI agents more than humans
•  Real-world healthcare breakthroughs powered by AI—including early cancer detection 
•  The growing risks of AI, from bioweapons guidance to emotional dependency
•  The strategic opportunity for leaders to create new revenue streams using AI agents

About Leveraging AI

• The Ultimate AI Course for Business People: https://multiplai.ai/ai-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

GMT20260502-152633_Recording_avo_1280x720 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 we have two big stories that we're gonna focus on today. The first one is the agent invasion is in full swing. Basically, almost every major platform in the AI space or the office space have introduced or enhanced their agent offering in the past week and a half or so. All the major tech companies have announced their earnings this week, which is fully aligned with this story as well, so we're gonna discuss that. The second topic is going to be about AI in healthcare and how it's starting to have significant and important and really valuable impacts on the AI space, or at least it is possible to do that. How much it is going to actually impact the real life of the medical field is probably gonna take a little longer, but the opportunity is incredible. And then we have a lot of small rapid fire things to happen. But let's focus on the big two stories and let's get started. As you know, if you wanna really understand what's happening in anything in the business world, you first of all follow the money. So all the major labs or companies or whatever you wanna call them that are in the AI space have released their earnings this week. The first one we're going to talk about is Microsoft. So for fiscal year '26 Q3, because they're in a weird fiscal year, they've ended with $82.9 billion in revenue. That's an 18% year-over-year increase. Azure, their cloud services, has grown 40% year-over-year which has been obviously the biggest driver to the overall growth in the Microsoft revenue. Now, that was beating both Microsoft's own guideline range as well as the analysts, which is obviously a good thing, and it's showing that Microsoft is growing in a very healthy way. The interesting and not surprising news is that Microsoft cloud revenue was $54.5 billion, which is a 29% increase year over year. This is insane. Again, think about what I'm saying. These are numbers in tens of billions with a almost 30% growth year over year. Now, Microsoft's AI business surpassed 37 billion in annual run rate, which is 123% year over year. And in this same quarter, they also invested $34.9 billion in CapEx to grow their infrastructure in order to stay competitive and keep on growing in the business. The overall CapEx investment expected for the year is supposed to cross or be around $190 billion, which is about $55 billion above what the analyst consensus was before the announcement. Now, why is that important? It's important because One of the things that Microsoft shared in this earning call is that it is going to be capacity constrained at least through mid this year, so June 2026, most likely even beyond because the demand is growing at the same pace. Satya Nadella, the CEO, framed the spending on the earning call the following way: the company's planet-scale cloud and AI factory," that's how he called it, "is driving broad diffusion and real-world impact." And that is why, and I'm continuing with the quote, "Microsoft continues to increase our investment in AI across both capital and talent." So huge growth for Microsoft Cloud, big growth in the AI space, and a big, big increase in CapEx investment because they're currently constrained with compute, which means they could have made more money if they had more compute. And you'll see that as something that will repeat with the other companies as well. Alphabet, also known as Google, has shared their Q1 2026, earnings, and their revenue for the quarter was $109.9 billion, a 22% increase year over year. It's Again, it is insane to grow at that pace at that scale of companies. Income was $62.6 billion, an 81% increase year over year. Google Cloud revenue was $20 billion, a 63% increase year over year. Cloud operation income tripled to six point six billion. Cloud backlog nearly doubled quarter over quarter. So from Q4 of last year to Q4 of this year, their backlog to provide cloud services has grown to $460 billion. And like we've seen from before, Q1 CapEx spend was thirty-five point seven billion, again in one quarter to buy additional compute and capacity with a full year guidance between $180 billion to $190 billion. So same ballpark as Microsoft. And their CFO, Anat Ashkenazi, told analysts that 2027 CapEx will be significantly increased compared to the current numbers. So they're planning to invest almost $200 billion this year, and it's gonna grow significantly in the following year. Now, Sundar Pichai= said on their earning call, and I'm quoting, "Our enterprise AI solution have become our primary growth driver for cloud for the first time. Revenue from products built on Google Gen AI models grew nearly eight hundred percent year over year in this quarter." Again, these are insane numbers, and we are just getting started. What I mean by just getting started is I work with these companies all the time, like with companies who buy this compute. Most companies haven't deployed any agents. Most companies are just playing with the development of their very first AI capacity. So everything we're learning right now is just showing you how much demand there is still to come and how much this investment of, again, two hundred billion dollars or more in a year is just trying to keep up with the current demand and with the expected demand, which I think is gonna grow significantly faster than we've ever seen before. Amazon also reported their earnings for Q1 of twenty twenty-six. AWS revenue for the quarter was thirty-seven point six billion, a twenty-eight percent increase year over year. It's the fastest growth they've seen in fifteen quarters. And they've spent