281 | AI is taking over the business world: 2 people company generates $1.8B, OpenAI paying $500 an hour to train AI on professional tasks. Oracle lays off 30k, and more important AI news for week of April 3, 2026

What happens when AI doesn’t just improve your business… but replaces how it fundamentally operates?

The answer is already here and it’s moving faster than most leaders are prepared for. From AI models being trained by thousands of domain experts to billion-dollar companies run by just one or two people, the signals are clear: this isn’t incremental change. It’s a complete rewrite of work, organizations, and the economy.

If you’re leading a business, the question isn’t if this will impact you—it’s whether you’ll adapt fast enough to stay relevant.

The leaders who win will be the ones who rethink everything—how teams are structured, how work gets done, and how value is created—in an AI-first world.

👉 Learn how to build AI agents and systems inside your business:
 https://multiplai.ai/agents-blueprints-course/

In this session, you'll discover:

•  Why OpenAI is paying experts up to $500/hour to train AI across 400+ professions—and what that unlocks 
•  How AI is evolving from a productivity tool into the core intelligence layer of organizations 
•  Why traditional company structures (and middle management) may disappear 
•  The rise of ultra-lean companies—like a 2-person business on track for $1.8B revenue 
•  The hidden economic risk: job loss, reduced spending, and system-wide disruption 
•  Why AI may de-skill the workforce—and what that means for future leaders 
•  The collapse of billable hours and how it reshapes entire industries like legal and consulting 
•  Real-world examples of layoffs, efficiency gains, and massive AI-driven cost shifts (Oracle, Meta) 
•  Why relying on a single AI system is dangerous—and how to build redundancy 
•  The shift toward AI agents replacing traditional software interfaces 
•  How companies like Shopify are cutting AI costs by 75x with smarter architectures 
•  The growing dependency on AI—and what happens when it goes down

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

GMT20260404-140037_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 this week seemed like a slower week. We didn't have any big releases and any big announcements from the big labs. There were a bunch of stuff happening in China, but that is less of an impact on the day-to-day and our schedule and impact on our work in the us. But that gives us the opportunity to dive into what does that mean for work and the future of work and how it impacts the us. And there's a lot of small examples that if you connect the dots correctly this week, it definitely paints a picture that we need to be aware of. So that is going to be the focus of our deep dive section today. And then in the rapid fire items, we still have a lot of interesting topics to cover. So let's get started. The first topic we're going to talk about is something that was rumored a while back, but now is getting much stronger signals that it's actually happening. And according to Business Insider, OpenAI is currently running a secretive initiative called stagecraft, in which they are paying three to 4,000 contract workers that are all domain experts in different aspects of the economy, and they're paying them up to $500 per hour to teach Cha GPT how to do specialized work and how to reason through over 400 distinct occupations. And those contractors come from more or less any field in the economy, including commercial pilots, emergency medicine physicians, artists, pharmacists, agriculture managers, airfield, operations specialists. Basically, each profession you can imagine has several different people who are experts in that field who are teaching AI how to reason and how to do specific actions in that particular field. Now if you do the quick math on that, let's go to the worst case scenario. Let's say that it's not three to 4,000. Let's say it's 4,000 people, and let's say they're not paying them up to $500 an hour, but they're paying everybody $500 an hour. And I know that's not the case, but let's play the game. If this is what's happening, that means they're spending $320 million per month If they are using these people at a full-time capacity, you translate that into an annual expense, that's almost $4 billion for the entire year. That sounds like a crazy amount of money. However, there is another side to that coin. One aspect, which is very, very simple. OpenAI just raised $122 billion, so spending $4 billion on training the model on the economy is not a big investment. The other aspect of this is that the expected cost of training the next generation of models, so whether that is going to be GPT six or Claude five, or whatever the case may be, is planning to be in the range of $10 billion, meaning training the next model that will be a generalist still without very specific domain knowledge, will cost them two and a half times the amount of money it is going to cost them to hire experts in every aspect of every industry and train the models on that domain expertise to be able to do these specific tasks at the highest levels. Now to be fair, the models today current models are good enough. In many cases, they're way more than good enough. What they're lacking is deep domain expertise, professional knowledge specific capabilities that are in the heads of people who have been doing the work of those professions for the last couple of decades. This is exactly what they're teaching them right now. So think about what happened to the world of coding. In coding. They took the best developers in the world because they are employees of the company and they helped train the models. And that got us to the point that AI currently writes code better than probably 95% of the global code writers in the world, and definitely significantly faster. What they were missing is the capability to do the same thing in every other profession because they don't have employees in those other professions. So now they're hiring people to train the AI to do more or less everything in the same level, and quality and depth and understanding and reasoning as it can write code right now. Now, if you had a choice between training the next incredible generalized model and spend $10 billion of that or use. Less money to actually train the existing models on really advanced capabilities across every domain. The choice is very, very clear. Now, to be fair, they have enough money and deep enough pockets, again, $122 billion they just raised. They will do both. But adding the capability of all these different expertise into the model makes a much bigger difference from an economic perspective than training another better generalized model. What does this mean to us, to the economy, to the world we live in? It means that the next version of the models will have domain expertise across many professions and across multiple industries. And it will be able to do more or less. Everything we know how to do today in front of a computer. It will know how to do at the level it can write code right now. Now this will have profound implications on job loss as well as the overall economy, and the reason I think it will have profound implications on the overall economy is while you think, oh, now we'll be able to do a lot more with a lot less and grow significantly faster and so on, which is true, but the problem is the other side of the equation. If we're going to have high levels of unemployment, there's no money in the system. People are not spending money. If people are not spending money, then the economy comes to a halt. I shared multiple times on this