277 | AI Agents Take Over: Nvidia’s NEMO Claw, Meta’s “My Computer,” and OpenAI’s "Superapp”. Stripe & Visa’s new Machine-Payment Protocol, More investment in data centers than office buildings, and more AI news for the week ending on March 20, 2026

📢 Want to thrive in 2026?
Join Isar’s flagship AI Business Transformation course and save $100 with code LEVERAGINGAI100.
🔗 Sign up here and secure your competitive advantage for 2026.  https://multiplai.ai/ai-course/

Learn more about Advance Course (Master the Art of End-to-End AI Automation): https://multiplai.ai/advance-course/

What happens when AI stops assisting… and starts acting?

This week makes one thing painfully clear: AI is no longer a tool, it’s becoming the workforce. From Nvidia’s enterprise push to Stripe enabling AI-to-AI payments, the shift toward autonomous agents isn’t coming—it’s already here. 

If you’re leading a business, the question is no longer if you’ll adopt AI but whether you’ll adapt fast enough to stay relevant. The companies winning right now are not the ones with the most resources but the ones with the clearest thinking, best requirements, and fastest execution. 

The opportunity? Massive.
 The risk? Also massive.
 This episode breaks down exactly what’s changing—and what you need to do about it.

In this session, you’ll discover:

•  Why AI agents are becoming the dominant force across every major tech company 
•  Nvidia’s bold move to become a full-stack AI infrastructure provider
•  How AI-to-AI payments could unlock a multi-trillion-dollar economy 
•  Why OpenAI is declaring a “code red” and shifting to enterprise focus 
•  The surprising data showing 73% of businesses choosing Anthropic first
•  Why traditional tools like Excel and PowerPoint are quietly becoming obsolete
•  The real reason most AI implementations fail (hint: it’s not the tech) 
•  The growing gap between companies that use AI vs. truly integrate it
•  What global AI adoption trends reveal about opportunity vs. fear
•  Why 2026 is shaping up to be the year of the agent—and the year of disruption

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

Speaker 0:00

Hello and welcome to a Weekend News episode of the Leveraging AI Podcast, a podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business, and advance your career. This Isar Metis, your host, and we have a very interesting episode. Today, we're going to start with a deep dive topic that is going to cover the identification of everything. Basically, the world's focus around AI is all going towards agents, and we're gonna touch on multiple aspects on how this materialized in this past week. The second topic is going to be some strategic focus and changes in some of the leading companies in the world around this topic of agents and how the focus on the major labs is shifting. And the third main topic is going to be how humans and AI work is impacted as they are learning to do this together. And then we have a lot of rapid fire items, including lots and lots of really cool and very important announcements, releases, and features from most of the major labs. Some big infrastructure news, some regulation and government changes and so on. So we have a lot to cover. So let's get started. This week Nvidia hosted their GTC conference and like every time they're running their GTC conference, there's a long and detailed and very interesting keynote from Jensen Huang, their CEO. And this one was in different, the keynote was about two and a half hours long, and it covered multiple topics from robotics to investments in data centers to changes in the industry to the focus of Nvidia. And I'm not going to cover everything that Jensen covered. It's gonna be again, just what he covered was two and a half hours. But I do wanna focus on a few specific things. First of all, Jensen said that 2026 is going to be an, I'm quoting an inflection point for inference. And the reason he's saying that is the fact that they're seeing a huge growth in demand for the usage of ai. Which aligns very well with everything they want to do and aligns very well with their announcement of their new chip that integrates the technology of Grok. Grok with a Q, which is a company that is building chips that specialize in faster, more efficient inference that they have recently acquired. So the new chip includes the regular GPUs and combined with the grok technology in order to enable faster, more efficient inference built to one unified solution. And so Jensen was pushing very aggressively on the fact that the growth in demand for AI is just going to continue and accelerate even more than it is today. And obviously he has to say that because he has a 4.4 trillion market cap to protect when there were several discussions about potential AI bubble, if anything, what you will see in the beginning of this episode is that I have a very strong feeling that he's not just promoting his own company, which is obvious, but also there's some very real facts behind what he's saying. Jensen also talked about potentially building the first space-based data center. He called it Vera Rubin, space one. Now, he did not provide any timelines or any real details, but he's definitely planning going that direction. If you remember, we talked about this, that Elon Musk is talking about doing this potentially this year. The biggest difference between Jensen and Musk is that Elon also has the launch capability through SpaceX, as well as a lot of experience in running computers that are interlinked in space through starlink. So Elon is definitely far ahead than anybody else in this vision of putting data centers in space. I'm personally still skeptic about this whole concept from two perspectives, the scale that is required to make this significant. So if you think about the size of data centers today, and obviously maintenance where you cannot maintain it, well, it's in space. But again, if both Elon and Jensen, some of the most successful entrepreneurs in history are saying it is doable and it makes sense from a business perspective, I'm probably wrong. The keynote also included robotics discussions as well as a robotic version of Olaf which is the snowman from frozen. I must admit this looked more like a Hollywood prop than something that is actually providing value to the economy or anybody right now. But it's very obvious that Nvidia is also going to playing very aggressively in the robotics market. But the most interesting thing that Nvidia and specifically Jensen has focused on is NEMO Claw. So if you remember just a couple of weeks ago. Jensen called Open Claw, the most important software release probably ever. That is a very strong statement by somebody who have seen more or less everything in the computer space in his career. Now, beyond the fact he's very excited about it, they actually took very aggressive steps in that direction with the launch of Nemo Claw, which is a software stack that integrates open claw agent platform into the entire agentic environment of Nvidia into a more robust enterprise ready solution that enjoys the benefits of Open Claw, while hopefully avoiding the pitfalls of mostly the security risks that it represents. So the platform bundles together, nitron open source models, which existed before the Open Shell runtime, Nvidia Agent Toolkit and delivering if you want, the missing piece that allows open law to become an enterprise level tool while maintaining security, reliability and so on, which was definitely the biggest hurdle of Open Claw so far. So Pneumoccal addresses that by providing policy-based security and network limits and privacy controls across everything that it is connected to. It is preventing access to sensitive data, and it is preventing the escalation of privileges unexpectedly, which definitely happens. We're gonna talk about this later on in this episode, so it builds a solid enterprise level security and data safety infrastructure around the very powerful capabilities of an ever connected ever on agent solutions. Now Nvidia worked on Nemo Claw together with Peter Steinberger, who is the creator