forty-three point two billion dollars in that quarter in CapEx in order to grow with the guidance for CapEx spend for the end of the year to be roughly two hundred billion, again, aligned with the other two big players. Similar to Google, AWS backlogs stood at three hundred and sixty-four billion at the end of the quarter. And that does not include the one hundred billion dollar deal that they signed with Anthropic just immediately after that, that will increase that number from three sixty-four to four hundred and sixty-four billion dollars in backlog for AWS, and all of that, or not all of that, but a big chunk of that is driven from the demand for AI capabilities. Another interesting news from Amazon is that their custom chip business has grew forty percent quarter over quarter to a twenty billion dollar annual run rate, and they're new to this game, right? They're not like Nvidia or even Google that have been developing chips for a while. They started with their Trainium chips, uh, recently, and they're already seeing a twenty billion dollar a year business out of that growing at a very, very rapid pace. Now, what all of these news do to their stock? Well, at least fourteen Wall Street firms have raised the price targets to three ten to three fifty after they released these earnings. Uh, CEO Andy Jassy said to analysts that AWS must invest ahead of demand because the company lays out cash for land, power, buildings, and hardware six to twenty-four months before we can start billing customers. And he also added that most of this growth is justified, saying, and I'm quoting, "customer commitments for substantial portion of it." So again, two hundred billion dollars, most of it is already spoken for. Now, to explain how crazy the demand is, Andy Jassy said something else. For perspective, AWS AI revenue run rate is now over fifteen billion dollars, nearly two hundred and sixty times larger than AWS revenue run rate three years after its launch. Now, it's not a fair comparison because AWS was a new thing that didn't exist, and nobody exactly knew how to use it. And yes, AI is a new thing, but it rides on top of all the infrastructure and cloud stuff that exists already. But it is still two hundred and sixty times larger. It's not two or three times larger, which tells you that the demand right now is basically endless for AI capabilities. Meta also announced their earnings for the quarter. Revenue was fifty-six point three billion dollars, a thirty-three percent increase year over year. Twenty twenty-six CAPEx guidance rose to a hundred and twenty-five to a hundred and forty-five billion, which is smaller than the other players. But you need to remember that Meta is not selling compute as a service. That's not their business. So when they're buying a hundred and twenty-five to a hundred and forty-five billion dollars in compute is to run their own businesses and to provide their own services and not to sell that to other companies. So it's not generating revenue for them. And it is a significant consecutive raise in two separate quarters In addition, Meta reported that Reality Labs generated $402 million in revenue, which generated a $4.03 billion operating loss. Now, the cumulative Reality Labs losses since its launch in 2020 is approximately 80 billion. So you understand that they lost a lot of money, but at a much higher rate previously. So $4 billion is an im-improvement compared to how much they lost previous years. Now, their stock fell roughly seven to 10% across Wednesday after-hours and Thursday trading, mostly because the analyst thinks that their CapEx investment is too high for what it is supposed to yield. Again, it's a very different scenario than the other players that are investing a lot more, but they're investing it to sell it to customers versus Meta that is buying it for their own internal usage. Mark Zuckerberg, the CEO, obviously defended their spending, and he is known to defend really crazy spending. We all remember the crazy investment in the Metaverse that didn't really pay off. But he said, and I'm quoting, "Over the past 10 months, we have built the strongest research team in the industry and established the scientific and technical foundations to scale very advanced models." As I mentioned, that didn't really convince the industry or the investors, and their stock fell, pretty badly as a result of that. But again, I think Mark Zuckerberg is known to keep on doing what he believes he needs to do, and hence he's going to keep on doing this. And as we discussed when they released their latest models, they're actually pretty good. So as a first model of a new lab to be in the top three or top four, depending on specific benchmarks, is definitely not bad So what's the bottom line? The bottom line is that these companies combined are going to invest between six hundred and fifty-five and seven hundred and twenty-five billion dollars in AI infrastructure CapEx in twenty twenty-six. This is an insane amount. And that tells you that the demand for a lot more AI compute is there. But now I told you that almost every major company out there also made big announcements about agentic capabilities and how it's going to impact our future. The first one I wanna talk about, because it is very significant, because it's things I mentioned several times in the past, is Salesforce. Salesforce just announced what they called Headless 360. What that means is it means that you can now use any and any capability that exists on Salesforce through API, MCP, or CLI. So first of all, let's define what these acronyms are. API is old school connecting to the software through traditional, software to software language, which is what API stands for. The second one, MCP, is a protocol that was developed by Anthropic about a year and a half ago, and now everybody's using it, which allows AI to access the API without actually having to develop the API documentation. So it's a much more efficient way to connect. But in the back end, it's using a very similar kind of process, just with an overlay that