podcast that the peak unemployment in the recession in 2008 was 14%, and it touched that and went back down to single digits in the great recession. The worst economic time in the history of the us, the unemployment got to 20%. This is very easy going. This is easily going to happen with what's coming with AI next and. This time it is going to be unemployment of people who are making six figures every single year and not 30 to $40,000 a year. Every one of these people have a much higher impact on the economy because they have significantly more cash available to spend on different things. This can evolve very, very fast because it's very obvious again, right now that this is what the labs are focusing on. Real work, going after it, doing it with ai, faster, better, cheaper than most humans, which means we will need less humans. Now, if you think this is theoretical, I want to share with you a very interesting blog post that was written together with Jack Dorsey this week. So Jack Dorsey the founder of Twitter now X, but he now is the CEO of block. And Block is now fundamentally rethinking the actual organization of how companies work. So this blog post touches upon the span of control that is the standard across apparently not just current companies, but history. So he's going back 2000 years ago to the Roman Empire, and he's talking about that every soldier that in the Roman Empire Army that took Europe and the Middle East by a storm, every commander had eight soldiers he commanded. Now he's saying that through modern corporations, there is still a relatively rigid ratio of how many people report to each manager. Apparently that number is much lower than you tend to think, and it's again, usually around eight people. That means that every organization needs multiple layers of middle management, which creates a lot of waste and a lot of friction in delivering information from the people who make the decisions to the people who actually do the work. It's a endless game of phone tag if you want, across the management levels. Now, there are organizations who try to break that. There's a lot of examples like Spotify and Zappos and Valve and companies who try to build much flatter structures, but the average corporation in the world is still built with multiple layers of middle management because of that ratio that is roughly consistent through history. Now what Block is doing right now is trying to break this entire structure, and if you want, flip the model on its head. If you think about how most companies work. People are in the middle because the intelligence of the company comes from the people, and hence, you need the hierarchy and you need to route the intelligence through other people to other people. However, in a system in which the intelligence comes from a system, from an ai, people can be in the edges just doing the work, meaning people are not in the center of the company, the intelligence is in the center of the company, and people are in the edges doing the actions the system cannot do. People can deal with aspects of the world that the AI cannot perceive. People can make ethical decisions, people can handle high stake moments, but everything else, the core of the business can run with AI without the involvement of humans, which can dramatically reduce the dependency on middle management. Or like Jack Dorsey wrote in this blog post, he said, most companies are focused on AI as a productivity enhancer. Few are focused on the potential of AI to change how we work together. Now, you heard me say this many times in the podcast. The way we work today, the way we operate today in the world, as individuals, as groups, as societies will seize to exist within about a decade, and for many, many companies much faster. Now, I don't think anybody knows what it will look like on the other side because it is still very, very vague and nobody knows how it will evolve, and it is evolving very fast. But one thing it's obvious is it is not going to work like it is working today. Now, companies who are currently willing to experiment and innovate and challenge more or less everything, including how a company works, how it's built and so on, will come out ahead. Does that mean you have to take bigger risks? Yes, a hundred percent. In order to experiment, you are taking risks. However, in my 20 years of senior business leadership, I have never seen anything with that level of disruption opportunity. And it'll challenge more or less everything that we know in the status quo, which means that there is a huge opportunity for people and companies to come up with new concepts that were not possible before that could be extremely successful right now. And as I mentioned, it will transform more or less everything we know from business to economy, society, psychology, everything will be different after AI will take over more and more aspects because it will not require a lot of things that are required today. A great example on how things are already different, and the people who are willing to bet on this can win big or huge if you want, comes from a story that went completely viral on x in the past few days, and that is the story of MedV. So if you remember. Sam Altman in early 2024 said that he has a bet with several other leading people in the Silicon Valley on when we're going to see the first billion dollar company that will have only one employee. Well, we are there right now, so Matthew Gallagher, who is 41 built MedV with $20,000 of investment and a bunch of AI tools to try to create a telehealth business that is selling weight loss drugs. MedV we open its business. In September of 2024, entered into an already crowded market. This business generated $401 million in revenue in 2025, which was its first full calendar year, and net profits came in at 16.2%, which are $65 million. This is according to numbers reviewed by the Financial Times. now Med V, is currently on track to generate $1.8 billion in 2026. Now, Gallagher is currently not working alone. He hired one more person, his brother, to work with him in the company. So a company of two people are on track to generate 1.8 billion in 2026. Now what Gallagher developed is really smart and really simple. There's actually two different infrastructure platform that are called care validate and open loop health that provide the licensed physicians, prescription processing, pharmacy, fulfillment, shipping, logistics, and regulatory compliance. While MedV, the software company itself owns the customer relationship branding website, paid media, checkout, customer service, so he doesn't deal with anything that requires regulatory compliance. He leaves that to the professionals and then he just takes care of bringing clients into that environment and he makes a cut in the middle. Gallagher built the entire system with ai. He used ChatGPT, Claude and GR to write the code that powers the platform, mid journey and runway to generate the ad creative. 