of Open Claw, which now works at OpenAI, and they jointly develop this solution that again, builds on the open claw ecosystem and over en wraps it with enterprise level readiness to allow it to connect to actual systems. There's two very interesting aspects of what I just said. One is nvidia has been pushing for a while into the software side of things, right, with their different tools for agent creations and so on, and I think this is their most aggressive move. So they are now making a very strong push into being a full stack AI solution provider and not just a hardware company. If they can push this aggressively into enterprises, they will become the core infrastructure from both hardware and software to run agents in the enterprise world. There's obviously a lot of competition in that field, but they're making very powerful moves into that direction. The second topic is something that I said several times on this podcast, and I'm gonna keep on saying A. So everybody who didn't listen before would listen now. And B, those of you who listen to it understand how critical this is. The world has shifted from the world where the most important thing for growth is resources to the world in which the most important things are great ideas, solid judgment, and very detailed requirements, documents and requirements, definition, capabilities. And the reason I'm saying it's related to this particular topic is the fact that a person like Peter Steinberger, who nobody heard of until a couple of months ago, that have built a solution in a weekend. Is now helping the largest company in the world from a market cap perspective and the fastest growing company in the world in the past few years to develop their future solution that they are pushing towards the enterprise world. Great ideas, great judgment, and solid requirements is what is pushing the world forward right now. Now, another interesting aspect about this product is that they have built it agnostic from a hardware and software perspective. So you do not have to run NEMO claw on NVIDIA's infrastructure. That being said, there are a lot of benefits in doing that because it has native connectivity to everything and it knows how to run on that hardware in the most effective way. But it is not a necessity. And in this push, Nvidia is allowing other people who are not using their rest of their infrastructure to still use this solution because they understand the power that it provides, staying on the identification of everything and the impact that Open Cloud had in the world. Meta just announced that the Manus platform, which they purchased late last year, is launching a core new feature called Manus, my computer, which is a desktop application that brings the general capabilities of the AI agents that were built by Manus into your own computer, connected to everything in your system in a very similar way to cloud code cloud cowork, and more importantly, open claw. So my computer is allowing AI agents to execute command line instructions, CLI. So very similar to the way these other tools work in a user local terminal, which allows it to read, analyze, edit, local files, launch and control applications, all without moving the data to a remote server, meaning the data stays on your computer and can run more securely. Previously, if you remember when Manus started and when they were purchased by Meta, it was purely a cloud-based generic agent. And now this new version allows you to operate very much like these other agentic tools. Now, the cool thing is that the, my computer feature integrates with the traditional Manus existing projects, agents and schedule tasks that were created before. So you can migrate things that you have developed in Manus from the web environment into your computer environment, and probably vice versa, which makes it a very powerful solution. Now to address the privacy concerns that obviously arise every time you let something run on your desktop and have access to all your system, every command requires an explicit approval from the user before execution. Those of you who have been running cloud code or cloud cowork, or any of the other tools, know exactly how that feels. So another strong contender that is now backed by another huge company. If you think about what we have just in the past few weeks in the desktop agent space, we came from having more or less nothing. I mean, yes, computer developers were using cloud code before that, but not on a broader sense of doing everything. And now we have more and more people using cloud code. We have cloud cowork from Anthropic that I've been using every single day for multiple hours. You have perplexity computer use, you have now Nvidia Nemo Claw, and now you have the Manus, my computer all competing in the local computer general agent space that is now burning hot. and all the companies, including OpenAI, which we're gonna talk a lot about in the next segment, are pushing more and more capabilities and features that will work more like open Claw while improving the safety and security aspect of it. So on the same lines, anthropic just launched what they called Cloud Code Channels, which is enabling developers to control anything that's happens in Cloud code through Discord, telegram, and any other messaging app. Mirroring one of the coolest features of Open Cloud that allows you to open your phone and just chat with your agent from anywhere, anytime, and have it always running and always on. So this is now available inside of the Anthropic Cloud code environment and in the Anthropic Cowork environment, which we're gonna talk about it in a minute. So this new channels feature runs on an MCP solution, which then allows you to connect to out of the box telegram, and telegram and Discord, but because it is an open source library that they have deployed on GitHub and anybody can get access to it, I'm sure that in the next few days we'll get connectors to Slack and WhatsApp and other messaging platforms because anybody can now develop it and provide it through the open source universe connecting to the closed source universe of anthropic. So this is actually a very cool way from Anthropic to increase their connectivity, to the rest of the world by open sourcing this MCP connector section that allows you to connect to cloud code from anywhere, any platform, any time. This is a direct shot in the direction of open Claw. And again, all the companies are moving in the same exact direction. In the same time, Anthropic also launched dispatch, which is allowing you to connect and control Claude Cowork sessions from the Claude Mobile app. So I already tested that. It's actually working amazingly well. It's a little slow and weird in the user interface, but it works perfectly fine. So you turn on dispatch on your desktop app, and then you go to dispatch on your mobile app, and you can then activate anything that you could have done from your computer you can now do from your mobile phone. Now, since I currently have 17 different projects that I'm working on that are either already in production or on the way to production on Claude Cowork, this is another superpower that I'm both excited about and really annoyed by. The reason I'm excited about it is because it allows me to continue working on each and every one of these projects wherever I am and whenever I want. The disadvantage is exactly the same thing. I have zero downtime because I can now access these agents and see what they're building and what the status is and how they're improving and what to do next from anywhere. So I personally have no time off because the temptation to go and check what is the status is very, very high. It's the ultimate fomo. Now, I had that FOMO before, but I couldn't do anything about it when I was away, and now I can, which is again, a blessing and a curse all bundled into one small and yet extremely powerful feature. Staying on the topic on how important the agentic world is. A very interesting announcement was made this week by Stripe and Paradigm and Visa. So Stripe, the payment Giant and venture firm Paradigm has