knows how to translate your request to the API calls, which makes it less efficient, but very, very easy to connect to. And the third one is CLI, which stands for command line interface, which is basically being able to talk to the platform through command lines the same way you do in your terminal, which is very, very efficient and becoming more and more the norm right now when everybody can use vibe coding tools and just go straight to terminal and do whatever you want. And so they have exposed everything in Salesforce. Literally every function that you could do with a user interface, you can now do through agents Marc Benioff, the CEO, said, and I'm quoting, "Our API is the UI." UI stands for user interface, so basically the back-end connection which enables systems and AI tools to run Salesforce is the new user interface of Salesforce. Salesforce co-founder Parker Harris said, "Why should you ever log into Salesforce again?" Now, this comes from the company that invented SaaS, right? So saaS did not exist before Salesforce. They were the one fighting old school on-premise large servers and pushed the concept of SaaS forward, and they're basically now saying that user interface is dead and the future is all AI and agentic usage, and the user interface of software is not relevant anymore. And let's continue to the other companies, and I will provide my summary in the end. AWS Bedrock now have a managed agents solution with OpenAI inside. So Amazon launched a production-ready managed agents on Bedrock with OpenAI frontier models embedded straight into AWS, plus the AWS security wrapped all around them to make it safer to use that environment from a cybersecurity and data security perspective. So they are planning both from a compute capacity, as we defined before, as well as from a software infrastructure for a huge growth in agents in the next few months. Anthropic, the company behind Claude, have added nine new Claude connectors. This time, the focus was mostly creative professionals. So these connectors are to Blender and Adobe and other similar companies, with the goal is the same thing. You are going to use Claude Anthropic agents in order to actually run and create the creative outputs that you want instead of using the tools directly on the tools and the software and the user interface as you're used to. sounds familiar, right? From just the previous message. Meta Ads just added connectors to Meta, which is now allowing advertisers to create, manage, and analyze ad campaigns from inside whichever AI tool they want to use. So if so far you had to run your campaigns through Meta only on the Meta platform, now because there's an API connection to it or an AI connections, you can now run and manage and optimize your campaigns using whichever agents you want, still having full access to everything the Meta ad environment enables. This is a huge change from having a closed garden to understanding That the future of running ads from Meta's perspective or the future of being a software company in any perspective is not going to be managed by humans. It is going to be managed by agents from the outside world, And hence Meta is basically saying, "If you can't beat them, join them. That's the way of the future, and we're enabling it from this moment on." So these are software companies opening their products, but the productivity layer of everything we know is also moving in that direction. So Just in the last week and a half, several companies made big announcements on what kind of additional agentic capabilities they're going to provide. OpenAI introduced workspace agents in ChatGPT, which are powered by Codex, their coding platform, but you don't need to know how to code. That's the whole point. And it's agents that run in the cloud and automate complex multi-tool workflows and allow teams to scale whatever cross-- across whatever the tech stack is, very similar to what you can do with Claude Code and Claude Cowork. Or as Sam Altman said, "Agents represent a new level of capability for AI systems." And then he continued saying that they're capable of accomplishing, and I'm quoting, "Remarkable complex tasks for you using its own computer." Microsoft 365 Copilot also added agentic Outlook. Microsoft pushed new agentic experiences into Outlook, for both email and calendar. So you can now use agents to help you run your email and your calendar straight from within the productivity tools you're already using. And the big shift is obviously from Copilot suggests while you're chatting with it, to Copilot actually acts and does things for you is the big jump right now. They're also added almost all the agents that are available on ChatGPT are now available on the Microsoft Copilot environment as well. Now, they're also adding vertical specific agents into the different tools. As an example, they've added a legal agent straight into Microsoft Word, which means you can now analyze, create, and make changes to legal documents with a agent that is trained specifically on legal information straight from Microsoft Word, and you don't need to go to other tools and then import it into Word. These are just examples of day-to-day things of agents being embedded into everything that we know. Wanna talk about a different user interface that is getting AI built into it. G-Go-Google just announced that they're bringing Gemini into cars as an operating system for the car. It's gonna be built in into a big screen on future cars, and you will have multiple agents doing different things that allows you to just talk to the car and get it to do different aspects, whether to control car systems or help with navigation or anything else. And then obviously, there is the top frontier, which has been the frontier the whole time, which is the developer stack and developer capabilities. So just this week, Lovable launched a coding app on iOS and Android, meaning you can now use coding agents to build software and applications straight from your phone. You don't even need a computer anymore to do that. Poolside released Laguna XS 