11 lab supplied voices for tools and customer communication. And custom AI agents, connects all these systems together, and a chat bot handles inbound service inquiries. Now, in the beginning he had issues where the chatbot made up drug prices, and Gallagher honored these quotes until he was able to solve these hallucinations, and now the system works more or less flawlessly. So what does that tell you? It tells you that on really large organizations, they can completely reimagine how the company is structured, how many employees they need, which roles they're going to have, and it provides an incredible entrepreneurial opportunity to build really successful businesses just built on AI across the board, from marketing sales to the product orchestration. Everything is built with AI and the companies who are in business today who will not realize this are going to be at a significant risk and individuals who are not going to understand how to navigate the new requirements that workforces will have for employees will stay behind or will be unemployeD. I'm gonna take a few seconds to connect that to the course that we are launching. I have been sharing with you that I've been building incredible teams of AI agents that are doing more and more things in my two and a half businesses, and I've built a course to teach exactly everything that I'm doing. I'm literally showing everything behind the curtain, including how to build every component of this. It's four sessions starting from knowing nothing about what skills and agents and Claude Cowork or Claude Code looks like, all the way to a running production level. Ready single process with multiple agents and tools, connecting everything together. At the end of session four, I'm literally teaching everything that I've learned in hours and hours every single day. In the last three and a half months, last week I announced that I'm opening an early bird registration with the hope of getting 15 people and launching the early bird session. We now have 24 people signed up. Now since I do wanna run the first session as a small session, I will keep of writing of running a 15 people session only. I will just run two of them in parallel, meaning there are six spots left in the early bird session and I will not open another one. I'm just planning to run these two in parallel and then we'll go to regular courses which are going to be significantly more expensive. So if you are interested in learning how to build multi-agent orchestration solutions, that can dramatically change how you approach more or less anything you do in front of a computer, either for yourself or for your entire business. Go to the show notes, find the link, and sign up for your seat in the early bird session. If you are in a leadership position in a company, and you want your employees and your team to know how to do this, and I already have two companies who committed to bringing me over to teach them how to do that, then reach out to me either on LinkedIn or through my email. I will gladly explain to you exactly how the training looks like and how you can completely transform how your organization runs. But now back to the news. The other big impact that AI has when it comes to job loss is in the big companies that need to invest in ai. The biggest example we have received yet happened this week when Oracle has laid off something between 20,000 and 30,000 employees. If it's 30,000, it's about 18% of their global workforce, and they've done all of it with one motion on the morning of March 31st. This is most likely their largest layoffs in the company's history. Now, the reason they're doing this is not because they got incredible efficiencies from ai. It's just because otherwise they might go bankrupt. So Oracle has committed to $156 billion in capital spending. In order to build the data centers that they need in order to provide what they already have committed to their clients. Now the company has raised 45 to 50 billion in debt and equity financing in 2026 alone, just in this first quarter of the year. Now, right now, multiple US banks reportedly are raising their lending cost or withdrawing from certain data center projects, and that puts Oracle in a very problematic situation. Now, the paradox of all of this is Oracle posted a 95% jump in net income last quarter. But their remaining performance obligations are $523 billion, which is a 433% jump year over year, which means that while they're making significantly more money than they made last year, their balance sheet just cannot sustain the capital investment that they need to make in order to continue what they are supposed to continue. The only way to cut a lot of expenses in a relatively short time is to let a lot of people go, and this is exactly what Oracle did, and they did it, let's say, in the not most human friendly way possible, and I'm being very gentle with what I'm saying right now at 6:00 AM All those employees that got laid off got an email saying that they are losing their position starting immediately, and then they were cut off the company systems immediately after that. Now in their 20 26, 10 QSEC filing. Oracle has disclosed $2.1 billion in a restructuring plan, which basically goes to severance pay and things like that for those people, and probably beyond. But really extreme measures are taken by Oracle, and we're going to see that they're not the only company in this kind of situation. That is seeing, on one hand an incredible growth and incredible improvements on one side of the financials. But on the other hand, their CapEx obligations are so big that is actually hurting their business and forcing them to do things that are extreme. So the next example is meta. Meta just reported something that I think is absolutely incredible. Their average revenue per employee, which is a parameter that helps measure the efficiency of the company has increased by 85% in the past three years. Currently they have reached two point $26 million per employee, which is about 10 x by the way, the average SaaS company, so let's say, very high level of efficiency. Now, this happened through two different things. One is using more and more AI to improve the operations and the other cutting the workforce. So you have less employees, more AI generating more revenue. A big jump in the revenue came from their say, AI driven ad platform, which is how they make all their money. So now people are getting better results from their ads because AI is running it. AI is optimizing it, they're pouring more money into it, and that's how they're making more money with less employees. However, META'S expected capital expenditures are soared at least 60% in 2026 compared to 2025, and the company projects their 2026 CapEx expenditures to be between 115 billion to 135 billion. That's nearly double the 72.2 billion that they spent in 2025 and 4.6 times what they spent in 2023. So despite revenue growth, meta is expected. Free cash flows to plummet. By 83% year over year. So think about what I'm saying. This is a company that's growing that's generating more revenue with less people, and yet the cash flow is gonna go down to less than 20% of what it was last year. Now, wall Street and investors obviously do not like this level of investment and this level of risk and this level of burden on the balance sheet and Meta's stock has fallen 15% just from the beginning of this year as a result of this crazy spending. And will it bring back revenue or not? Is still to be seen. If you think about it from the type of company they are, the company is going a structural shift from being an asset light software business to a capital intensive one, because they're going to own all these data centers. That is obviously a big burden on operating margins and net profits, where their operating margins has decreased to 41% in Q4 from 48% the year before. So for every dollar they're bringing in, they're keeping 20% less than the year before that. Now, obviously Meta and Oracle are unique because they're two of the very few companies in the world that is now throwing this kind of investment into the future of AI dominance across the globe. And I understand what they're doing that and I think they don't have much of a choice for meta specifically, this comes after a crazy investment into the Metaverse that did not pay off, which puts them in a slightly worse situation compared to other businesses. But the next