officially launched what they call the MPP, which is the Machine Payment Protocol and the Tempo blockchain, which is an open standard designed to enable AI agents to conduct autonomous transactions across the internet. Now, visa has connected to this in order to extend MPP to facilitate card payments and not just to facilitate credit card payments. The idea is that AI agents will need to pay one another for access to services, computing, power data models, et cetera. McKinsey estimate that this part of the economy could reach three to $5 trillion by 2030. Just the intra agent commerce of exchanging value for money is going to be a huge part of the economy because agent will do a bigger and bigger share of what humans actually do right now. So MPP basically function like an OAuth for money. So it allowing AI agents to authorize a spending cap, once, meaning now the agent knows what it is allowed to spend, and then it can stream micropayments continuously to consume different services and consolidate thousands of micro transactions for everything that you're doing, like paying for tokens or tools and stuff like that. Including payments across all the traditional payment mechanisms such as Bitcoin and other virtual coins, stable coins and credit cards through Visa into one unified solution. Now, they already integrated the protocol into more than 100 services, including model providers, developer tools, compute platforms, and data vendors. So examples include right now Anthropic, DoorDash, MasterCard, Nubank, OpenAI, ramp Revolt, Shopify, and many others in support of Agent eCommerce. Now there were attempts to do similar things previously from Coinbase and CloudFlare, uh, in 2025, but it seems that Stripe and Paradigm, first of all, are significantly bigger players, and also the protocol that they created is much broader and multi-form and multi-platform, which I feel has a much higher chances of actually being successful. Visa specifically, in addition to being a part of this initiative, has also launched a Visa CLI, again, a command line interface. You're gonna hear that a lot in the next few months, which is an experimental command line tool allowing AI agents and developers to execute card payments from terminal without managing API keys and pre-funding accounts. So again, a whole shift in the way transactions in the world happen is happening right now in order to support the agentic universe. This is no longer a geeky software writing experiment. This is changing the way global commerce is happening, and what we're seeing right now is the infrastructure to enable that in the immediate and long future. It has been very clear to me since the beginning of 2026 that this is the year where agents are going to take over everything. We are just in March, so not even the end of Q1, and it seems that everything we knew about AI in 2025 seems like ancient history. The entire focus of all my businesses has shifted to agents. The entire focus of all the major labs have shifted to agents. The entire focus on the underlying models and infrastructure has shifted to support agents across the board, again, including the commerce side of it. If you are not focusing your attention in that direction, you must start making that shift because otherwise this year is going to catch you very much unprepared. Now I want to take this opportunity to mention three different courses that are coming up this coming Monday, March 23rd. We're launching the second cohort of the AI Business Automation course. It is the perfect middle point between basic chat usage and advanced agent implementation. It builds on basic knowledge on how to use chat and how to develop things like custom GPTs, and it's gonna show you how to integrate that into your entire company's tech stack. So how to connect it to your CRM, to your ERP, to your marketing solutions, to your email, and so on. And it's providing the infrastructure for a really wide range of sophisticated automations that can work across more or less every aspect of your business. Connecting to most of the tools you're already using, it is an extremely powerful functionality. And knowing how to do it will allow you to automate a lot of the things that are currently completely manual in your business, relieving your employees and or yourself from a lot of tedious work. As I mentioned, this course starts this Monday, March 23rd, so if you're listening to this podcast on either Saturday or Sunday, you can still join this course, and there's a link to that in the show notes. If you don't believe you ready it for this kind of course, and you need to first build your AI foundations, which I totally agree, you can join our AI Business Transformation course. The course will give you solid foundations across things like prompting and data analysis, image and video generation, and strategic thinking around AI implementation in a company. This course has been the flagship of Multiply My Company. I have been teaching it for over three years. Thousands of people have been through it and have completely transformed their businesses and career based on the foundations in AI usage that they built through this course. I continuously update the course for the latest and greatest. So despite the fact I'm teaching it, once a month and sometimes more, I always update to the latest things that happen. And the next cohort of this course is starting on April 13th, the third. Course is our Advanced Agentic implementation course. I'm in the final stages of putting all of it together, and it is most likely going to be the most important course you ever took in your life, and I'm serious about it. It teaches in eight to 10 hours, everything I've learned in three months of multiple hours a day, building multi-agent orchestration solutions for myself and for my clients. And it will give you everything you need in order to develop age agent solutions for everything in your business and in your personal life as well. If you want, if you're interested in this course, please reach out to me on LinkedIn or via email and let me know that you're interested. I've said that in several previous shows, and we already have a solid list of people who showed interest, and we are going to include all of these people in the early bird session, which will be significantly discounted compared to the regular course. I'm anticipating launching this course probably by the end of April or beginning of May. And so if you want to join it again, reach out to me. I will include you in the early bird option. And I have all these courses because I want to give you options depending on where you are in your journey. And it's perfectly fine to be a total beginner. It's totally fine being halfway and it's totally fine being more advanced. If you want, you can obviously combine and take the basic course and then take the more advanced course, however you wish, but you have links to all of this in the show notes, and I would gladly meet with you in the next few weeks in one of these courses. And now back to the episode. The next main topic we're going to talk about is the shift, the focus of the major companies into this new world where agents and enterprise are the main focus. So OpenAI, CEO of applications. Fiji cmo, just delivered an internal message to the staff this week declaring that the company is treating the competitive gap with tropic, with Anthropic as a code red situation, and redirecting all major resources towards a focused approach on coding tools and enterprise customers. Smo shared with employees in the all hands meeting that her and Sam Altman and Chief Research Officer Mark Chen, are actively identifying areas for deep prioritization and that staff are going to learn soon, which projects will be cut in the next coming weeks. The company is aggressively pivoting into focusing on two core pillars, advanced coding tools, including a new version of Codex app, and scalable enterprise AI solutions. CO explicitly characterized Anthropics recent enterprise advances as a wake up call for OpenAI. If you think about the very different