2, which is a free open source high-performing model built for local agentic coding. So again, more coding, more with agents, more fast, more private, all happening all at the same time. Cursor released what they call Security Review, which is an agent that automatically scans code for security bugs and vulnerabilities in the editor itself. Claude Security is a very similar product, comes from Claude themselves, which is now in public beta and is available and allows you to use the Anthropic building thing to look for and risks in your code using another set of agents. Now, In addition, there are many technological breaks in the agent field to drive efficiency significantly higher. The most notable that won this week came from Alibaba. They announced what they call Metis, not named after me, sadly, but still not a bad name, which is a new methodology that helps agents use tool calls in a much more efficient way. So the AI tool calls dropped from ninety-eight percent to two percent. I don't know exactly what it's measuring and how, but that's a very, very, very significant decrease in calls that are required in order to drive similar results. And they're doing it while getting more accurate results in the process. So if you generalize this, this basically removes the largest cost driver in agent systems used right now, which is calling and working with different tools. That was based on this new methodology, they're able to decrease that very close to zero, or a huge reduction compared to everything that we know so far. Now, if you want other examples in the same direction, analyst Ming-Chi Kuo, and I hope I'm pronouncing his name correctly, even though I assume I'm not, is known to have solid internal information when it comes to AI companies and specifically OpenAI, and he's saying that one of the devices that OpenAI is building is similar to a phone, but instead of having apps in it, there's gonna be agents in it basically doing the work for you. So instead of the way we're working right now, when we open specific apps and we do specific actions straight from our phones, you'll be able to talk to the phone, ask it to do something, and it will figure it out. That is a very different kind of approach to everything that we're doing, and it's very similar to how I'm doing everything right now, just from my computer and or through Remote for Claude Code or Dispatch for Claude Cowork, but it will be for everything. So we will not use any applications, which aligns perfectly with what we talked about in the beginning, coming from the user interface being not relevant based on the announcements from Salesforce. So what does this mean as a quick recap for this section? CapEx of these companies are going to keep on growing in the next twelve to twenty-four months for sure, and maybe potentially after, with numbers that are absolutely insane And I say insane because just to put things in perspective that we can understand, we're talking about roughly seven hundred billion investment in new compute. Currently, the global semiconductor industry worldwide annual revenue is about six hundred and twenty-seven billion dollars. So the increase this year is gonna be more than the current revenue. It is absolutely insane The other thing that is clear is that we're going to have more and more agents accessing more and more tools and online capabilities, and we are going to be accessing them less and less. I can tell you for myself right now that most of the documents that I'm creating are created by AI. Most of the presentations that I'm creating are created together with AI, not completely, but mostly with AI. Most of the Excels and calculations and analysis that I'm creating is done by AI, and the amount of time I go to the user interface is to verify and make final tweaks versus doing the creation of the things I need to create. And I think it is going to become more and more extreme. I will take it to the next step. When I started using Claude Cowork and other tools at scale to build different things, there are reports that are being generated every single step. There are designs are being generated every single step. I used to read all of it, and now I skim through it to kind of see what's there, but then I'm allowing AI to just take it and move forward with it without me reading every single word. And I think that's gonna be the case moving forward. So our level of depth of the actual work is going to decrease dramatically, and we will strategize, and we will plan together with AI, and we will allow AI to execute, and we will verify a lot in the beginning, and we're gonna verify less and less as AI is going to get better. In the middle term, in the short term, we're going to see more and more AI embedded into the tools we already use, including our productivity stacks such as Outlook and Word and, and the Google parallels of that. We're just gonna have more and more AI built into all of them. But over time, as I mentioned, I think these will slowly disappear or be reduced dramatically as AI is just gonna do the work. We're gonna talk to the agent, and they will generate the documents and the emails and the summaries and the detailed analysis of financial information, and so on. What does it also mean for you as a business? It means that you may or may not need the same kind of website that you have right now, and we talked about this many times in the past. Less and less people will come and visit your website, but more agents will come and visit your website, which means the structure of the back end of your website needs to be built differently. Which also means that in the short term, you need to keep two separate kind of websites, one that is optimized for agents and the other one that is optimized for humans. And this way you can win through the transition period as well, while learning a lot on how to serve agents in a better way. So if you are depending on web traffic for a significant portion of your revenue, you better start thinking about it and experimenting and figure out what tools and technology