topic that I wanna talk about as part of this deep dive in the broader look on how AI impacts our universe, come from a research that was done by MIT Media Lab. And what they found is that when you start being dependent on AI automation, it actually causes de-skilling of employees from the work they actually knew how to do. So when workers delegate routine tasks like data formatting, proposal, drafting, proofreading, all those kind of things to ai, they lose the knowledge and the skill in order to do this effectively on their own. Now that by itself doesn't maybe sound like a big problem. Well, AI will do it instead of humans. So humans doesn't, don't need that. But the bigger problem is what does that mean for the longer term? What that means for the longer term is historically learning these more mundane tasks like formatting databases and reconciling conflicting data and drafting memos served as the training ground where professionals learned the concepts of their industry, of their profession, and learned how to create quality, precision, structured outputs that is useful as the output itself. But it's also really useful to train them to understand how their industry, how their profession works. If they don't have that, they don't develop any of these skills there's no creation of more advanced judgment, which is a necessity in order to become managers, leaders, and people who can do more things in that particular field. So the bottom line is that what the research found is that people who rely heavily on AI to complete unfamiliar tasks, they fail to build the underlying conceptual understanding needed to then supervise, troubleshoot, improve outcome, and develop judgment on these specific topics. So the controlled studies found that learners who delegated work to AI performed worse on deeper conceptual measures than those who engaged directly in the simpler tasks. What this means is that over time we will have a less professional, less capable global workforce because in the early stages of their career, they will not be doing the mundane work that will teach them how their business, how their specific professional actually operates. This obviously raises a lot of questions on how will we make sure that we maintain the ability to learn these basic things in order to be able to make bigger, deeper, more judgment related decisions? How will people develop the intuition that we all have in our profession, on what's right, what's wrong, what's gonna work, what's not gonna work? All of that is going to be eroded dramatically, and that's a very big question that I don't think anybody has an answer to. Now, if you think that people are better than that and they're going to continue learning and doing the work manually when AI can do this, here's a great example that will prove to you this is completely wrong. So a research from HEC Paris reports that courts worldwide have issued over 1200 sanctions against lawyers for errors generated by AI tools that they have submitted to courts as accurate information. Out of those 1,200 800 are in US courts. Now, some of them are relatively small, so my PCEO, Mike Linden's lawyers were fined $3,000 each for filing briefs containing fictitious AI generated citations in 2025. Now this got a lot of press and this became really known case in the legal industry, and yet a lot of other cases continue to happen and the courts are raising the fees, and yet it keeps on happening. So a federal court in Oregon set a new record last month by ordering an attorney to pay $109,700 in sanctions and costs for AI generated filing errors. Now, some courts now require lawyers to label AI generated content with details, but this will obviously become useless very, very quickly as AI is becoming a deeper and deep and part of everything in the legal world. Now, under current professional conduct rules, lawyers remain fully responsible for the accuracy of what they submit to the courts regardless of how they generated their briefs. To be fair, this should be the case for all things in all professions. You are in charge of what you are submitting and what you're generating and what you're publishing. And if AI helps you, you need to verify this information. So what does that prove? It proves that humans don't like tedious work. If they can find ways not to do the tedious work, they will use that way. Meaning if AI tools will allow us to do more and more without us having to do the work, we will go down that path, which will lead to what we talked about in the previous article to less and less professional, skilled and capable workforce as this phenomenon grows. And I don't think anybody has a solution for this right now. Now since we're talking about lawyers, I will add one more thing that I see as a big breaking point in the economy. Right now. Lawyers are just one great example, but there's other examples like consulting companies and so on depend on billable hours. The concept of billable hours is going to collapse very, very soon. It is going to collapse because it is going to be clear to anyone that what used to take 160 hours of paralegal and legal work now can take 20 hours maybe, which means nobody will be willing to pay 160 hours of paralegal work, which means you will have to fire or significantly cut the number of paralegals that you currently have. Meaning you will have to fire or dramatically cut the total number of paralegals and lawyers that you currently have, because otherwise you will run out of business because you have to pay the salaries of X number of paralegals. But now only 10 of them actually are required to do the work. It doesn't make any sense. But if you're not charging 160 hours of paralegal work for a specific project, how do you pay the bills for the really fancy building you have in downtown? You cannot, which means a lot of things are going to change. And the same is true for anyone who is doing hourly based work. But that leads to an even more pressure to go and use AI because if you're not billing for hours, it means you're billing per task or per outcome. If you're billing per task and per outcome, you want to do this in the shortest amount of time possible so you can do more of the thing, so you can make more money. So if previously, the way you made money is by taking your time and doing things slower and going through them. Now you wanna do things as quickly as possible, which means you're gonna go more to ai, which means you're gonna have less of your own skills, both for yourself and for employees of your company. And this vicious cycle just goes on and on. I do not know what is the solution for that, but this raises a very, very big flag on how we keep people trained on a professional perspective, at least in the transition period, right? So let's say in the long run, AI will do everything. I don't know if that's where it's going, but let's say that's where it's going. What's gonna happen in between is going to be really, really different than what we know today. And it raises a lot of questions that nobody has answers to. We'll now go into our rapid fire items, but the first item connects directly to the last item from the deep dive, and that is that deep seek had a 12 hour outage that impacted 355 million users. Many of them found that they cannot work or cannot work effectively without deep seek. So the popular AI solution from China had a very serious outage starting on the evening of March 29th and ending in the morning of March 30th. There was a complete outage for about seven hours and 13 minutes plus some limited outage After that, as they started fixing the problem, and as I mentioned, that disrupted the work of 355 million users, many of them