approach of these two companies, OpenAI has been focused on trying to be everything for everyone, including the development of so. Video generation and image generation capabilities and the Atlas web browser and the new hardware device and e-commerce capabilities inside of Chachi pt. And this has created a lack of focus and a lot of internal confusion versus anthropic that has been focused on one thing, which is delivering a very solid coding platform. And then building on that the rad scaffolding that enables it to become an age agent solution for everyone in the shape of Claude Cowork. The other problem that this vast distribution of solutions inside of OpenAI means that every department has less compute to use in order to develop these solutions and train the models around it. And employees describe the compute resources shifting as unpredictable across teams and that it creates a lot of disorder and stress, not knowing exactly how much compute you are going to get in order to develop the solution that you need. CMOs said that the company must avoid distraction and concentrate on delivering productivity solutions, especially for business clients. In addition to the competition with Anthropic, one of the very strong signals that OpenAI has seen in the past few weeks is the huge ride of Codex adoption. So that is their coding solution that has grown to over 2 million users, more or less quadrupling since January. And OpenAI is trying to capitalize on that in several different ways. One, they're deploying actual software engineers into large corporates, helping them implement and adopt their Codex technology. And as I mentioned, focusing on that in order to develop solutions that will be aligned with that versus spreading across multiple different things. Now, despite this very clear pivot and focus on enterprise, SEMA also emphasized that open AI's 900 million. Chachi PT users remain a major strategic asset and the goal is to convert them into high compute, high productivity users rather than abandon them completely. That sounds like a lot of tongue in cheek for me. It is very obvious where the focus is going to be and it's not going to be on consumers that are just trying to chat with chat GPT. So major shift in OpenAI. If you remember end of last year, Sam Altman said, and I'm quoting betting on a series on startups, and he was talking about startups inside of open ai. And all of that is going away to a much more disciplined and enterprise focused strategy. As part of that strategy, there is a plan inside of OpenAI to merge its desktop solutions, so the Chachi PT application, the Codex Platform, and the Atlas browser into a single super app that will provide a user access to all those different capabilities under one framework. This obviously reduces the amount of resources inside of OpenAI that needs to address this and will also provide a more robust single user interface for users to actually work with. This sounds like a great solution to me. Again, I'm now currently working on my screen at any given moment on my large screen, I have Claude Cowork on the right and Claude Code on the left, and I'm working in both of them in tandem. And having them unified into a single environment could be a great idea and to related to the Atlas integration, anthropic Cowork already has access to the browser, and even it's not unified with their browser solution. It still has access to the browser and can execute on the browser, which has been extremely powerful tool for me in more or less everything that I'm doing. So the approach from OpenAI makes a lot of sense to integrate all these three universes into one application. Again, the idea is obviously to merge everything that you can do on a computer into one unified environment, including having access to files and applications, but also the ability to write code and connect to APIs as necessary. It makes any user a super user for anything they want to create. Going back to what I said before, great ideas, solid judgment, and good requirements, definitions. Now, we heard a lot in the past few weeks that Anthropic has been growing like crazy, way faster than OpenAI and that they've been catching up to their revenue numbers. But there was another parameter that was released this week that actually blew my mind to show how extreme this is. So ramp, which is a FinTech company that provides corporate card and expense management solutions have shared just recently that on their platforms, companies that are spending the first spend on AI solutions, 73% of that goes to anthropic. I'm gonna say that again, of companies who are starting to spend on AI using the RAM platform. So this is not the broader population, but they are serving over 50,000 businesses. So in those 73% of them spend their first money on AI on Anthropic. That is insane. It's three out of every four company that is up from 50% just in January. So the sense of urgency inside of OpenAI is highly justified because right now they're losing the battle for business consumers in a landslide, staying on the refocusing of companies around these new agentic solutions in microsoft has done a significant reshuffling within the copilot AI division. They have appointed Jacob Andrew to the new executive Vice President of copilot, who is tasked with unifying consumer and commercial copilot experiences and reporting directly to CEO Sachin Nadela. In parallel to that, Mustafa Suleman, the CEO of Microsoft AI will now exclusively focus on the company's super intelligence mission to develop their own frontier models and become independent of open AI and other providers such as Anthropic. So this new shift in focus is structured around four major pillars, co-pilot experience, co-pilot platform, Microsoft 365 apps, and how they're connected to AI and AI models. The development of the underlying models themselves so far co-pilot has been a complete failure compared to the potential of Microsoft with its size and distribution. So they currently, the app has 150 million monthly active users. That is nothing compared to the 900 million of OpenAI and 750 million of Gemini. But the even more disappointing numbers. Only 15 million out of Microsoft. 450 million business customers are currently utilizing the paced Microsoft 365 copilot service. Again, this is 3%, so a very bad adoption rate right now for Microsoft 365 copilot, and the recent changes are very much to change that approach. Before I continue to the third company and the major changes they are making, I wanna touch on this topic of Microsoft and Google as well when it comes to that in the latest AI Friday Hangouts, which happens every Friday, and any one of you is welcome to join. If you want to join the Hangouts. There's a link in the show notes. It's a community-based conversation that we're having every Friday and it's absolutely amazing. There are dozens of people from all around the world to just join and talk about anything from practical AI to concepts to where the world is going. And it's basically like an open mic for AI related current conversations. But we had a interesting conversation about where's the world of office applications going? There was a debate around this topic, and I will tell you what I believe is happening. I believe the power and the need for the office application, so the excel, the world, the PowerPoint, and it doesn't matter which universe you're in, is becoming less and less necessary. From my very personal experience, I would say that right now the vast majority of documents that I create, whether Excel files, word documents, PDFs, and PowerPoints are created by Claude Cowork and not by me. I then edit them and add like some finesse to them and minor changes, but the creation of the documents is not done by me and the vast majority of the work is not done by me. And to be fair, if Anthropics Artifact had even decent editing capabilities, I will not open Microsoft Office or Google Workspace applications at all because I'll be able to do all my work inside of the cowork space. Now, does this take everything away from Google and or