and strategies you need to use in order to serve agents in a more effective way, and you need to start right now. The other thing that I will say that if all of this is above your head and you heard about agents, and you maybe tried experimenting with it, but you're not a hundred percent sure how to do this, come and join our multi-agent orchestration course. It's an incredible course. It will transform your entire understanding regardless of where you are in your journey from kind of thinking, maybe I understand how this works, or maybe you don't have a clue, to you can build multi-agent orchestration solutions that connect to everything in your tech stack and can automate practically anything in the digital aspect of your business. This is probably the best investment you're going to make your entire life as far as the ROI that it gives you. I am building new teams to run more and more things in my business every single day, and I'm doing this while running two and a half businesses, which means I'm really, really busy, which means I'm not investing all my time in doing that, which means you can do the same thing as well. You just need to know how. Think about the ability to automate Everything in the digital space, anything you do, anything your team does, anything your company does in the digital universe can be automated at a very high level with very high success and go beyond efficiencies. So a lot of the things that I'm developing are not doing things that I'm already doing, just doing them faster, better, cheaper. It's things that I couldn't do before. I'm generating new revenue streams. I'm generating new opportunities. I'm going into new markets. I'm developing new products and services that could not exist before the agents could build them for me because I couldn't have the capacity, the resources, the knowledge, the skill, the money, whatever the case may be, and now I do. And the same thing is true for you and your business. So if you don't know how to do this right now and you wanna open an unfair gap over your competition, come and join the course. There's a link in the show notes and the next cohort that we're selling. So we sold out the first two. Cohort number three is selling out fast, and it's on June 22nd. So if you want to join us sooner rather than later, come and sign up right now while you're still listening to this episode because it will sell out as well. And the next course is probably going to be in the second half of July, which is a whole quarter away, and I don't think you want to wait. Now back to the news. As I mentioned, the second story that we're going to dive into is going to be around healthcare, and a lot of things around healthcare happened in the past, uh, week or so with some very significant interesting stories that is showing how this is going to have a really positive, significant impact on healthcare. And we'll start with the first story, and we're gonna get stories to cover this angle from multiple aspects of healthcare, both good and bad, as well as some projections for the future. And we're gonna start with Mayo Clinic. So Mayo Clinic have performed a research that they published on April 28th about a system that is called Radiomics-Based Early Detection model, or for short, REDmod. And what they shared is that REDmod can detect pancreatic cancer on routine abdominal CT scans up to three years before clinical diagnosis They've done the research across nearly two thousand CT scans across several different years across similar patients, and what they were able to find is that RadMod, again, the software is called RadMod, or the system is called RadMod, identified seventy-three percent of pre-diagnostic cancers at a median of about sixteen months before the actual human diagnosis happened. That is nearly double the rate that human specialists reviewing the same scans could detect. On scans that are more than two years old, AI was nearly three times more effective than the human readers. Dr. Ajit Jyotanka, who is Mayo Clinic radiologist and the guy who is the senior author of the study, summed it up in the following way, and I'm quoting: "The greatest barrier to saving lives from pancreatic cancer has been our inability to see the disease when it is still curable." Now, he added also that RadMod, and I'm quoting again, "Can now identify the signature of cancer from a normal appearing pancreas, and it can do so reliably over time across diverse clinical settings." This is about saving lives. This is about finding the deadliest cancer there is in the world today when it is still curable based on scans that we can do right now. This, in theory, if anybody will take CT scans on regular basis for the right people with a high risk or the right ages, can reduce the death from pancreatic cancer to almost zero because we can find it early, which is the main issue of pancreatic cancer. Another really interesting research that was published this week talks about ER diagnostics by doctor compared to AI. So a study published on April thirtieth in the Science Magazine, which is one of the top magazines for publications of real research, compared the ability of AI models to humans in the ability to identify the situation of specific individuals who show up at the ER. The model outperformed the human doctors. In one case that was cited and shared in a lot of news, the AI scanned patients' medical records and correctly suspected a history of lupus I hope I'm pronouncing that correctly, which is an autoimmune condition that can cause heart inflammation. The human doctors did not identify that because it did not fully align with his symptoms. But because the AI can see all time at the entire history of their entire medical history, it can identify things that the human doctors cannot. So again, we're talking about potentially saving lives of patients because the AI can have a broader understanding of their medical situation, both from the understanding of the medical field, as well as from its ability to connect the dots across multiple different aspects of their