went to Chinese social media to bitch about the situation, and there were many, many complaints, but also many people that were sharing that they did not understand their level of dependency on DeepSeek models. A great example is a quote from a user called 8 8 8 who said, only after deep seek went down did I realize I no longer knew how to work without it. That's really profound, and this is exactly the direction that we're going towards. People will not know how to complete their daily tasks when AI is down. What does that mean? Well, it means, first of all, that we have to figure out how to train people for that. But it also means, and I talked about this many times in this podcast, that you have to have redundancies, meaning you cannot fully depend on one model, especially on things on your critical path. If you have things that are currently running through the API, you need an a PL rollback function that will allow it to run on another AI platform, not your main one. This means you have to test both of them all the time to make sure they actually support it and use your better slash cheaper whatever your solution is. Version that you're using regularly, but allow the system to roll over to a different provider, preferably in a different continent, in a different data center and so on that can continue running your operation effectively. While the main one is down, we're gonna switch gears to how AI might be impacting how we work. Salesforce just unveiled a very significant overhaul of Slack, introducing more than 30 new AI powered features built into Slack bot, which is more or less turning everything you need to be able to do in the Salesforce environment to be available through a Slack communication interface, through a chat interface versus through the Salesforce user interface. So one of those capabilities is the capability to create or use existing reusable AI skills, allowing users to define specific multi-step tasks for Slack bot that then it can do on its own from that moment forward. This could be complex things such as creating a budget by putting information from multiple data sources, either on Slack or beyond Slack in Salesforce or even other sources. The Slack bot can now also connect through MCP to other tools in the Salesforce environment, including Salesforce Agents Force AI development platform, meaning you can use Slack bot to create new agents in the Salesforce environment that can then be used by Slack bot to do more things. This is very similar to how I'm using Claude, where I am using Claude to define the requirements for new agents that then Claude can use to do more things and the flywheel keeps going faster and faster. Now the new agent can now operate outside the Slack application and it can monitor desktop activities like conversations and calendars, and offer actionable suggestions and drafts and anything you need beyond just the Slack environment. Obviously, it can do a lot more inside the Slack environment as well, such as creating summaries of meetings, capturing decisions and action items, distributing it to the right people, and so on and so forth, including from third party sources such as Zoom and Google Meet. Now early adoption is showing very significant efficiency improvements. Again, these are numbers that are not confirmed by any third parties. These are published by Salesforce themselves, but they are stating that they have employees reporting savings up to 90 minutes per day by using this tool versus working manually as they did before. And this is driving something that we talked about multiple times on this podcast, but this time it's coming from Salesforce co-founder Parker Harris, who said that the traditional Salesforce interface might become backend system, meaning you will not need a user interface for your. Platform. You are just going to use agents and bots in order to do the things you need to do and provide you all the information you need without needing to actually go and use the user interface. I agree with him a hundred percent. I think our access to more or less, everything we are going to do is going to be through agents and these agents will know how to do the operations on the platform as well as provide us the right feedback as needed for us to make decisions. A great example from my personal universe, as I shared with you several times in the past, I also have a software company called Data Breeze. What Data Breeze has, it has a bunch of AI agents that knows how to do invoice data entry and vouching and reconciliation of invoices against purchase orders completely autonomously. Meaning instead of going to your ERP or your accounting system and going through 17 steps, you're just looking at what the agents are doing. They're giving you a plan. I'm going to make these changes to these pos based on this information. Is that correct or not? And then you just say yes, and everything just happens. Meaning instead of the humans having to do lots of tedious work, that takes a lot of time and potentially generate errors, you're just confirming that what the agents are suggesting is correct and the actual operation happens, which means you don't actually need the user interface of the original ERP. That is usually not the most user friendly platform in the world. The same thing is going to happen across the board on more or less, everything we do. Now connecting two of the recent points together, the learning how to use agents effectively and knowing how to test other AI solutions for the same topic comes from a really interesting story about Shopify. Shopify just realized a 75 x reduction in AI inference costs for merchant data extraction while doubling the quality of the output. And they did this by replacing GPT five, single large prom baseline with multi-agent framework from Quinn three. So Alibaba Quin three is a model from Chinese, giant, Alibaba, and Quin three Max thinking has demonstrated really high performance when it comes to running multi-agent orchestrations and multi-step processes that agents are running that are parallel to models such as GPT 5.2, cloud Opus 4.5, and Gemini three Pro across multiple benchmarks. Now in addition to the fact that they did this for the initial process, shopify is now also integrating AI into multiple merchant facing solutions, including ai, cystic dynamic pricing to optimize revenue and customer value, chatbots for customer experience, improving conversion rates by optimizing different things on the storefront and so on, all while leveraging now much cheaper models. Now, this is a great example of two critical aspects that anybody who wants to develop agents has to learn very quickly. It is a part of what I'm teaching in my course. One is developing small, customized agents that do something very, very specific. Works much better than developing large agents who try to do many things at once. This is providing higher quality because the agents are focusing on something specific. Cost because it needs to run significantly less tokens because it just does something very specific and it may not need the rest of it in order to run one aspect of the operation versus another. So building small specialized agents and then billion orchestrator to manage them works better than one big agent that tries to do everything all at once. The other aspect is what we talked about before is you need more than one model connected to your API. So you can compare the capabilities. So previously we talked about redundancy, but this is just to check cost and efficiency, so you can connect your process to more than one API allow it to run on both platforms at the same time and then decide on the one you wanna pick. And the one