Microsoft? The answer is absolutely not, because they still hold all the documents. So the data, whether you are on SharePoint or on Google Drive, is held in that environment, but the power of the office distribution and user interface is going away, and in my eyes, it is going away very, very fast, and I'm shocked. I'm completely surprised that both Microsoft and Google haven't yet developed a strong enough functionality inside their universe with AI in order to eliminate the need to go to Claude and or open AI or any other platforms. I think it is actually putting what they're selling right now at a very, very high risk because the functionality inside of OpenAI and Anthropic has grown tenfold. While the AI capabilities inside their workspace ecosystems has grown slowly and it's getting better, definitely on the Gemini and Google front, but it is far from being aligned with what's happening in the alternative. And it will be very interesting to see where this is going now, back to changes in focus and now to the Chinese side of the world. But it is the same kind of focus. Alibaba group is now going through a very significant shift in the focus of the company to becoming a token driven business model that is centered around delivering AI agents. So Alibaba is establishing a new specialized AI unit, separating them from the cloud computing arm and establishing what they call Alibaba Token Hub or a TH, which is a business group that will report directly to the C-E-O-A-D-W. And this new group consolidates key AI initiatives like IL Laboratories and the Quinn AI model unit and mass platform and Wang and the AI Innovation Business Unit. So multiple business units are now rolling into one. The company as a whole is pivoting towards what they're called a token omics, which is basically a token economics models which is driven from the understanding that autonomous AI agents which consume tokens at 40 to 60 x higher than traditional chatbots is becoming the core business of the company, and hence a huge revenue opportunity. In parallel to that, they are significantly increasing their pricing on its AI services and AI chips at around a 30% hike, which is very significant, but because the demand is there, they can actually do this. Now the overall shift towards AI focus is not new. Alibaba's headcount dropped 34% in the year of 2025. A lot of it driven by the sale of sun art retail group and the exit from the in time department store chain. And the push is trying to reverse their huge decline in revenue. Now this is coming in the wake of their Q4 results, which were, I will be very gentle, not great. So the company's profitability has been down 67%. Their stock has fell 6%. When they share their earnings, their headcount has dropped. But the good news is that they are divesting from things that are not their core business. So their overall headcount has dropped 34% in 2025, mostly by selling. The sun art retail group and exiting from the in time department store chain. So they're focusing a lot more on their AI deliveries. And again, now all under one umbrella, reporting directly to the CEO And the CEO Eddie Wu has disclosed that their five-year target for cloud and AI revenue should exceed $100 billion annually. A summary of this section, AI agents are taking over everything they're capable of doing, more or less. Everything we're doing in knowledge work right now, learning how to do this is becoming critical and nothing short of that. And everything in the world is focusing towards that from the creation of data centers to the focus of companies like Nvidia to the reshuffling and reshifting of strategies in the largest companies in the world. And you should do the same. Our third deep dive has to do with the relationship of humans and ai, and we're going to talk about two different findings. Both are very interesting. One is from Adi Pollack, who is the director of Advocacy and developer experience Engineering at Confluent, who has been working with multiple organizations on AI implementation, and she identified that most AI implementations fail because of the human side and not because of the technology not being mature or capable enough. Pollack identified three critical practices that successful enterprises implement, expanding AI literacy beyond engineering teams, establishing clear autonomy frameworks for AI systems, and creating cross-functional playbooks that. That codify how departments collaborate with ai. I'm aligned with her 100%. This is exactly what I've been focusing on in all the training that I'm doing with companies who hire me to help them in their AI journey. She's saying that the core problem is that initiatives fails when engineering teams build models and solutions that are not connected to what the actual needs are. And the data scientists create prototypes. Operations cannot maintain and that applications sit unused because the users were not involved in defining what useful actually means. Again, a disconnect between the people in the company and going back to the three things that I mentioned multiple times, defining the requirements correctly and the prioritizations correctly. She's also talking about the AI literacy gap and how the lack of understanding across multiple layers of the company, including decision making, is driving to making the wrong decisions, and I couldn't agree more. Knowing what AI can do, even if you're not doing it yourself, is extremely important as a manager to be able to make the right decisions right now. And what Pollack is suggesting is that organization that treat cultural transformation and workflow design as seriously as the technical implementation will dramatically outperform those who will focus solely on the technology side. And again, I agree 100%. and I will summarize with a quote from Pollak, the question isn't whether AI technology is sophisticated enough. It is whether the organization is ready to work with it. The other aspect that I want to talk about when it comes to humans working with AI is a recent survey that Anthropic just released, which is maybe the biggest of its kind so far. So Anthropic has interviewed over 80,000 cloud users across 159 countries and 70 different languages on their AI adoption and journey. The survey covers a lot of interesting aspects, and I recommend you read it again. The link to it is gonna be in the show notes, but I'm gonna touch on the key findings. The top aspirations for people to use AI is professional excellence at number one with 18.8% of answers followed by personal transformation at 13.7%, and life management at 13.5%. Time freedom at 11% with roughly one third seeking AI to elevate current burdens and another quarter wanting to do better, more fulfilling work than they're doing right now. Now the things people find that AI does well is productivity gains led by actual AI delivery is at 32% in number one, followed by cognitive partnership at 17.2%, learning at 9.9%, and technical accessibility at 8.7%. However, on the flip side, 18.9%, so almost one in every five people reported AI failed to deliver expectations due to inaccuracy or insufficient capabilities. I must say that right now, these cases, I would guess are more tied to not knowing how to use the technology well than anything else. I'm not saying AI can do anything. Absolutely not. Does it still have issues? Absolutely, yes. Other limitations for sure. I think 20% is way higher than what it needs to be. Another interesting aspect from that, the survey finds is the tension or if you want the dissonance between the good and the bad of ai, so it is creating benefits and harm at the same time. You're getting time savings on one hand, but you're getting an illusory productivity. On the other hand, you're learning on one hand, but you are losing your cognitive skills because you're not losing them. On the other hand, you're getting emotional support from ai, but you're developing dependency in this AI companion. You're getting decision making aid, but you are basing some of your decision on unreliable data. AI is generating economic empowerment, but