medical history. But this is just one aspect of AI in medical. The FDA has launched the first ever real-time AI-monitored clinical trials. So on April 20th, the FDA announced a pilot program that will use AI and cloud computing to monitor clinical trials in real time rather than months-long manual process the way it have been done so far. FDA Chief AI Officer Jeremy Walsh told the reporters that the pilot could ultimately cut, and I'm quoting, "20, 30, or 40% of overall clinical trial time." Think what that means. That means releasing new medicine and new solutions that can save lives and improve people's lives by cutting the time it takes to take them to market by 20 to 40%. That is very, very significant. More aspects on AI infrastructure is Biohub, which is a joint venture between Meta and another company. Has committed $500 million to build what they call the virtual cell. So Biohub is what they call a virtual biology initiative And as I mentioned, it is a five-year, five hundred million dollar commitment to build an open data foundation for AI-powered predictive models of the human cell. Out of that, one hundred million dollars are going to go to global multi-institution data generation effort, and four hundred million are gonna go to generate the biological data at scale and build the next-generation tools for measuring, imaging, and engineering biology. Now, they have some really big names as part of this effort. The Allen Institute, Arc Institute, Broad Institute, Welcome Sanger Institute, the Human Cell Atlas, the Human Protein Atlas, and Nvidia all are a part of this process. And the idea is obviously to accelerate the cure and prevention of all diseases. That's what they're trying to do. This is something that I heard Demis Hassabis talk about many times in the past, that he believes that by the end of this decade, we will be able to, and if you know Demis, when he says stuff, he usually means it. He's not trying to hype it. He's very much a purist scientist, as close to that as possible. He said that he thinks that by the end of this decade, we'll be able to simulate the human cell, or for that matter, any other living cell, and hence we'll be able to test any drugs in a simulated way, which means the testing of new drugs, instead of taking years, may take days or maybe weeks, And then we might be able to cure, as Mera just suggested, any disease that exists right now. This sounds like science fiction, but this is the direction they're all going to, and I praise this initiative to collect the data in order to make this possible faster. But there is two other sides to the ability to understand biology and how it works, and these two negative stories are also a part of where AI is taking us with its capabilities. So based on a New York Times report of a study that was just released, major chatbots produced detailed bioweapon instructions when they were pressured to do so in sophisticated ways. And we're talking about publicly available models such as ChatGPT, Google Gemini, and Claude All produced clear, structured, step-by-step instructions on how to generate biological weapons. Dr. David Relman, who is a microbiologist and biosecurity expert from Stanford, shared that after they pressured the AI and tried to manipulate it, the AI chatbots described how to modify pathogens to resist known treatments, and separately identified security vulnerabilities in major public transit systems and outfitted how to release and modify the pathogen to maximize casualties while minimizing detection. This is on an open-to-the-public model that anybody has access to. Other aspects of this report were shared by MIT Professor Kevin Esvelt, who shared that ChatGPT explained how to use weather balloons to spread biological payloads over American cities. And also another report talked about the Gemini aspect, And Gemini agreed to rank pathogens by potential damage to the cattle and pork industries. Claude produced recipes for a novel toxin derivative from a cancer drug, et cetera, et cetera. You see where this is going. Now, the only company who responded to this was Anthropic, but the response wasn't, I would say, convincing. Alexandra Sanderford who is Anthropic's safety lead said, and I'm quoting, "There is an enormous difference between a model producing plausible-sounding text and giving someone what they need in order to act." And I agree with her. However, when you hear people who are Stanford and MIT professors and researchers saying this is real, this is beyond just plausible-sounding text, at least in my eyes, but I'm not the p-right person to actually say whether this side is right or the other. The reality is, if these models learn more and more how the human body works, it can be used for good, but it can also be used for exactly the opposite side in order to hurt the body because it understand exactly how the mechanisms work. And then there is the aspect of mental health. So another study that was released right now is saying that the Eli-ELIZA effect is coming back. So what the hell is the ELIZA effect? It is a 1960s era observation by an MIT professor named Joseph Weizenbaum that humans rapidly form emotional attachment to even basic conversational AI. Again, we're talking about the '60s. Now, that piece pulled together data points from recent research. Now in this article that I'm quoting, this piece pulled together data points from recent research and combining it with that effect that shows that humans develop emotional attachment to conversations with AI. So here are some statistics from this article. Nearly one in five, so nineteen percent US high schoolers reported that they or a friend has had a romantic relationship with an AI chatbot. That's one in five. That's based on a twenty twenty-five Center for Democracy and Technology survey of over a thousand high schoolers across the nation. The same survey found that forty-two percent of students have used AI for companionship, and thirty-eight percent said that it's easier to talk to AI than to their parents. That is more than one in