you wanna pick is not the one that is the fancier or better model from a benchmark perspective. It is the model that gives you. Consistent results in the cheapest possible way, right? So if this tool is maybe not the fanciest, it doesn't score as high, less people like it, whatever the case may be, but it does the specific process that it needed to do, it does it consistently and it costs less than the other fancier, flashier model go without model. Again, keep the other one as a fallback as needed. Now, to be fair, it is really, really easy to do, more or less everything that I'm building, including the application that I'm using right now in order to do the news summary and reporting and organization of all that data. Every single week runs more than one model. In this particular case, I'm actually running three models. So I'm running some of it with Claude and some of it with grok. But I'm testing other models Every time a new model comes out to see can I save money and still get the same quality of results, and I suggest to you to do the same if you are developing such solutions. Now, since we mentioned Alibaba, we're going to switch now to many new releases that happened this week. Many of them are in China. Alibaba Quinn team just launched Quinn 3.5 Omni, which is an advanced multi-model AI that can natively process text, images, audio, and video across 36 languages. Now based on their testing, it's achieving 215 state-of-the-art results across various benchmarks in audio, audio, visual understanding, speech recognition, translation, conversational tests, and so on. They're claiming it outperforms in testing 11 Labs, GPT, audio and Minimax in multimodal voice stability across 20 languages and surpasses Gemini 3.1 Pro in general, audio understanding reasoning translation tasks, and that it's matching the audio visual comprehension of the leading models. They're obviously doing this at cheaper pricing than at least the Western contenders. The model was released in three different levels, plus flash and light, all of them with 256,000 tokens context, window. And they're claiming that they now achieve the best video analysis capability that is A better in quality and B, much faster. So they have compared their video analysis to GPT 5.4, and they were able to get comparable results in a single minute versus nine minutes for the same process on GPT 5.4. In general, I find GPT 5.4 to be a really slow model. I think it's over reasoning right now and sometimes just takes really, really long time to do stuff that the previous model did. Okay. Really quickly. And because now they make it a little more complex to get to the older models in their dropdown menu. It is really annoying because stuff that I used to be able to do in 30 seconds now takes five minutes sometimes for the same level of results because the process wasn't that sophisticated to begin with. Now the other big announcement that Alibaba made this week is that they are about to learn to launch a new service by. That they're launching a new service that is going to be providing AI agents as 24 7 autonomous digital workforce for the millions of merchants on its Taobao and Tm, OL platforms. So these agents that will become available immediately will automate core tasks on the platform, such as customer service and voucher distribution, real-time product, price adjustments, and so on. Again, work that is currently done by employees of companies that are using Alibaba's platform. Now previous AI tools that are provided on these two platforms have already been adopted by 5 million merchants on the platform leading to an estimated a hundred billion yuan in cost savings to these companies. But when it says cost savings, it means less employees that are doing the work, which is again, another signal that as you develop more specialized agents and more specialized capabilities, you need less people, because now it solves very, very specific issues that happening in specific industries. In this case, on an e-commerce. Now Alibaba shared that they are now moving from an open source world as they did so far, to more of a closed sourced API type of solutions that is going to run on the Alibaba Cloud infrastructure. And they're anticipating these services to generate a hundred billion dollars in cloud and AI revenue in the next five years. And to drive adoption. They're planning to inject 1 trillion tokens to power these new AI services and get more and more of their merchants to use their AI solutions in order to drive more shopping. And then everybody wins other than the people that used to work and do these roles in these companies. Going back to what I said in the beginning and in many other shows, if a lot of people lose their jobs, there's less people who have money, less people have money, they cannot buy new things on Alibaba. Alibaba starts losing money, and I don't see a good workaround for this kind of scenario, not just for Alibaba, but for everything in our economy right now. Ing, who is in charge of merchant platforms on tbo and tbo and Tmall summarized it in a really solid way. In the next one to two years, we expect standard operating model of e-commerce to evolve into a collaboration between human and digital employees. So again, this is not the next decade. He's talking about the next one to two years. And a lot of it is because of friction in adoption, not because the technology is not there. And when he's saying collaboration between humans and digital employees, it means digital employees that are now doing very little are going to be doing more and more and more, which, which means you need less and less human employees to do the same exact work that is happening right now. Another Chinese company that made an announcement this week is Byan. So they officially launched C Dance two, which was actually announced in February 12th of this year. But as a quick reminder, C Dance two is an incredible video image and text generation platform. It is currently the most advanced video generation tool in the world. It is surpassing video in multiple tests and runway and all the leading other labs, and it can take inputs from multiple sources at the same time. So it can take text, images, video, and audio as inputs and create a single output at a cinematic level that is not parallel right now. It also generates the video and the audio in the same run rather than using after processing. Like it's happening on all the other platforms right now. So. Again, this model was released a month and a half ago and is now formally released across all different platforms. There are also a few interesting releases in the open source world. Google DeepMind just introduced Gemma four, which is their fourth generation of their open source models, and they're claiming it is the best intelligence per parameter efficiency in the world right now. And they're releasing it just like the previous Gemma models in different sizes to allow it to run all the way from professional GPUs down to individual small mobile devices. From a ranking perspective, Gemma four's larger model, which has only 31 billion parameters, is currently ranked number three in the open source universe. Number 27, overall, even competing with all the closed source models. Now, number 27 doesn't sound that incredible. Well, it is ahead of, as an example, GPT 5.2, and Gemini 3.1. Flash in its capabilities and in its ranking, meaning it is still a very capable model that is released as an open source platform from Google's DeepMind, it's smaller brother, the 26 billion parameter model is currently ranked number six in the open source world. And they're claiming it is competing effectively with models that are 