also job displacement. So all of these are contradicting conceptually, but they're not because all of this is the outcome of ai. There's also a very big difference in the perception of AI across different areas of the world. As an example, Sub-Sahara Africa, central Asia and South Asia, show the highest AI adoption, 71 to 76% net positive sentiment in these areas. With the lowest concern about job displacement, only 15 to 18% people in these countries think that AI is gonna drive job displacement. On the other hand, north America, Australia, and Western Europe, register a lower sentiment of only 64 to 65% with the highest economic anxiety of 22 to 25%. My gut feeling tells me is that in less developed economies, AI provides a huge opportunity and less of a risk. And it's exactly the other way around. In more developed economies where people already have a job and they're afraid AI is gonna come for it, while people in less developed economies can now do things that they couldn't dream of before, and one of the quotes that they're sharing from one of the participants in the survey is saying, I'm in a tech disadvantaged country and I can't afford many failures with AI of reached professional level in cybersecurity, ux design, marketing, and project management simultaneously. So that tells you that in these countries, the opportunity to suddenly compete on the global stage is something that was not attainable previously or was very hard, and now anybody with internet access can do it. That's it for the deep dive today. Now we're gonna go into rapid fire, and we're gonna start with Cursor. Cursor, the AI development platform. They just launched Composer two, which is their in-house coding model, which provides a significant performance improvement over their previous model and also aligns and in some cases surpasses the latest models from open AI and Anthropic. I'm not gonna dive into the specific benchmark scores, but right now they are second to only GPT 5.4 on some of these benchmarks. Surpassing Andros Opus 4.6, which is the model that I've been using for most of the work that I'm doing, which is highly impressive. Cursor is pricing their new models very aggressively with the regular model at a dollar 50 per million input tokens and $7 50 cents for the output tokens and the faster model at 50 cents per million tokens and two and a half dollars for million output tokens. To compare this opens 4.6 is at $5 for million input and $25 for output. And GPT 5.4 is at $2 50 for input tokens and $15 for output tokens. So they're cheaper than Cha Chachi PT 5.4 and much cheaper than Opus 4.6. That being said, the only way to use it is to pay per token, and I'll give you my personal experience with this. I tried their new model. It is actually very impressive. It runs very well and it works very efficiently. and I've done a similar experiment with Rept as well. When I do this, when I spend a whole day of vibe coding in the background, as I'm doing several different things using Rept and or cursor in-house models, I'm spending 20 to $40 a day in token charges, despite the fact that tokens are quote unquote cheaper. At the same time, with my $200 a month plan in Anthropic, I'm running multiple agent solutions that are already running, consuming these tokens of the $200 while I'm developing new agents and new projects while I'm VAB coding on two or three different projects at the same time, all capped at $200 a month. so I think the biggest problem that cursor has right now and replicate the same way, is that yes, they're providing very powerful capabilities, but the anthropic plan is just way more economical and still provide very solid results. Staying on the topic of vibe, coding capabilities, Google just announced a very important step forward in their war for dominance. In the vibe coding universe, they have just upgraded AI Studio to include anti-gravity coding agent, together with a Firebase backend integration. And the idea here is, is in addition to the vibe coding, the platform understands what backend capability you need and it creates that for you automatically. So if you're now using their new solution, it detects when apps need databases or login systems or things like that. Provisions Cloud Firestone for storage and Firebase authentication for secure Google sign in and connects to productivity tools and other aspects of the Google universe, such as Google Maps and so on. This is obviously a very powerful capability that takes away a lot of the manual work that you have to do on the other vibe coding platforms, while simplifying it to the point you just need to know English, and even all the backend orchestration databases, APIs gets done for you autonomously by this new platform while being fully integrated into the Google environment. Another thing that Google has announced this week is a tool for vibe design. So they have completely Remanage Stitch, and they have evolved it into an AI native software design canvas. So it's an endless canvas that allows you to create, iterate, and collaborate on high fidelity UI designs. Straight from natural language. And to make it even more straightforward, they are using the Gemini Live, which is an incredible voice interface that you can just chat with your design environment in order to make changes or create designs from scratch. In a very similar approach to everything else that Google is generating, it's building on their incredible capabilities of Gemini image capabilities with nano banana and combining it with everything back in Google that allows you to create these designs in a way that are a very pleasing but also built correctly from a component perspective so you can use them and couple them with any other programming platform such as Cursor or obviously Anti-Gravity inside of Google. So a new way to create designs competing directly with Figma built into the Google universe. Now staying on the topics of new releases, OpenAI just released GPT 5.4 Mini and GPT 5.4 Nano, which are the smaller versions of GPT 5.4 which they just recently released. And like all these smaller models, they're always almost as good as the big models and significantly faster and cheaper and significantly better than the previous fast models. So GPT 5.4 mini is not different than that. It runs twice faster than GPT five mini, which was the previous mini version, while scoring significantly higher than it. As an example on the SWE Bench Pro, which is a software engineering task benchmark, the 5.4 mini is scoring 54.4% to compare it. The regular, not mini, the regular G PT 5.4 is at 57.7, so it's relatively close to it. GPT five, mini scored 45%. So 10 points below what the new mini version provides. Now the interesting thing, because this model is significantly more efficient using GPT 5.4 mini ENC codex only consumes 30% of the limits compared to the full GPT 5.4, meaning while it is not as capable, only close to it, you can run tasks 3.3 times longer by using the mini version versus using the full version before hitting your limits. So very similar approach to what we see from all the other labs together with sonnet and haiku and so on. The same kind of approach, a smaller model that is faster, significantly cheaper, and more efficient. That is almost as good as the leading model. And in parallel to this announcements, OpenAI released another thing for Codex, which is Codex Subagent features that is now available inside of their development environment, which allows the user to request the environment to spin up subagents, to do tasks in parallel. So subagents, as an example, can work in parallel on tasks like pull request reviews. So pull request is when you are merging your code into the general code and you need to verify it before you do that. And there's multiple steps in doing this, such as frontend, debugging, workflows where agents need to map and code and reproduce bugs and implement fixes and batch processing multiple files at the same time and so on. So in many cases, doing things in parallel could be significantly more efficient. And now you can do that inside