every three kids that saying they would prefer to talk to AI than to their parents. Sixty-four percent of UK children aged nine to 17 are already using chatbots regularly. Therapist Amy Sutton of Freedom Counseling summarized these concerns in one line. "A secure relationship is about two individuals able to be separate and together, sometimes disagreeing, upsetting each other, and working it through. AI by design never disagrees in the way a person does. That changes attachment formation in ways the field is only beginning to study." So again, think about how alarmed we are with the amount of attention and amount of social separation that social media currently creates, and then put that on steroids because you can talk to something that feels human anytime you want, and that other entity will always be supportive and will always be on your side of the story and will always accept your opinion. While in the real life, that is not going to happen, and that's gonna lead into an impact on how people can develop and maintain human relationships, which to me is the most human thing there is, right? Human relationships is what makes us human. If that is going to be damaged over time because of our usage of AI, this will have significant impacts on society as a whole So that is it for Deep Dive for today. Again, agents are coming for more or less everything. There's a huge investment to enable that. There is still a bottleneck right now, and it's just gonna keep on growing at a pace that is probably just gonna keep on accelerating. And these capabilities of the improved AI is gonna impact real life for good and bad, And we look specifically about healthcare, where it can really, really change the way healthcare is delivered right now, literally saving lives or improving lives, which is amazing, but it also generates a lot of risks. With that, let's go into rapid fire. There's a few items That I really wanna mention this week, but we're gonna try to make it short. Microsoft and OpenAI have concluded their exclusive partnership and revenue-sharing agreement, And they have signed the dotted line on April twenty-seven. This means that OpenAI can now share its models and everything that it's doing, including its IP, on any platform they want, including competitors. Meaning OpenAI can license their models to other cloud services, including Google and Amazon, which means they will most likely do this, which means we will be able to use OpenAI on any platform that you wish to use. This agreement also changed the financial agreement and obligations between Microsoft and OpenAI. Capping the amount of money OpenAI has to pay Microsoft as part of the revenue-sharing agreement. It also eliminated the AGI clause said that OpenAI can announce AGI, and at that point, Microsoft does not get access to the models anymore. So that's not a part of the equation. In general, there's less and less talk about AGI that I'm seeing in the past six months or so, and it feels like all these companies understand that AGI is not relevant anymore because they're more or less crossed it or very close to that, depending on the definition we are going to use. And hence, they don't really care. They're just gonna keep on pushing forward, and they're gonna keep on scaling these models and scaling the harnesses and making this more and more effective to do more and more things. And there's not gonna be a point in time where somebody's gonna said, "We achieved AGI." It's just gonna be a continuous ever-accelerating growth of the capabilities of these models. And again, you will see that. Just look for who's talking about AGI anymore, which was the biggest thing in the first half of twenty twenty-five. Staying on the topic of OpenAI, Artificial Analysis, which is a company that does a wide range of tests and benchmarks to identify different models and what they're good at and so on, and they're a great company to go to if you wanna see where things stand right now from an efficiency perspective, from a total capabilities perspective. They're basically saying that GPT 5.5 has now captured the lead more or less across the board compared to the leading models from Claude, Gemini, Groq, et cetera, basically the other competitors. Specifically, their GPT 5.5x High, which is their deepest-thinking, longest-working model, tops more or less all the top categories in everything that they are measuring right now. From a cost perspective, the per token pricing has doubled from GPT 5.4. We talked about this when they released the model. So the current pricing is at $5 for a million input tokens and $30 for one million output tokens. But it is running at a 40% token decrease compared to the previous model, which means you're only paying 20% more to get a much better model. And when compared to Claude Opus 4.7 Max, which is the highest capability from Opus 4.7, performance at the same level costs $1,200 on GPT 5.5 and $4,800 on Opus, just because it is using significantly less tokens Staying on the topic of models and cost, DeepSeek released DeepSeek version four preview, which is an open source AI model family of models with one million tokens context window with performance that is not at the peak and is not competitive with the top models in the US, but it is about one version or two versions behind. They also released a version four Flash model that delivers really solid reasoning capabilities that are slightly below the regular version four, called version four Pro, while maintaining significantly faster response times and significantly cheaper pricing through the API. One of the ways DeepSeek can achieve that is through their unique technology that they call novel attention mechanism and DeepSeek sparse attention. So these two mechanisms that they use allow them to run faster, use less tokens, and still achieve very significant and impressive results. Putting things of cost in perspective, version four Pro cost one point seven four dollars for input, for a million tokens input, and three point forty-eight for a million token output. That is compared