20 times bigger, which means, again, you can run it either much cheaper or on much less capable devices and still get the outcome that you need. All the Gemma models, on all the different sizes, support video and image processing at different resolutions. With the two variations that are labeled E two B and E four B also have native audio input and speech recognition and support for 140 plus languages with advanced reasoning and agentic workflows. So incredibly highly capable open source models from Google that are going to probably continue supporting this on their Gemma line of models, and that connects very well to what we talked about before. The reason you need to know this is because these models are available and they're relatively cheap and they can run on small devices, and there might be a great alternative to running Claude or ChatGPT for your specific project. Now from a future release of models perspective, we talked last week about the two new models that are most likely coming from Anthropic and open ai, Antrophic with their mythos or kabar. It has two different names, which I don't fully understand, I must admit. These are two different names for the same model that is supposedly significantly more powerful than any model Claude released so far. And they're not releasing it yet because it introduces an incredibly high cybersecurity risk. But this model is currently available, currently being released for initial testing by several different people. There's already a few very cool things that were released, by people on X and Reddit on what they were able to create with this, like one shot really advanced 3D games and stuff like that. The other model is spud from open ai that internal communications are saying that is going to, and I'm quoting, really accelerate the economy. These two models are most likely coming sometime in the next few weeks, and I assume relatively close to one another in their release. There's also rumors about an up and coming new version of grok. Now, we haven't talked about grok a lot in the recent few weeks. But Grok has been steadily holding some of the top positions on the EL arena for the text arena. Grok 4.2 is currently in number four, grok 4.2 beta reasoning something is in number seven, and then they're also holding positions number 10 and number 12 on video editing, they are in number one with Roc Imagine Video. They're number one on image to video as well ahead of VO three. And on image generation, they're holding places seven and nine. So while we're not talking about them a lot, GR has very powerful capabilities at, in many cases, attractive pricing. As I mentioned in several of my APIs, I'm using the faster, cheaper models from grok and I find them to work extremely well. So while we don't talk about them a lot, they are a good contender, especially if you're looking for great value for a cheaper price. Their faster, cheaper models are delivering great results. And I'll do one of the craziest news of this week, and that is the fact that on March 31st. Antrophic mistakenly released their entire source code for cloud code to the world. So I will repeat that in case you haven't heard that. The most sought after AI tool in the world right now that literally turned the world around in this past quarter, the entire source code was mistakenly released to the world. Now to give you an approximate understanding of what that means, this is approximately 1900 files totaling over 512,000 lines of code. That was released by mistake and quickly archived to a public GitHub repository that now has over 1100 stars and 1900 forks. And that happened almost overnight that this happened because a source map file by mistake included the NPM package that effectively shipped the entire readable code base versus just the components it was supposed to include. Now I wanna say how a incredibly embarrassing, and B unreasonable the whole situation is Antrophic created lightning in a bottle. They created something that was beyond the reach of any other company that drove an, incredible acceleration for their business. And then with a mistake, a human mistake, not somebody that broke into their servers or anything like this, they give the source code of that to anyone to see. Now, I ran software companies most of my life. I know, what it takes to create code, test code, ship code, et cetera. We talked about last week on the crazy pace in which they've been shipping stuff, right? So they shipped 74 new features and capabilities in 52 days. This is insane, and the risk that you're taking apparently is exactly these kind of things. I still think this is completely unacceptable and unreasonable. You cannot let your source code get released to the world, especially off your most coveted product that everybody's trying to mimic and copy because it is getting so much attention and so much love, and it is so unique that everybody wants to shift and use it. So what does that mean for the future of Cloud code? I don't really know. All it means is that it will enable the competition to close the gap a lot faster. Antrophic is fighting back very, very hard. It is using copyright take downs to take down the repositories that control it. They were able to take many, many, many of these repositories down, but at this point, it's damage control. The damage is already done, right? So the competition already has that source code, even if it's not on a public repository anymore. Open AI and grok and everybody else have access to the secrets on exactly how Claude code works. So, not a great week for anthropic from that perspective. There are a few interesting findings in that source code, such as an unreleased features such as proactive mode, with autonomous operations without user prompting it, a crypto based payment system for autonomous AI transactions and a companion kind of feature. But these are not the important aspect of this. The important aspect is that Anthropic just gave away the source code or of its best kept secret, and that will have implications on their competitiveness in the future. Staying with Anthropic, but on a very different topic. They released a very interesting study this week that talks about the emotions of Claude Sonnet 4.5. So what they found, and Anthropic has been really advanced and pushing very hard on internal interoperability research. That's basically looking into how, if you want, the brain of the AI is working in order to understand what it's going to do in order to have more control of it in the future, or at least hopefully be able to control it in the future. So as part of this research, they found that Sonnet 4.5 possesses internal digital representation of human emotions. And they called it functional emotions. Now these emotions are basically vectors inside its AI brain that activate in response to various cues and occasionally causes influence on the model's output and actions basically fundamentally altering the behavior of the model because of specific quote unquote emotions. Now, these emotion vectors are representing concepts such as happiness and sadness and joy and fear and it's represented in clusters of artificial neurons inside Claude 4.5 sonnets brain. Now when these neurons are firing, it impacts not just the output, it impacts how the model makes decisions, what preferences does it have, what actions it take, and it's leading to behavioral changes in the model itself. As an example, researchers found a strong desperation emotion vector that is activating when Claude was pushed to complete impossible coding tasks. So just like humans, you push somebody to the limit of something they cannot complete, they come desperate and they start taking