of Codex simply by asking the agent to do it. In addition, developers can define their own custom agents in config files, which then can have specific instructions to be able to do specific tasks and even use specific models. Why does that matter? Because smaller, less complex tasks can be completed by 5.4 mini nano, as an example, which is significantly cheaper rather than spending your tokens on the more expensive model when it is not necessary. So this is a very powerful capability to delegate and save money at the same time while making the process faster because you're running it in parallel. The other huge benefit of subagents is the fact that they're not consuming the context window of the main chat. So you need to remember that all these reasoning models, which is more or less everything we're using right now, definitely when you're writing code, a huge amount of the tokens that takes space in your context window are used by the models to quote unquote think, right? So it's not the output that you see is what the model is doing in the backend to think about the problem, to analyze the problem, to read the code, and then to eventually come up with the output by using subagents. Every subagent has its own context window where it's doing all its thinking in that other context window and the main. Agent that runs the process just gets the output, which could be a single digits or up to 20% of the overall context that is being used to generate that output. So it is a very powerful capability, not just from the speed and saving in money, but also in the usage of context, window of the main process. Staying on the topic of context window, Claude Opus 4.6 and Sonnet 4.6 now have a 1 million tokens context window, which was announced previously, but now is available to the general public. To put things in perspective, if you don't know what 1 million tokens means, it means around 750,000 words. Which is a few really large books, or if you want a very big portion or your entire code base in a single chat, and it is a very significant jump from the previous context window limit of the previous model. That was 200,000 tokens, so it's five x the amount of data you can now put into Claude Opus and Claude Sonnet inside of a single chat. Anthropic did not change the max output limits of both models. The max output is the total length of a single output that the model can generate. So Opus 4.6, it's still at 128,000 tokens for max output and sonnet 4.6 is at 64,000, which is still a lot. And definitely enough I can tell you that as a heavy user of cloud code and cloud cowork, I'm seeing significantly less compacting of the conversations compared to the previous version, which is a huge benefit from my perspective and obviously from the perspective of anybody else who's using these platforms. Combine that with what we mentioned earlier with Anthropic channels. So now you can connect to the same project, the same conversation from anywhere with your phone and talk to it through Telegram or WhatsApp or whatever tells you that. The combination of this makes it a very, very powerful solution that can run for hours and you can still have access to what's going on and participate and give feedback anywhere, anytime. Another company that made a big splash this week is Minimax, the Chinese companies who introduced M 2.7, which is the first model that has used autonomous optimization cycles in order to get to the capabilities it has right now. We talked about this in depth last week when we talked about the loop that was released by Andre Carpathy. So a similar concept was used here. 2.7 autonomously executed over 100 iterations of its own optimization cycles, achieving a 30% performant improvement working on its own in order to make the model better. This model currently scores 56.22 on the SWE Pro benchmark, which is almost as good as Opus 4.6. It has also scored 1495 on the GDP Val benchmark, which OpenAI has came up with two compare models to a real world use cases. Placing it as the highest, most capable open source model on that benchmark, showing solid proficiency across tasks on Word, Excel, PowerPoint, and so on, including generating financial models, research reports, and more or less everything you need in the real world environment for a fraction of the cost of the Western models. Misra, the French company that we didn't hear a lot about in the recent months have released Lynn trial, which is a coding agent that is built on open source, lean programming language. It is not as powerful as the capabilities in these other tools that we just talked about, but it comes in significantly cheaper for doing the same tasks. So if you have tasks that are not as sophisticated, going to these open source cheaper models might be a very good idea. Another interesting release this week came from Microsoft. So Microsoft just launched a I Image two, which is the latest generation of their image generation tool, which currently ranks number three on the chatbot arena, leaderboard for image generation behind only Google with their nano banana and open ai. So this model is a homegrown model from Microsoft that now competes at the highest level of image generation, including highly photorealistic or any kind of style that you want. It handles complex scenes with superior spatial positioning and body proportions. It looks really, really solid. It knows how to generate texts accurately in multiple languages. But it comes with multiple limitations compared to the competition right now, first of all, there is a 32nd cool down between each generation to the next, which is really annoying, uh, for somebody who needs to generate multiple variations until you get exactly what you want. That will just take you a lot longer. You have a 15 image daily limit right now, and it's limited to a one-to-one resolution. So square, no landscape or portrait capabilities. And it comes with very aggressive content moderation that rejects lots of requests. It also lacks features that now becoming standards such as image to image editing in painting out painting, and reference image support. So these are now more or less given in most of the platforms, not having them is definitely not a good option if you want to use professional things, but it is the first really professional capability coming out from Microsoft in its homegrown model capabilities. This is now available in the limited deployment on the MAI playground, and it's going to be integrated into everything Microsoft soon. Switching gears to a different topic. This week has been the first week in US history where the investment in new data centers has surpassed the investment in new office buildings. So the value of US data centers currently under constructions has reached $45.1 billion, which is a 29% year over year increase. Which is now higher than the 43.5 billion for office building, which is a 13% decrease from last year overall, since the launch of Chachi PT in 20 in the end of November of 2022, data center constructions has surged 228%, while office constructions at the same time declined 38%. So right now, the US economy is investing more in data centers than it is in office buildings. And if that's not a sign of where the economy is going, I don't know what is. And now to some government developments across the US and the world, the Department of War or Department of Defense have submitted a 40 page filing against the anthropic lawsuits that has designated them a supply chain risk. So if you haven't been following what's been going on in the past few weeks, and we spent a lot of time reporting on this in the past few week and we've spent a lot of time reporting it, you can go back and hear more details in the past two episodes. But the Anthropic declined to align with the Department of Raw requirements to allow the government to use Anthropic models for generating autonomous weapons and for spying on Americans. And as a result in the attempt to bend them to their will, the government. Threatened that, if they will not allow them to do that, they will designate them a supply chain