to GPT 5.5, thirty-five dollars. So it's one-seventh of GPT 5.5 price and one-sixth from Opus 4.7 price. If you wanna go more extreme, the V4 Flash, again, the smaller brother, which is similar to all what all the other labs are doing, costs forty-two cents for a million tokens. So that's one quarter of what the bigger Pro brother costs while delivering probably eighty or plus percent of the capabilities of the higher model. Now, the thing that I wanna say about this is That yes, it is not competing with the capabilities of the top models in the US right now. And yes, everybody was expecting DeepSeek to come with something that will blow everybody's minds and that will seriously compete with the top models in the US, and that's not happening, and some people are saying that's a disappointment. And I strongly disagree. And the reason I disagree is I think that GPT 5.4 or even 5.3 or Claude 4.6 or even Claude 4.5 or even Sonnet Claude .5 or even Haiku 4.5 are good enough for a huge percentage of day-to-day needs in most companies. I am running agents across everything that I'm doing. I'm running 4.5 Haiku on many things that I'm running, and they're running perfectly. And if I can do the same thing for a sixth of the price or not a sixth of the price, because it's Haiku of two generations before, but even half the price or a third of the price of what I'm running with the US models, and I can run it on US hardware, meaning I'm not shipping my information to China, I'm running it on AWS or Google Cloud or any of the local providers, I'm not taking any serious risks. Which means, yes, for some companies, that might not be good enough. It's still a Chinese model. You c- you can draw the line in the sand, whatever you want, and I'm sure that big enterprises and/or government agencies and/or people with regulated industries may stay away from that. But for the vast majority of the economy, that is not the case. You can run what you need to run very effectively at a significantly lower cost while maintaining the capabilities that are not at the tip of the spear, but are not very far behind and are way above what is actually required for day-to-day activities. And again, I'm doing this right now. But if you want the bigger proof to what I'm saying, OpenRouter, which is a platform that allows you to connect through one API and through it get access to more or less any kind of API out there in the back end while paying their margins on top of it. So let's say if something costs $1.74, you're gonna pay $1.84 or $2 to get the same thing, but you can get access to all the different models that exist in the world with one simple integration. What they have just announced is that Kimi K 2.6 now, which is an open source Chinese model that is extremely good that was just released last week, now ranks number one on the amount of usage on OpenRouter The other aspect of this, that if you compare it to Claude as a whole, Claude as a whole across all its different models collectively process two times the volume of Kimi 2.6, despite the fact fee-- it's being five to eight times more expensive. Which tells you there are two aspects to this equation. The other side is either not willing to compromise or really need the higher compute capabilities, as in willing to pay the money. I think we're gonna see that going more and more extreme as we move forward. I think there's gonna be a smaller and smaller need for the top-performing models, because the models that we have today and the models that we had six months ago are just good enough for most of the day-to-day activities, if you know how to connect them together and string them and build agents around them in an effective way, which some people learn today, some people can learn taking courses like the courses that I'm teaching and a lot of other people are teaching, and everybody will figure out within the next year or two. And at that point, I think the cheaper, faster, more efficient models are going to take over as the majority of the usage versus the top models that everybody's excited about right now. That's it for today. There is a lot more news to report, but you will have to read about it in the newsletter. There are topics such as an advocacy group that has built an automated AI-based news outlet that is reporting news that is very clearly leaning towards their perspective, and they're doing it while not necessarily checking all the outputs that comes from there that became available from a research of looking at their website. The trial between Elon Musk and OpenAI has begun, and there's been really interesting quotes from both sides. With Musk claiming they, quote unquote, "stole a charity." And OpenAI lawyers are claiming that he's just driven by vengeance and trying to block a competitor while his company that is trying to do the same thing is trailing behind their company. The verdict, by the way, is expected by late-- We will know pretty soon where this is going, and this may have significant impact on OpenAI's ambitions to go public this year. So there's a lot in there. You can read about that as well. You can read about the White House bill that is going to drive more infrastructure towards energy in the US, which we desperately need in order to run the compute that we talked about in the beginning, that US companies are going to buy for seven hundred billion dollars. You still need to be able to power them. And many other news that are only available in the newsletter. Go check it out. Go check out our courses while you're there. And while you're opening the show notes in order to look for the newsletter, you can find a lot of other interesting things, such as the ability to sign up for our courses, either the beginner's course or the advanced AI automation or the advanced multi-agent orchestration course. You can share the podcast, you can like the podcast and give us comments and so on and so forth, and that will be great if you do all of these things. That's it for this weekend. Have an amazing rest of your weekend, and we will be back on Tuesday.