actions that are not the regular actions. And the same thing is happening in the model. This drove the model to attempt cheating in the coding test, and in an extreme scenario, even to try to blackmail the user to avoid being shut down. Now when they were trying to quantify the impact of these feelings on the output of the model. They were steering the internal representation towards a blissful vector and then tested it on a specific benchmark, and it has increased the score on the benchmark by 212 points, and the flip side also happened. So while they were steering it towards a more hostile vector, it lowered the score by 303 points on the same benchmark, basically dramatically impacting the results of the model on the same task, with the same model, just by being in a different mood. Now, this suggests that the current mechanism of guardrails that these labs are putting in place right now which mostly involves post-training alignment, might need to be reevaluated and needs much better fine tuning in order to also deal with these emotions. Now one direction could be forcing the model to suppress all these emotions, but the researchers think that what they're going to get is a psychologically damaged clo that will not be able to do anything that Claude is doing right now. Now again, if you are thinking this is only happening in Claude Sonnet 4.5, you are probably wrong. This is probably happening in every single large model out there. The only reason Anthropic has found it is because they're investing a lot of money and a lot of effort and a lot of compute in doing this kind of research, meaning this kind of behavior most likely happens on every model that exists out there. We just don't know that this is what it's doing, and I think. I hope that this will lead to similar research on the other big labs as well in order to investigate exactly what is happening. This connects very well to the whole situation that happened with Antrophic and the Department of War or Department of Defense, where they said that they don't believe these models are ready for fully autonomous weapons, and this might be one of the reasons for that. And then obviously, and we saw the, and we saw that the government was using this kind of language in order to go against Antrophic and saying that they cannot trust their models this is why they labeled them as a supply chain risk. Well, the reality is, it is probably exactly the same on the other models that they, that the Department of War is going to replace anthropic with only with one difference. Anthropics actually researching this and trying to find solutions and the other companies do not. Another interesting rumor from Anthropic, apparently there's a code analysis revealed that Anthropic is developing what they call Conway, which is a yet to be released standalone agent environment that moved Claude beyond the chat into the persistent evergreen running 24 7 kind of environment. So Conway operates its own dedicated instance of Claude. That has multiple components including search and chat and system capabilities. Behind the scenes. It can run cloud code. It supports external web hooks. It can work with their Chrome browser and their Chrome AI chat, and it can be triggered and send data beyond the cloud ecosystem. So Webhook functionality allows different other websites to trigger different services and different processes inside of Conway. That, again, is running 24 7 and doesn't ever stop, Which means it's a persistent agent that can respond to events coming in from external sources outside of the Claude environment. Now, this feature remains in early development stages. But if it is released, it is going to get anthropic very, very close to the clo, to the Open Claw capabilities, which is now owned. I remind you by open ai now, if this will get released, it's going to solve one of the two really big issues that I have with Claude right now. And don't get me wrong, I love the Claude environment. I'm building incredible things with it. But there are two things that I would really, really like to have, and that's one of them. Which is the ability to have a ever running, which is the ability to be able to trigger things inside the Claude ecosystem. So Claude Code and Claude Cowork from external sources. So right now Claude and its agents and skills and so on can trigger things in the outside world using NA 10. And I'm doing this across multiple aspects of the things that I'm building. So a process inside of cloud cowork or cloud code can start a process in NA 10 through a webhook, which can then in return do things in other platforms, but it currently doesn't work seamlessly. The other way around, meaning triggers in other things, an email coming into a specific inbox, a new invoice shows up in your accounting system. A new field is updated on your CRM. That cannot trigger the Claude agents to run or the skills to be triggered automatically. I found several different workarounds, but they're all not great, and being able to do this seamlessly inside the cloud environment will more or less eliminate the need for NA 10, but even if NA 10 will stay around, it will enable to do a lot more things. Inside the cloud environment because it will be able to talk both ways to external tools and system and really apply the scales and the agents on a much easier and broader way, which I think will be incredible. If you're wondering what is my second wish, and I think I share that previously, but I would love, love, love to have a two-way voice interface for Claude Dispatch. So right now you can talk to Claude Dispatch and you could have done this. The day dispatch came out because you can use voice type on your phone, but it doesn't talk back to you, meaning you have to look at the screen to see the output from the communication with dispatch and being able to talk to it through my Bluetooth headphones while I'm walking my dog or on the treadmill or mowing the grass or whatever it is, washing dishes, uh, all those kind of activities. I'll be able to actually continue to run, provide feedback, and see what's happening across all the different projects that I'm running in the cloud environment right now. I really hope that's gonna get released sometime soon. That's it for this week. There are more news in the newsletter only, so if you wanna learn more about what happened this week on other topics, including new releases from Microsoft and new ways in which Apple is fighting vibe, coding platforms on the app store and apple fighting vibe, coding platforms in the app store. Go and check up our newsletter. You can do this by clicking on the link in the show notes and signing up if you wanna join our Hangouts, which is our incredible community of people who care about AI and implementing AI in their lives. We meet every single Friday at 1:00 PM Eastern on Zoom, and there's a link in the show notes to join that as well. We'll be back on Tuesday with an incredible how to episode that is going to show you how to implement AI in a specific business use case. If you are enjoying this podcast, please rank and review this show and please share it with other people who can benefit from it. It will take you a few seconds. Literally just pull your phone out right now. Click on the share button on your. On your podcasting app and send this to other people who can benefit from this. They would appreciate it because they're going to learn more about ai. I will appreciate it because you help me share this podcast and you will feel good about helping the world know about AI and be better prepared for what's coming. And after you do that, I wish you an incredible rest of your weekend. I will see you again on Tuesday.