risk, which is something that was never done to a US company ever before. And they actually moved forward and designated them as such, which means government agencies are not allowed to use Anthropic tools and potentially many suppliers to the government are not allowed to use that technology as well. If they want to keep their government contracts Anthropic filed two lawsuits in California federal courts, and in DC district courts to appeal this designation. And now in a reply, the DOD has said that anthropic are unacceptable risk to national security. What they said is that Anthropic might attempt to disable the technology preemptively after. And what they blamed is the Constitution. Basically, the rules behind the anthropic tools that they're saying may not align with the Constitution of the United States. They also said that the fact that the model has its own constitution or its own internal red lines may put the lives of war fighters at risk. I don't know if the people who wrote this 40 page memo do not understand how AI works, or they're just trying to protect the government's standing in this particular situation. But what they said makes absolutely no sense. It makes absolutely no sense from two different reasons. Reason number one, this model has been seriously evaluated and was selected to be the only model that was connected into a highly secure defense environments so far. So no other model still right now is running on such secure environments. They will add OpenAI in the near future and potentially X as well. But as of right now, the only model that is allowed to run on secure environments is anthropics models. But the other reason, which makes even less sense to me is the fact that all other models has the same exact thing. They have an internal set of rules that tells the model how to operate and what to do and what not to do and what the red lines are. And it's not different in any other model than it is in the Anthropic model. So I don't know how this will evolve in court. I'm not a lawyer. I don't pretend to be or understand exactly how this is going to evolve. But if this is what they're going to stand on, they cannot use any AI models in similar environments because of the same exact justifications that they're using in order to push anthropic out Staying on governments. But this time to the UK government just reversed their proposals to permit AI companies to use copyrighted work to training their models with an opt-out mechanism for creators to be able to opt out of the training data. This happened after a very significant rejection and pushback from the creative sector. And I must admit this is one of the biggest problems of AI right now. So first of all, I'll give you two different opinions coming from two different people in the uk. Tom Cahill, the Chief executive of UK music, called The Backtrack of the Government, A Major victory for campaigners with artists like Sir Elton John and Dua Lipa, having publicly opposed the plan to allow this and to use the opt-out mechanism as the guiding lines. The reality is there is a very serious collusion between two different concepts here. On one hand, there is what the UK defined, and I'm quoting, the UK culture is a world leading national asset, but at the same time, the AI industry is growing 23% faster than the rest of the economy in the uk. As I mentioned, this is one of the most problematic issues with AI right now, is how can it use training data and its sources to generate all the things that it can generate. And as a creator and an AI expert, I understand both sides of the argument. The question that I keep asking myself is, what is the expected outcome in the long run that the artists are trying to achieve? And I don't have a good answer for that. I believe that in the long run, a very large portion of the creative we are going to consume will be AI generated. I think in the beginning, in the near future, humans will be willing to pay a premium for human generated creative. So whether it's images or videos or graphics or anything else, but after a short while, people won't care. People will want to listen to music. They enjoy listening to. They want to watch TV shows and movies that move them and entertain them and excites them. They don't care how it is created. I said that loud and clear when the strikes happened last summer in Hollywood from the script writers and the actors and while I'm empathetic to creators in their push to protect their creation, I think the long run it makes absolutely no difference. Now, while what I'm gonna say may sound really strange, I think sometime in the not too far future there's gonna be ai music that will be more popular than the Beatles that will be AI generated. I think it is inevitable, and I'm not saying this because I'm happy about it. I'm a huge fan of creators. I love The Beatles. I love following multiple bands. I love watching movies of specific artists and so on. but I do not see this evolving any other way. And we're gonna end with a warning on security, which again, is not new, but this is just another example. A meta agent recently caused a major security incident by autonomously posting flawed technical advice on an internal forum leading to the exposure of sensitive company and user data in an unauthorized ways to unauthorized employees for nearly two full hours. Now, the incident began when a software engineer used an in-house AI agent tool similar to Open Claw. If you want to analyze a technical question, the AI agents then posted the response in an internal discussion forum, offering advice without approval leading to another employee to act on the flawed guidance and inevitably trigger a critical data exposure. So the thing here that you need to understand is that some of these tools have access to more or less every company system. It is supposed to keep that data. It is supposed to keep the answers siloed for specific individuals versus having it available to anybody who has access to the agent. But the agents sometimes mis follow their guidance or don't understand the exposure that they're generating. This is a very significant risk that is available right now in any company that implements agents that are connected to multiple data sources. A recent ciso, chief AI Security Officer AI risk report found that 47% of CISOs observed AI agents exhibiting unintended or unauthorized behavior with only 5% of the CISOs confident they could contain the compromised agent. So half of CISOs have seen that behavior from agents and only 5% think they can prevent it in the future. Now beyond data exposure. Internally, there are other risks as well. As an example, AWS Amazon's Web Services experienced a 13 hour outage of its cost explorer tool in China in December of 2025, after an internal AI coding tool, kiro with elevated permissions, decided to delete and reactivate the environment because it's thought it's the right thing to do. So why am I sharing all of this with you? While everybody's so bullish, myself included, on using agents across everything that we're doing, there are risks that you need to take into account and try to mitigate and minimize as much as possible. That is it for today. I remind you of the different courses that are coming up, including the one that's coming up on Monday. You can find information about all of them in the show notes. If you are enjoying this podcast and you're finding it valuable, please share it with other people who can benefit from it. It will take you five seconds. Literally just click on the app right now, click on the share button, add people's names, and hit send. And these people will be grateful for you for helping them learn what's going on with ai. I will be grateful to you for sharing my work, and you'll feel good about yourself as participating in the global effort of making everybody more aware and hopefully better prepared to what is coming. We'll be back on Tuesday with another fascinating how-to episode that's gonna teach you how to use AI for a specific business use case. And for now, have an amazing rest of your weekend.